Networking, Intelligent Systems and Security (Proceedings of NISS 2021)

Smart Innovation, Systems and Technologies 237
Mohamed Ben Ahmed

Horia-Nicolai L. Teodorescu
Tomader Mazri
Parthasarathy Subashini
Anouar Abdelhakim Boudhir Editors
Networking,
Intelligent Systems
and Security
Proceedings of NISS 2021
Smart Innovation, Systems and Technologies
Volume 237
Series Editors
Robert J. Howlett, Bournemouth University and KES International,
Shoreham-by-Sea, UK
Lakhmi C. Jain, KES International, Shoreham-by-Sea, UK
The Smart Innovation, Systems and Technologies book series encompasses the topics
of knowledge, intelligence, innovation and sustainability. The aim of the series is to
make available a platform for the publication of books on all aspects of single and
multi-disciplinary research on these themes in order to make the latest results avail-
able in a readily-accessible form. Volumes on interdisciplinary research combining
two or more of these areas is particularly sought.
The series covers systems and paradigms that employ knowledge and intelligence
in a broad sense. Its scope is systems having embedded knowledge and intelligence,
which may be applied to the solution of world problems in industry, the environment
and the community. It also focusses on the knowledge-transfer methodologies and
innovation strategies employed to make this happen effectively. The combination
of intelligent systems tools and a broad range of applications introduces a need
for a synergy of disciplines from science, technology, business and the humanities.
The series will include conference proceedings, edited collections, monographs,
handbooks, reference books, and other relevant types of book in areas of science and
technology where smart systems and technologies can offer innovative solutions.
High quality content is an essential feature for all book proposals accepted for the
series. It is expected that editors of all accepted volumes will ensure that contributions
are subjected to an appropriate level of reviewing process and adhere to KES quality
principles.
Indexed by SCOPUS, EI Compendex, INSPEC, WTI Frankfurt eG, zbMATH,
Japanese Science and Technology Agency (JST), SCImago, DBLP.
All books published in the series are submitted for consideration in Web of Science.
More information about this series at http://www.springer.com/series/8767

Mohamed Ben Ahmed ·
Horia-Nicolai L. Teodorescu · Tomader Mazri ·
Parthasarathy Subashini ·
Anouar Abdelhakim Boudhir
Editors
Networking, Intelligent
Systems and Security
Proceedings of NISS 2021
Editors
Mohamed Ben Ahmed Horia-Nicolai L. Teodorescu
Faculty of Sciences and Techniques Technical University of Iasi
of Tangier Ias, i, Romania
Abdelmalek Essaadi University
Tangier, Morocco Parthasarathy Subashini
Department of Computer Science
Tomader Mazri Avinashilingam University
National School of Applied Sciences Coimbatore, Tamil Nadu, India
Ibn Tofail University
Kénitra, Morocco
Anouar Abdelhakim Boudhir

Department of Computer Sciences
Faculty of Sciences and Techniques
of Tangier
Abdelmalek Essaadi University
Tangier, Morocco
ISSN 2190-3018 ISSN 2190-3026 (electronic)

Smart Innovation, Systems and Technologies
ISBN 978-981-16-3636-3 ISBN 978-981-16-3637-0 (eBook)
https://doi.org/10.1007/978-981-16-3637-0
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Singapore Pte Ltd. 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Committee
Conference Chair
Tomader Mazri, ENSA, Ibn Tofail University, Kenitra, Morocco
Conference General Chairs
Mohamed Ben Ahmed, FST, Tangier UAE University, Morocco

Anouar Abdelhakim Boudhir, FST, Tangier UAE University, Morocco
Bernadetta Kwintiana Ane, University of Stuttgart, Germany
Conference Technical Programme Committee Chair
Wassila Mtalaa, Luxembourg Institute of Science and Technology, Luxembourg
Keynote and Panels Chair
Domingos Santos, Polytechnic Institute Castelo Branco, Portugal
Publications Chair
İsmail Rakıp Karas, o, Karabuk University
v
vi Committee
Special Issues Chair
Senthil Kumar, Hindustan College of Arts and Science, India
Local Organizing Committee
Tomader Mazri, ENSA, UIT, Morocco

Hassan Mharzi, ENSA, UIT, Morocco
Mohamed Nabil Srifi, ENSA, UIT, Morocco
Tarik Jarou, ENSA, UIT, Morocco
Benbrahim Mohammed, ENSA, UIT, Morocco
Abderrahim Bajjit, ENSA, UIT, Morocco
Rachid Elgouri, ENSA, UIT, Morocco
Imane Sahmi, ENSA, UIT, Morocco
Loubna El Amrani, ENSA, UIT, Morocco
Technical Programme Committee
Ismail Rakip Karas, Karabuk University, Türkiye

Abdel-Badeeh M. Salem, Ain Shams University, Egypt
Abderrahim Ghadi, FSTT, UAE, Morocco
Accorsi Riccardo, Bologna University, Italy
Aftab Ahmed Khan, Karakoram International University, Pakistan
Ahmad S. Almogren, King Saud University, Saudi Arabia
Ahmed Kadhim Hussein, Babylon University, Iraq
Alabdulkarim Lamya, King Saud University, Saudi Arabia
Alghamdi Jarallah, Prince Sultan University, Saudi Arabia
Ali Jamali, Universiti Teknologi Malaysia
Alias Abdul Rahman, Universiti Teknologi Malaysia
Anabtawi Mahasen, Al-Quds University, Palestine
Anton Yudhana, Universitas Ahmad Dahlan, Indonesia
Arioua Mounir, UAE, Morocco
Assaghir Zainab, Lebanese University, Lebanon
Astitou Abdelali, UAE, Morocco
Aydın Üstün, Kocaeli University, Türkiye
Aziz Mahboub, FSTT, UAE, Morocco
Barış Kazar, Oracle, USA
Bataev Vladimir, Zaz Ventures, Switzerland
Behnam Alizadehashrafi, Tabriz Islamic Art University, Iran
Behnam Atazadeh, University of Melbourne, Australia
Committee vii
Ben Yahya Sadok, Faculty of Sciences of Tunis, Tunisia

Bessai-Mechmach Fatma Zohra, CERIST, Algeria
Biswajeet Pradhan, University of Technology Sydney, Australia
Berk Anbaroğlu, Hacettepe University, Türkiye
Bolulmalf Mohammed, UIR, Morocco
Boutejdar Ahmed, German Research Foundation, Bonn, Germany
Chadli Lala Saadia, University Sultan Moulay Slimane, Morocco
Cumhur Şahin, Gebze Technical University, Türkiye
Damir Žarko, Zagreb University, Croatia
Dominique Groux, UPJV, France
Dousset Bernard UPS, Toulouse, France
Edward Duncan, The University of Mines and Technology, Ghana
Eehab Hamzi Hijazi, An-Najah University, Palestine
El Kafhali Said, Hassan 1st University, Settat, Morocco
El Malahi Mostafa, USMBA University, Fez, Morocco
El Mhouti Abderrahim, FST, Al-Hoceima, Morocco
El Haddadi Anass, UAE University, Morocco
El Hebeary Mohamed Rashad, Cairo University, Egypt
El Ouarghi Hossain, ENSAH, UAE University, Morocco
En-Naimi El Mokhtar, UAE, Morocco
Enrique Arias, Castilla-La Mancha University, Spain
Tolga Ensari, Istanbul University, Türkiye
Filip Biljecki, National University of Singapore
Francesc Anton Castro, Technical University of Denmark
Ghulam Ali Mallah, Shah Abdullatif University, Pakistan
Habibullah Abbasi, University of Sindh, Pakistan
Haddadi Kamel Iemn, Lille University, France
Hanane Reddad, USMS University, Morroco
Hazim Tawfik, Cairo University, Egypt
Huseyin Zahit Selvi, Konya Necmettin Erbakan University
Ilker Türker, Karabuk University, Türkiye
Iman Elawady, Ecole Nationale Polytechnique d’Oran, Algeria
Indubhushan Patnaikuni, RMIT—Royal Melbourne Institute of Technology,
Australia
Ismail Büyüksalih, Bimtaş A. Ş., Türkiye
Ivin Amri Musliman, Universiti Teknologi Malaysia
J. Amudhavel, VIT Bhopal University, Madhya Pradesh, India
Jaime Lioret Mauri, Polytechnic University of Valencia, Spain
Jus Kocijan, Nova Gorica University, Slovenia
Kadir Ulutaş, Karabuk University
Kasım Ozacar, Karabuk University
Khoudeir Majdi, IUT, Poitiers University, France
Labib Arafeh, Al-Quds University, Palestine
Laila Moussaid, ENSEM, Casablanca, Morocco
Lalam Mustapha, Mouloud Mammeri University of Tizi Ouzou, Algeria
viii Committee
Loncaric Sven, Zagreb University, Croatia

Lotfi Elaachak, FSTT, UAE, Morocco
Mademlis Christos, Aristotle University of Thessaloniki, Greece
Miranda Serge, Nice University, France
Mohamed El Ghami, University of Bergen, Norway
Mohammad Sharifikia, Tarbiat Modares University, Iran
Mousannif Hajar, Cadi Ayyad University, Morocco
Muhamad Uznir Ujang, Universiti Teknologi Malaysia
Muhammad Imzan Hassan, Universiti Teknologi Malaysia
My Lahcen Hasnaoui, Moulay Ismail University, Morocco
Mykola Kozlenko, Vasyl Stefanyk Precarpathian National University, Ukraine
Omer Muhammet Soysal, Southeastern Louisiana University, USA
Ouederni Meriem, INP—ENSEEIHT Toulouse, France
R. S. Ajin, DEOC, DDMA, Kerala, India
Rani El Meouche, Ecole Spéciale des Travaux Publics, France
Sagahyroon Assim, American University of Sharjah, United Arab Emirates
Saied Pirasteh, University of Waterloo, Canada
Senthil Kumar, Hindustan College of Arts and Science, India
Siddique Ullah Baig, COMSATS Institute of Information Technology, Pakistan
Slimani Yahya, Manouba University, Tunisia
Sonja Grgić, Zagreb University, Croatia
Sri Winiarti, Universitas Ahmad Dahlan, Indonesia
Suhaibah Azri, Universiti Teknologi Malaysia
Sunardi, Universitas Ahmad Dahlan, Indonesia
Tebibel Bouabana Thouraya, ESI, Alger, Algeria
Xiaoguang Yue, International Engineering and Technology Institute, Hong Kong
Yasyn Elyusufi, FSTT, UAE, Morocco
Youness Dehbi, University of Bonn, Germany
Yusuf Arayıcı, Northumbria University, UK
Zigh Ehlem Slimane, INTTIC, Oran, Algeria
Zouhri Amal, USMBA University, Fez, Morocco
Preface
In an age of explosive worldwide growth of electronic data storage and communica-

tions, effective protection of information has become a critical requirement.
With the exponential growth of wireless communications, Internet of Things,
and cloud computing, and the increasingly dominant roles played by electronic
commerce in every major industry, safeguarding the information in storage and trav-
elling over the communication networks is increasingly becoming the most critical
and contentious challenges for the technology innovators.
This trend opens up significant research activity for academics and their partners
(industrialists, governments, civil society, etc.) in order to establish essential and
intelligent bases for developing the active areas of networking, intelligent systems
and security.
This edited book aims to present scientific research and engineering applications
for the construction of intelligent systems and their various innovative applications
and services. The book also aims to provide an integrated view of the problems
to researchers, engineers, practitioners and to outline new topics in networks and
security.
This edition is the result of work accepted and presented at the Fourth International
Conference on Networks, Intelligent Systems and Security (NISS 2021) held on
April, 1–2, 2020, in Kenitra, Morocco. It brings together original research, work
carried out and proposed architectures on the main themes of the conference.
The goal of this book edition is constructing and building the basics and essentials
researches, innovations and applications that can help on the growth of the future
next generation of networks and intelligent systems.
We would like to acknowledge and thank Springer Nature staff for their support,
guidance and for the edition of this book.
ix
x Preface
Finally, we wish to express our sincere thanks to Prof. Robert J. Howlett, Mr.
Aninda Bose and Ms. Sharmila Mary Panner Selvam for their kind support and help
to promote and develop research.
Tangier, Morocco Mohamed Ben Ahmed

Ias, i, Romania Horia-Nicolai L. Teodorescu
Kénitra, Morocco Tomader Mazri
Coimbatore, India Parthasarathy Subashini
Tangier, Morocco Anouar Abdelhakim Boudhir
Contents
Artificial Intelligence for Sustainability

Detection of Human Activities in Wildlands to Prevent
the Occurrence of Wildfires Using Deep Learning and Remote
Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Ayoub Jadouli and Chaker El Amrani
The Evolution of the Traffic Congestion Prediction and AI
Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Badr-Eddine Soussi Niaimi, Mohammed Bouhorma, and Hassan Zili
Tomato Plant Disease Detection and Classification Using
Convolutional Neural Network Architectures Technologies . . . . . . . . . . . . 33
Djalal Rafik Hammou and Mechab Boubaker
Generative and Autoencoder Models for Large-Scale Mutivariate
Unsupervised Anomaly Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Nabila Ounasser, Maryem Rhanoui, Mounia Mikram, and Bouchra El Asri
Automatic Spatio-Temporal Deep Learning-Based Approach
for Cardiac Cine MRI Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Abderazzak Ammar, Omar Bouattane, and Mohamed Youssfi
Skin Detection Based on Convolutional Neural Network . . . . . . . . . . . . . . . 75
Yamina Bordjiba, Chemesse Ennehar Bencheriet, and Zahia Mabrek
CRAN: An Hybrid CNN-RNN Attention-Based Model for Arabic
Machine Translation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Nouhaila Bensalah, Habib Ayad, Abdellah Adib,
and Abdelhamid Ibn El Farouk
Impact of the CNN Patch Size in the Writer Identification . . . . . . . . . . . . . 103
Abdelillah Semma, Yaâcoub Hannad, and Mohamed El Youssfi El Kettani
xi
xii Contents
Network and Cloud Technologies

Optimization of a Multi-criteria Cognitive Radio User Through
Autonomous Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Naouel Seghiri, Mohammed Zakarya Baba-Ahmed,
Badr Benmammar, and Nadhir Houari
MmRPL: QoS Aware Routing for Internet of Multimedia Things . . . . . . 133
Hadjer Bouzebiba and Oussama Hadj Abdelkader
Channel Estimation in Massive MIMO Systems for Spatially
Correlated Channels with Pilot Contamination . . . . . . . . . . . . . . . . . . . . . . . 147
Mohamed Boulouird, Jamal Amadid, Abdelhamid Riadi,
and Moha M’Rabet Hassani
On Channel Estimation of Uplink TDD Massive MIMO Systems
Through Different Pilot Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Jamal Amadid, Mohamed Boulouird, Abdelhamid Riadi,
NarrowBand-IoT and eMTC Towards Massive MTC: Performance
Evaluation and Comparison for 5G mMTC . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Adil Abou El Hassan, Abdelmalek El Mehdi, and Mohammed Saber
Integrating Business Intelligence with Cloud Computing: State
of the Art and Fundamental Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
Hind El Ghalbzouri and Jaber El Bouhdidi
Distributed Architecture for Interoperable Signaling Interlocking . . . . . . 215
Ikram Abourahim, Mustapha Amghar, and Mohsine Eleuldj
A New Design of an Ant Colony Optimization (ACO) Algorithm
for Optimization of Ad Hoc Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Hala Khankhour, Otman Abdoun, and Jâafar Abouchabaka
Real-Time Distributed Pipeline Architecture for Pedestrians’
Trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
Kaoutar Bella and Azedine Boulmakoul
Reconfiguration of the Radial Distribution for Multiple DGs
by Using an Improved PSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Meriem M’dioud, Rachid Bannari, and Ismail Elkafazi
On the Performance of 5G Narrow-Band Internet of Things
for Industrial Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Abdellah Chehri, Hasna Chaibi, Rachid Saadane, El Mehdi Ouafiq,
and Ahmed Slalmi
A Novel Design of Frequency Reconfigurable Antenna for 5G
Mobile Phones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Sanaa Errahili, Asma Khabba, Saida Ibnyaich, and Abdelouhab Zeroual
Contents xiii
Smart Security
A Real-Time Smart Agent for Network Traffic Profiling
and Intrusion Detection Based on Combined Machine Learning
Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
Nadiya El Kamel, Mohamed Eddabbah, Youssef Lmoumen, and Raja Touahni
Privacy Threat Modeling in Personalized Search Systems . . . . . . . . . . . . . 311
Anas El-Ansari, Marouane Birjali, Mustapha Hankar,
and Abderrahim Beni-Hssane
Enhanced Intrusion Detection System Based on AutoEncoder
Network and Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
Sihem Dadi and Mohamed Abid
Comparative Study of Keccak and Blake2 Hash Functions . . . . . . . . . . . . 343
Hind EL Makhtoum and Youssef Bentaleb
3
Cryptography Over the Twisted Hessian Curve Ha,d . . . . . . . . . . . . . . . . . . 351
Abdelâli Grini, Abdelhakim Chillali, and Hakima Mouanis
Method for Designing Countermeasures for Crypto-Ransomware
Based on the NIST CSF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Hector Torres-Calderon, Marco Velasquez, and David Mauricio
Comparative Study Between Network Layer Attacks in Mobile Ad
Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
Oussama Sbai and Mohamed Elboukhari
Security of Deep Learning Models in 5G Networks: Proposition
of Security Assessment Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
Asmaa Ftaimi and Tomader Mazri
Effects of Jamming Attack on the Internet of Things . . . . . . . . . . . . . . . . . . 409
Imane Kerrakchou, Sara Chadli, Mohammed Saber,
and Mohammed Ghaouth Belkasmi
H-RCBAC: Hadoop Access Control Based on Roles and Content . . . . . . . 423
Sarah Nait Bahloul, Karim Bessaoud, and Meriem Abid
Toward a Safe Pedestrian Walkability: A Real-Time Reactive
Microservice Oriented Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Ghyzlane Cherradi, Azedine Boulmakoul, Lamia Karim, and Meriem Mandar
Image-Based Malware Classification Using Multi-layer Perceptron . . . . . 453
Ikram Ben Abdel Ouahab, Lotfi Elaachak, and Mohammed Bouhorma
Preserving Privacy in a Smart Healthcare System Based on IoT . . . . . . . . 465
Rabie Barhoun and Maryam Ed-daibouni
xiv Contents
Smart Digital Learning

Extracting Learner’s Model Variables for Dynamic Grouping
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
Noureddine Gouasmi, Mahnane Lamia, and Yassine Lafifi
E-learning and the New Pedagogical Practices of Moroccan
Teachers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
Nadia El Ouesdadi and Sara Rochdi
A Sentiment Analysis Based Approach to Fight MOOCs’ Drop Out . . . . 509
Soukaina Sraidi, El Miloud Smaili, Salma Azzouzi, and My El Hassan Charaf
The Personalization of Learners’ Educational Paths E-learning . . . . . . . . 521
Ilham Dhaiouir, Mostafa Ezziyyani, and Mohamed Khaldi
Formulating Quizzes Questions Using Artificial Intelligent
Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
Abdelali El Gourari, Mustapha Raoufi, and Mohammed Skouri
Smart Campus Ibn Tofail Approaches and Implementation . . . . . . . . . . . 549
Srhir Ahmed and Tomader Mazri
Boosting Students Motivation Through Gamified Hybrid Learning
Environments Bleurabbit Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
Mohammed Berehil
An Analysis of ResNet50 Model and RMSprop Optimizer
for Education Platform Using an Intelligent Chatbot System . . . . . . . . . . 577
Youness Saadna, Anouar Abdelhakim Boudhir, and Mohamed Ben Ahmed
Smart Information Systems

BPMN to UML Class Diagram Using QVT . . . . . . . . . . . . . . . . . . . . . . . . . . 593
Mohamed Achraf Habri, Redouane Esbai,
and Yasser Lamlili El Mazoui Nadori
Endorsing Energy Efficiency Through Accurate Appliance-Level
Power Monitoring, Automation and Data Visualization . . . . . . . . . . . . . . . 603
Aya Sayed, Abdullah Alsalemi, Yassine Himeur, Faycal Bensaali,
and Abbes Amira
Towards a Smart City Approach: A Comparative Study . . . . . . . . . . . . . . 619
Zineb Korachi and Bouchaib Bounabat
Hyperspectral Data Preprocessing of the Northwestern Algeria
Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635
Zoulikha Mehalli, Ehlem Zigh, Abdelhamid Loukil, and Adda Ali Pacha
Contents xv
Smart Agriculture Solution Based on IoT and TVWS for Arid

Regions of the Central African Republic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653
Edgard Ndassimba, Nadege Gladys Ndassimba,
Ghislain Mervyl Kossingou, and Samuel Ouya
Model-Driven Engineering: From SQL Relational Database
to Column—Oriented Database in Big Data Context . . . . . . . . . . . . . . . . . . 667
Fatima Zahra Belkadi and Redouane Esbai
Data Lake Management Based on DLDS Approach . . . . . . . . . . . . . . . . . . . 679
Mohamed Cherradi, Anass EL Haddadi, and Hayat Routaib
Evaluation of Similarity Measures in Semantic Web Service
Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691
Mourad Fariss, Naoufal El Allali, Hakima Asaidi, and Mohamed Bellouki
Knowledge Discovery for Sustainability Enhancement Through
Design for Relevance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705
Abla Chaouni Benabdellah, Asmaa Benghabrit, Imane Bouhaddou,
and Kamar Zekhnini
Location Finder Mobile Application Using Android and Google
SpreadSheets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 723
Adeosun Nehemiah Olufemi and Melike Sah
Sign Language Recognition with Quaternion Moment Invariants:
A Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 737
Ilham El Ouariachi, Rachid Benouini, Khalid Zenkouar,
Arsalane Zarghili, and Hakim El Fadili
Virtual Spider for Real-Time Finding Things Close to Pedestrians . . . . . 749
Souhail Elkaissi and Azedine Boulmakoul
Evaluating the Impact of Oversampling on Arabic L1 and L2
Readability Prediction Performances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
Naoual Nassiri, Abdelhak Lakhouaja, and Violetta Cavalli-Sforza
An Enhanced Social Spider Colony Optimization for Global
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 775
Farouq Zitouni, Saad Harous, and Ramdane Maamri
Data Processing on Distributed Systems Storage Challenges . . . . . . . . . . . 795
Mohamed Eddoujaji, Hassan Samadi, and Mohamed Bohorma
COVID-19 Pandemic
Data-Based Automatic Covid-19 Rumors Detection in Social
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815
Bolaji Bamiro and Ismail Assayad
xvi Contents
Security and Privacy Protection in the e-Health System: Remote

Monitoring of COVID-19 Patients as a Use Case . . . . . . . . . . . . . . . . . . . . . . 829
Mounira Sassi and Mohamed Abid
Forecasting COVID-19 Cases in Morocco: A Deep Learning
Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845
Mustapha Hankar, Marouane Birjali, and Abderrahim Beni-Hssane
The Impact of COVID-19 on Parkinson’s Disease Patients
from Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 859
Hanane Grissette and El Habib Nfaoui
Missing Data Analysis in the Healthcare Field: COVID-19 Case
Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 873
Hayat Bihri, Sara Hsaini, Rachid Nejjari, Salma Azzouzi,
and My El Hassan Charaf
An Analysis of the Content in Social Networks During COVID-19
Pandemic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 885
Mironela Pirnau
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 899

About the Editors
Mohamed Ben Ahmed is an associate professor of computer sciences at Abdel-

malek Essaâdi University, Morocco; he received the Ph.D. degree in computer
sciences and telecommunications, in 2010 from Abdelmalek Essaâdi University.
His researches are about smart and sustainable cities, data mining and routing in
wireless sensor networks. He is currently a supervisor of several thesis and an inves-
tigator in several international research projects about smart cities. He is the author
of more than fifty papers published in international journals and conferences. He is
the co-editor of Springer Innovations in Smart Cities Applications book. He is a chair
and a committee member of several international conferences.
Prof. Horia-Nicolai L. Teodorescu teaches intelligent systems at Gheorghe

Asachi Technical University of Iasi and language technology at Alexandru Ioan Cuza
University of Iasi; in addition, he is the director of the Institute of Computer Science
of the Romanian Academy. He served for extended periods as a visiting and invited
professor at Swiss Federal Institute of Technology, Lausanne, University of South
Florida, Tampa, and Kyushu Institute of Technology and FLSI, Iizuka, Japan, among
others. He served as a co-director of postgraduate and doctoral studies in Lausanne
and in University of Leon, Spain. He also served as a vice-rector of Gheorghe Asachi
Technical University of Iasi. When he was included in the Romanian Academy, he
was the youngest member of the learned body. Dr. Teodorescu occupied several
positions in national and international societies and institutions, including member
of the independent expert group and vice-chair of the group for Computer Science of
NATO. Dr. Teodorescu served as a member of the editorial boards of several major
journals issued by publishers as IEEE, Francis & Taylor, Elsevier, and the Romanian
Academy. He authored about 250 conference and journal papers and more than 25
books; he holds 24 national and international patents.
Prof. Tomader Mazri received her HDR degree in Networks and Telecommunica-
tion from Ibn Tofail University, Ph.D. in Microelectronics and Telecommunication
from Sidi Mohamed Ben Abdellah University and INPT of Rabat, Master’s in Micro-
electronics and Telecommunication Systems, and Bachelor’s in Telecommunication
from the Cadi Ayyad University. She is currently a professor at the National School
xvii
xviii About the Editors
of Applied Sciences of Kenitra, a permanent member of Electrical and Telecom-

munications Engineering Laboratory, and an author and a co-author of 15 articles
journals, 40 articles in international conferences, 3 chapters, and 5 books. Her major
research interests are on microwave systems for mobile and radar, smart antennas,
and mobile network security.
Parthasarathy Subashini has also received Ph.D. in Computer Science in 2009

from Avinashilingam University for Women, Tamil Nadu, India. From 1994, she
is working as a professor in the Computer Science Department of Avinashilingam
University. Concurrently, she contributed to several fields of mathematics, espe-
cially nature-inspired computing. She has authored or co-authored 4 books, 6 book
chapters, 1 monograph, 145 papers, including IEEE, Springer’s in various inter-
national, national journals, and conferences. She has held positions as a reviewer,
chairpersons for different peer-reviewed journals. Under her supervision, she has ten
research projects of worth more than 2.7 crores from various funding agencies like
Defence Research and Development Organization, Department of Science and Tech-
nology, SERB, and University Grants Commission. She has visited many countries
for various knowledge sharing events. As a member of IEEE, IEEE Computational
Intelligence Society, and IEEE Computer Society of India, she extended her contri-
bution as IEEE Chair for Women in Computing under IEEE Computer Society of
India Council in the year 2015–2016.
Anouar Abdelhakim Boudhir is currently an associate professor at the Faculty

of Sciences and Technique of Tangier. Actually, he is the president of the Mediter-
ranean Association of Sciences and Technologies. He is an adviser at the Moroccan
union against dropping out of school. He received the HDR degree from Abdelmalek
Essaadi University; he is the co-author of several papers published in IEEExplorer,
ACM, and in high indexed journals and conference. He co-edited a several books
published on Springer series, and he is a co-founder of a series of international
conferences (Smart health17, SCIS’16, SCA18, SCA19, SCA20, NISS18, NISS19,
NISS20, DATA21) till 2016. His supervise several thesis about artificial intelligence,
security, and E-healthcare. His key research relates to ad hoc networks, VANETS,
WSN, IoT, big data, computer healthcare applications, and security applications.
Artificial Intelligence for Sustainability
Detection of Human Activities
in Wildlands to Prevent the Occurrence
of Wildfires Using Deep Learning
and Remote Sensing
Ayoub Jadouli and Chaker El Amrani
Abstract Human activities in wildland are responsible for the largest part of wildfire
cases. This paper presents a work that uses deep learning on remote sensing images
to detect human activity in wildlands to prevent fire occurrence that can be caused
by humans. Human activities can be presented as any human interaction with wild-
lands, and it can be roads, cars, vehicles, homes, human shapes, agricultural lands,
golfs, airplanes, or any other human proof of existence or objects in wild lands.
Conventional neural network is used to classify the images. For that, we used three
approaches, in which one is the object detection and scene classification approach,
the second is land class approach where two classes of lands can be considered which
are wildlands with human interactions and wildland without human interaction. The
third approach is more general and includes three classes that are urban lands, pure
wildlands, and wildlands with human activities. The results show that it is possible
to detect human activities in wildlands using the models presented in this paper. The
second approach can be considered the most successful even if it is the simpler.
1 Introduction
1.1 Wildfire and Machine Learning
Machine learning (ML) is the term for techniques that allow the machine to find a way
to resolve problems without being specially programmed for that. ML approaches
are used in the data science context, relating: data size, computational requirements,
generalizability, and interpretability of data. In the last two decades, there is a big
increase in using ML methods in wildfire fields. There are three main types of ML
methods:
A. Jadouli (B) · C. El Amrani

LIST Lab, Faculty of Sciences and Techniques of Tangier, Abdelmalek Essaâdi University,
Tangier, Morocco
e-mail: ajadouli@uae.ac.ma
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 3
M. Ben Ahmed et al. (eds.), Networking, Intelligent Systems and Security,
Smart Innovation, Systems and Technologies 237,
https://doi.org/10.1007/978-981-16-3637-0_1
4 A. Jadouli and C. El Amrani
• Supervised ML: The goal is learning a parametrized function or model that maps
a known input (i.e., predictor variables) to a known output (or target variables).
So, an algorithm is used to learn the parameters of that function using examples.
Supervised learning can solve two types of problems, and it can be a classification
problem when the target variables are categorical or a regression problem when
the target variables are continuous. There is a lot of methods that can be cate-
gorized as supervised ML: Naive Bayes (NB), decision trees (DT), classification
and regression tree (CART), random forest (RF), deep neural network (DNN),
Gaussian processes (GP), artificial neural networks (ANN), genetic algorithms
(GA), recurrent neural network (RNN), maximum entropy (MAXENT), regres-
sion trees (BRT), random forest (RF), K-nearest neighbor (KNN), support vector
machines (SVM) [Hearst, Dumais, Osuna,Platt, Scholkopf, 1998], and K-SVM.
Supervised ML can be used in these fields (fire spread/burn area prediction, fire
occurrence, fire severity, smoke prediction, climate change, fuels characterization,
fire detection, and fire mapping) [1].
• Unsupervised Learning: It is used when the target variables are not available,
and generally, the goal is understanding the patterns and discovering the output,
dimensionality reduction, or clustering. The relationships or patterns are extracted
from the data without any guidance as to the right answer. A lot of methods
can be considered in that field (K-means clustering (KM), self-organizing maps
SOM, autoencoders, Gaussian mixture models (GMM), iterative self-organizing
DATA algorithm (ISODATA), hidden Markov models (HMM), density-based
spatial clustering of applications with noise (DBSCAN)), T-distributed stochastic
neighbor embedding ( t-SNE), random forest (RF), boosted regression trees (BRT)
[Freund, Shapire, 1995], maximum entropy (MaxEnt), principal component anal-
ysis (PCA), and factor analysis). Unsupervised ML can be used for fire detection,
fire mapping, burned area prediction, fire weather prediction landscape controls
on fire, fire susceptibility, and fire spread/burn area prediction [1].
• Agent-Based Learning: A single or a group of autonomous agents interact with
the environment following specific rules of behavior. Agent-based learning can
be used for optimization and for decision making. The next algorithms can be
considered as agent-based: genetic algorithms (GA), Monte Carlo tree search
(MCTS), Asynchronous Advantage Actor-Critic (A3C), deep Q-network (DQN),
and reinforcement learning (RL) [Sutton, Barto, 1998]. The agent-based learning
can be useful for optimizing fire simulators, fire spread and growth, fuel treatment,
planning and policy, and wildfire response [1].
1.2 Deep Learning and Remote Sensing
In the last decade, deep learning models can be considered as the most successful
ML methods. It is considered to be an artificial neural network that involves multiple
hidden layers [2]. Because of the large successful use of these methods by big compa-
nies in production, the research interest in the field has increased, and more and more
Detection of Human Activities in Wildlands to Prevent … 5
applications have been used to solve a large scale of problems including remote
sensing problems.
Remote sensing is a technique that uses reflected or emitted electromagnetic
energy to get information about the earth’s land and water surfaces, obtaining quan-
titative measurements and estimations of geo-bio-physical variables. That is possible
because every material in the scene has a special interaction with electromagnetic
radiation that can be emitted, reflected, or absorbed by these materials depending on
their shapes and their molecular composition. With the increase of the spatial reso-
lution of satellite images created by merging their data with information collected at
a higher resolution, it is possible to achieve resolution up to 25-cm [3].
1.3 Problem
Wildfires are mostly caused by humans. So, to prevent the occurrence of fire, the
monitoring of human activities in wildlands is an essential task. A lot of technical
challenges are involved in that field. Deep learning applied to high-resolution spatial
images can be considered to solve a part of the problem. So, this work focuses on the
usage of convolutional neural network (CNN) to detect human activities in wildlands
based on remote sensing images. The results of this work will be used in future work
to achieve a prediction of fire occurrence that can be caused by humans. The details
can be found in Sect. 3.
2 Related Work
Deep neural networks are the most successful methods used to solve problems linked
to the interpretation of remote sensing data. So, a lot of researchers are interested in
these methods:
Kadhim et al. [4] presented useful models for satellite image classification that
are based on convolutional neural networks, the features that are used to classify
the image extracted by using four pre-trained CNN models: Resnet50, GoogleNet,
VGG19, and AlexNet. The Resnet50 model achieves a better result than other models
for all the datasets they used. Varshney and Debvrat [5] used a convolutional neural
network and fused SWIR and VNIR multi-resolution data to achieve pixel-wise clas-
sification through semantic segmentation, giving a good result with a high snow-and-
cloud F1 score. They found that their DNN was better than the traditional methods
and was able to learn spectro-contextual information, which can help in the semantic
segmentation of spatial data. Long et al. [6] propose a new object localization frame-
work, which can be divided into three processes: region proposal, classification, and
accurate object localization process. They found that the dimension-reduction model
performs better than the retrained and fine-tuned models and the detection precision
of the combined CNN model is much higher than that of any single model. Wang
et al. [7] used mixed spectral characteristics and CNN to propose a remote sensing
recognition method for the detection of landslides and obtained an accuracy of 98.98
and 97.69%.
With the success of CNN in image recognition, more studies are now interested to
find the best and appropriate model that can be used to solve a specific problem. This
is the case of Zhong et al. [8] proposing RSNet a remote sensing DNN framework.
Their goal is to automatically search and find the appropriate network architecture
for image recognition tasks based on high-resolution remote sensing images (HRS).
3 Proposed Solution and Methods
3.1 Architecture and Implementation of the Design
Our research focuses on the detection of human activities on wildlands using CNN
on satellite high-resolution Images. Even if this subject is very general, it is very
much related to our project because human activities are the primary cause of the
wildfire. The goal is to use the result of the models to be able to predict the areas
where the occurrence of fire in wildlands will start with the help of weather data (See
Fig. 1). We proposed three approaches to solve this problem:
• A simple CNN model trained by UC Merced dataset with five classes of 21 classes
as output. And based on the 21 classes, a conclusion can be made.
• A simple CNN model for a simple classification used by two classes (wildland
with human activities and pure wildlands)
• ResNet50 pre-trained model with transfer learning to output with three classes
(urban lands, wildlands with human interactions, and pure wildland without
human interaction) (See Fig. 2).
3.2 Conventional Neural Network Model
Convolutional neural networks (CNN) are a type of deep neural network which have
one or more convolutional layers. This network can be very useful when there is a
need to detect a specific pattern in data and make sense of them. CNN is very useful
for image analyses and matches the requirement of our study.
In our case, we have used the same sequential model for the first approach with
five classes and the second approach with two classes (See Fig. 3). Where ten layers
can be found described in Table 1 for the first approach and Table 2 for the second
approach.
The models are built thanks to Python 3 and Keras library, and the training and
tests are made with the parallelization on the top of NVIDIA GeForce 840M single
GPU device, with 384 CUDA cores, 1029 MHz in frequency, and 2048 MB DDR3.
Fig. 1 Process of wildfire prediction using DL CNN and LSTM/GRU based on human activity
detection
3.3 Transfer Learning
Transfer learning is a technique where information learned in a primary model of

machine learning is transferred to a secondary machine learning model to achieve a
new task. This technique is used in the third approach for the classification of three
land classes (pure wildland, wildland with human activities, and urban lands) See
Fig. 2.
In this model, we try to use the pre-trained model of RestNet50 without the
top layer “resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5” with transfer
learning to classify the three classes that we work on based on the pre-trained network
(Table 3; Fig. 4).
3.4 Datasets
This study uses UC Merced dataset as the principal source of data because it is widely
utilized in land use cases of studies and showed good results in machine learning
classification problems.
Fig. 2 Diagram shows how ResNet50 pretreated network is used with transfer learning methods
to classify three categories of wildlands images
UC Merced land use dataset is introduced by Yang and Newsam [9]. It is a land use
image dataset with 21 classes, each class has 100 images, and each image measures
256 × 256 pixels, with a spatial resolution of 0.3 m per pixel. The images are
extracted from United States Geological Survey National Map Urban Area Imagery.
The images were manually cropped to extract various urban areas around the United
States [9].
We have subdivided the dataset into three other datasets to match our case study.
In the first dataset, we have extracted five classes that can be linked to wildlands
which are as follows: forest because it is the main study area, freeways because they
are mostly built in wildland and are proof of human activities in wildlands, golf
course because the images match the forest images and we need our model to find
the differences, river because sometimes it looks like a freeway or a road and we
need our model to find the differences, and finally, sparse residential areas because
we found buildings that are mostly built near wildlands and that can be considered
as another proof of human activities in wildlands (See Fig. 6).
In the second dataset, we have split the first dataset into two classes which are
Human Activity Wildland (wildland or images that look like wildlands with human
activity proofs) and Pure Wildland (clean forest and rivers) (See Fig. 7).
The third UC Merced dataset images are subdivided into three classes which
are Urban Land (images of urban areas), Human Activity Wildland (wildlands or
wildland like images where a trace of human activities can be found), and Pure
Wildland (images of wildlands with no trace of human activity) (See Fig. 8). The
Fig. 3 Approach 1 and

approach 2 model’s layers
Table 1 Approach 1 model’s

Layer (type) Output shape Param #
details
conv2d (Conv2D) (None, 254, 254, 32) 896
max_pooling2d (None, 127, 127, 32) 0
(MaxPooling2D)
conv2d_1 (Conv2D) (None, 125, 125, 32) 9248
Max_pooling2d_1 (None, 62, 62, 32) 0
(MaxPooling2D)
Dropout (Dropout) (None, 62, 62, 32) 0
Flatten (Flatten) (None, 123,008) 0
Dropout_1 (Dropout) (None, 128) 15,745,152
Dropout_1 (Dropout) (None, 128) 0
Dense_1 (Dense) (None, 5) 645

details
conv2d (Conv2D) (None, 254, 254, 32) 896
max_pooling2d (None, 127, 127, 32) 0
(MaxPooling2D)
conv2d_1 (Conv2D) (None, 125, 125, 32) 9248
max_pooling2d_1 (None, 62, 62, 32) 0
(MaxPooling2D)
Dropout (Dropout) (None, 62, 62, 32) 0
Flatten (Flatten) (None, 123,008) 0
Dropout_1 (Dropout) (None, 128) 15,745,152
Dropout_1 (Dropout) (None, 128) 0

details
Resnet50 (Functional) (None, 2048) 89,623,587,712
Dropout (Dropout) (None, 2048) 0
Dense (Dense) (None, 128) 262,272
Dropout (Dropout) (None, 128) 0
goal of this dataset is to have a global class type that can match any image obtained
using visual bands remote sensing (Fig. 5).
For the third approach, we have used the pre-trained model of ResNet50 (trained
using ImageNet dataset). The retained model can be downloaded at this link: https://
www.kaggle.com/keras/resnet50
Fig. 4 Approach 3 model’s

layers
Fig. 5 Sample images from UC merced land dataset

Fig. 6 Sample images from the approach 1 dataset
3.5 Results
To avoid overfitting, we have split the data into two parts, namely the training set
and the validation set. We watch the results based on validation accuracy that we
obtained based on the validation set to evaluate the model’s performance. We also
watch the training time and epochs which can be considered as batches or sets that
are fed one by one, sequentially for training to avoid saturation of the memory.
3.5.1 Approach 1
See Table 4; Fig. 9.
3.5.2 Approach 2
See Table 5; Fig. 10.
3.5.3 Approach 3
With the third approach, we have obtained the result of training in 100 s with the
same environment and a validation accuracy of 63%.
Fig. 7 Sample images from

the approach 2 dataset
3.6 Discussions
Both the approaches one and two have a maximum accuracy with 50 epochs: 70%
for the fst approach and 75% for the second approach which means that 50 epochs
are enough to train the model.
The results show that the second approach performs better than the first approach
in both accuracy and time which prove that using a maximum of two classes is better
than using multiple classes.
The accuracy of 75% seems to be low compared to the accuracy in other researches
and studies, but the goal of this research is not to achieve the best possible accuracy
but just to prove that we can use deep learning and remote sensing images to detect
human activities in wildlands. With 75% accuracy applied to large areas that have
multiple images, we can effectively detect human activities.
With the third approach, the training time is faster than the two previous
approaches, but the accuracy is inferior. We can explain the low accuracy because
Fig. 8 Sample images from the approach 3 dataset
Table 4 Approach 1 results details based on training validation accuracy and training time
Epoch 5 10 20 50 100
Time (s) 123.77 244.19 489.50 1200.34 2406.10
Acc 0.48 0.60 0.54 0.70 0.52
70%
60%
54%
52%
48%
5 10 20 50 100
Fig. 9 Approach 1 validation accuracy by number of epochs
Table 5 Approach 2 results details based on training validation accuracy and training time
Epoch 5 10 20 50 100
Time (s) 50.86 103.16 201.76 495.31 992.35
Acc (%) 60 65 55 75 75
75%
75%
65%
60%
55%
5 10 20 50 100
Fig. 10 Approach 2 validation accuracy by number of epochs

the ResNet50 pre-trained layers are trained using the ImageNet dataset. Which is a
general dataset not spatialized in remote sensing. Better results might be produced
with a larger remote sensing dataset. The three approaches have proven that we can
use the DL methods to watch human activities in wildlands because all results have
a test accuracy that is over 63%.
4 Conclusion
We have proved that CNN can be used to classify wildland with human activities.
So while there is still a lot of work to implement the full ideas, we can be optimistic
nonetheless when we think of the results of the work. The idea may seem very
intuitive, so there is a big probability that other researchers are working on the same
problem even if we cannot find any work that used the same ideas in the field of our
research.
The detection of human activities in wildland may be used for other purposes if
we can obtain a better accuracy to watch wildlands. But for our purpose, more than
70% is enough to detect human activities in wildland because the same areas have a
lot of images and the purpose is to calculate the probabilities of fire occurrence with
the help of weather data and the history of fire.
A larger dataset may be introduced in future works with better DL models to
increase the accuracy and let our research be more efficient. The hope is to help the
professionals better do their jobs and in doing so help reduce wildfires caused by
humans by increasing the efficiency of the monitoring.
References
1. Jain, P., Coogan, S.C.P., Subramanian, S.G., Crowley, M., Taylor, S., Flannigan, M.D.: A review
of machine learning applications in wildfire science and management (2020)
2. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document
recognition. Proc. IEEE (1998)
3. Gomez-Chova, L., Tuia, D., Moser, G., Camps-Valls, G.: Multimodal classification of remote
sensing images: a review and future directions. Proc. IEEE 103(9), 1560–1584 (2015)
4. Kadhim, M.A., Abed, M.H.: Convolutional neural network for satellite image classification.
In: Studies in Computational Intelligence, vol. 830, Issue January. Springer International
Publishing (2020)
5. Varshney, D.: Convolutional Neural Networks to Detect Clouds and Snow in Optical Images
(2019). http://library.itc.utwente.nl/papers_2019/msc/gfm/varshney.pdf
6. Long, Y., Gong, Y., Xiao, Z., Liu, Q.: Accurate object localization in remote sensing images
based on convolutional neural networks. IEEE Trans. Geosci. Remote Sens. 55(5), 2486–2498
(2017)
7. Wang, Y., Wang, X., Jian, J.: Remote sensing landslide recognition based on convolutional
neural network. Mathematical Problems in Engineering (2019)
8. Wang, J., Zhong, Y., Zheng, Z., Ma, A., Zhang, L.: RSNet: the search for remote sensing deep
neural networks in recognition tasks. IEEE Trans. Geosci. Remote Sens. (2020)
9. Yang, Y., Newsam, S.: Bag-of-visual-words and spatial extensions for land-use classifica-
tion. In: GIS: Proceedings of the ACM International Symposium on Advances in Geographic
Information Systems (2010)
10. Waghmare, B., Suryawanshi, M.: A review- remote sensing. Int. J. Eng. Res. Appl. 07(06),
52–54 (2017)
11. Li, T., Shen, H., Yuan, Q., Zhang, L.: Deep learning for ground-level PM2.5 prediction from
satellite remote sensing data. In: International Geoscience and Remote Sensing Symposium
(IGARSS), 2018-July (November), 7581–7584 (2018)
12. Tondewad, M.P.S., Dale, M.M.P.: Remote sensing image registration methodology: review and
discussion. Procedia Comput. Sci. 171, 2390–2399 (2020)
13. Xu, C., Zhao, B.: Satellite image spoofing: Creating remote sensing dataset with generative
adversarial networks. Leibniz Int. Proc. Inf. LIPIcs 114(67), 1–6 (2018)
14. Zhang, L., Xia, G. S., Wu, T., Lin, L., Tai, X.C.: Deep learning for remote sensing image
understanding. J. Sens. 2016 (2015)
15. Rodríguez-Puerta, F., Alonso Ponce, R., Pérez-Rodríguez, F., Águeda, B., Martín-García,
S., Martínez-Rodrigo, R., Lizarralde, I.: Comparison of machine learning algorithms for
wildland-urban interface fuelbreak planning integrating ALS and UAV-Borne LiDAR data
and multispectral images. Drones 4(2), 21 (2020)
16. Li, Y., Zhang, H., Xue, X., Jiang, Y., Shen, Q.: Deep learning for remote sensing image
classification: a survey. Wiley Interdisc. Rev. Data Min. Knowl. Discov. 8(6), 1–17 (2018)
17. Khelifi, L., Mignotte, M.: Deep learning for change detection in remote sensing images:
comprehensive review and meta-analysis. IEEE Access 8(Cd), 126385–126400 (2020)
18. Alshehhi, R., Marpu, P.R., Woon, W.L., Mura, M.D.: Simultaneous extraction of roads
and buildings in remote sensing imagery with convolutional neural networks. ISPRS J.
Photogramm. Remote. Sens. 130(April), 139–149 (2017)
19. de Lima, R.P., Marfurt, K.: Convolutional neural network for remote-sensing scene classifica-
tion: Transfer learning analysis. Remote Sens. 12(1) (2020)
20. Liu, X., Han, F., Ghazali, K.H., Mohamed, I.I., Zhao, Y.: A review of convolutional neural
networks in remote sensing image. In: ACM International Conference Proceeding Series, Part
F1479 (July), 263–267 (2019)
21. Goodfellow, I.: 10—Slides—Sequence Modeling: Recurrent and Recursive Nets (2016). http://
www.deeplearningbook.org/
22. Semlali, B.-E.B., Amrani, C.E., Ortiz, G.: Adopting the Hadoop architecture to process satellite
pollution big data. Int. J. Technol. Eng. Stud. 5(2), 30–39 (2019)
The Evolution of the Traffic Congestion
Prediction and AI Application
Badr-Eddine Soussi Niaimi, Mohammed Bouhorma, and Hassan Zili
Abstract During the past years, there were so many researches focusing on traffic
prediction and ways to resolve future traffic congestion; at the very beginning, the goal
was to build a mechanism capable of predicting the traffic for short-term; meanwhile,
others did focus on the traffic prediction using different perspectives and methods,
in order to obtain better and more precise results. The main aim was to come up with
enhancements to the accuracy and precision of the outcomes and get a longer-term
vision, also build a prediction’s system for the traffic jams and solve them by taking
preventive measures (Bolshinsky and Freidman in Traffic flow forecast survey 2012,
[1]) basing on artificial intelligence decisions with the given predictions. There are
many algorithms; some of them are using statistical physics methods; others use
genetic algorithms… the common goal was to achieve a kind of framework that will
allow us to move forward and backward in time to have a practical and effective traffic
prediction. In addition to moving forward and backward in time, the application of
the new framework allows us to locate future traffic jams (congestions). This paper
reviews the evolution of the existing traffic prediction’s approaches and the edge
given by AI to make the best decisions; we will focus on the model-driven and data-
driven approaches. We start by analyzing all advantages and disadvantages of each
approach to reach our goal in order to pursue the best approaches for the best output
possible.
1 Introduction
Nowadays, we are noticing that our cities are becoming overpopulated very fast,
which leads to a greater number of vehicles as well as a considerable number of
deaths caused by traffic accidents. Therefore, our cities need to become smarter
in order to deal with the risks that come with these evolutions. As a matter of fact,
becoming smarter requires a lot of improvements to be made in the related sectors. In
B.-E. Soussi Niaimi (B) · M. Bouhorma · H. Zili

Faculty of Sciences and Techniques of Tangier, Abdelmalek Essaâdi University, Tangier, Morocco
e-mail: bsoussiniaimi@uae.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_2
20 B.-E. Soussi Niaimi et al.
the hope of reducing the number of incidents and the time/money waste, also having
a better monitoring of our cities’ roads as well as implementing the best preventive
measures to the infrastructure to have the optimal structure possible. Therefore,
building features that allow us to control our infrastructure should be our number
one priority to overcome the dangers we are facing every day on our roads. In
other words, taking our road’s management to the next level, using all that we have
today; technologies, frameworks and sources of data that we can gather. Furthermore,
exploiting the advantages of the traffic congestion prediction algorithms will save
a lot of human lives as well as the time and money, to have a brighter and smarter
future. However, at this moment, to precisely reroute the right amount of traffic can
be developed in the future [2].
Regarding the high speed of the evolution in the transportation sector, the use of
these algorithms became crucial to keep up with the impact that affects our cities,
given that our cities are becoming bigger and more crowded than ever. Moreover,
applying other concepts such as artificial intelligence (AI) and big data… seems to
be an obligation to have an edge in the future, because the traffic jams are causing a
huge time/money loss nowadays.
Moreover, in Morocco, there were more than 3700 deaths and over 130,00 injuries
in 1 year (2017) caused by road accidents (89,375) [3], alongside with the occur-
rence of so many traffic jams over the years in the populated areas. As well as,
during the special occasions (Sport events, holidays …), we cannot help noticing
that the accidents’ counter is increasing rapidly year after another with more than
10% between 2016 and 2017. As we know, many road accidents are caused by the
traffic congestion, the road capacity and management, as well as, the excess speed
of the vehicles while traveling and not respecting the traffic signs and the road’s
marks. We should concentrate our efforts to reduce these accidents, given that traffic
prediction algorithms can prevent future congestion. As a result, we practically will
have the ability to reduce the number of accidents and save lives, time and money.
Furthermore, making traveling through the road is easier, safer and faster.
We will discuss in this paper the different approaches of traffic prediction
approaches. Moreover, how to exploit those results using AI to make enhancements
to the current roads and the new ones. Furthermore, we will shed light on some rele-
vant projects in order to have an accurate overview of the utilities of these predictions
in real life simulated situations. Also, we will answer some questions such as how
can we predict short/long-term traffic dynamics using the real-time inputs? What are
the required tools and algorithms in order to achieve the best traffic management?
The data have a huge impact on the output results, when it comes to transportation
research, the old traffic models are not data driven. As a result, handling modern
traffic data seems to be out of hand; in order to analyze modern traffic data from
multiple sources to cover an enormous network, the data source could be retrieved
from sensors or analyzing the driving behavior and extracting patterns basing on the
trajectory data, as well as, transit schedule and airports, etc.
What do we need? A technology that allows us to teleport in order to save all our
problems, but this is unlikely to happen, also we do not really need it, what we need is
an improvement of what we have, a technology breakthrough to enhance our vehicles
The Evolution of the Traffic Congestion Prediction … 21
and our road’s infrastructure. In the matter of facts, the existing road’s element can
be extremely powerful and efficient with a little adaptation by using mathematics,
information technologies and all the available sources of information. Because nowa-
days, we are living the peak of communication evolution and AI breakout, regarding
all the inventions happening in almost every sector, information became available
with a huge mass, more that we can handle. Therefore, processing those data is the
biggest challenge and extracting the desired information and making decisions is the
ultimate goal to achieve the best travel experience, we do have all the requirements
to move forward and go to the next step, as well as, making the biggest revolution
in traffic management and road’s infrastructure.
2 Method
In contemplation of pointing out the advantages and the weakness of each approach,
we conducted a systematic literature review [4], as we know every existing approach
has its own strengths as well as its limitations or weaknesses; in our detailed review,
we will focus on the strength spots and the possibility to combine multiple approach in
order to find a way to overcome the limitations that come with the existing approaches.
The first goal was to compare the existing approaches to come out with the best
one of them, but after conducting a global review of every approach, we realized
that every single approach is unique in its own way. As a result, the approaches
cannot be compared because they handle a different aspect or area of expertise.
Therefore, to achieve our goal, which is building the ultimate traffic congestion
prediction mechanism, we should combine multiple approaches, but before that, we
have to analyze the weaknesses and the strengths to choose what will work the best
for us.
3 Data-Driven Approach
The data-driven approach is a relatively new approach, because it is a data-based

approach. Therefore, we will shed the light in this part on the most common source
of data used in this approach which are the weather, the traffic intensity, road sensors,
GPS, social media, etc., because of the evolutions happening around the world
regarding the communication tools, sharing data became easier and faster, as well
as, available to everyone around the globe. Thanks to the smartphones and their new
amazing features nowadays, it becomes easier to collect traffic’s information basing
on the publicly shared location and the integrated GPS along with mobile network
that can give us a the approximated coordinate of the phone also the vehicle estimated
location; as a result [5], we can also extract the traveling speed and the accuracy of
the data, and the Wi-Fi also can be used to enhance the accuracy of the previous
methods; in the final analysis, we can say the data are not an issue because of the
various ways that we can use in order collect the desired specimen, but the quality
and the accuracy of these data are crucial to have an accurate output or prediction;
in our case, the first step to handle the incoming data is the storage problem because
of the enormous size of the inputs. Thanks to the appearance of the new storage
technologies with the capacity to handle huge amount of data in an efficient way
(big data), also the evolutions happened to the computer capacities to process a huge
amount of inputs in short duration; because of all the previous breakouts, it was time
for data-driven approach to rise; using this approach, we are capable now of finding
the link between the traffic’s conditions and in incoming information, and we use
this link in order to predict future traffic’s jam and preventing it.
3.1 Real-Time Data Sources
As stated before, the main source of data is the vehicles GPS, road sensors, surveil-
lance system and phone’s location mostly combining GPS, mobile network and Wi-Fi
to enhance the accuracy. As well as, historic data, there are other useful information
that we could use, such as the taxis and bus station and trajectory that can be integrated
to the collected data, to have a wide vision and more accurate results, the timing and
quality of these data are crucial for the outcome of any data-driven approach; in order
to locate every vehicle on the grid in real time, we mostly corporate all the previous
sources of data to enhance the accuracy of the coordinates.
Many methods can be used such static observation using traffic surveillance to
locate all vehicles using sensors and camera, but it requires a lot of data processing
to end up with high-quality results. On the other hand, there is the route observation
using GPS-enabled vehicles and phone’s coordinate on the road, which allow us to
get more information about the targeted vehicle such the current speed and exact
position in real-time without the necessity to process a huge amount of data. The
vehicles’ position is not the only data required as input to the data-driven approach
the weather reports (because the rain only can have a huge impact on the traffic [6]),
the road’s state, the special events, the current date and time (holidays, working days
and hours) have a huge impact on the traffic flow and the vehicle travel’s time.
3.2 Gravity Model
In order to focus on the mobility patterns, the gravity model was created inspired
by Newton’s law of gravitation [7], the gravity model is commonly used in public
transportation management system [8], geography [9], social economics [10] and
telecommunication [11], and the gravity model is defined with the following
equation:
β
xi∝ x j
Ti,J =
f (di, j )’
Originally, Ti, j is the bilateral trade volume between the two locations i and j;
in our case, it is the volume of the people flowbetween the two given locations in
direct proportion of the population size, and f di, j is the function of the distance
di, j between them [12, 13], and the model measures the travel costs between the two
given locations in the urban area.
However, the model considers the travel between the source the destination which
is the same for both directions. In reality, this is far from being accurate. Also, the
model needs in advance to estimate some parameters based on the empirical data
[14, 15].
4 Model-Driven Approach
This approach is mostly used for the long-term traffic prediction. Therefore, the goal
is most likely changing the infrastructure, because we will be modeling months or
years of future traffic regardless of the real-time information [16], and we can use the
long-term predictions in order to simulate events such as conferences, concerts and
football matches. Regarding the last technological evolutions that happened in the
last few years, we are capable of analyzing the current traffic conditions and handling
multiple factors and parameters in order to end up with an accurate prediction (hours
in the future). It is mostly used for road evaluation that helps to determine the best
signals to use, the speed limits, junctions, round bounds, lanes … it is mostly used
throughout simulators to see the behaviors according to the giving model.
There are many simulators that use this model in order to give real time and future
traffic information and conditions; we will shed the light on the following ones:
4.1 VISTA (Visual Interactive System for Transport

Algorithms)
It is a very powerful framework that combines multiple transportations analysis

tools, the framework unified the data interface in order to build a user interface
that could be exploited using many programming languages, with the capacity to
be ran over a network in many OS. The user interface to VISTA functions as a
geographic information system (GIS) with zooming panning, with the possibility to
execute queries to retrieve the desired data. Model can also be used to obtain traffic
estimation and prediction in roads with partially loop detector coverage [17]. It was
developed in 2000 by Ziliaskopoulos and Waller [18].
4.2 Dynamic Network Assignment for Management

of Information to Travelers (DynaMIT)
The project is based on a dynamic traffic assignment (DTA) system 1 for estimation
of network conditions, real-time traffic prediction and generation of drivers’ guidance
developed at MIT’s Intelligent Transportation Systems Laboratory [18].
In order to work properly, the system needs real-time information and offline ones.
The offline information is the representation of the network topology using a set of
links, nodes and loading elements [19]. As well as the traveler’s socioeconomic data
such as gender, age, vehicle ownership, income and the reason for the trip that we can
get using polls and questionnaires. For the real-time information, those are mostly
the road sensors and cameras data, also the traffic control strategies and the incident
properties such as the coordinates, the date, time, the expected results on the traffic
flow and the road capacity. After the integration of the previous data to the system,
the system will be capable of providing prediction, estimation and travel information
[18], as well as flow speed, link density and driver characteristics (travel time, route
choice and departure time) [20]
5 AI Application
Every city across the world is using or planning to use AI in order to build the optimal
traffic management system, a system that is capable of solving complex issues, real
time and future ones, as we can notice that most of the real-time approaches can
solve most of the problems. However, in real life situations and in case of complex
traffic jams, there will be some side-effects, and the same goes with the long-term
traffic prediction because of the unpredictable changes that could cause anomalies
that makes the accuracy of the model questionable. Therefore, we are far away
from having a perfect traffic prediction model. The main goal of having an accurate
prediction is to take conclusive and efficient decisions in any situation, because
each model has its own flaws, as well as its strengths, there is where the role of
artificial intelligence comes in. By harnessing the capability of the AI in hand, we
will be able to give solutions in real-time. As well as, suggestions with the goal to
improve the roads infrastructure to avoid future problems. As a result, achieving
an advanced traffic management system that can also provide to the traveler some
valuable information such as:
The best transportation mode to use.
The best route to take.
The predicted traffic jams.
The best parking spots and its availability.
The services provided along the roads.
This information can be also used to manipulate or even solve congestion and
control the flow in order to have a better travel experience for everyone as well as
avoiding traffic incidents and all the losses that come with it.
If we want an overview of the system, we should have three important levels. The
first level is the data collection (raw data). The second level is the data analysis and
interpretation; at this level, we could use any approach or even combine the model-
driven and data-driven approach to have more accuracy and precision. And the last
level is the decision making and traffic control basing on the previous level output,
improving the final result depends on the quality of every level, from data collection
and analysis of the decision making and the control actions.
In this part, we will focus on the decision making of traffic control actions manage-
ment. Currently, the method used is the signal plan (SP) selection. The SP is selected
after computing offline data in a library; all the SP of the network is a result of an
optimization program executed on a predefined traffic flow structure. The selected
SP is based on a set of values representing traffic flow situations. Therefore, there is
no insurance that the selected set is appropriate for the actual situation. As a result,
our choice will be based on the similarities [21], the selected SP is supposed to be
the best match for the current traffic flow situation. By analyzing the output results
of trying multiple combinations of signal plans in order to choose the most suitable
one.
6 Discussion
The goal of our review is to build a new perspective on the existing approaches of
the traffic congestion prediction. Instead of comparing the strengths and the weak-
nesses, we concluded that combining the strengths is the way to achieve the ultimate
intelligent transport system. A system whose capable of predicting the future traffic
congestions, as well as, the capability of proposing solution and handling real-time
changes with high efficiency. Furthermore, AI usage is a crucial part to accomplish
our goal, while it is challenging to harmonize the data-driven approach to work
along with the model-driven approach. Because the foundation of every approach is
completely different. Hence, the data-driven approach consists of analyzing data’s
masses in real-time to come out with predictions. On the other hand, the model-driven
approach is mostly based on the historical data to propose changes for the current
infrastructure. In order to reach our goal, we must use the model-driven approach to
build an efficient infrastructure and then comes the data-driven approach to handle
the variable factor of the road and the traffic flow. We will also need an AI capable of
making decisions in any critical situation. By combining these approaches, we will
be empowered by an efficient source of information, so we can notify the drivers of
the road’s state, the best route to take, the possible traffic jams and the existing ones,
as well as, the optimal transportation mode and all the existing services provided
along the roads.
7 Related Works
The cognitive traffic management system (CTMS): Basing on the Internet of

Things, the development and deployment of smart traffic lights are possible. The
e-government system proposes smart traffic lights as replacements for the current
traditional ones, by taking advantages proposed by the cloud computing power and
the unlimited scaling opportunity that we have nowadays, the build of the smart traffic
lights is possible using the wireless technologies in order to interconnect different
things without human interaction.
The real-time traffic management system (RTMS): In order to adapt with the
rapid growth with the population of India, and the unbalance between the roads
and the vehicles’ number, the proposed solution is based on real-time monitoring
composed of mobile units. These units work alongside with small networks of road-
side units, the goal is to calculate dynamically the time for each traffic light, and the
data have to be processed in real time in order to update dynamically each traffic
light time accordingly.
8 Proof of Concept
The new system will cover all the bases and provide futuristic features with advanced
capabilities. The first and the most important part is the long-term prediction; in order
to construct the road model, this step allows us to detect the exact points and locations
that required the intervention, which leads up to reduce the cost and the time required
to set up the features. The new system contains much more features than any other
existing or proposed ones; we will shed the light on them in the next paragraph.
The new features allow us to control the traffic flow in the most efficient way, along
with a very high accuracy. Furthermore, the cost of the installations will be at its
minimum comparing to other projects thanks to our road’s model that allows us to
have a deeper understanding of the roads flaws and the possible enhancements.
The new system will include many imbedded elements that allow us to manage the
road and shape-shift the infrastructure depending on the need and to perfectly adapt
to any given situation. Also, the road smart signs and the road marks will be placed
strategically to avoid any possible congestion, the preventive measure will allow us
to meticulously be ready for any traffic flow and control it with high efficiency thanks
to the integrated system, and every part of the system’s units is connected to the same
network to exchange information and roads state. Empowered by AI, the system will
be capable of making critical decisions and solve complex road’s situation with the
optimal way possible.
9 Proposed Architecture
The first and the most important part is the data analysis to extract information about
the future traffic hints. Using the model driving approach, we managed to move
forward in time and predict the future traffic of the selected area. For the first trial,
we moved 24 h in time in order to verify the accuracy of our model. Afterward, we
moved 2 months, then 2 years. The congestion points were consistent in few roads,
which makes the road selection for our smart features easier and more accurate in
matter of result and traffic congestion wise.
The location of each smart sign is selected based on the road’s model, the goal is
to prevent the congestions and redirect the traffic flow when it is needed; each feature
(smart traffic light, smart sign and smart road mark) is a stationary microcomputer
connecter to the main server by a transceiver.
In order to solve traffic congestion problems, we should start by establishing a
new road model, a new architecture for our road, giving the capability to our road
to adapt itself to the traffic flow. We propose a smart road, a shape shifting road, in
other words, a road that can change its properties depending on the situation, and the
challenge is to make it efficient, economize the costs, because building a smart road
occupied with a lot of smart features such as:
• Smart signs: Electrical signs that can be changed by analyzing the traffic flow,
as well as, the short-term prediction in order to avoid any future congestions.
• Smart traffic light: The duration of the traffic light depends on the current traffic
situation [22], but the real goal is to avoid any future traffic jam.
• Smart road marking: The road can change its lines in order to relieve the conges-
tion on the congested direction; the road marking should be based on the historical
data to be set on the road and the real-time data to get prediction thus handling
the real-time changes.
Every part of our system is enhanced with AI capability in order to make decisions
based on the current situation, according to the given predictions (Fig. 1).
Processing the historical data consists of analyzing the traffic road situation for
the past years, and pointing out all the previous traffic jam situations, the causes
(Sport events, holidays, schools, weather…). By processing all those data, we will
be able to locate all the possible congestion points, which are the locations that are
known for congestions with a constant frequency. In order to locate the congestion
points, we used these historical data: Weather reports, emergency calls, police/public
reports, incident reports, traffic sensors, traffic cameras, congestion reports, social
media of interest, transit schedules and status.
After processing all previous sources of information in order to extract patterns,
we were able to locate the congestion points shown below (Fig. 2).
In the previous figure, we can see the roads in red which are the most known for
congestion. Therefore, they have the highest priority for the smart features’ instal-
lation, but it is not the only criteria in order to pick the right, the impact of the new
Fig. 1 Process to set up the road’s model
Fig. 2 Example of historical data processing results for roads

Congestions
20
15
10
Fig. 3 Daily traffic congestion for the selected area
Congesons Regular road

30
25
20
15
10
5
0
Fig. 4 Daily traffic congestions for the same area after setting up the smart features
installation on the other road should be considered as well, the cost of the installations
and efficiency of the road’s features regarding in possible event (Fig. 3).
After running a traffic road simulation on a road in order to observe the congestion
variation during a typical day (without any special events), we can notice that during
periods of time the congestion reached the pick. Those results are obtained before
the addition of the smart features (Fig. 4).
The figure above shows us the same road congestions statistics during the same
day, the blue line displays the congestion after the integration of the smart signs,
traffic light and smart markers, we were able to reduce the traffic congestions by
50.66%, and the percentage can be much higher if we made the surrounding roads
smart as well.
10 Conclusion and Future Research
In this paper, we presented different approaches of predictions. As well, we will be

focusing on the output and the inputs of each method used by those approaches.
Moreover, the impact of the input data in the result regardless of the approach used.
After the data collection from multiple sources, because nowadays, there are so many
real times and historical source of information thanks to the ongoing evolution in
the communication and the technologies. But the accuracy of the output depends
directly on the quality of the input and the data processing and analysis. After the
data collection, then comes the data’s analyzing and processing in order to make a
sense out of the mass of data regarding the road infrastructure and the traffic flow
concerned.
The second step is the decision making using the prediction and the output infor-
mation from the appropriate approach. Using those predictions, we will be able to
take actions in order to prevent any future congestion or potential accidents, more-
over, by doing the aftermath of each action and the consequences in the short and
long term, we will have a clear path ahead. Because some changes should be made to
the road itself, by changing the current infrastructure to make it better and smarter,
in addition of having a better chance to avoid and solve future jams. As well as,
some other decision should be made to solve congestions, with the constraints of the
road infrastructure. In order to move forward in time, to analyze the traffic flow and
choose the best set of decision possible. Furthermore, the real-time decision based
on the real-time data input collected and analyzed on the spot to solve the instant
congestions efficiently.
The traffic prediction system can be used in so many ways, such as changing roads
infrastructure decisions. It can be used even before the construction of a road, with
the goal of having an edge in the future. When it comes to traffic jams, as well as,
solving congestion for the existing roads to avoid accidents and giving the travelers
the best experience possible. Also, these approaches can be applied to reduce time
of an ambulance to reach a certain destination with the most efficient way and the
optimal time and save life as a result, or simply allowing for a regular traveler to
travel safer, faster and more comfortable, and having the best traveling experience
possible.
References
1. Bolshinsky, E., Freidman, R.: Traffic Flow Forecast Survey. Technion—Computer Science
Department, Tech. Rep. (2012)
2. Matthews, S.E.: How Google Tracks Traffic. Connectivist (2013)
3. Ministry of Equipment, Transport, Logistics and Water (Roads Management) of Morocco
(2017)
4. vom Brocke, J., Simons, A., Riemer, K., Niehaves, B., Plattfaut, R., Cleven, A.: Standing on
the Shoulders of Giants: Challenges and Recommendations of Literature Search in Information
Systems Research (2015)
5. Barbosa, H., Barthelemy, M., Ghoshal, G., James, C.R., Lenormand, M., Louail, T., Menezes,
R., Ramasco, J.J., Simini, F., Tomasini, M.: Human mobility: models and applications. Phys.
Rep. 734, 1–74 (2018)
6. Saberi, K.M., Bertini, R.L.: Empirical Analysis of the Effects of Rain on Measured
Freeway Traffic Parameters. Portland State University, Department of Civil and Environmental
Engineering, Portland (2009)
7. Zipf, G.K.: The p1p2/d hypothesis: on the intercity movement of persons. Am. Sociol. Rev.
11(6), 677–686 (1946)
8. Jung, W.S.: Gravity model in the korean highway. 81(4), 48005 (2008)
9. Feynman, R.: The Brownian movement. Feynman Lect. Phys. 1, 41–51 (1964)
10. Matyas, L.: Proper econometric specification of the gravity model. World Econ. 20(3), 363–368
(1997)
11. Kong, X., Xu, Z., Shen, G., Wang, J., Yang, Q., Zhang, B.: Urban traffic congestion estimation
and prediction based on floating car trajectory data. Futur. Gener. Comput. Syst. 61, 97–107
(2016)
12. Anderson, J.E.: The gravity model. Nber Work. Papers 19(3), 979–981 (2011)
13. Barth´elemy, M.: Spatial networks. Phys. Rep. 499(1), 1–101 (2011)
14. Lenormand, M., Bassolas, A., Ramasco, J.J.: Sys- tematic comparison of trip distribution laws
and models. J. Transp. Geogr. 51, 158–169 (2016)
15. Simini, F., Gonz´alez, M.C., Maritan, A., Baraba´si, A.L.: A universal model for mobility and
migration patterns. Nature 484(7392), 96–100 (2012)
16. INRIX.: Who We Are. INRIX Inc. (2014)
17. Lopes, J.: Traffic prediction for unplanned events on highways (2011)
18. Ziliaskopoulos, A.K., Waller, S.: An Internet-based geographic information system that inte-
grates data, models and users for transportation applications. Transp. Res. Part C: Emerg.
Technol. 8(1–6), 427–444 (2000)
19. Ben-akiva, M., Bierlaire, M., Koutsopoulos, H., Mishalani, R.: DynaMIT: a simulation-based
system for traffic prediction. DACCORD Short Term Forecasting Workshop, pp. 1–12 (1998)
20. Milkovits, M., Huang, E., Antoniou, C., Ben-Akiva, M., Lopes, J.A.: DynaMIT 2.0: the
next generation real-time dynamic traffic assignment system. In: 2010 Second International
Conference on Advances in System Simulation, pp. 45–51 (2010)
21. Li, Q., Zheng, Y., Xie, X., Chen, Y., Liu, W., Ma, W.Y.: Mining user similarity based on location
history. In: ACM Sigspatial International Conference on Advances in Geographic Information
Systems, page 34. ACM (2008)
22. Wheatley, M.: Big Data Traffic Jam: Smarter Lights, Happy Drivers. Silicon ANGLE (2013)
Tomato Plant Disease Detection and
Classification Using Convolutional
Neural Network Architectures
Technologies
Djalal Rafik Hammou and Mechab Boubaker
Abstract Agriculture is efficient from an economic and industrial point of view. The
majority of countries are trying to be self-sufficient to be able to feed their people.
But unfortunately, several states are suffering enormously and are unable to join the
standing up to satisfy their populations in sufficient quantities. Despite technological
advances in scientific research and advances in genetics to improve the quality and
quantity of agricultural products, today we find people who die of death. In addition to
famines caused by wars and ethnic conflicts and above all plant diseases that can dev-
astate entire crops and have harmful consequences for agricultural production. With
the advancement of artificial intelligence and vision from computers, solutions have
brought to many problems. Smartphone applications based on deep learning using
convolutionary neural network for deep learning can detect and classify plant diseases
according to their types. Thanks to these processes, many farmers have solved their
harvesting problems (plant diseases) and considerably improved their yield and the
quality of the harvest. In our article, we propose to study the plant disease (tomato)
using the PlantVillage [1] database with 18,162 images for 9 diseased classes and
one seine class. The use of CNN architectures DenseNet169 [2] and InceptionV3 [3]
made it possible to detect and classify the various diseases of the tomato plant. We
used transfer learning technology with a batch-size of 32 as well as the RMSprop and
Adam optimizers. We, therefore, opted for a range of 80% for learning and 20% for the
test with a period number of 100. We evaluated our results based on five criteria (num-
ber of parameters, top accuracy, accuracy, top loss, score) with an accuracy of 100%.
D. R. Hammou (B)
Faculty of Exact Science, EEDIS, Department of Computer Sciences, Djillali Liabes University,
BP 89, 22000 Sidi Bel Abbes, Algeria
e-mail: r_hammou@esi.dz
M. Boubaker
Faculty of Exact Science, LSPS, Department of Probability and Statistics, Djillali Liabes
University, BP 89, 22000 Sidi Bel Abbes, Algeria
https://doi.org/10.1007/978-981-16-3637-0_3
34 D. R. Hammou and M. Boubaker
1 Introduction
Plants represent an enormous economic stake in the world of industrialized agricul-

ture. The development of agriculture has enabled many countries to combat famine
and eradication from the face of the earth. Many necessitous states are suffering, and
their people have not eaten anything. Despite technological progress and advanced
scientific research that had increased production and improved yields, there are still
people dying of hunger. Agriculture is a large and rich field. Since the dawn of time,
human beings have produced their food by cultivating the land. Agriculture offers
a very varied diversification for human food such as cereals (wheat, rice, corn, bar-
ley, starch, etc.), fruits (banana, apple, strawberry, kiwi, pear, etc.), and vegetables
(potatoes, tomatoes, carrot, zucchini, onion, etc.).
Plants are the staple of our diet and represent a large field of research. They are
living organisms and consist of complex plant cells. They are part of the eukaryotic
culture. There are several specialties in the field of plants (medicinal plants, botanical
plants, etc.). The plant classification depends on different criteria such as climate,
temperature, size, type of stem, and geographical area. We can find plants of the
polar regions, high mountains, tropical, cold, and hot (classification according to
climate). We cannot determine the exact number of plant varieties, but a scientific
study in 2015 determined that more than 400,000 plant species existed [4]. The weak
point of plants is diseases that can kill them or demolish an entire crop. Among the
efficient diseases that attack vegetables and fruit trees: early blight, anthracnose,
blight, mildew, moniliasis, Mosaic Virus, blossom end necrosis, phytophthora, rust,
and virosis.
In our article, we are interested in tomato plants (vegetables). The name of tomato
comes from the word Solanum Lycopersicum [5], and it is a vegetable in the classi-
fication of the agricultural world. It is part of the family Solanaceae that originally
comes from South America in the northwest (Peru, Ecuador). It was cultivated for
the first time in Mexico. It is famous all over the world, and it has become a staple
in our daily life. It is a specific plant cultivated in the Mediterranean countries such
as Algeria, Morocco, Tunisia, Spain, Italy, etc.
Algeria is among the most efficient countries in the production and export of
tomatoes. During the 2017–2018 agricultural season (see Table 1), the annual pro-
duction reached 2.91 million tons, 1.37 million tons for household consumption, and
1.54 million tons for industrial processing (see Fig. 1) [6].
The structure of our article is represented according to the following plant.
Section 1 gives a general introduction to the importance of agriculture and the culti-
vation of plants. Section 2 describes a literature review on machine learning, and deep
learning techniques in the detection and classification of plant diseases. Section 3 is
devoted to the choices of CNN architectures. Section 4 presents the strategy to follow
for the deployment of CNN deep learning architecture. Section 5 gives a general idea
of the hardware and software tools to be used for the experiments. Section 6 describes
the results obtained from the experiments on the PlantVillage [1] database. Finally,
the last section describes the conclusion and future research perspectives.
Tomato Plant Disease Detection and Classification … 35
Fig. 1 Tomato production in Algeria [6]
2 Related Work
In 2012, Hanssen et al. [5] described the tomato plant in detail from its origin to its
implementation in the Mediterranean region. They explain the different diseases that
tomato production can affect and the solutions adopted to deal with this kind of dis-
ease. In December 2013, Akhtar et al. [7] implemented a three-part method. First, the
segmentation to locate the diseased region of the plant. Then it extracts the segmented
region image to be able to code the features. Then these characteristics are classified
according to the type of disease. They obtained an accuracy of 94.45% by comparing
with the techniques of state of the art (K-nearest neighbor (KNN), Naïve Bayes clas-
sifier, support vector machine (SVM), decision tree classifier (DTC), recurrent neural
networks (RNN)). In December 2015, Kawasaki et al. [8] proposed an innovative
method based on convolutional neural networks (CNN) with a custom architecture.
The experiments were performed on a cucumber image database with a total of 800
images. They used a cross-validation (fourfold ) strategy by classifying the plants
Table 1 The 2017–2018 Algerian annual yield of tomatoes in the different wilayas of the country
concerning industrial production and household consumption [6]
Wilaya Household Wilaya Industrila processing
consumption (tons) (tons)
Biskra 233,000 Skikda 465,000
Mostaganem 133,000 ElTarf 350,000
Tipaza 106,000 Guelma 206,000
Ain Defla 73,000 Ain Defla 168,000
into two classes (diseased cucumber class, seine class). The results gave an average
accuracy of 94.9%. In June 2016, Sladojevic et al. [9] developed a system for identi-
fying plant disease of 13 different types. The method is based on a deep convolution
network with the help of a Caffe framework. The agricultural database used for the
experiments, which contain 4483 images with 15 different classes of fruit. The results
reached an accuracy of 96.30%. In September 2016, Mohanty [10] proposed a system
for classification and recognition of plant disease based on convolutional neuron net-
works. They tested their system on a corpus of images (54,306 images) with two types
of CNN architectures: AlexNet and GoogleNet. They employed learning and testing
strategy of different rates ([80–20%], [60–40%], [50–50%], [40–60%], [20–80%]),
and they obtained a good result with an accuracy of 99.34%. In November 2016,
Nachtigall et al. [11] used a system for detecting and classifying apple plant diseases
using convolutional neural networks. They carried out experiments on a database of
1450 images with 5 different classes. They used AlexNet architecture and achieved
97.30% accuracy. In December 2017, Lu et al. [12] proposed an approach to solving
the problem of plant pathology of plant disease (rice stalk leaf). The CNN archi-
tecture used in the experiments is AlexNet. They used a database of 500 rice stem
plant images with 10 disease classes. Finally, they were able to obtain an accuracy
of 95.48%. In July 2017, Wang et al. [13] submitted an idea of detecting diseases
in apple plants using deep learning technology. They proceeded to use the follow-
ing CNN architectures: VGG16, VGG19, InceptionV3, ResNet50. The operation of
the experiments is done with a rate of 80% for learning and 20% for testing. They
used the technology of transfer learning. The PlantVillage database was used, with an
image count of 2086 for 4 classes of apple plant disease. The best result was obtained
with the VGG16 architecture for an accuracy of 90.40%. In 2018, Rangarajan et al.
[14] proposed a system to improve the property and quantity of tomato production
by trying to detect plant diseases. The system involves using deep and convolutional
neural networks. They experimented with 6 classes of diseased tomatoes and a seine
from the PlanteVillage database (number of images is 13 262). The CNN archi-
tectures deployed for the tests are AlexNet and VGG16, with a result of 97.49%
accuracy. In September 2018, Khandelwal et al. [15] implemented an approach for
classification and visual inspection of the identification of plant diseases in general.
They used a large database (PlanteVillage, which contains 86 198 images) of 57
classes from 25 different cultures with diseased plants and seines. The approach
is based on deep learning technology using CNN architectures (InceptionV3 and
ResNet50). They used transfer learning with different rates for learning and testing
([80–20%], [60–40%], [40–60%], [20–80%]) as well as a batch-size of 25 and 25
epochs. They reached an accuracy of 99.374%. In February 2020, Maeda-Gutiérrez
[16] proposed a method, which consists of using 5 CNN deep learning architectures
(AlexNet, GoogleNet, InceptionV3, ResNet18, ResNet34) for the classification of
tomato plant disease. They carried out a learning rate of 80 and 20% for the test.
They also used the learning transfer with the following hyper-parameters: batch-size
of 32, 30 epochs. They used the PlantVillage database (tomato plant with 9 different
disease classes and one seine class) with 18 160 images. The results are evaluated
based on five criteria (accuracy, precision, sensitivity, specificity, F-score) with an
accuracy of 99.72%.
Aproch proposed: our approach is based on the following points with our modest
contribution from our article:
• First, we will study the methods used in the literature of machine learning and
deep learning for the detection and classification of plant diseases.
• We will use specific and particular convolutional neural network architectures for
this type of problem.
• Next, we will test our approach on a corpus of images.
• We will evaluate our results obtained according to adequate parameters (accuracy,
number of parameters, top accuracy, top loss, score).
• We will establish a comparative table of our approach with those of state of the
art.
• Finally, we will end with a conclusion and research perspectives.
3 CNN Architecture
We have chosen two CNN architectures for our approach:
3.1 DenseNet169
Huang et al. [2] invented the DenseNet architecture based on convolutional neural
networks. The specificity of this architecture is that each layer is connected directly
to all the other layers. The DenseNet contains L (L + 1) 2 direct connection and
is an enhanced version of the ResNet [3] network. The difference between the
two is that the DenseNet architecture contains fewer parameters, and it computes
faster than the ResNet architecture. The DenseNet network has certain advantages:
such as the principle of reuse of features and alleviates the gradient problem. The
DenseNet architecture has been evaluated in benchmark object recognition compe-
titions (CIFAR100, ImageNet, SVHN, CIFAR-10) and achieved significant results
with other architectures in the bibliographic literature. Among the variants of this
architecture is DenseNet169. It has a depth of 169 layers and an input image of 224
× 224 pixels.
3.2 InceptionV3
Over the years, InceptionV3 architecture has emerged as a result of several researchers.
It is built on the basis in the article by Szegedy et al. [17] in 2015. They designed the
inception module (network complex). Its design depends on the depth and width of
the network. For InceptionV3 to emerge, it was necessary to go through InceptionV1
and InceptionV2. Inception V1 (GoogleNet) was developed in the ImageNet visual
recognition competition (ILSVRC14) [18]. GoogleNet is a deep 22-layer network,
and it uses size convolution filters (1×1, 3×3, and 5×5) as input. The trick was to
use a 1×1 size filter before 3×3 and 5×5 because 1×1 convolutions are much less
expensive (computation time) than 5×5 convolutions. InceptionV2 and InceptionV3
created by Szegedy and Vanhoucke [19] in 2016. InceptionV2 has the particularity
of factoring the 5×5 convolution product into two 3×3 convolution products. It has
a significant impact on the computation time. This improvement is important (a 5
× 5 convolution is more costly than a 3 × 3 convolution). InceptionV3 used the
InceptionV2 architecture with upgrades in addition to the RMSProp optimizer, 7×7
convolution factorization, and BatchNorm in the auxiliary classifier. InceptionV3 is
a deep 48-layer architecture with an input image of 299 × 299 pixels.
4 Deployment Strategy for the Deep Learning Architecture
We have adopted a strategy for training and testing CNN architectures. The goal
is to optimize the neural network and avoid the problem of over fighting by using
mathematical methods with the following criteria:
4.1 Data-Collection
The data collection consists of preparing the dataset for the neural network. Consider
the technical characteristics of the CNN architecture. The database must be large
enough for CNN to function correctly. The size of the input image must be compatible
with the input of the neural network. The image database used in our experiments is
PlantVillage [1]. It contains over 18,000 plant images and is the best dataset.
4.2 Transfert Learning
Learning transfer is an optimized mathematical method of machine learning. It is a

method that consists of using pre-trained weights from the ImageNet [20] database
(it is a database that contains over 1.2 million images with over 1000 different object
classes) in another CNN architecture to solve a well-defined problem. The interest
of the process is in using the information and knowledge gained from ImageNet to
feed it back into another network to solve another classification problem.
4.3 Data Augmentation
It is a mathematical technique that increases the dimension of the image database.

The process consists of using mathematical operations (such as rotations (rotate the
image 90◦ , 180◦ , 270◦ ), translations (scaling), changing the image size, decreasing
the clarity of the picture, blur the image with different degrees, change the color
and effect of the image, geometric transformation of the picture, flip the picture
horizontally or vertically, etc.).
4.4 Fine-Tuning
Fine-tuning is a technique for optimizing convolutional neural networks. The prin-

ciple is to modify the last layer of the neural network or to add an intermediate layer
before the output. The goal is to adapt the new classification problem so that it can
have good accuracy.
5 Hardware and Software Tool
To test our approach on CNN architectures, we used the material which is described
in Table 2.
6 Result and Discussion
Dataset:
PlantVillage [1] is a plant image database that contains pictures of healthy and
diseased plants. It is dedicated to agriculture so that they can get an idea of the type of
Table 2 Hardware and software characteristics

Hard and soft Technical characteristics
Processor (CPU) Intel (R) Xeon (R) @ 2.20 GHz.
Graphic card (GPU) GeForce GTX 1080 X, 8Gb
Memory (RAM) 25 GB
Operating system Windows 8, 64 bits
Programming language Pyhton 3.6
Architecture keras 2.3
Fig. 2 Pictures of tomato plants from the PlantVillage database [1]
Table 3 Characteristic of hyper-parameters

Hyper-parameters Values
Batch-size 32
Epochs 100
Optimizer Adam
Learning rate 0.0001
Beta 1 0.9
Beta 2 0.999
Epsilon 1e−08
Number class 10
disease in the plant. It contains 54,309 images of 14 different fruit and vegetable plants
(Strawberry, Tomato, Soybean, Potato, Peach, Apple, Squash, Blueberry, Raspberry,
Pepper, Orange, Corn, Grape, Cherry). The database contains images of 26 diseases
(4 bacterial, 2 virals, 17 fungal, 1 mite, and 2 molds (oomycetes)). There are also
12 species of seine plant images, making a total of 38 classes. A digital camera
type (Sony DSC - Rx100/13 20.2 megapixels) was used to take the photos from the
database at Land Grant University in the USA.
In our article, we are interested in the tomato plant (healthy and sick). The PlantVil-
lage [1] database contains 10 classes of tomato plants (see Fig. 2) with a total of 18,162
images. The different classes of tomatoes are Alternaria solani, Septoria lycoper-
sici, Corynespora cassiicola, Fulvia fulva, Xanthomonas campestris pv. Vesicato-
ria, Phytophthora infestans, Tomato Yello Leaf Curl Virus, Tomato Mosaic Virus,
Tetranychus urticae, healthy (see Table 4).
Concerning the hyper-parameters, they are described in Table 3.
Regarding the dataset partitioning method, we used a rate of 80% for training and
20% for evaluation using the cross-validation method. The strategy for dividing the
dataset is to separate the dataset into three parts (training, validation, test). Since the
Table 4 The different characteristics of tomato plant diseases (PlantVillage database) [1]
Name Nb images Fungi Bacteria Mold Virus Mite Healthy
classe
Tomato 2127 – Xanthomonas – – – –
bacterial campestris
spot pv.
Vesicatoria
Tomato 1000 Alternaria – – – – –
early blight solani
Tomato 1592 – – – – – Healthy
healthy
Tomato late 1910 – – Phytophthora – – –
blight infestans
Tomato leaf 952 Fulvia fulva – – – – –
mold
Tomato 1771 Septoria – – – – –
septoria lycopersici
leaf spot
Tomato 1676 – – – – Tetranychus –
spider urticae
mites
Tomato 1404 Corynespora – – – – –
target spot cassiicola
Tomato 373 – – – Tomato – –
mosaic mosaic
virus virus
Tomato 5357 – – – Tomato – –
yellow leaf yello leaf
curl virus curl virus
Total 18,162 – – – – – –
Fig. 3 Result of the experiments with the DenseNet169 architecture for loss and accuracy
Fig. 4 Result of the experiments with the InceptionV3 architecture for loss and accuracy
Table 5 Comparison table of the results of the experiments the tomatoes database of PlantVillage
of the different CNN architecture
Architecture Parameters Top accuracy Accuracy (%) Top loss Score
(%)
DenseNet169 12,659,530 99.80 100 1.2665 e−07 0.0178
InceptionV3 21,823,274 99.68 100 3.5565 e−05 0.0002
database contains 18 162 images of plants, we took 11,627 for training, 2903 for
validation, and 3632 for testing.
The experiments on the tomato plant image database (PlantVillage) gave good
results. We were able to obtain an accuracy of 100% with the DenseNet169 archi-
tecture (see Fig. 3) and the same thing with the InceptionV3 architecture (see Fig. 4).
Table 5 reflects the comparison the evaluation of the results of the CNN
DenseNet169 and InceptionV3 architecture according to the following points: Num-
ber of parameters, top accuracy, accuracy, top loss, score.
Table 6 represents a comparison of the results we obtained with those of the
literature.
7 Conclusion and Perspectives Research
Computer vision and artificial intelligence technology (deep learning) have solved
a lot of plant disease problems. Thanks to the architectures of convolutional neuron
networks (CNN), detection, and classification have become accessible. Farmers who
use deep learning applications on SmartPhone in remote areas can now detect the
type of plant disease and provide solutions that can improve the productivity of their
crops. The results of our experiments on tomato plant disease reached an accuracy
of 100%. We plan to use autoencoder architecture (such as U-Net) for visualization
Table 6 Comparison chart between different methods

Author Year Nbr of class Nbr of CNN Accuracy
images architecture (%)
Kawasaki et al. [8] 2015 3 800 Customized 94.90
Sladojevic et al. [9] 2016 15 4483 CaffeNet 96.30
Mohanty et al. [10] 2016 38 54,306 AlexNet 99.34
GooglexNet
Nachtigall et al. [11] 2016 5 1450 AlexNet 97.30
Lu et al. [12] 2017 10 500 AlexNet 95.48
Wang et al. [13] 2017 4 2086 VGG16 90.40
VGG19
InceptionV3
ResNet50
Rangarajan et al. [14] 2018 7 13,262 AlexNet 97.49
VGG16
Khandelwal et al. [15] 2018 57 86,198 InceptionV3 99.37
ResNet50
Maeda-Gutiérrez [16] 2020 10 18,160 AlexNet 99.72
GoogleNet
InceptionV3
ResNet18
ResNet34
Our aproch. 2020 10 18,162 DenseNet169 100
InceptionV3
of plant leaves, which can improve the detection and segmentation of the diseased
region and facilitate classification work.
Acknowledgements I sincerely thank Doctor Mechab Boubaker from the University of Djillali
Liabes of Sidi Bel Abbes for encouraging and supporting me throughout this work and also for
supporting me in hard times because it is thanks to him that I was able to do this work.
References
1. Hughes, D., Salathé, M.: An open access repository of images on plant health to enable the
development of mobile disease diagnostics through machine learning and crowdsourcing. arXiv
preprint arXiv:1511.08060 (2015): n. pag
2. Huang, G., Liu, Z., K. Q. Weinberger, and L. van der Maaten, “Densely connected convolutional
networks,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
Jun 2017, pp. 4700–4708
3. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceed-
ings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV,
USA, 26 June–1 July 2016, pp. 770–778 (2016)
4. Bachman, S.: State of the World’s Plants Report. Royal Botanic Gardens, Kew, p. 7/84 (2016)
(ISBN 978-1-84246-628-5)
5. Hanssen, I.M., Lapidot, M.: Major tomato viruses in the Mediterranean basin. In: Loebenstein,
G., Lecoq, H. (eds.) Advances in Virus Research, vol. 84, pp. 31–66. Academic Press, San
Diego (2012)
6. Market developments in Fruit and Vegetables Algeria [https://meys.eu/media/1327/market-
developments-in-fruit-and-vegetables-algeria.pdf], MEYS Emerging Markets Research
7. Akhtar, A., Khanum, A., Khan, S.A., Shaukat, A.: Automated plant disease analysis (APDA):
performance comparison of machine learning techniques. In: Proceedings of the 11th Interna-
tional Conference on Frontiers of Information Technology, pp. 60–65 (2013)
8. Kawasaki, Y., Uga, H., Kagiwada, S., Iyatomi, H.: Basic study of automated diagnosis of
viral plant diseases using convolutional neural networks. In: Advances in Visual Computing:
11th International Symposium, ISVC 2015, Las Vegas, NV, USA, December 14–16, 2015.
Proceedings, Part II, 638–645 (2015)
9. Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D., Stefanovic, D.: Deep neural networks
based recognition of plant diseases by leaf image classification. Comput. Intell, Neurosci (2016)
10. Mohanty, S.P., Hughes, D.P., Salathé, M.: Using deep learning for image-based plant disease
detection. Front. Plant Sci. 7, 1419 (2016)
11. Nachtigall, L.G., Araujo, R.M., Nachtigall, G.R.: Classification of apple tree disorders using
convolutional neural networks. In: Proceedings of the 2016 IEEE 28th International Conference
on Tools with Artificial Intelligence (ICTAI), pp. 472–476. San Jose, CA 6–8 November 2016
12. Lu, Y., Yi, S., Zeng, N., Liu, Y., Zhang, Y.: Identification of rice diseases using deep convolu-
tional neural networks. Neurocomputing 267, 378–384 (2017)
13. Wang, G., Sun, Y., Wang, J.: Automatic image-based plant disease severity estimation using
deep learning. Comput. Intell. Neurosci. 2917536 (2017)
14. Rangarajan, A.K., Purushothaman, R., Ramesh, A.: Tomato crop disease classification using
pre-trained deep learning algorithm. Procedia Comput. Sci. 133, 1040–1047 (2018)
15. Khandelwal, I., Raman, S.: Analysis of transfer and residual learning for detecting plant dis-
eases using images of leaves. Computational Intelligence: Theories. Applications and Future
Directions-Volume II, pp. 295–306. Springer, Singapore (2019)
16. Maeda-Gutiérrez, V., Galván-Tejada, C.E., Zanella-Calzada, L.A., Celaya-Padilla, J.M.,
Galván-Tejada, J.I., Gamboa-Rosales, H., Luna-García, H., Magallanes-Quintanar, R., Guer-
rero Méndez, C.A., Olvera-Olvera, C.A.: Comparison of convolutional neural network archi-
tectures for classification of tomato plant diseases. Appl. Sci. 10, 1245 (2020)
17. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke,
V., Rabinovich, A.: Going deeper with convolutions. In: 2015 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), pp. 1–9 (2015)
18. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recog-
nition. CoRR, vol. abs/1409.1556 (2014)
19. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architec-
ture for computer vision. In: 2016 IEEE Conference on Computer Vision and Pattern Recog-
nition (CVPR), pp. 2818–2826 (2016)
20. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional
neural networks. CACM (2017)
Generative and Autoencoder Models
for Large-Scale Mutivariate
Unsupervised Anomaly Detection
Nabila Ounasser , Maryem Rhanoui , Mounia Mikram ,

and Bouchra El Asri
Abstract Anomaly detection is a major problem that has been well studied in various
fields of research and fields of application. In this paper, we present several meth-
ods that can be built on existing deep learning solutions for unsupervised anomaly
detection, so that outliers can be separated from normal data in an efficient manner.
We focus on approaches that use generative adversarial networks (GAN) and autoen-
coders for anomaly detection. By using these deep anomaly detection techniques,
we can overcome the problem that we need to have a large-scale anomaly data in the
learning phase of a detection system. So, we compared various methods of machine
based and deep learning anomaly detection with its application in various fields.
This article used seven available datasets. We report the results on anomaly detec-
tion datasets, using performance metrics, and discuss their performance on finding
clustered and low density anomalies.
1 Introduction
Anomaly detection is an important and classic topic of artificial intelligence that has
been used in a wide range of applications. It consists of determining normal and
abnormal values when the datasets converge to one-class (normal) due to insuffi-
cient sample size of the other class (abnormal). Models are typically based on large
amounts of labeled data to automate detection. Insufficient labeled data and high
labeling effort limit the power of these approaches.
N. Ounasser (B) · M. Rhanoui · B. E. Asri

IMS Team, ADMIR Laboratory, Rabat IT Center ENSIAS, Mohammed V University in Rabat,
Rabat, Morocco
M. Rhanoui · M. Mikram
Meridian Team, LYRICA Laboratory, School of Information Sciences, Rabat, Morocco
M. Mikram
Faculty of Sciences, LRIT Laboratory, Rabat IT Center, Mohammed V University in Rabat,
Rabat, Morocco
https://doi.org/10.1007/978-981-16-3637-0_4
46 N. Ounasser et al.
While it is a problem widely studied in various communities including data min-

ing, machine learning, computer vision, and statistics, there are still some challenges
that require advanced approaches. In recent years, deep learning enabled anomaly
detection has emerged as a critical direction toward addressing these challenges.
Generative models [12] are used in various domains such as Person Identification
[10], Image Synthesis [8], Image Generation (WGAN) [1], Face Aging [4, 17], etc.
Autoencoders are a very interesting group of neural network architectures with
many applications in computer vision, natural language processing, and other fields.
Applications of autoencoders include also compression, recommender systems, and
anomaly detection.
We aim to provide a comparative study of the research on deep anomaly detec-
tion. We have grouped existing techniques into different categories based on the
underlying approach adopted by each technique. For each category, we have identi-
fied key assumptions, which are used by the techniques to separate between normal
and anomalous behavior. When we apply a technique to a particular domain, these
assumptions can be used as guidelines to assess the effectiveness of the technique
in that domain. Further, we identify the advantages and disadvantages of the tech-
niques. Then, we report results, using performance metrics, and we discuss their
performance to decide who is the most efficient in anomalies detection.
2 Background and Context: Anomaly Detection
An anomaly is an observation which deviates so much from other observations as

to arouse suspicion, it was generated by a different mechanism [13]. Anomalies are
also referred to as abnormalities, deviants, or outliers in the data mining and statistics
literature. Anomaly detection is the problem of determining if a given point lies in a
low-density region [20].
Anomalies or outliers are extreme values that differ from other observations on
data. They may be due to variability in a measure, experimental errors or novelty.
Thus, the detection of anomalies highlights the importance of the diversity of the
fields it covers and the advantageous results it brings. In particular, anomaly detection
is used to identify fraudulent banking transactions. In this context, companies in the
banking sector, for example, try to identify abnormal customer behavior, detect fake
cards, etc. In addition, anomaly detection also applies in the detection of network
intrusions. In fact, cyber attacks are currently on the rise. These attacks mainly target
information theft and system malfunctions. Generally, the detection of these attacks
can be achieved through the control and monitoring of atypical behaviors of all
information system entities. In addition, web-based anomaly detection applications
are used to detect malicious and malicious users, including spammers and scammers
who publish false news, such as false recommendations in e-commerce sites such as
Amazon, Wish, etc.
Generative and Autoencoder Models for Large-Scale … 47
3 Anomaly Detection Techniques
This section illustrates the different anomaly detection techniques (supervised, unsu-
pervised, semi-supervised), with a focus on unsupervised detection. This approach
is the most flexible method that does not require any labeled data. Usually super-
vised models allow labeled data, which is not always available, hence the use of
unsupervised models (Fig. 1).
3.1 Unsupervised Machine Learning for Anomaly Detection
Anomalies detection is the process of identifying outliers. It is based on the assump-

tion that the behavior of the intruder that generates an anomaly is significantly dif-
ferent from normal or legitimate behavior.
Unsupervised anomaly detection includes several approaches, and we can cate-
gorize these approaches as:
Linear Models A linear model is specified as a linear combination of features. Based
on the training data, the learning process calculates a weight for each entity to train
a model that can predict or estimate the target value. We will explain two algorithms
that are part of this approach; PCA and OC-SVM.
Kernel PCA [21] is an anomaly detection method based on kernel PCA [25] and
reconstruction error. The method consists of assigning scores to the points based on
their projection in the space generated by the core PCA. The greater the reconstruction
error of these projections, the more likely the points are to be anomalies.
Fig. 1 Anomaly detection techniques

OC-SVM: One-Class Support Vector Machines Schoelkopf et al. [24] present an

anomaly detection method based on SVMs, particularly the one-class SVM. This
method estimates the support vertor SV of a distribution by identifying the regions
in the input space where most of the cases occur. In this purpose, data are projected
nonlinearly in a feature space and separated of their origin by a margin as wide as
possible. All data points outside this range are considered anomalies.
Proximity Proximity models observe the spatial proximity of each object in the
data space. If the proximity of an object differs significantly from the proximity of
other objects, it is considered an anomaly. For this approach, we will look at three
algorithms: LOF, K-NN, and HBOS.
LOF: Local Outlier Factor Breunig et al. [5] propose LOF, an anomaly detection
algorithm. LOF is the most widely known algorithm for detecting local anomalies
and has introduced the concept of local anomalies.
The LOF score is therefore essentially a ratio of local density. This means that
normal instances, whose density is as high as the density of their neighbors, obtain a
score of about 1.0. The anomalies, which have a low local density, will have a higher
score. At this point, we also see why this algorithm is local: it is based only on its
direct neighborhood, and the score is a ratio mainly based on the k neighbors only.
It is important to note that, in anomaly detection tasks, when local anomalies are
not of interest, this algorithm can generate many false alarms.
K-NN: k-Nearest Neighbors The K-NN method can be summarized as follows:
1. For each record in the dataset, the k-nearest neighbors must be selected
2. An anomaly score is calculated using these k neighbors and this according to two
possibilities: either the distance to a single K-th nearest neighbor or the average
distance of all k-th nearest neighbors.
In Ramaswamy et al. [22], a new formula for scoring distance-based anomalies is

proposed. This scoring is based on the distance from a point in its nearest neighbor.
Each point is then ranked according to its distance from its nearest neighbor. Finally,
the top % points in this ranking are declared as outliers, therefore anomalies.
HBOS: Histogram-based Outlier Score The outlier value score based on the HBOS
histogram is a simple statistical anomaly detection algorithm that assumes the inde-
pendence of the variables.
Goldstein et al. [11] present a histogram-based anomaly detection (HBOS)
method, which models densities of univariate variables (features) using histograms
with a fixed or dynamic bin width. Thereafter, all histograms are used to calculate
an anomaly score for each instance of data.
• HBOS performs well on global anomaly detection problems but cannot detect
local anomalies.
• HBOS is faster for larger datasets.
Nevertheless, HBOS appears to be less effective on problems of local outliers.

Outlier Ensembles and Combination Frameworks Outlier ensembles and com-
bination frameworks consist of combining the results of different models in order
to create a more robust model. We will detail in this subsection; isolation forest and
feature bagging.
Isolation Forest Isolation forest [16] explicitly identifies anomalies instead of profil-
ing normal data points. Isolation forest, like any other method of tree “Set”, is built
on the basis of decision trees. In these trees, partitions are created by first randomly
selecting an element, then selecting a random split value between the minimum and
maximum values of the selected function.
As with other anomaly detection methods, an anomaly score is required to make
decisions.
Feature Bagging Feature bagging is a method that consists of using several learning
algorithms to achieve the best predictive performance that could come from any
learning algorithm used alone. Lazarevic et al. [14], through tests on synthetic and
real datasets, have found that the combination of several methods gives better results
than each algorithm used separately, and this on datasets with: different degrees of
contamination, different sizes, different dimensions, benefiting from different output
combinations and the diversity of individual predictions.
Clustering Clustering models classify data into different clusters and count points
that are not part of any of the clusters known as outliers. We mention here: K-means
and DBSCAN.
K-means Syarif et al. [26] present a benchmark between the k-means algorithm, as
well as three other variants (improved k-means, k-medoids, EM clustering). K-means
is a clustering method used for the automatic detection of similar data instances. K-
means starts by randomly defining k centroids.
• Methods based on the K-means algorithm are relatively fast

• On the other hand, they have a high rate of FP (False Positive)
DBSCAN: density-based spatial clustering of applications with noise DBSCAN [9],?

is a clustering algorithm that detects clusters of arbitrary shapes and sizes, relying on a
notion of cluster density: clusters are high-density regions in space, separated by areas
of low density. It is not necessary to specify parameters that are generally difficult to
define a priori, such as the number of k clusters, unlike K-means. It addresses different
clustering and correlation analysis methods for unsupervised anomaly detection. For
the DBSCAN algorithm, the following points were discussed :
• DBSCAN provides better results in small spaces, because high-dimensional spaces

are usually rare, making it difficult to distinguish between high and low density.
• DBSCAN has certain parameters that limit its performance; in particular, two
parameters that define the notion of cluster density: the minimum number of mod-
els that define a cluster and the maximum neighboring distance within models
(Table 1).
Table 1 Models synthesis

Approach Model Strengths Weaknesses
Linear Models PCA Suitable for large data, Can not model
sensitive to noise complex data
distributions
OC-SVM Do not make When clusters become
assumptions about more complex,
data distribution can performance decreases
characterize a complex
boundary
Proximity models LOF Easy to use (only one Based in the
parameter k) calculation only on the
nearest neighbors
K-NN Simple and intuitive As the dataset grows,
memory-based adapts the efficiency and
to new training data speed of the algorithm
declines poor
performance in
unbalanced data
HBOS Faster than clustering Less efficient on local
and nearest neighbors outlier problems
models
Outlier ensembles Isolation forest Efficient for large and Does not handle
high dimensionality categorical data
dataset slow temporal
complexity
Feature bagging Valid for different Sensitive to the size of
degrees of sampled datasets
contamination, sizes
and dimensions
Clustering models K-means Easy to implement, Requires numerical
fast data requires number
of cluster k when
clusters become more
complex, performance
decreases
DBSCAN No need to set cluster Some parameters limit
number k performance
3.2 Unsupervised Deep Learning for Anomaly Detection
Autoencoder Anomaly Detection Autoencoders are deep neural networks used to

reproduce the input at the output layer; i.e., the number of neurons in the output layer
is exactly the same as the number of neurons in the layer entry.
The architecture of the autoencoders may vary depending on the network applied
(LSTM, CNN, etc.). A deep autoencoder is made up of two symmetrical deep arrays
Fig. 2 Autoencoder: loss

function
used to reproduce the input at the output layer. One network takes care of the encoding
of the network and the second of decoding (Fig. 2).
Deep Autoencoding Gaussian Mixture Model (DAGMM) proposed by Zong et al. [27]
is a deep learning framework that addresses the challenges of unsupervised anomaly
detection from several aspects. This paper is based on a critique of existing methods
based on deep autoencoding. First of all, the authors state the weakness of compres-
sion networks in anomaly detection, as it is difficult to make significant modifications
to the well-trained deep autoencoder to facilitate subsequent density estimation tasks.
Second, they find that anomaly detection performance can be improved by relying
on the mutual work of compression and estimation networks. First, with the regular-
ization introduced by the estimation network, deep autoencoder in the compression
network learned by the end-to-end training can reduce the reconstruction error as
low as the error of its pre-processed counterpart. This can be achieved only by per-
forming end-to-end training with deep autoencoding. Second, with the well learned
low-dimensional representations of the compression network, the estimation network
is capable of making significant density estimates.
Chen et al. [6] For unsupervised anomaly detection tasks, the GMAA is a model
that aims to jointly optimize dimensionality reduction and density estimation. In this
paper, the authors’ attention was focused on the subject of confidentiality. In this new
approach which aims at improving model performance, we aggregate the parameters
of the local training phase on clients to obtain knowledge from more private data. In
this way, confidentiality is properly protected. This work is inspired by the work we
discussed before. Therefore, this paper presents a federated deep autocoded Gaussian
federated mixture model (DAGMM) to improve the performance of DAGMM caused
by a limited amount of data.
Matsumoto et al. [19] This paper presents a detection method of chronic gastritis
(an anomaly in the medical field) from gastric radiographic images. Among the
constraints mentioned in this article and that traditional methods of anomaly detection
cannot overcome is the distribution of normal and abnormal data in the dataset.
The number of non-gastritis images is much higher than the number of gastritis
images. To cope with this problem, the authors of this article propose the DAGMM
as a new approach to detect chronic gastritis with high accuracy. DAGMM allows
also the detection of chronic gastritis using images other than gastritis. Moreover,
as mentioned above, the DAGMM differs from other models by the simultaneous
learning of dimensionality reduction and density estimation.
Fig. 3 GANs: generator + discriminator
GAN-Based Anomaly Detection In addition to the different approaches mentioned,

mainly in the machine learning domain, there are also other anomaly detection meth-
ods that prefer the use of neural networks, with deep learning, and in particular the
GAN model.
Generative adversarial networks (GAN) is a powerful member of the neural net-
work family. It is used for unsupervised deep learning. It is made up of two competing
models, a generator and a discriminator. The generator takes care of creating the real-
istic synthetic samples from the noise, the z-latent space, and the discriminator is
designed to distinguish between a real sample and a synthetic sample (Fig. 3).
AnoGAN proposed by Schlegl et al. [23]is the firstly proposed method using GAN
for anomaly detection. It is a deep convolutional generative adversarial network
(DCGAN) that has been trained in synthetics data then it can detect anomalies in
new images. The objective of AnoGAN is the standard GAN exploitation, formed
on positive samples, to learn a mapping of the representation of the latent space z to
the real sample G (z), and the goal is to use this representation learned to map new
samples to latent space.
BiGANs Donahue et al. [7] present BiGAN that extends the classic GAN architecture
by adding a third component: The encoder, which learns to map from data space x to
latent space z. The objective of the generator remains the same, while the objective
of the discriminator is modified to classify between a real sample and a synthetic
sample and in addition between a real coding, i.e., given by the coder, and a synthetic
coding, i.e., a sample of the latent space z.
DOPING The introduction of the generative antagonist network (GAN) allowed the
generation of realistic synthetic samples, which were used to expand the training sets.
In this paper [15], they focused on unsupervised anomaly detection and proposed a
new generative data augmentation framework optimized for this task. Using a GAN
variant known as the contradictory auto-analyzer (CAA), they imposed a distribution
on the latent space of the dataset and systematically sample the latent space to generate
artificial samples. This method is the first data augmentation technique focused on
improving the performance of unsupervised anomaly detection.
GANomaly Ackay et al. [2, 3] introduce an anomaly detection model, GANomaly,
including a conditional generative contradictory network that “jointly learns the
generation of a high-dimensional image space and the inference of latent space.”
The GANomaly model is different from AnoGAN and BiGANs because it compares
the encoding of images in latent space rather than the distribution of images. The
generator network in this model uses sub-networks encoder-decoder-encoder.
GAAL Liu et al. [18] present a new model that brings together GAN and active
learning strategy. The aim is to train the generator G to generate anomalies that will
serve as an input to the discriminator D, together with the real data to train him to
differentiate between normal data and anomalies in an unsupervised context.
4 Datasets Description and Performance Evaluation
4.1 Datasets
Experimentally, we perform experiments on synthetic and real datasets. Several

aspects are taken into consideration in the choice of these datasets. First, the nature
of the unlabeled data. In addition, the influx of data from a multitude of sources
requires the development of a proactive approach that takes into account the volume,
variety, and velocity of the data.
The choice of method and model depends completely on the context, the intended
objective, the available data, and their properties. In this study, we investigate the
problem of anomaly detection in its global sense. We are going to try to process diver-
sified datasets, each dataset represents a specific domain. Our study therefore con-
cerns all data and information production sectors, regardless of the type of anomaly:
failure, defect, fraud, intrusion ...
We used for this study anomaly detection several available datasets. These datasets
used are all labeled, but the labels is only be used in the evaluation phase, where
comparison measures between the predicted labels and the real labels is applied
(Table 2).
Table 2 Datasets descriptions

Dataset Speciality Size Dimension Contamination
Credit cards Bank fraud, 7074 14 4.8
financial crime
KDDCup99 Intrusion/cyber 4,898,431 41 20
security
SpamBase Intrusion/cyber 4207 57 39.9
security
Waveform Sport 5000 21 2.9
Annthyroid Medicine 7129 21 7.4
WDBC Medicine 367 30 2.7
OneCluster – 1000 2 2
4.2 Results and Discussion
The table below lists the models and the categories to which they belong, the datasets
and their characteristics, i.e., specialty, size, dimension, and contamination rate, and
finally, the measurement metrics chosen for the evaluation of the models (Table 3).
The goal of this study is to be able to detect anomalies using unlabelled datasets.
To do so, we used several methods: detection using machine learning algorithms
(one-class SVM, LOF, isolation forest, and K-means) and deep learning SO-GAAL,
MO-GAAL, and DAGMM approaches.
In this section, we will evaluate these elaborated methods by comparing the per-
formance of several techniques allowing the detection of anomalies.
To evaluate the models, we have used the metrics of AUC, precision, F1 score,
and recall. This combination of measures is widely used in classification cases and
allows a fair comparison and a correct evaluation.
We applied the different algorithms on the seven datasets. From the table above,
several observations can be obtained:
While in general, DAGMM, MO-GAAL, and SO-GAAL demonstrate superior
performance to machine learning methods in terms of F1 score on all datasets. Espe-
cially on KDDCup99, DAGMM achieves a 14 and 10% improvement in F1 score
compared to other methods. OC-SVM, K-means and isolation forest suffer from
poor performance on most datasets. For these machine learning models, the curse
of dimensionality could be the main reason that limits their performance. For LOF,
although it performs reasonably well on many datasets, the deep learning models out-
perform it. For example, at DAGMM, the latent representation and the reconstruction
error are jointly taken into account in the energy modeling.
One of the main axes that affects method performance is the contamination rate, so
contaminated training data negatively affects detection accuracy. In order to achieve
better detection accuracy, it is important to form a model with high quality data, i.e.,
clean or keep the contamination rate as low as possible. When the contamination
rate does not exceed 2% of the mean accuracy, the recall and F1 score decreases
for all methods except GAAL (SO-GAAL and MO-GAAL). During this time, we
observe that the DAGMM is more sensitive to the contamination rate, we notice that
it maintains a good detection accuracy with a high contamination rate.
In addition, the size of the datasets is an essential factor affecting the performance
of the methods. For MO-GAAL, as the number of dimensions increases, superior
results are more easily obtained.
In particular, MO-GAAL is better than SO-GAAL. SO-GAAL does not perform
well on some datasets. It depends if the generator stops the training before falling
into the problem of mode collapse. This demonstrates the need for several generators
with different objectives, which can provide more user-friendly and stable results.
MO-GAAL directly generates informative potential outliers.
In summary, our experimental results show that the GAN and DAGMM models
suggest a promising direction for the detection of anomalies on large and complex
datasets. On the one hand, this is due to the strategy of the GAAL models, which
Table 3 Models evaluation

Model Category Dataset AUC Precision F1 Recall
DAGMM AE WDBC 0.9201 0.932 0.937 0.942
Annthyroid 0.9013 0.9297 0.944 0.936
KddCup 0.9746 0.9683 0.961 0.954
SpamBase 0.9236 0.9370 0.938 0.937
Credit card 0.8777 0.8478 0.765 0.804
Waveform 0.8611 0.8762 0.833 0.820
Onecluster 0.4981 0.5078 0.5078 0.4983
SO-GAAL GAN WDBC 0.981 0.9411 0.9486 0.9561
Annthyroid 0.6269 0.6371 0.6147 0.6248
KddCup99 0.691 0.6804 0.6501 0.6822
SpamBase 0.645 0.6311 0.6389 0.6577
Credit card 0.7066 0.7234 0.7109 0.7278
Waveform 0.8581 0.8302 0.8356 0.8411
Onecluster 0.9984 0.9640 0.9412 0.9715
MO- GAN WDBC 0.9885 0.9714 0.9681 0.9704
GAAL
Annthyroid 0.6972 0.7002 0.7212 0.7326
KddCup99 0.7688 0.7717 0.7703 0.7656
SpamBase 0.6864 0.6745 0.6945 0.6812
Credit card 0.5682 0.5504 0.5324 0.5579
Waveform 0.8526 0.8479 0.8456 0.8681
Onecluster 0.9994 0.9811 0.9808 0.9739
Isolation Classification WDBC 0.0545 1 0.028 0.0545
forest
Annthyroid 0.1495 0.9981 0.0808 0.1495
KddCup99 0.4113 0.9868 0.2598 0.4113
SpamBase 0.5338 0.6679 0.4445 0.5338
Credit card 0.2111 1 0.1180 0.2111
Waveform 0.0581 1 0.0299 0.0581
Onecluster 0.040 1 0.0204 0.040
LOF Density based WDBC 0.9155 0.9939 0.9188 0.9549
Annthyroid 0.8509 0.9311 0.9058 0.9183
KddCup99 0.7657 0.8096 0.9205 0.8615
SpamBase 0.5583 0.5882 0.8815 0.7056
Credit card 0.8398 0.9080 0.9135 0.9107
Waveform 0.8911 0.9790 0.9073 0.9418
Onecluster 0.9155 0.9939 0.9788 0.9549
(continued)
Table 3 (continued)
Model Category Dataset AUC Precision F1 Recall
KMeans Clustering WDBC 0.9046 0.9969 0.9048 0.9486
Annthyroid 0.3494 0.9474 0.3142 0.4719
KddCup99 0.2083 0 0 0
SpamBase 0.460 1 0.0107 0.0212
Credit card 0.1508 0.9036 0.0566 0.1065
Waveform 0.5193 0.9694 0.5214 0.6781
Onecluster 0.4640 0.9805 0.4622 0.6283
One-Class Classification WDBC 0.4741 0.9457 0.4874 0.6433
SVM
Annthyroid 0.5176 0.928 0.5095 0.6615
KddCup99 0.4868 0.7870 0.4822 0.5980
SpamBase 0.3530 0.4534 0.3776 0.4120
Credit card 0.1055 0 0 0
Waveform 0.5086 0.97977 0.9043 0.6659
Onecluster 0.4940 0.9740 0.4969 0.6581
do not require the definition of a scoring threshold to separate normal data from
anomalies, and the architecture of the sub-models, generator G and discriminator D,
which give the possibility to set different parameters in order to obtain the optimal
result: activation function, number of layers and neurons, input and output of each
model, optimizer as well as the number of generators. On the other hand, the end-
to-end learned DAGMM achieves the highest accuracy on public reference datasets
and provides a promising alternative for unsupervised anomaly detection.
Among the constraints we faced, it is in the data collection phase that we do
not find valid databases for anomaly detection. When dealing with datasets that
contain a high contamination rate, we will converge to binary classification instead
of anomaly detection. As discussed, anomaly detection aims to distinguish between
“normal’ and “abnormal” observations. Anomalous observations should be rare, and
this also implies that the dataset should be out of balance. Unlike classification, class
labels are meant to be balanced so that all classes have almost equal importance.
Also, GAN and AE are powerful models that require high performance materials
which are not always available.
5 Conclusion
In this article, we have compared various machine and deep learning methods for
anomaly detection along with its application across various domains. This paper has
used seven available datasets.
In the experimental study, we have tested four machine learning models and three
deep learning models. One of our findings is that, with respect to performance metrics,
DAGMM, SO-GAAL, and MO-GAAL were the best performers. They had demon-
strated superior performance over state-of-the-art techniques on public benchmark
datasets with up to over 10% improvement on the performance metrics and sug-
gests a promising direction for unsupervised anomaly detection on multidimensional
datasets.
Deep learning-based anomaly detection is still active research, and a possible
future work would be to extend and update this article as more sophisticated tech-
niques are proposed.
References
1. Adler, J., Lunz, S.: Banach wasserstein gan. In: Advances in Neural Information Processing
Systems, pp. 6754–6763 (2018)
2. Akçay, S., Abarghouei, A.A., Breckon, T.P.: Ganomaly: semi-supervised anomaly detection
via adversarial training. In: ACCV (2018)
3. Akçay, S., Atapour-Abarghouei, A., Breckon, T.P.: Skip-ganomaly: Skip connected and adver-
sarially trained encoder-decoder anomaly detection (2019). arXiv preprint arXiv:1901.08954
4. Antipov, G., Baccouche, M., Dugelay, J.L.: Face aging with conditional generative adversarial
networks. In: 2017 IEEE International Conference on Image Processing (ICIP), pp. 2089–2093.
IEEE (2017)
5. Breunig, M.M., Kriegel, H.P., Ng, R.T., Sander, J.: Lof: identifying density-based local outliers.
In: ACM Sigmod Record, vol. 29, pp. 93–104. ACM (2000)
6. Chen, Y., Zhang, J., Yeo, C.K.: Network anomaly detection using federated deep autoencoding
gaussian mixture model. In: International Conference on Machine Learning for Networking,
pp. 1–14. Springer (2019)
7. Donahue, J., Krähenbühl, P., Darrell, T.: Adversarial feature learning (2016). arXiv preprint
arXiv:1605.09782
8. Dong, H., Liang, X., Gong, K., Lai, H., Zhu, J., Yin, J.: Soft-gated warping-gan for pose-guided
person image synthesis. In: Advances in Neural Information Processing Systems, pp. 474–484
(2018)
9. Ester, M., Kriegel, H.P., Sander, J., Xu, X., et al.: A density-based algorithm for discovering
clusters in large spatial databases with noise. Kdd. 96, 226–231 (1996)
10. Ge, Y., Li, Z., Zhao, H., Yin, G., Yi, S., Wang, X., et al.: Fd-gan: pose-guided feature distilling
gan for robust person re-identification. In: Advances in Neural Information Processing Systems,
pp. 1222–1233 (2018)
11. Goldstein, M., Dengel, A.: Histogram-based outlier score (hbos): a fast unsupervised anomaly
detection algorithm. In: Poster and Demo Track of the 35th German Conference on Artificial
Intelligence, pp. 59–63 (2012)
12. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville,
A., Bengio, Y.: Generative adversarial nets. In: Advances in Neural Information Processing
Systems, pp. 2672–2680 (2014)
13. Hawkins, D.M.: Identification of Outliers, vol. 11. Springer (1980)
14. Lazarevic, A., Kumar, V.: Feature bagging for outlier detection. In: Proceedings of the Eleventh
ACM SIGKDD International Conference on Knowledge Discovery in Data Mining, pp. 157–
166. ACM (2005)
15. Lim, S.K., Loo, Y., Tran, N.T., Cheung, N.M., Roig, G., Elovici, Y.: Doping: Generative
data augmentation for unsupervised anomaly detection with gan. In: 2018 IEEE International
Conference on Data Mining (ICDM), pp. 1122–1127. IEEE (2018)
16. Liu, F.T., Ting, K.M., Zhou, Z.H.: Isolation-based anomaly detection. ACM Trans. Knowl.
Discov. Data (TKDD) 6(1), 3 (2012)
17. Liu, S., Sun, Y., Zhu, D., Bao, R., Wang, W., Shu, X., Yan, S.: Face aging with contextual
generative adversarial nets. In: Proceedings of the 25th ACM International Conference on
Multimedia, pp. 82–90. ACM (2017)
18. Liu, Y., Li, Z., Zhou, C., Jiang, Y., Sun, J., Wang, M., He, X.: Generative adversarial active
learning for unsupervised outlier detection. IEEE Trans. Knowl, Data Eng (2019)
19. Matsumoto, M., Saito, N., Ogawa, T., Haseyama, M.: Chronic gastritis detection from gas-
tric x-ray images via deep autoencoding gaussian mixture models. In: 2019 IEEE 1st Global
Conference on Life Sciences and Technologies (LifeTech), pp. 231–232. IEEE (2019)
20. Menon, A.K., Williamson, R.C.: A loss framework for calibrated anomaly detection. In: Pro-
ceedings of the 32nd International Conference on Neural Information Processing Systems, pp.
1494–1504. Curran Associates Inc. (2018)
21. Mika, S., Schölkopf, B., Smola, A.J., Müller, K.R., Scholz, M., Rätsch, G.: Kernel pca and
de-noising in feature spaces. In: Advances in Neural Information Processing Systems, pp.
536–542 (1999)
22. Ramaswamy, S., Rastogi, R., Shim, K.: Efficient algorithms for mining outliers from large data
sets. In: ACM Sigmod Record, vol. 29, pp. 427–438. ACM (2000)
23. Schlegl, T., Seeböck, P., Waldstein, S.M., Schmidt-Erfurth, U., Langs, G.: Unsupervised
anomaly detection with generative adversarial networks to guide marker discovery. In: Inter-
national Conference on Information Processing in Medical Imaging, pp. 146–157. Springer
(2017)
24. Schölkopf, B., Platt, J.C., Shawe-Taylor, J., Smola, A.J., Williamson, R.C.: Estimating the
support of a high-dimensional distribution. Neural Comput. 13(7), 1443–1471 (2001)
25. Schölkopf, B., Smola, A., Müller, K.R.: Kernel principal component analysis. In: International
Conference on Artificial Neural Networks, pp. 583–588. Springer (1997)
26. Syarif, I., Prugel-Bennett, A., Wills, G.: Unsupervised clustering approach for network anomaly
detection. In: International Conference on Networked Digital Technologies, pp. 135–145.
Springer (2012)
27. Zong, B., Song, Q., Min, M.R., Cheng, W., Lumezanu, C., Cho, D., Chen, H.: Deep autoencod-
ing gaussian mixture model for unsupervised anomaly detection. In: International Conference
on Learning Representations (2018)
Automatic Spatio-Temporal Deep
Learning-Based Approach for Cardiac
Cine MRI Segmentation
Abderazzak Ammar, Omar Bouattane, and Mohamed Youssfi
Abstract In the present paper, we suggest an automatic spatio-temporal aware, deep

learning-based method for cardiac segmentation from short-axis cine magnetic reso-
nance imaging MRI. This aims to help in automatically quantifying cardiac clinical
indices as an essential step towards cardiovascular diseases diagnosis. Our method
is based on a lightweight Unet variant with the incorporation of a 2D convolutional
long short-term memory (LSTM) recurrent neural network based layer. The 2D con-
volutional LSTM-based layer is a good fit for dealing with the sequential aspect
of cine MRI 3D spatial volumes, by capturing potential correlations between con-
secutive slices along the long-axis. Experiments have been conducted on a dataset
publically available from the ACDC-2017 challenge. The challenge’s segmentation
contest focuses on the evaluation of segmentation performances for three main car-
diac structures: left, right ventricles cavities (LVC and RVC respectively) as well as
left ventricle myocardium (LVM). The suggested segmentation network is fed with
cardiac cine MRI sequences with variable spatial dimensions, leveraging a multi-
scale context. With less overhead on preprocessing and no postprocessing steps, our
model has accomplished near state-of-the-art performances, with an average dice
overlap of 0.914 for the three cardiac structures on the test set, alongside good cor-
relation coefficients and limits of agreement for clinical indices compared to their
ground truth counterparts.
1 Introduction
The World Health Organization (WHO) repeatedly reports its concern about increas-
ing cardiovascular diseases threats counting amongst the leading cause of death glob-
ally [16]. Cardiovascular diseases have attracted the attention of researchers in an
attempt to early identifying heart diseases and predicting cardiac dysfunction. As it
is generally admitted by the cardiologists community, this goes necessarily by quan-
tifying ventricular volumes, masses and ejection fractions (EF) also called clinical
A. Ammar (B) · O. Bouattane · M. Youssfi

ENSET Mohamedia, Hassan II University, Casablanca, Morocco
https://doi.org/10.1007/978-981-16-3637-0_5
60 A. Ammar et al.
parameters or indices. On the other hand, cardiac cine MRI among other modalities
is now recognized as one of the favorite tools for cardiac function analysis. Generally
acquired as 3D spatial volumes evolving over time from diastole to systole then back
to diastole, cine MRI sequences present 2D short-axis images at each slice level
aggregated on the long-axis as a third spatial dimension and the temporal dimen-
sion or frame index. According to the cardiologists, evaluation of cardiac indices
at only two frames: End Diastole and End Systole respectively (ED and ES) are
sufficient for a reliable cardiac analysis. Given the spatial resolutions, calculation
of these volumetric-based parameters could be achieved by first delineating cardiac
chambers cavities and walls boundaries. However, manual delineation by experts of
such contours is tedious and time consuming; this is why a fully automatic cardiac
segmentation is highly sought after.
Over time, researchers have attempted to perform the cardiac segmentation task
by adopting one of two main approaches: image-based methods such as thresholding,
clustering and deformable models with no prior knowledge, or model-based methods
such as statistical shape models, active appearance models and deformable models
with prior knowledge. Recently, with the advances in deep learning techniques, par-
ticularly with convolutional neural networks (CNNs) [9] and fully convolutional net-
works (FCNs) [10], it turned out to be a most promising tools for high performances
in segmentation tasks. CNNs advocate weight sharing and reduced connectivity to
only a restricted receptive field to leverage spatial relationships within images.
Most of the research methods dealing with medical images segmentation relied
on the FCN use, either solely or in combination with other methods. However, many
of them attempted the combination of FCN-based models with recurrent neural net-
works (RNNs) architectures [1, 3, 12] as it is the case for our suggested model.
In the following section, we present the dataset, the related data preparation steps
along with detailed description of our segmentation method. Subsequently, the seg-
mentation metrics, loss functions, training settings and hyperparameters tuning are
presented. Discussions and results presentation along with comparisons with the
state-of-the-art methods are dealt with thereafter. Finally, a conclusion section sum-
marizes the work in this paper and gives indications on future research work to further
enhance accomplished results.
2 Materials and Methods
2.1 Dataset Presentation
The dataset on which our experiments have been conducted has been made publically
available by the Automated Cardiac Diagnosis Challenge ACDC-2017 organizers.
It comprises 150 real clinical exams of different patients evenly divided in five
pathological classes: NOR (normal), MINF (previous myocardial infarction), DCM
(dilated cardiomyopathy), HCM (hypertrophic cardiomyopathy) and RV (abnormal
Automatic Spatio-Temporal Deep Learning-Based … 61
right ventricle). The dataset has been acquired by means of cine MRI short-axis slices
with two MRI scanners of different magnetic strengths (1.5–3.0 T). The cine MRI
short-axis slices go through the long-axis from base (upper slice) to apex (lower slice),
each slice is of 5–8 mm thickness, 5 or 10 mm inter-slice gap and 1.37–1.68 mm2 /px
for spatial resolution [2]. The dataset was divided into two separate subsets: the train
set with 100 cases (20 for each pathological category) and a test set with 50 cases (10
for each pathological category). For each patient, the 3D spatial volumes at the two
crucial instants ED and ES were provided separately. For the training set, images
alongside their respective manually annotated ground truth GT masks, drawn by
two clinical experts, were also provided for training purposes. For the test set, only
cine MRI images were provided while their GTs counterparts were kept private for
evaluation and participant methods ranking purposes.
2.2 Data Preprocessing
From the provided cine MRI sequences of the ACDC-2017 dataset, there are notice-
able differences in both images spatial dimensions and intensity distributions. While
CNN-based classification-oriented applications need to standardize spatial dimen-
sions to a common size, this is not mandatory for FCN architectures such as Unet.
We thus, choose to keep the original dimensions for two main reasons: this offers a
multi-scale context in the learning process and mainly because, as we are planing the
use of an LSTM-based RNN module, we need to proceed by handling one patient
volume at the time where the sequential character makes sense. A small adjustment
though has been carried out on images spatial dimensions which consisted in align-
ing down to the closest multiple of 32px for both height H and width W. On the
other hand, before feeding the segmentation network, images intensities need to be
normalized, we choose to operate on a per slice-based normalization:
prep X i, j − X min
X i, j = (1)
X max − X min
where X i, j is image intensity at pixel (i, j), X min , X max are minimum and maxi-
mum intensities of image X , respectively, given the assumption of independancy
and identical distribution iid of image intensity.
2.3 Data Augmentation
As it is a common practice, to cope with training data scarcity leading to overfitting

models, we resort to the use of a data augmentation technique as a means of regular-
ization. Based on the provided 100 patients data in the training set, we proceeded to
create 100 other virtual patients. This is achieved by small shifts both horizontally
62 A. Ammar et al.
and vertically, small rotations and small zooms on the original training set images.
Input images and their GT counterparts need to be jointly transformed.
2.4 2D Convolutional LSTM Layer
RNNs are basically designed to inherently exhibit a temporal dynamic behavior.

Thus, they are a good fit for learning-based models for sequential data handling.
Early RNNs implementations, also called vanilla RNN, faced the well-known van-
ishing and/or exploding gradient issues; thus, training such architecture-based mod-
els is harder. Long short-term memory (LSTM)-based RNN [6], one of the most
popular implementations, precisely comes with the idea to overcome these issues
by implementing gates controlling the contributions to the cell memory state Ct . As
such, an LSTM unit features the ability to maintain its cell memory state Ct —in a
learning-based way—from relevant contributions of previous observations through-
out sequential inputs, while being able to discard irrelevant informations too. A 2D
convolutional-based extension to the 1D LSTM unit has been suggested by [14] and
allows FCNs-like architectures to benefit from this structure while preserving spa-
tial correlations. Figure 1 shows an 2D convolutional LSTM cell’s architecture with
peephole [4] and the following equations (Eq. 2) summarize its working principle.
i t = σ (Wi x ∗ Xt + Wi h ∗ Ht−1 + Wic Ct−1 + bi )

f t = σ (W f x ∗ Xt + W f h ∗ Ht−1 + W f c Ct−1 + b f )
C˜t = tanh(Wcx ∗ Xt + Wch ∗ Ht−1 + bc )
Ct = f t Ct−1 + i t C˜t (2)
ot = σ (Wox ∗ Xt + Woh ∗ Ht−1 + Woc Ct + bo )

Ht = ot tanh(Ct )
where subscript t denotes the current time step or frame. Wi j , bi learnable convo-
lutional weights between input j and output i before activation and related bias,
respectively. In our application, the sequential aspect is not of a temporal nature but
is rather sought after between consecutive slices along the long-axis. Indeed, the
cardiac structures should presumably show some kind of shape’s variability pattern
along the long-axis, in that they are getting smaller starting from the base towards
the apex while keeping some shape similarity. However, this is not evenly true in
the same way for all structures especially for the RV structure and particularly for
pathological cases.
Fig. 1 2D convolutional LSTM cell with peephole: i t , f t and ot are input, forget and output gates
activations, respectively. Xt , Cˆt , Ct , and Ht are input, cell input, cell memory state and hidden state,
respectively
2.5 Unet-Based Segmentation Network
The segmentation network is a lightweight variant of the well-known Unet archi-

tecture [13]. It is fed with images from the cine MRI sequences at their original
dimensions slightly cropped at the lowest dimension: min (H, W ) aligned down
to 32px. As it is shown in Fig. 2, the Unet-based segmentation network follows an
encoder/decoder pattern. A contracting path (on the left, in Fig. 2) chains up with a
series of encoding blocks, starting from input down to a bottleneck. Each encoding
block comprises two stacked subblocks 3 ×3 kernel convolutional layers, followed
by a dropout DP layer as a means of regularization [5], a batch normalization layer
BN [7] then a rectified linear unit (Relu) activation layer [11]. At each level in
the contracting path, the number of feature maps increases (×2), while the spatial
dimensions decrease (/2) by a prior downsampling (D) block, through 2×2 strided
maxpooling layer. In The expanding path [in the middle of Fig. 2], at each ascend-
ing level, upsampling blocks (U) increase the spatial dimensions (×2) by means
of transposed convolutions or deconvolutions [17], while feature maps number is
reduced by the half. The particularity of the Unet architecture is the reuse of earlier
feature maps of the contracting path at their corresponding level in the expanding path
64 A. Ammar et al.
Fig. 2 Segmentation network architecture
where spatial dimensions match. This is achieved by simple channel-wise concate-

nation operations. As shown on the right of Fig. 2, our proposed architecture is a Unet
variant where the output is constructed by aggregating all the expanding path levels
outputs by upsampling with appropriate projections to perform pixel-wise additions.
The aggregating output path ends with a four outputs softmax layer to predict pixel-
wise class for the background, RVC, LVM and LVC as a raw one-hot encoded four
valued vector. In the training phase, this is enough to guide the learning process; how-
ever in the inference phase, we need to retrieve the one channel mask to compare
against the GT counterpart; this is achieved simply by an argmax operator applied
to the softmax outputs. We choose to introduce the 2D convolutional LSTM layer in
the middle of the aggregating path to keep the overall architecture as lightweight as
possible while keeping a solid enough contribution of the 2D convolutional LSTM
layer in the learning process. It is noteworthy that because of the 2D nature of the
convolutional blocks in the construction of the Unet derived architecture and as this
is fed with temporal sequences of images, these blocks need to be wrapped within
time distributed layers referred to as Time-dist in Fig. 2.
3 Experiments and Model Training
3.1 Segmentation Evaluation Metrics
Let Ca the predicted or automatic contour delineating the object boundary in the image
to segment, Cm its ground truth counterpart and Aa , Am sets of pixels enclosed by
these contours, respectively. In the following, we recall the definitions of two well-
known segmentation evaluation metrics:
Hausdorff Distance (HD) This is a symmetric distance between Ca and Cm :

H(Ca , Cm ) = max max min d(i, j) , max min d(i, j) (3)
i∈Ca j∈Cm j∈Cm i∈Ca
where i and j are pixels of Ca and Cm respectively and d(i, j) the distance between
i and j. Low values of HD indicate both contours are much closer to each other.
Dice Overlap Index Measures the overlap ratio between Aa and Am . Ranging from
0 to 1, high dice values imply a good match:
2 × |Am ∩ Aa |
Dice = (4)
|Am | + |Aa |
3.2 Loss Functions
Supervised learning-based models training is achieved by the minimization of a

suitable loss function. In our suggested method, which is actually acting as a pixel-
wise classification model to achieve a semantic segmentation, we choose a tandem
based on a dice overlap related and a crossentropy-based terms.
66 A. Ammar et al.
Cross Entropy Loss The categorical or multi-class crossentropy loss is defined as:

C

− y(c, x) log ŷ(c, x) (5)
c=1
at the pixel level. C denotes the number of classes, y is the ground truth one-hot
encoded label vector for pixel x and
ea(c,x)
ŷ(c, x) = C (6)
a(i,x)
i=1 e
its estimated softmax score counterpart applied to activation functions a. As the

ground truth is one-hot encoded, only the positive label is retained and the sum of C
terms reduces to only one term. The crossentropy for the whole image sample yields:

Lce = − log ŷ( p, x) p = arg max (y(c, x)) (7)
x∈ c
where is the image spatial domain.

Dice Overlap Based Loss The dice overlap index we recalled the definition above
as a performance metric could also serve for the definition of a loss function. Seen
as a metric, a good segmentation is achieved by maximizing the dice overlap index,
or similarly by minimizing the deducted loss function:
Ldice = − log(dice) (8)
3.3 Total Loss
We choose as a total loss function for training our segmentation network a combi-
nation of the above mentioned individual loss terms (Eqs. 7 and 8) plus an L 2 based
weights decay penalty as a regularization term.
Ltot = αLce + βLdice + γ W 2 (9)
where W represents network weights. We choose to set both the crossentropy and
dice-based loss terms contribution weights to 1, the L 2 based regularization contri-
bution weight γ is adjusted to 2 × 10−4 (see 3.4).
3.4 Training Hyperparameters Tuning
Our experiments have been conducted with the following hyperparameters:

• # of initial filters: N = 16, 2.061 M as total number of learnable parameters.
• Dropout [15]-based regularization probability: 0.1.
• L 2 based regularization: γ = 2 × 10−4 .
• Relu activation function [11].
• Adam optimizer [8], learning rate = 1 × 10−4
• Number of iterations: 100 epochs, with variable batch size depending on the num-
ber of slices in the patient’s volume.
• five fold stratified cross-validation.
Figure 3 shows the evolution, over training epochs, of both total loss and heart overall
dice overlap training curves.
4 Results and Discussions
After training the suggested model in a five fold stratified cross-validation way and
before going in the inference phase on the test set, we first gather validation results
and try to analyze them. It is noteworthy that the achieved segmentation results are
raw predictions, without any postprocessing actions.
4.1 Validation Results
From Fig. 4 it can be seen that

1. Figure 4a, LVC presented the highest dice score among the three structures, ED
frames scores are higher than ES ones. Except for the HCM pathology at ES
frames where there is a substancial dispersion, most of the obtained scores are
less spread and keep a median above 0.95.
2. Figure 4b While RVC dice scores at ED frames are better than ES frames ones
as for the LVC, the distributions are noticeably more spread, especially for ES
frames.
3. Figure 4c At the opposite of the previous two cavities, LVM wall segmentation
tends to report good performances in the ES frames, while globally its scores are
the lowerest among the three structures.
This can be explained by the fact that the LVC component presents the most regular
shape close to a circular one along the long-axis, except for the very early basal
slices. The RVC component however presents the most larger variability in shape
from basal to apical slices. LVM suffers from relying on two boundaries delineation
(endocardium and epicardium), which is responsible for prediction cumulative errors.
68 A. Ammar et al.
Fig. 3 Training and validation curves, in orange (training) and in blue (validation)
Finally, due to the shrinking state of the heart at the end of systole phase ES, the
LVC and RVC structures see their predictions performances decrease, while the LVM
performance is rather increasing as the cumulated errors gets minimized with small
structures.
4.2 Test Results
Our segmentation results on the test set (unseen data), along with clinical indices
are reported in Tables 1, 2 and 3 for LVC, RVC and LVM structures, respectively.
Compared to the top ranking participant methods on the same challenge’s test set, our
method achieved rather good results while being lightweight in that it requires a few
parameters. Highlighted results indicate either first or second rank. This agrees with
Fig. 4 Dice overlap index

validation results
70 A. Ammar et al.
Table 1 Challenge results for LVC structure (on the test set)
LVC
Method DICE HD (mm) EF (%) Vol. ED (ml)
ED ES ED ES Corr. Bias Std. Corr. Bias Std.
Simantiris 2020 0.967 0.928 6.366 7.573 0.993 −0.360 2.689 0.998 2.032 4.611
Isensee 2018 0.967 0.928 5.476 6.921 0.991 0.49 2.965 0.997 1.53 5.736
Zotti 2019 0.964 0.912 6.18 8.386 0.99 −0.476 3.114 0.997 3.746 5.146
Painchaud 2019 0.961 0.911 6.152 8.278 0.99 −0.48 3.17 0.997 3.824 5.215
Ours 0.966 0.928 7.429 8.150 0.993 −0.740 2.689 0.995 −0.030 7.816
Table 2 Challenge results for RVC structure (on the test set)
RVC
Method DICE HD (mm) EF (%) Vol. ED (ml)
Isensee 2018 0.951 0.904 8.205 11.655 0.91 −3.75 5.647 0.992 0.9 8.577
Simantiris 2020 0.936 0.889 13.289 14.367 0.894 −1.292 6.063 0.990 0.906 9.735
Baldeon 2020 0.936 0.884 10.183 12.234 0.899 −2.118 5.711 0.989 3.55 10.024
Zotti 2019 0.934 0.885 11.052 12.65 0.869 −0.872 6.76 0.986 2.372 11.531
Ours 0.924 0.871 10.982 13.465 0.846 −2.770 7.740 0.955 −6.040 20.321
Table 3 Challenge results for LVM structure (on the test set)
LVM
Method DICE HD (mm) Vol. ES (ml) Mass ED (g)
Isensee 2018 0.904 0.923 7.014 7.328 0.988 −1.984 8.335 0.987 −2.547 8.28
Simantiris 2020 0.891 0.904 8.264 9.575 0.983 −2.134 10.113 0.992 −2.904 6.460
Baldeon 2020 0.873 0.895 8.197 8.318 0.988 −1.79 8.575 0.989 −2.1 7.908
Zotti 2019 0.886 0.902 9.586 9.291 0.98 1.16 10.877 0.986 −1.827 8.605
Ours 0.890 0.906 9.321 10.029 0.972 5.420 12.735 0.980 2.080 10.199
the observations on the validation results, in that LVC, RVC and LVM dice overlap
scores ranking is preserved, the same can be said for HD metric. From the same
tables, the clinical indices results: correlation coefficients and limits of agreement
(bias and std) show that the RVC is the structure where the network is less performing.
This is expected as it is the structure which presents the high shape variability along
the long-axis; thus, it is likely that the recurrent LSTM-based convolutional layer
captures less relevant correlations in the related input sequences. An example of a
successful segmentation is shown in Fig. 5.
Fig. 5 Example of a successful volume segmentation from the test set a ED frame, b ES frame.
Showing images from basal (top left) to apical (bottom right) slices for each frame. In overlay are
predicted masks annotations in red, green and blue for LVC, LVM and RVC, respectively
5 Conclusion
In this paper, we suggested an automatic spatio-temporal deep learning-based

approach for cardiac cine MRI segmentation. This has been implemented by incorpo-
rating a convolutional LSTM layer into a lightweight Unet variant to capture potential
correlations between consecutive slices along the long-axis. The suggested model
has been trained, validated and tested on a public dataset provided by the ACDC-
2017 challenge and achieved an average dice overlap score of 0.947, 0.898, 0.899
and 0.914 for LVC, RVC, LVM and overall heart bi-ventricle chambers respectively
on the challenge’s segmentation contest. Given the context of medical images seg-
mentation challenging task, our method has rather achieved good performances and
even outperformed, for some metrics, the state-of-the-art participant methods to the
72 A. Ammar et al.
challenge. Our method could benefit from further postprocessing operations to refine
the obtained predicted masks, seeking for coupling with other established methods,
adding a multi-scale approach to the architecture. These are some of the directions
we will head to, in future work, to extend the suggested model and enhance the
obtained results.
References
1. Alom, M.Z., Hasan, M., Yakopcic, C., Taha, T.M., Asari, V.K.: Recurrent residual convolu-
tional neural network based on U-Net (R2U-Net) for medical image segmentation (2018).
arXiv:1802.06955
2. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir,
K., Camara, O., Gonzalez Ballester, M.A., Sanroma, G., Napel, S., Petersen, S., Tziritas,
G., Grinias, E., Khened, M., Kollerathu, V.A., Krishnamurthi, G., Rohe, M.M., Pennec, X.,
Sermesant, M., Isensee, F., Jager, P., Maier-Hein, K.H., Full, P.M., Wolf, I., Engelhardt, S.,
Baum- gartner, C.F., Koch, L.M., Wolterink, J.M., Isgum, I., Jang, Y., Hong, Y., Patravali,
J., Jain, S., Humbert, O., Jodoin, P.M.: Deep learning techniques for automatic MRI cardiac
multi-structures segmentation and diagnosis: is the problem solved? IEEE Trans Med Imaging
37, 2514–2525 (2018). https://doi.org/10.1109/TMI.2018.2837502
3. Chakravarty, A., Sivaswamy, J.: RACE-Net: a recurrent neural network for biomedi-
cal image segmentation. IEEE J. Biomed. Health Inform. 23, 1151–1162 (2019). doi:
10.1109/JBHI.2018.2852635
4. Gers, F., Schmidhuber, J.: Recurrent nets that time and count. In: Proceedings of the IEEE-
INNS-ENNS International Joint Conference on Neural Networks. IJCNN 2000. Neural Com-
puting: New Challenges and Perspectives for the New Millennium, pp. 189–194, vol.3. IEEE
(2000). https://doi.org/10.1109/IJCNN.2000.861302
5. Hinton, G.E., Srivastava, N., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.R.: Improving neu-
ral networks by preventing co-adaptation of feature detectors, 1–18 (2012). arXiv:1207.0580
6. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9, 1735–1780
(1997). doi: 10.1162/neco.1997.9.8.1735
7. Ioffe, S., Szegedy, C.: Batch normalization: accelerating deep network training by reducing
internal covariate shift. In: 32nd International Conference on Machine Learning, ICML 2015
1, 448–456 (2015). arXiv:1502.03167
8. Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization. In: 3rd International Con-
ference on Learning Representations, ICLR 2015—Conference Track Proceedings abs/1412.6
(2014). arXiv:1412.6980
9. Lecun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature 521, 436–444 (2015). DOI
10.1038/nature14539, arXiv:1807.07987
10. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic
segmentation. IEEE Trans. Pattern Anal. Mach. Intell. 39, 640–651 (2014). doi:
10.1109/TPAMI.2016.2572683
11. Nair, V., Hinton, G.E.: Rectified linear units improve restricted Boltzmann machines. In: ICML
2010—Proceedings, 27th International Conference on Machine Learning, 807–814 (2010).
URL https://icml.cc/Conferences/2010/papers/432.pdf
12. Poudel, R.P.K., Lamata, P., Montana, G.: Recurrent fully convolutional neural networks for
multi-slice MRI cardiac segmentation. In: Lecture Notes in Computer Science (including sub-
series Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), volume
10129 LNCS, pp. 83–94 (2017). DOI 10.1007/978-3-319-52280-7_8
13. Ronneberger, O., Fischer, P., Brox, T.: U-Net: convolutional networks for biomedical image
segmentation 9351, 234–241 (2015). https://doi.org/10.1007/978-3-319-24574-4_28
14. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.C.: Convolutional LSTM
network: a machine learning approach for precipitation nowcasting. In: Advances in Neural
Information Processing Systems, pp. 802–810 (2015). arXiv:1506.04214
15. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.:. Dropout: a simple
way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15, 1929–1958 (2014).
URL http://jmlr.org/papers/v15/srivastava14a.html
16. World Health Organization.: Cardiovascular diseases (CVDs) (2017). URL https://www.who.
int/en/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds)
17. Zeiler, M.D., Krishnan, D., Taylor, G.W., Fergus, R.: Deconvolutional networks. In: Proceed-
ings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition
2528–2535 (2010). https://doi.org/10.1109/CVPR.2010.5539957
Skin Detection Based on Convolutional
Neural Network
Yamina Bordjiba, Chemesse Ennehar Bencheriet, and Zahia Mabrek
Abstract Skin detection is an essential step in many human–machine interaction

systems such as e-learning, security, communication… etc., it consists of extracting
regions containing the skin in a digital image. This problem has become the subject
of considerable research in the scientific community where a variety of approaches
has been proposed in the literature; however, few recent reviews exist. Our principal
goal in this paper is to extract skin regions using a Convolutional neural network
called LeNet5. Our framework is divided into three main parts: At first, a deep
learning is performed to Lenet5 network using 3354 positive examples and 5590
negative examples from SFA dataset, then and after a preprocessing of each arbitrary
image the trained network will classify image pixels into skin/non-skin. Lastly, a
thresholding and prost-processing of classified regions is carried out. The tests were
carried out on images of variable complexity: indoor, outdoor, variable lighting,
simple and complex background. The results obtained are very encouraging, we
show the qualitative and quantitative results obtained on SFA and BAO datasets.
1 Introduction
Skin is one of the most important parts of the human body, so it is logical to consider
it as the main element to be detected in many artificial vision systems operating on
human beings such as medicine for disease detection and recognition, security for
intrusion detection, people identification, facial recognition, gesture analysis, hand
tracking, etc.
Although considered an easy and simple task to be performed by the human, the
recognition of human skin remains an operation of high complexity for the machine
Y. Bordjiba · Z. Mabrek
LabStic Laboratory, 8 Mai 1945- Guelma University, BP 401, Guelma, Algeria
e-mail: bordjiba.yamina@univ-guelma.dz
C. E. Bencheriet (B)
Laig Laboratory, 8 Mai 1945- Guelma University, BP 401, Guelma, Algeria
e-mail: bencheriet.chemesseennehar@univ-guelma.dz
https://doi.org/10.1007/978-981-16-3637-0_6
76 Y. Bordjiba et al.
despite the technological progress of the sensors and processors used, for several
reasons such as lighting and shooting conditions of the captured image, background
variation (indoor/outdoor), skin color variation (different ethnicity), etc.…
The main objective of our work is to design a model with a deep learning archi-
tecture, and to implement a convolutional neural network model for skin detection,
for these we propose an approach based on LeNet 5 network.
Our contribution is divided into three main parts: At first, a deep learning is
performed to Lenet5 network using 3354 positive examples and 5590 negative exam-
ples from SFA dataset, then and after a preprocessing of each arbitrary image the
trained network will classify image pixels into skin/non-skin. Lastly, a thresholding
and prost-processing of classified regions are carried out.
The remainder of this paper is structured as follows: Sect. 2 gives the development
of principal steps of our proposed framework. Section 3 provides the experimental
results using two different datasets and Sect. 4 concludes the paper with discussions
and future research directions.
2 Related Work
Skin detection is a difficult problem and has become the subject of considerable study,
to improve the skin detection process [1], but this requires a high rate of accuracy due
to the noise and complexity of the images. In this context, the research community
is divided into two parts: Conventional research and deep learning-based research
[2]. Conventional methods can be divided into different categories. They can be
based on pixel classification [3, 4] or region segmentation [5, 6], while other studies
have selected a hybrid of two or more methods. Among researches based on region
segmentation, authors of [7] propose a technique purely based on region for skin color
detection, they cluster similarly colored pixels, based on color and spatial distance.
First, they use a basic skin color classifier, then, they extract and classify regions called
superpixel. Finally, a soothing procedure with CRF (Conditional Random Field) is
applied to improve result. This proposed method reaches 91.17% true positive rate
and 13.12% false-positive rate. Authors indicate that skin color detection has to be
based on regions rather than pixels.
Many studies have also investigated the effects of color space selection [8, 9];
they confirm that RGB color space is not the best one for this task. In [10], authors
use Cb-Cr color space and extract Skin regions using the Gaussian skin color model.
The likelihood ratio method is used to create a binary mask. To design skin color
model, they also use a combination of two different databases to encompass larger
skin tones. For performance evaluation, a total of 165 facial images from the Caltech
database were randomly selected; the achieved accuracy is about 95%.
Color spaces have been widely used in skin detection. In [11], the authors present
a comparative study of skin detection in two color spaces HSV and YCbCr. The
detection result is based on the selection of a threshold value. The authors concluded
that HSV-based detection is the most appropriate for simple images with a uniform
Skin Detection Based on Convolutional Neural Network 77
background. However, the YCbCr color space is more effective and efficient to be
applied for complex color images with uneven illumination.
The authors of [12] propose to model skin color pixels with three statistical func-
tions. They also propose a method to eliminate the correlation between skin chromi-
nance information. For this method’s tests, they used the COMPAQ skin data set for
the training and testing stages, with different color spaces. The accuracy achieved
was 88%, which represents, according to the authors, an improvement over previous
statistical method.
Many researchers have used neural networks to detect skin color, and recently deep
learning methods have been widely used and have achieved successful performance
for different problems of classification in computer vision. However, there are few
researches on human skin detection based on deep learning (especially convolutional
neural networks) and they limited their studies to diagnosing skin lesions, disorders
and cancers only [13].
In [13], authors propose a sequential deep model to identify the regions of the
skin appearing on the image. This model is inspired by the VGGNet network, and
contains modifications to treat finer grades of microstructures commonly present in
skin texture. For their experiments, they used two datasets: Skin Texture Dataset and
FSD dataset, and compared their results with conventional texture-based techniques.
Based on the overall accuracy, they claim to obtain superior results.
Kim et al. [14] Realize one of the most interesting work in skin detection using
deep learning, where they propose two networks based on well-known architectures,
one based on VGGNet, and the second based on the Network in Network (NiN)
architecture. For both, they used two training strategies; one based on full image
training, and the other based on patch training. Their experiences have shown that
NiN-based architectures provide generally better performance than VGGNet-based
architectures. They also concluded that full image-based training is more resistant
to illumination and color variations, in contrast to the patch-based method, which
learns the skin texture very well, allowing it to reject skin-colored background when
it has different texture from the skin.
3 Proposed Method
The aim of this work is to propose a new approach to skin detection based on deep
learning. The detection will be done in two steps. the first is a learning phase of the
CNN, once its weights are found, they are used in the second phase which is the
segmentation, based on patches; where the input image has to be pre-processed, then
it is divided into overlapping patches obtained by a sliding window. These patches are
classified as skin or non-skin by the CNN already trained in the first phase. Finally, a
post-processing stage is applied. The global architecture of our skin detection system
is illustrated in Fig. 1.
Fig. 1 Global architecture of our skin detection system
3.1 The Used CNN Architecture
Recently, “convolutional neural networks” (CNNs) have emerged as the most popular
approach to classification and computer vision problem and several convolutional
neural network architectures were proposed in literature. One of the first successful
CNNs was LeNet by LeCun [15], which was used to identify handwritten numbers
on checks at most banks in the United States. Consisting of two convolutional layers,
two maximum grouping layers, and two fully connected layers for classification, it
has about 60,000 parameters, most of which are in the last two layers.
Later, LeNet-5 architecture (one of the multiple models proposed in [15]) was
used for handwritten character recognition [16]. It obtained a raw error rate of 0.7%
out of 10,000 test examples. As illustrated in Fig. 2 and Table 1, the network defined
the basic components of CNN, but according to the hardware of the time, it required
high computational power. This makes it unable to be as popular and used as other
algorithms (such as SVM), which could obtain similar or even better results. One of
the main reasons for choosing “Lenet-5” is its simplicity, and this feature allows us
to preserve as many characteristics as possible, because the large number of layers
Fig. 2 Used CNN architecture

Table 1 [12–16] A detailed description of different layers of used CNN

Convolution Average Convolution Average Convolution Fully Fully
pooling pooling connected connected
Input 17 × 17 × 3 17 × 17 16 × 16 × 6 12 × 12 6 × 6 × 16 480 84
size ×6 × 16
Kernel 5 × 5 5×5 5×5 2×2 5×5 – –
size
Stride 1 1 1 2 1 – –
Pad Same Valid Valid Valid Valid – –
# 6 – 16 – 120 – –
filters
Output 17 × 17 × 6 16 × 16 12 × 12 × 6 × 6 × 2 × 2 × 120 84 2
size ×6 16 16
in our experience destroys the basic characteristics, which are color and texture. The
database contains small sample sizes, which makes the use of a large number of
convolution or pooling layers unnecessary or bad.
LeNet5 model has been successfully used in different application areas, such
as facial expression recognition [17], vehicle-assisted driving [18], traffic sign
recognition [19] and medical application like sleep apnea detection [20] …etc.
3.2 Training Stage
The training phase is a very important step; it is carried out to determine the best
weights of all CNN layers. Our network is trained with image patches of positive and
negative examples where inputs are skin/non-skin patches and outputs correspond
to the label of these patches, sized 17 × 17, are manually extracted by us from the
training images of the database (Fig. 3). Training is achieved by optimizing a loss
function using the stochastic gradient descent approach (the Adam’s optimizer). The
loss function in our case is simply cross entropy. Finally, a low learning rate is set at
0.001 to form our CNN.
3.3 Segmentation Stage
Our proposed skin detection system is actually a segmentation algorithm, which

consists of scanning the entire input image through a 17 × 17 window and a 1:16
delay step.
Then, each of these thumbnails is classified by the CNN network, previously
formed, so each pixel of these thumbnails is replaced by its probability of belonging
to the skin or non-skin class. The obtained result is grayscale probability image
(Fig. 3) then a thresholding is applied to obtain a skin binary image (Fig. 3).
In order to clean the binary image resulting from noise, we applied morpholog-
ical operators as post-processing: the closure is used to eliminate small black holes
(Fig. 4), and the aperture is used to eliminate small white segments of the image
(Fig. 5).
The last step is the displaying of skin image performed by a simple multiplication
between the original image and the binary image, to give as a result an RGB image
with only skin detected regions.
It is necessary to note that we foresee a pre-processing phase to improve the
quality of images that are too dark or too light, because the lighting can seriously
affect the color of the skin.
Fig. 3 Segmentation stage a original image, b likelihood image, c binary image
Fig. 4 Application of closure to eliminate small black holes

Fig. 5 Application of aperture to eliminate small white segments
4 Results and Discussion
4.1 Dataset and Training
This section reports the result of skin detection using LeNet5 Convolutional Neural
Network. In the training stage, we used the SFA dataset [21] that was constructed
on the basis of face images from FERET dataset [22] (876 images) and AR dataset
[23] (242 images) databases, from which skin and non-skin samples were retrieved
differently from 18 different scales (Fig. 6).
The dataset contains over 3354 manually labeled skin images and over 5590
non-skin images. The dataset is divided into 80% for training and 20% for testing
(validation).
The testing phase is a crucial step in the evaluation of training of the CNN network.
It consists of evaluating the network on a complete scene (indoor/outdoor) without
any conditions on the shots. We select images from both SFA [13] (Fig. 7) and BAO
Fig. 6 SFA dataset used for training. a Non-skin examples. b Skin examples
Fig. 7 Examples from SFA database used in testing stage
Fig. 8 Examples from SFA database used in testing stage
[24] datasets (Fig. 8). Different lighting conditions and complex scenes make these
datasets suitable for evaluating our skin detection system.
4.2 Experiments and Discussion
For quantitative analysis of the obtained results, accuracy and error rate was used,
shown by the accuracy rates called respectively training accuracy (train-accuracy) and
testing or validation accuracy (val-accuracy), and the error rates called respectively
training loss (train-loss) and testing or validation loss (val-loss). Figure 9 shown the
results obtained with precision training rate of 93%. Figures 10 and 11 shows some
tests performed on SFA and BAO datasets where the precision tests obtained are
consecutively 96 and 95%.
5 Conclusion
Our principal goal in this paper is to extract skin regions using a Convolutional neural
network called LeNet5. Our framework is divided into three main parts: At first, a
deep learning is performed to Lenet5 network using 3354 positive examples and
Fig. 9 Training results. a Training and validation loss. b Training and validation accuracy
Fig. 10 Tests on BAO dataset. a Original image. b Skin image results
5590 negative examples from SFA dataset, then and after a preprocessing of each
arbitrary image the trained network will classify image pixels into skin/non-skin.
Lastly, a thresholding and post-processing of classified regions are carried out. The
tests were carried out on images of variable complexity: indoor, outdoor, variable
lighting, simple and complex background. The results obtained are very encouraging,
we show the qualitative and quantitative results obtained on SFA and BAO datasets
where the precision tests obtained are consecutively 96 and 95%.
Fig. 11 Tests on SFA dataset. a Original image. b Skin image results
Acknowledgements The work described herein was partially supported by 8 Mai 1945 University
and PRFU project through the grant number C00L07UN240120200001. The authors thank the staff
of LAIG laboratory, who provided financial support.
References
1. Naji, S., Jalab, H.A., Kareem, S.A.: A survey on skin detection in colored images. Artif. Intell.
Rev. 52, 1041–1087 (2019). https://doi.org/10.1007/s10462-018-9664-9
2. Zuo, H., Fan, H., Blasch, E., Ling, H.: Combining convolutional and recurrent neural networks
for human skin detection. IEEE Sig. Process. Lett. 24, 289–293 (2017). https://doi.org/10.1109/
LSP.2017.2654803
3. Zarit, B.D., Super, B.J., Quek, F.K.H.: Comparison of five color models in skin pixel classi-
fication. In: Proceedings International Workshop on Recognition, Analysis, and Tracking of
Faces and Gestures in Real-Time Systems. In Conjunction with ICCV’99 (Cat. No. PR00378).
pp. 58–63 (1999). https://doi.org/10.1109/RATFG.1999.799224
4. Phung, S.L., Bouzerdoum, A., Chai, D.: Skin segmentation using color pixel classification:
analysis and comparison. IEEE Trans. Pattern Anal. Mach. Intell. 27, 148–154 (2005). https://
doi.org/10.1109/TPAMI.2005.17
5. Ashwini, A., Murugan, S.: Automatic skin tumour segmentation using prioritized patch based
region—a novel comparative technique. IETE J. Res. 1, 12 (2020). https://doi.org/10.1080/037
72063.2020.1808091
6. Li, B., Xue, X., Fan, J.: A robust incremental learning framework for accurate skin region
segmentation in color images. Pattern Recogn. 40, 3621–3632 (2007). https://doi.org/10.1016/
j.patcog.2007.04.018
7. Poudel, R.P., Nait-Charif, H., Zhang, J.J., Liu, D.: Region-based skin color detection. In:
VISAPP (1) VISAPP 2012-Proceedings of the International Conference on Computer Vision
Theory and Applications 1, pp. 301–306. VISAPP (2012)
8. Kolkur, S., Kalbande, D., Shimpi, P., Bapat, C., Jatakia, J.: Human skin detection using RGB,
HSV and YCbCr Color Models. In: Presented at the International Conference on Communi-
cation and Signal Processing 2016 (ICCASP 2016) (2016). https://doi.org/10.2991/iccasp-16.
2017.51
9. Brancati, N., De Pietro, G., Frucci, M., Gallo, L.: Human skin detection through correlation
rules between the YCb and YCr subspaces based on dynamic color clustering. Comput. Vis.
Image Underst. 155, 33–42 (2017). https://doi.org/10.1016/j.cviu.2016.12.001
10. Verma, A., Raj, S.A., Midya, A., Chakraborty, J.: Face detection using skin color modeling
and geometric feature. In: 2014 International Conference on Informatics, Electronics Vision
(ICIEV). pp. 1–6 (2014). https://doi.org/10.1109/ICIEV.2014.6850755
11. Shaik, K.B., Ganesan, P., Kalist, V., Sathish, B.S., Jenitha, J.M.M.: Comparative study of skin
color detection and segmentation in HSV and YCbCr color space. Procedia Comput. Sci. 57,
41–48 (2015)
12. Nadian-Ghomsheh, A.: Pixel-based skin detection based on statistical models. J. Telecommun.
Electron. Comput. Eng. (JTEC) 8, 7–14 (2016)
13. Oghaz, M.M.D., Argyriou, V., Monekosso, D., Remagnino, P.: Skin identification using deep
convolutional neural network. In: Bebis, G., Boyle, R., Parvin, B., Koracin, D., Ushizima, D.,
Chai, S., Sueda, S., Lin, X., Lu, A., Thalmann, D., Wang, C., Xu, P. (eds.) Advances in Visual
Computing, pp. 181–193. Springer International Publishing, Cham (2019). https://doi.org/10.
1007/978-3-030-33720-9_14
14. Kim, Y., Hwang, I., Cho, N.I.: Convolutional neural networks and training strategies for skin
detection. In: 2017 IEEE International Conference on Image Processing (ICIP), pp. 3919–3923
(2017). https://doi.org/10.1109/ICIP.2017.8297017
15. Lecun, Y., Jackel, L.D., Bottou, L., Cartes, C., Denker, J.S., Drucker, H., Müller, U., Säckinger,
E., Simard, P., Vapnik, V., et al.: Learning algorithms for classification: a comparison on
handwritten digit recognition. In: Neural Networks: The Statistical Mechanics Perspective,
pp. 261–276. World Scientific (1995)
16. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document
recognition. Proc. IEEE 86, 2278–2324 (1998). https://doi.org/10.1109/5.726791
17. Wang, G., Gong, J.: Facial expression recognition based on improved LeNet-5 CNN. In: 2019
Chinese Control and Decision Conference (CCDC), pp. 5655–5660 (2019). https://doi.org/10.
1109/CCDC.2019.8832535
18. Zhang, C.-W., Yang, M.-Y., Zeng, H.-J., Wen, J.-P.: Pedestrian detection based on
improved LeNet-5 convolutional neural network. J. Algorithms Comput. Technol. 13,
1748302619873601 (2019). https://doi.org/10.1177/1748302619873601
19. Zhang, C., Yue, X., Wang, R., Li, N., Ding, Y.: Study on traffic sign recognition by optimized
Lenet-5 algorithm. Int. J. Patt. Recogn. Artif. Intell. 34, 2055003 (2019). https://doi.org/10.
1142/S0218001420550034
20. Wang, T., Lu, C., Shen, G., Hong, F.: Sleep apnea detection from a single-lead ECG signal
with automatic feature-extraction through a modified LeNet-5 convolutional neural network.
PeerJ7, e7731 (2019) https://doi.org/10.7717/peerj.7731
21. Casati, J.P.B., Moraes, D.R., Rodrigues, E.L.L.: SFA: a human skin image database based on
FERET and AR facial images. In: IX workshop de Visao Computational, Rio de Janeiro (2013)
22. Phillips, P.J., Moon, H., Rizvi, S.A., Rauss, P.J.: The FERET evaluation methodology for face-
recognition algorithms. IEEE Trans. Pattern Anal. Mach. Intell. 22, 1090–1104 (2000). https://
doi.org/10.1109/34.879790
23. Martinez, A., Benavente, R.: The AR face database. Tech. Rep. 24 CVC Technical Report.
(1998)
24. Wang, X., Xu, H., Wang, H., Li, H.: Robust real-time face detection with skin color detection
and the modified census transform. In: 2008 International Conference on Information and
Automation, pp. 590–595 (2008). https://doi.org/10.1109/ICINFA.2008.4608068
CRAN: An Hybrid CNN-RNN
Attention-Based Model for Arabic
Machine Translation
Nouhaila Bensalah, Habib Ayad, Abdellah Adib,

and Abdelhamid Ibn El Farouk
Abstract Machine Translation (MT) is one of the challenging tasks in the field of
Natural Language Processing (NLP). The Convolutional Neural Network (CNN)-
based approaches and Recurrent Neural Network (RNN)-based techniques have
shown different capabilities in representing a piece of text. In this work, an hybrid
CNN-RNN attention-based neural network is proposed. During training, Adam opti-
mizer algorithm is used, and then, a popular regularization technique named dropout
is applied in order to prevent some learning problems such as overfitting. The exper-
iment results show the impact of our proposed system on the performance of Arabic
machine translation.
1 Introduction
MT is an intricate process that uses a computer application to translate text or speech

or even capture from one natural language to another [2]. Many approaches from
traditional rule-based approaches to the recent neural methods have been applied
since the introduction of MT [4, 7, 8, 25]. Due to the excellent performance that
achieves Deep Learning (DL) on difficult problems such as question answering [3,
6], sentiment analysis [3, 9], and visual object recognition [14, 20] for a small
amount of steps, Google has investigated the use of DL to develop its own MT
system. In the same context, Linguee team have developed DeepL based on CNNs
that support a various number of languages such as French, Spanish, English, and
N. Bensalah (B) · H. Ayad · A. Adib

Team Networks, Telecoms & Multimedia, University of Hassan II Casablanca, Casablanca 20000,
Morocco
e-mail: nouhaila.bensalah@etu.fstm.ac.ma
A. Adib
e-mail: abdellah.adib@fstm.ac.ma
A. Ibn El Farouk
Teaching, Languages and Cultures Laboratory Mohammedia, Mohammedia, Morocco
https://doi.org/10.1007/978-981-16-3637-0_7
88 N. Bensalah et al.
others. MT systems based on DL often use a sequence-to-sequence model in order

to map between the input and the target sequences directly.
The whole sequence-to-sequence process is described by Sutskever et al. in [25].
In short, the first step is to compute a representation of the source sentence using
an encoder which can be Long Short Term Memory (LSTM), or Gated Recurrent
Unit (GRU). In view to extract the relevant features from that encoder, an attention
module can be used. Finally, the obtained vectors are transferred to the decoder which
generates the output sentence. The aim of this research is to exploit the full advantages
of CNN and RNN in order to map the input sentence to a low dimensional vector
sequence. Specifically, in the first stage, a conventional CNN is used. Hence, the
encoding of the input sentence will be processed in parallel to properly manipulate
and optimize GPU hardware during training. And due to the extensive attention
that has gained RNN in recent years, an improved architecture of RNN, namely
Bidirectional GRU (BiGRU), is performed on the same input sequence. Finally, a
mechanism of self-attention is applied to merge the features generated by both BiGRU
and CNN, and then, the obtained vectors will be utilized as inputs to a GRU layer in
order to generate the translation of the input sentence. The used attention mechanism
could be considered as an instance of the widely known attention mechanism [4], as
well as its recent variants; i.e. the self attention [21] and the inner attention [11]. In
this case, the attention mechanism is performed within the same input sentence rather
than the alignment of the output and the input sentences. The remainder of this paper
is organized as follows. Section 2 describes the proposed model for implementing
the Arabic MT system. Section 3 details the experimental setup and results. Finally,
the conclusion is summarized in Sect. 4.
2 CRAN Arabic MT Model
Selecting the best features, from an input sequence, lies at the core of any MT system.
Most of the state-of-the-art MT systems employ neural network-based approaches
such as the CNN and the RNN-based architectures. In spite of their easy deploy-
ment and generally their capabilities in representing a piece of text, they represent
some disadvantages [19]. In the CNN-based approaches, each source sentence is
presented as a matrix by concatenating the embedding vector sequence as columns.
Then, the CNN is applied to identify the most influential features about the input
sentence. Nonetheless, these techniques could uniquely learn regional features, and
it is not straightforward to handle with the long-term dependency between the fea-
tures extracted from the source sentence. On the other hand, employing the (GRU
or LSTM)-based approaches allow the model to generate effective sentences rep-
resentation using temporal features since they capture the long-term dependencies
between the words of a source sentence. However, these approaches represent some
weaknesses, most significantly, their inability to distinguish between the words that
contribute to the selection of the best features. This is due to the fact that they manip-
ulate each word in a source sentence equally. Since the RNN and the CNN can
CRAN: An Hybrid CNN-RNN Attention-Based Model … 89
complement each other for MT task, various solutions have been proposed [1, 17].
Most of the existing methods that combine these two models focus on applying the
LSTM or GRU on the top of the CNNs. Consequently, they could not be applied
directly on the source sentence, and hence, some features will be lost. In order to
incorporate the full strength of these two groups of architectures, we present in this
paper a novel architecture based on the use of the CNN and BiGRU architectures
applied both on the input data. The proposed model is depicted in Fig. 1, and it is
summarized as following:
1. The input sentence is preprocessed and then decomposed into words; each one is
represented as a fixed-dimension vector using FastText model [10]. The obtained
vectors are concatenated to generate a fixed-size matrix.
2. Several convolutional filters are applied on the resulting matrix. Each convolution
filter has a view over the entire source sequence, from which it picks features. To
extract the maximum value for each region determined by the filter, a max-pooling
layer is used.
3. In order to deal with the long-term dependency problem and extract the temporal
features from the same input sentence, a BiGRU is applied on the whole input
sentence.
4. An attention mechanism is then performed with the objective of merging the useful
temporal features and the regional ones obtained, respectively, by the BiGRU layer
and the CNN model acting on the whole input sequence.
5. To generate the output sentence from the obtained vector sequence, a GRU layer
is used, and a Softmax layer is then applied to generate the translation of the input
sequence.
Hereafter, we will detail the different layers through which the whole process
passes.
2.1 Input Layer
This is the first layer in our model, and it is used to represent each word in a sentence
as a vector of real values. First, the input sentence is decomposed into words. Then, to
obtain the same length for all the input sentences, a padding technique is performed
on the sentences which are short (length < n) where n is the maximum length of
the source sentences. Then, every word is embedded as g = (g1 , . . . , gn ) where gi
(i ∈ (1, 2, .., n)) represents a column in the embedding matrix. Finally, the obtained
vectors, called the embedding vector sequence, will be fed into the BiGRU layer and
the CNN model.
Fig. 1 Block diagram of the overall architecture of the proposed approach
2.2 Conventional CNN
The CNN architecture was developed by LeCun et al. [20] and has risen to prominence
as a state of the art in MT. In this study, a conventional CNN is investigated to
extract the most influential features from the embedding vector sequence. Generally,
a conventional CNN model consists of the following layers:
• Convolutional layer: utilizes a set of filters (kernels) to convert the embedding
vector sequence g into feature maps.
• Nonlinearity: between convolutional layers, an activation function, such as tan h
or ReLU which represents, respectively, a tangent hyperbolic and rectified Linear
Unit functions, is applied to the obtained feature maps to introduce nonlinearity
into the network. Without this operation, the network would hence struggle with
complex data. In this paper, ReLU was adopted, which is defined as:
f (x) = max(0, x) (1)
• Pooling layer: Its main role is to reduce the amount of parameters and computation
in the network by decreasing the feature maps size. Two common methods used
in the pooling operation are:
– Average pooling: outputs the average value in a region determined by the filter.
– Maximum pooling (or Max pooling): The output is the maximum value over a
region processed by the considered filter.
In this paper, we used the max pooling to preserve the largest activation in the
feature maps.
• Dropout layer: its role is to randomly drop units (along with their connections)
from the neural network during training to avoid overfitting.
2.3 BiGRU Layer
Given the embedding vector sequence g = (g1 , . . . , gn ), a standard RNN [15] gen-
erates the hidden vector sequence e = (e1 , . . . , en ). The main objective of RNN is
to capture the temporal features from the source sentence. The output of the network
is calculated by iterating the following equation from t = 1 to t = n:
et = f (Wge gt + Wee et−1 + be ) (2)
where the W terms are weight matrices, i.e., for example, Wge denotes the input-
hidden weight matrix, the be denotes the hidden bias vector, and f is the hidden
layer activation function. To avoid the issue of the vanishing gradient that penalizes
the standard RNN, GRU [12] is proposed to store the input information without a
memory unit. A single GRU cell is illustrated in Fig. 2, and is defined as follows:
rt = σ (Wgr gt + Wer et−1 + br ) (3)
u t = σ (Wgu gt + Weu et−1 + bu ) (4)

eet = tan h(Wge gt + Wee (rt et−1 ) + be ) (5)

et = (1 − u t ) et−1 + u t eet (6)
where tan h is the tangent hyperbolic function, σ the element-wise sigmoid activation
function, is the element-wise Hadamard product, r and u are, respectively, the
reset gate and update gate, all have the same size as the hidden vector e and the b
terms are bias vectors, i.e., for example, bu denotes the update bias vector.
In MT, the GRU architecture is motivated by two main reasons. First, such archi-
tecture has been shown to represent the sequential data by taking into account the
previous data. Second, it is better at exploiting and capturing long-range context due
to its gates that decide which data will be transferred to the output. However, the GRU
is only able to make use of the amount of information seen by the hidden states at the
previous steps. In order to exploit the future information as well, bidirectional RNNs
Fig. 2 Architecture of GRU gt
up date gate
σ tanh
ut eet
σ
1− ×
rt
×
reset gate
et−1 × + et
Fig. 3 Bidirectional RNN
(BRNNs) [24] are introduced. They process the data in two opposite directions with
two separate hidden layers as illustrated in Fig. 3.
In this case, the forward hidden vector sequence −→
e t and the backward hidden
←−
vector sequence e t are computed by iterating the backward layer from t = 1 to
t = n and the forward layer from t = 1 to t = n.
−
→
e t = tan h(Wg−
e gt + U−
→ →
e−e et−1 + b−
→ e )
→ (7)
←
e−t = tan h(Wg←
e− gt + U← e− et−1 + b←
e−← e− ) (8)
Fig. 4 Attention mechanism
In this way, a sentence could be represented as e = (e1 , e2 , .., en ) where et =

[−
→e t, ←
e−t ]. Combining BRNN with GRU gives bidirectional GRU [18] that could
exploit long-range context in both input directions.
2.4 Attention Layer
The attention mechanism is one of the key components of our architecture. In this
context, there are several works that can help to locate the useful words from the
input sequence [4, 22]. Motivated by the ability of the CNN model to capture regional
syntax of words and the capacity of the BiGRU model to extract temporal features
of words, we aim to use the output vector sequence generated by the CNN, h =
(h 1 , h 2 , . . . , h n ), and the hidden vector sequence calculated by the BiGRU layer
e = (e1 , e2 , . . . , en ) during the attention mechanism. In the proposed approach, at
each step t (t ∈ (1, 2, . . . , n)), a unique vector h t and the hidden vector sequence
e = (e1 , e2 , . . . , en ) are used to calculate the context vector z t . The detail of the
proposed attention mechanism computation process is illustrated in Fig. 4.
In short, the first step is to measure the similarity denoted as m t j between the hid-
den vector e j ( j ∈ (1, 2, ..n)) generated by the BiGRU layer and the vector h t
(t ∈ (1, 2, ..n)) produced by the CNN. Three different methods could be used to
calculate m t j :
1. Additive attention:
m t j = Wa tan h(We e j + Uh h t ) (9)
2. Multiplicative attention:

m t j = e j Wm h t (10)
3. Dot product:

mt j = e j ht (11)
where: Wa , We , Uh , Wm are the weight matrices. Then, a Softmax function is used

to get a normalized weight st j of each hidden state e j . Finally, the context vector z t
is computed as a weighted sum of the hidden vector sequence e.
In our case, the following equations are iterated from j = 1 to j = n:
m t j = Wa tan h(We e j + Uh h t ) (12)
exp(m t j )
st j = n (13)
k=1 exp(m tk ))
The context vector sequence z = (z 1 , z 2 , . . . , z n ) is generated by iterating the fol-

lowing equation from t = 1 to t = n:

n
zt = st j e j (14)
j=1
2.5 The Output Layer
In summary, the goal of our model is to map the input sentence to a fixed sized
vector sequence z = (z 1 , z 2 , . . . , z n ) using the CNN-BiGRU and a mechanism of
attention. Then, a GRU layer is applied on the obtained vector sequence. Finally, we
add a fully connected output layer with a Softmax activation function which gives, at
each time step, the probability distribution across all the unique words in the target
language. The predicted word at each time step is selected as the one with the highest
probability.
2.6 Training and Inference
The Arabic MT process involves two main stages: training and inference. During the
training stage, features extraction is performed after the built of the training Arabic-
English sentences. It aims at providing a useful representation of an input sentence
in such a way that it can be understandable by the model. Then, a CNN-BiGRU with
a mechanism of attention followed by a GRU layer are applied on these obtained
features. Finally, a Softmax layer is performed with the objective of optimizing the
parameters of the neural network by comparing the model outputs with the target
sequences (what we should achieve). After the training is done, the Arabic MT model
is built and can be used to translate an input sentence with any help from the target
sentences. The output sequence is generated word by word using the Softmax layer.
Its main role during inference stage is to generate at each time step the probability
distribution across all unique words in the target language.
3 Results and Evaluation
The experiments were conducted over our own Arabic-English corpus. It contains
a total of 266,513 words in Arabic and 410,423 ones in English, and the amount
of unique words was set to 23,159 words in Arabic and 8323 ones in English. The
database was divided randomly into a training set, a validation set, and a testing set.
20,800 sentences for both Arabic and English languages were used for training, 600
sentences for validation and 580 for testing.
To build our corpora, we select 19,000 sentences from the UN dataset from the
Web site.1 In order to improve the performance of our model, we have used two other
datasets. First, we select manually the best English-Arabic sentences from the Web
site2 which contains blog-posts, tweets in many languages. Finally, we have used the
sentences in the English-Arabic pair which can be found in this Web site.3
In the following, we present a series of experiments for Arabic MT analysis to
understand the practical utility of the proposed approach. As an evaluation metric,
we compute the BLEU score [23], the GLEU score [27], and the WER score [26]
which are the most commonly used in MT task.
3.1 Hyperparameters Choices
In this part, we investigated the impact of each hyperparameter on Arabic MT system

to select the best ones to use during training. The performance of Arabic MT was
first evaluated by varying the batch size to 32, 64, 128, and 256. Table 1 shows the
BLEU, GLEU, and WER scores for different batch sizes.
From Table 1, we can see that the minimal WER score (WER score = 0.449) is
reached using 64 as batch size. Examining the results, it seems that using a too low
or too high batch size does not result in better performance.
1 http://opus.nlpl.eu/.
2 http://www.cs.cmu.edu.
3 http://www.manythings.org.
Table 1 Arabic MT performance for different batch sizes

Batch size 32 64 128 256
BLEU score 1-gram 0.571 0.575 0.512 0.479
2-gram 0.577 0.578 0.518 0.470
3-gram 0.592 0.590 0.523 0.473
4-gram 0.603 0.602 0.527 0.478
GLEU score 1–4 gram 0.445 0.463 0.371 0.317
1–3 gram 0.471 0.487 0.404 0.352
1–2 gram 0.513 0.526 0.451 0.424
WER score 0.473 0.449 0.529 0.567
Table 2 Arabic MT performance with respect to the length of the input sentences
Sentence length 10 20 30
BLEU score 1-gram 0.485 0.529 0.575
2-gram 0.483 0.526 0.578
3-gram 0.484 0.536 0.590
4-gram 0.489 0.538 0.602
GLEU score 1 to 4 gram 0.324 0.401 0.463
1–3 gram 0.369 0.426 0.487
1–2 gram 0.445 0.472 0.526
WER score 0.589 0.546 0.449
Then, we evaluated the performance of Arabic MT by varying the sentence length

to 10, 20, and 30 which is the maximum length of the source sentences. Table 2
illustrates the influence of the sentence length on the BLEU, GLEU, and WER
scores.
The reported results show clearly that we get better performance as the sentence
length increases. In our work, the optimal value of the sentence length is set to 30.
The performance of Arabic MT was evaluated for different optimization tech-
niques. The results are reported in Table 3.
We can clearly observe that, globally, the best WER score has been reached for
Adam optimizer.
Next, we increase the number of units in the BiGRU and the output layers so as
to compare the performance of Arabic MT achievable by our model in this case.
Table 4 reports the results obtained using different values of units.
We can notice that changing the number of units affects the performance of Arabic
MT and the best results are obtained using 400 units.
In the following, we analyze the impact of increasing the number of layers on the
Arabic MT quality. The results are reported in Table 5.
Table 3 Arabic MT performance for different optimization techniques

Optimization SGD RMSprop Adam
technique
BLEU score 1-gram 0.263 0.573 0.575
2-gram 0.225 0.580 0.578
3-gram 0.223 0.597 0.590
4-gram 0.248 0.610 0.602
GLEU score 1–4 gram 0.123 0.460 0.463
1–3 gram 0.151 0.483 0.487
1–2 gram 0.194 0.522 0.526
WER score 0.758 0.455 0.449
Table 4 Arabic MT performance for different number of units

Number of 100 200 300 400
units
BLEU score 1-gram 0.507 0.501 0.547 0.575
2-gram 0.493 0.492 0.546 0.578
3-gram 0.492 0.493 0.556 0.590
4-gram 0.491 0.496 0.562 0.602
GLEU score 1–4 gram 0.348 0.342 0.406 0.463
1–3 gram 0.384 0.377 0.437 0.487
1–2 gram 0.437 0.429 0.482 0.526
WER score 0.529 0.542 0.497 0.449
Table 5 Arabic MT performance for different numbers of layers

Number of layers 1 2 3
BLEU score 1-gram 0.568 0.575 0.564
2-gram 0.574 0.578 0.569
3-gram 0.588 0.590 0.582
4-gram 0.598 0.602 0.593
GLEU score 1–4 gram 0.448 0.463 0.452
1–3 gram 0.474 0.487 0.476
1–2 gram 0.514 0.526 0.514
WER score 0.465 0.449 0.462
Table 6 Arabic MT performance for different CNN filter sizes

Size 2 3 4 5
BLEU score 1-gram 0.567 0.569 0.577 0.573
2-gram 0.569 0.572 0.579 0.575
3-gram 0.579 0.584 0.591 0.587
4-gram 0.591 0.594 0.602 0.599
GLEU score 1–4 gram 0.450 0.455 0.462 0.455
1–3 gram 0.476 0.479 0.487 0.480
1–2 gram 0.516 0.519 0.526 0.521
WER score 0.458 0.452 0.448 0.454
The reported results in Table 5 show that using a too low or too high number of layers
does not result in better performance. For next experiments, we choose the number
of layers to be 2.
Finally, based on a manual tuning, we initialize the learning rate with a value
of 0.001, and with the beginning of overfitting, we start to multiply this value by
0.4 at each 2 epochs until it falls below 10−7 . If the performance of the model on
the validation set stops to grow, the early stopping technique based on the validation
accuracy is performed. It aims to stop the training process after 5 epochs. More details
of these techniques and other tips to reach better training process are described in
[16].
The performance of Arabic MT was also evaluated by varying the size of CNN
filters to 2, 3, 4, and 5. Table 6 shows the BLEU, GLEU, and WER scores for different
CNN filter sizes.
From Table 6, we can see that the best Arabic MT scores come from the CNN
filter of size 4 and 5 which have been concatenated in this work.
3.2 The Impact of Different RNN Variants on Arabic MT
To study the performance of the Arabic MT under different RNNs, four different
combinations of RNNs have been used:
1. BiLSTM for the encoding process (discussed in 2.3) and GRU in the output layer.
2. BiGRU for the encoding process (discussed in 2.3) and LSTM in the output layer.
3. BiLSTM for the encoding process (discussed in 2.3) and LSTM in the output
layer.
4. BiGRU for the encoding process (discussed in 2.3) and GRU in the output layer.
It is clear from Table 7 that the combination 4, which is the proposed approach,
gives better performance in terms of BLEU, GLEU, and WER scores. Furthermore,
one of the attractive characteristics of our model is its ability to train faster than the
combinations 1, 2, and 3 (Total Time = 610 s).
Table 7 Arabic MT performance for different combinations of RNNs

Combination 1 2 3 4
BLEU score 1-gram 0.544 0.380 0.344 0.575
2-gram 0.549 0.357 0.328 0.578
3-gram 0.564 0.359 0.336 0.590
4-gram 0.577 0.371 0.358 0.602
GLEU score 1–4 gram 0.414 0.202 0.173 0.463
1–3 gram 0.423 0.239 0.207 0.487
1–2 gram 0.467 0.295 0.258 0.523
WER score 0.511 0.663 0.704 0.449
Total time (s) 1270 2027 1908 610
Table 8 Comparison between the Arabic MT performance of RCAN and our approach
RCAN Our approach
BLEU score 1-gram 0.535 0.575
2-gram 0.545 0.578
3-gram 0.555 0.590
4-gram 0.565 0.602
GLEU score 1–4 gram 0.407 0.463
1–3 gram 0.434 0.487
1–2 gram 0.476 0.523
WER score 0.511 0.449
3.3 Comparison with RCAN Model
In this part, another model denoted as RCAN is proposed for comparison. In this
case, the context vector z t (t ∈ (1, 2, .., n)) is calculated using the vector sequence
h = (h 1 , h 2 , . . . , h n ) and the hidden vector et .
Table 8 illustrates the results of Arabic MT using the proposed architecture and
compares them to the results of RCAN.
It can be seen, from Table 8, that our approach achieved relatively ideal overall
performance using our corpus and improved the performance by 6.2% in terms of
WER score. These findings may be explained by the use of the temporal vector
sequence generated by the BiGRU, instead of the regional vector sequence produced
by CNN, to calculate the context vector. In this case, the model becomes able to
automatically search for parts of a source sentence that are relevant to predict a
target word.
Table 9 Comparison with state-of-the art works using our own corpus
[4] [13] Our approach
BLEU score 0.493 0.485 0.575
WER score 0.492 0.515 0.449
3.4 Comparison with Previous Related Works and Qualitative

Evaluation
Because this work is inspired by the approaches proposed by Bahdanau et al. [4] and
Cho et al. [13], the performance of Arabic MT is evaluated in terms of BLEU, GLEU,
and WER scores reached using our model and these works. Table 9 summarizes the
obtained results for Arabic MT task on our corpus with the considering literature
works.
We can clearly observe from Table 9 that in all the cases the best performance is
achieved using our approach with a limited vocabulary. This is likely due to the fact
that our model does not encode the whole input sentence into a single vector. Instead,
it focus on the relevant words of the source sentence during the encoding process.
As an example, consider this source sentence from the test set:
Our model translated this sentence into:

paris is pleasant during November but it is usually beautiful in September
The truth is:
paris is nice during November but it is usually beautiful in September
The proposed approach correctly translated the source sentence, but it replaced nice
with pleasant.
Let us consider another sentence from the test set:
Our model translated this sentence into:
these books are my books
The truth is:
these books belong to me
These qualitative observations demonstrates that the proposed approach does not
translate the input sentence as the truth, but instead it preserves the original meaning
of the source sentence.
4 Conclusion
In this paper, we proposed the use of both CNN and BiGRU with the mechanism of
attention system for the task of MT between English and Arabic texts. The motiva-
tion for introducing such a system is to improve the performance of Arabic MT by
capturing the most influential words in the input sentences using our corpora. In this
context, we described first how the used corpus is produced. A comparative perfor-
mance analysis of the hyperparameters is performed. As expected, the experimental
results show that the proposed method is capable of providing satisfactory perfor-
mance for Arabic MT. As part of future work, we aim to use saliency to visualize
and understand neural models in NLP [5].
References
1. Alayba, A.M., Palade, V., England, M., Iqbal, R.: A combined cnn and lstm model for arabic
sentiment analysis. In: International Cross-Domain Conference for Machine Learning and
Knowledge Extraction, pp. 179–191 (2018)
2. Alqudsi, A., Omar, N., Shaker, K.: Arabic machine translation: a survey. Artif. Intell. Rev.
42(4), 549–572 (2014)
3. Antoun, W., Baly, F., Hajj, H.M.: Arabert: transformer-based model for arabic language under-
standing (2020) . CoRR abs/2003.00104
4. Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by jointly learning to align and
translate. In: Bengio, Y., LeCun, Y. (eds) 3rd International Conference on Learning Represen-
tations, ICLR (2015)
5. Bastings, J., Filippova, K.: The elephant in the interpretability room: why use attention as
explanation when we have saliency methods? In: Proceedings of the Third BlackboxNLP
Workshop on Analyzing and Interpreting Neural Networks for NLP, pp. 149–155. Association
for Computational Linguistics (2020)
6. Bensalah, N., Ayad, H., Adib, A., el farouk, A.I.: Combining word and character embeddings in
Arabic Chatbots. In: Advanced Intelligent Systems for Sustainable Development, AI2SD’2020,
Tangier, Morocco (2020)
7. Bensalah, N., Ayad, H., Adib, A., Farouk, A.I.E.: LSTM or GRU for Arabic machine transla-
tion? Why not both! In: International Conference on Innovation and New Trends in Information
Technology, INTIS 2019, Tangier, Morocco, Dec 20–21 (2019)
8. Bensalah, N., Ayad, H., Adib, A., Farouk, A.I.E.: Arabic machine translation based on the
combination of word embedding techniques. In: Intelligent Systems in Big Data, Semantic
Web and Machine Learning (2020)
9. Bensalah, N., Ayad, H., Adib, A., Farouk, A.I.E.: Arabic sentiment analysis based on 1-D
convolutional neural network. In: International Conference on Smart City Applications, SCA20
(2020)
10. Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with subword infor-
mation. Trans. Assoc. Comput. Linguist., 135–146 (2017)
11. Cheng, J., Dong, L., Lapata, M.: Long short-term memory-networks for machine reading. In:
Su, J., Carreras, X., Duh, K. (eds) Proceedings of the 2016 Conference on Empirical Methods
in Natural Language Processing, EMNLP, pp. 551–561 (2016)
12. Cho, K., van Merrienboer, B., Bahdanau, D., Bengio, Y.: On the properties of neural machine
translation: encoder-decoder approaches. In: Proceedings of SSST@EMNLP 2014, Eighth
Workshop on Syntax, Semantics and Structure in Statistical Translation, pp. 103–111. Asso-
ciation for Computational Linguistics (2014)
13. Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio,
Y.: Learning phrase representations using rnn encoder-decoder for statistical machine transla-
tion (2014). arXiv preprint arXiv:1406.1078
14. Ciresan, D.C., Meier, U., Schmidhuber, J.: Multi-column deep neural networks for image clas-
sification. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence,
pp. 3642–3649. IEEE Computer Society (2012)
15. Elman, J.L.: Finding structure in time. Cognit. Sci. 14(2), 179–211 (1990)
16. Feurer, M., Hutter, F.: Hyperparameter optimization. In: Automated Machine Learning, pp.
3–33. Springer (2019)
17. Gehring, J., Auli, M., Grangier, D., Dauphin, Y.N.: A convolutional encoder model for neural
machine translation. In: Proceedings of the 55th Annual Meeting of the Association for Com-
putational Linguistics, ACL 2017, pp. 123–135. Association for Computational Linguistics
(2017)
18. Graves, A., Schmidhuber, J.: Framewise phoneme classification with bidirectional lstm and
other neural network architectures. Neural Netw. 18(5–6), 602–610 (2005)
19. Guo, L., Zhang, D., Wang, L., Wang, H., Cui, B.: Cran: a hybrid CNN-RNN attention-based
model for text classification. In: International Conference on Conceptual Modeling, pp. 571–
585. Springer (2018)
20. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P., et al.: Gradient-based learning applied to docu-
ment recognition. Proc. IEEE 86(11), 2278–2324 (1998)
21. Lin, Z., Feng, M., dos Santos, C.N., Yu, M., Xiang, B., Zhou, B., Bengio, Y.: A structured self-
attentive sentence embedding. In: 5th International Conference on Learning Representations,
ICLR (2017)
22. Luong, M., Pham, H., Manning, C.D.: Effective Approaches to Attention-based Neural
Machine Translation. CoRR abs/1508.04025 (2015)
23. Papineni, K., Roukos, S., Ward, T., Zhu, W.-J.: Bleu: a method for automatic evaluation of
machine translation. In: Proceedings of the 40th Annual Meeting on Association for Compu-
tational Linguistics, pp. 311–318. Association for Computational Linguistics (2002)
24. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Trans. Sig. Process.
45(11), 2673–2681 (1997)
25. Sutskever, I., Vinyals, O., Le, Q.: Sequence to sequence learning with neural networks.
Advances in NIPS (2014)
26. Wang, Y.-Y., Acero, A., Chelba, C.: Is word error rate a good indicator for spoken language
understanding accuracy. In: 2003 IEEE Workshop on Automatic Speech Recognition and
Understanding (IEEE Cat. No. 03EX721), pp. 577–582. IEEE (2003)
27. Wu, Y., Schuster, M., Chen, Z., Le, Q.V., Norouzi, M., Macherey, W., Krikun, M., Cao, Y.,
Gao, Q., Macherey, K., Klingner, J., Shah, A., Johnson, M., Liu, X., Kaiser, L., Gouws, S.,
Kato, Y., Kudo, T., Kazawa, H., Stevens, K., Kurian, G., Patil, N., Wang, W., Young, C., Smith,
J., Riesa, J., Rudnick, A., Vinyals, O., Corrado, G., Hughes, M., Dean, J.: Google’s neural
machine translation system: bridging the gap between human and machine translation. CoRR
abs/1609.08144 (2016)
Impact of the CNN Patch Size in the
Writer Identification
Abdelillah Semma, Yaâcoub Hannad, and Mohamed El Youssfi El Kettani
Abstract Writer identification remains a very interesting challenge where many

researchers have tried to find the various parameters which can help to find the right
writer of a handwritten text. The introduction of deep learning has made it possible
to achieve unprecedented records in the field. However, the question to ask, what
size of patch to use to train a CNN model in order to have the best performance? In
our paper, we try to find an answer to this question by investigating the results of
the use of several patch sizes for a Resnet34 model and two languages Arabic and
French from the LAMIS-MSHD dataset.
1 Introduction
Writing remains one of the great foundations of human civilization for communica-
tion and the transmission of knowledge. Indeed, many objects that are around us are
presented in the form of writing: signs, products, newspapers, books, forms ...
Allowing the machine to read and capture more information will surely help
in the process of identifying the authors of handwritten documents in a faster and
more efficient manner. Indeed, with the advent of new information technologies, and
with the increase in the power of machines, the automation of processing operations
(editing, searching, and archiving) seems inevitable. Therefore, a system that enables
the machine to understand human handwriting is needed.
Writer recognition is used to identify the author or writer of a handwritten text
using a database containing samples of training writing. This challenge is not easy
A. Semma (B) · M. E. Y. El Kettani

Ibn Tofail University, Kenitra, Morocco
e-mail: abdelillah.semma@uit.ac.ma
M. E. Y. El Kettani
e-mail: elkettani@univibntofail.ac.ma
Y. Hannad
National Institute of Statistics and Applied Economics, Rabat, Morocco
e-mail: y.hannad@insea.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_8
104 A. Semma et al.
because a person’s writing style depends on several factors such as their mental and
physical state, their pen, the position of the paper, and the writer at the time of writing
and the font size, which can vary according to the need.
The identification of the writers of handwritten documents touches several fields
such as criminology which in some cases seeks to identify the exact person who
produced a handwritten document. Writer identification also helps to recognize the
name of the author of a book whose writer is not known. For a long time, human
experts has tried to guess the writer of a manuscript manually, something which is not
easy, that is why researchers have tried to design automatic systems for identifying
writers.
For the writer identification, we proceed generally by three main steps: The pre-
processing phase to prepare the handwritten images to the processing phase. The
feature extraction phase that allows the extraction of the characteristics of images or
parts of images in vectors. The last phase is that of classification where one seeks to
calculate the distance between the test features and those of training in order to know
the minimum distance which corresponds to the images of the requested author.
In writer recognition, we distinguish between two types: writer identification and
writer retrieval. In the process of writer identification, the system must find the right
writer of a handwritten document through a training database. While in the process of
writer retrieval, we must find all handwritten documents similar to a test document.
The key highlights of our study are:
• Study of the performance induced by different patch sizes.

• Investigation of the probable relationship between the language of the dataset and
the performance of diverse patch sizes.
• Study of the impact of the choice of the patch size on the rate of attribution of test
patches to the right writer.
• Test the performance of several patch sizes by conducting experiments on two
Arabic and French datasets.
The content of our paper is presented in four sections. In the following section, we
present the old research works that have been interested in the writer identification,
and we focus on the works that have used deep learning. In Sect. 3, we explain the
methodology adopted as well as the dataset and the CNN used. The representation
of the tests performed will be in Sect. 4. Finally, we end with a brief conclusion.
2 Related Work
Among the earliest work in the field of offline writer identification is that of [18] who
employed the multichannel Gabor filtering technique to identify 1000 test scripts
from 40 writers. He obtained an accuracy of 96.0%. Srihari et al. [20] tried to
identify the writings of 1500 people in the USA by taking the overall characteristics
of their writings such as line separation, slant, and character shapes. Other works
Impact of the CNN Patch Size in the Writer Identification 105
have focused on the use of descriptors based on LBP [12], LPQ [4], GLCM [5],
OBIF [16], or HOG [11]. While other researchers have proposed systems based on
codebook-based graphemes [3] or codebook-based small fragments [19].
AlexNet [15] success in ImageNet competition of large-scale visual recognition
challenge (LSVRC) in 2012 allowed the entry of the era of deep learning in the field
of image recognition. Thus, in 2015, [10] used patches of 56 × 56 to train a neorones
network comprising three convolutional layers and two others fully connected. The
study achieved an accuracy of 99.5% in the ICDAR-2011 database, 88.5% in ICDAR-
2013, and 98.9% in CVL.
Xing and Qia [21] tested a deepWriter with several patches sizes 227 × 227, 131
× 131, and 113 × 113. Their experiments carried out in IAM and HWDB datasets
achieved an accuracy: 99.01% on 301 IAM writers and 93.85% on 300 HWDB
writers.
Yang et al. [22] proposed a deep CNN DCNN with patches size 96 × 96 which
gave results with an accuracy of 95.72% on NLPR handwriting database (Chinese
text) and 98.51% on NLPR (English text).
Christlein et al. [6] used a GMM encoding vector of a CNN layer to identify the
writers of the ICDAR13 and CVL datasets and exemplar SVM for classification.
The same author published another study [7] in which he uses Resnet as CNN and
the cluster indices of the clustered SIFT descriptors of each image as the targets
and the SIFT 32 × 32 patches as input data. The VLAD encoding of activations
of the penultimate layer was considered as local feature descriptors. Classification
with an SVM exemplar gave an accuracy of 88.9% on Historical-WI and 84.1% on
CLaMM16. In [8], the same author conducted experiments in three datasets KHATT,
CVL, and ICDAR13 where he used a Resnet-34 with 32 × 32 patches and the
encoding of the penultimate layer as the local descriptor.
Rehman et al. [17] extract local descriptors from CNN activations features. They
used QUWI as a test database. The width of each image is resized to 1250 pixels
with respect for the width/height ratio. The CNN used is Alexnet with patches of
227×227 and a data-augmentation of the sharped, contoured, sharped contours, and
the negatives form version of these patches. They conduct their experiments using the
outputs of each of the seven layers to achieve scores of 92.78% on English, 92.20%
on Arabic, and 88.11% on Arabic and English.
As we have seen in these works, several sizes of patches were used. So, the question
to ask is what is the best patch size to train a CNN and have higher identification
rates.
We will try to answer this question in this paper by testing several patch sizes for
a Resnet-34 and a bilingual dataset LAMIS-MSHD [9].
3 Proposed Approach
In this section, we present the methodology adopted to verify the impact of patch sizes
on the performance of convolutional networks in the field of writer identification.
106 A. Semma et al.
Fig. 1 CNN classification
This methodology is based on the following points:

• Apply a preprocessing step to different images
• Extract random patches of size 32×32 and save the locations of the patch centers
to take them into consideration when extracting other patch sizes.
• Train the convolutional neural network (CNN) on the patches extracted from the
training images.
• Classify the test images by passing the test patches through the CNN to predict
the corresponding writer.
Figure 1 represents the main steps of the methodology adopted.
3.1 Preprocessing
Before segmenting an image and extracting the CNN input patches, a preprocessing
phase is necessary. In this phase, all the images are converted to gray mode. Then, we
try to correct the skew at the level of each image using the skew algorithm developed
by Abecassis [1].
3.2 Extraction of Patches
Since our current study is based on the types or sizes of patches, then we opt for
seven main sizes (32 × 32, 64 × 64, 100 × 100, 125 × 125, 150 × 150, 175 × 175,
and 200 × 200).
As we know, the CNN input patches can have several sizes in terms of their
width and height, and each type of CNN has a minimum size that must absolutely
be respected, and this depends on the types of layers contained in the CNN. For
example, a max-pooling or average-pooling layer of pool size of 2 × 2 implies that
the size of the inputs of the next layer will be that of the previous layer divided by
two. Knowing that the size of the input of the last convolutional layer must not be
less than (1,1). In our study, we use a Resnet-34 which requires a minimum size of
32 × 32.
So we opted for patches of size greater than or equal to 32 × 32. In our study, we
limit ourselves to square patches where the width is equal to the height.
The patches were extracted randomly the first time for each dataset with the
condition that the center of each patch is a black pixel (containing text). The centers
of each image are then saved to a file.
To be able to make a fair comparison of the performance resulting from the use of
each patch size, we extracted the different patch sizes from the same centers saved
during the first extraction.
We took 400 patches from each image, which gave us 1800 training patches, 200
validation patches, and 400 test patches for each writer.
3.3 CNN Training
In our study, we employ a convolution network which proved its worth in the Ima-
geNet competition of ILSVRC 2015 by taking the first place. The residual networks
are known by residual blocks which allow using the skip connection technique to
skip 2 or 3 layers and thus save the identity of the input of the block. In Resnet-34,
the residual block allows the skip connections of two layers. In the original version
of ResNet-34 [14], the addition of the identity is done before the application of the
activation function, but in the modified version [13] and which is used in our study,
the activation function is applied before the addition of identity.
We trained the CNN with batch sizes ranging from 500 for the patch size 32 × 32
to 20 for the patch size 200 × 200. For the value of the learning rate, we start with a
value of 0.001 which would be divided by 10 after the 6th, 9th and 12th epoch.
3.4 Database
Our study was carried out in the bilingual LAMIS-MSHD database. This dataset
contains 1300 signatures, 600 handwritten documents in Arabic, 600 others in French,
and 21,000 digits. To design this database, 100 Algerians including 57% female and
43% male of different age, and level of education were invited to complete 1300
forms. The main purpose of the database is to provide scientific researchers with a
multi-script platform. The average line height is approximately 139 pixels and 127
for the Arabic and French version, respectively.
To train our CNN we took 75% of the data, for the validation phase, we took 8%
and the rest 17% for the test phase which corresponds to one image per writer.
We can see some samples of the LAMIS-MSHD dataset in Fig. 2
108 A. Semma et al.
Fig. 2 Handwriting samples

from Lamis-MSHD dataset a
Arabic sample, b French
sample
4 Experiments and Results
In this section, we present the results of the experiments carried out. We start with a
presentation of the values obtained from the accuracy and loss of the training patches,
and then, we continue with the results obtained in the test images, after we present
the accuracy and loss of the patches test followed by a description of the various
results obtained.
4.1 Training Patches
As can be seen in Fig. 3 which describes the evolution of accuracy with epochs and
patch sizes, more and more the patch size is increased more and more CNN converges
faster. The CNN trained by 200 × 200 patches of the Arabic version of the Lamis-
MSHD dataset for example reached from the first epoch an accuracy of 60.23% and
ended with an accuracy of 99.70% at the end of the 12th epoch. Unlike the small
Fig. 3 Patch training accuracy for different patch size of Lamis-MSHD Arabic database
Fig. 4 Patch training accuracy for different patch size of Lamis-MSHD French database
32 × 32 patches which reached 18.28% at the first epoch and ended up 48.70% at
the 14th epoch.
The evolution of the accuracy compared to the epochs and different patch sizes of
the French version of the Lamis-MSHD dataset which is represented by Fig. 4 looks
like that described previously for the Arabic version of the Lamis-MSHD.
Since the CNN accuracy converges faster for large patches, then the best values
for the loss parameter are those recorded for large patches. The same observation
can be shared between the Arabic and French version of the LAMIS-MSHD dataset
as can be seen in Figs. 5 and 6.
110 A. Semma et al.
Fig. 5 Patch training loss for different patch size of Lamis-MSHD Arabic database
Fig. 6 Patch training loss for different patch size of Lamis-MSHD French database
4.2 Test Patches
After having trained our CNN on the training patches, we proceed to the test phase.
In this phase, we extract the test patches in the same way as in the training phase with
400 patches per image. This phase allows us to provide us with three main values:
Table 1 Top-1 classification of test images

Patch size Lamis-MSHD Arabic Lamis-MSHD French
32 × 32 93 98
64 × 64 97 98
100 × 100 99 98
125 × 125 99 98
150 × 150 100 98
175 × 175 99 91
200 × 200 40 68
• Top-1 ranking rate for identifying the right writer for test images presented in
Table 1.
• The percentage of test patches that have been assigned to the real writer is presented
by Fig. 7.
• The average probability that the last fully connected layer gives to a test patch in
the classification vector for the correct writer’s box (see Table 2).
As we can see, the best performance for the Arabic version of the LAMIS-MSHD
dataset corresponds to that of the patches of size 150 × 150 where the top-1 ranking
rate is 100% followed by the patches of size 125 × 125 and 175 × 175 with 99%.
While for the French version the best performance regarding the image-level classi-
fication rate is recorded for patch sizes less than or equal to 150 × 150 with a score
of 98%.
The second remark concerns the large sizes 200 × 200 where the classification
rate records very low values with 40 and 68% for the Arabic and French version,
respectively. This shows that for very large patch sizes, the performance of CNN
deteriorates rapidly.
In addition, if we look at the values relating to the probability of assigning test
patches to the right writer and the percentage of test patches that were assigned to the
good writer, we can see that the best scores are recorded for patches of size 125 × 125
for the French version and 150 × 150 for the Arabic version.
As can be seen, the best performance for the Arabic version corresponds to a patch
size of 150 × 150 which is close to the average line height of the Arabic version which
is around 139 pixels. Likewise, the correct values for the French version correspond
to the patch size 125 × 125 which is very close to the average height of the French
version of the LAMIS-MSHD dataset which is approximately 127 pixels.
Another observation can be deduced, it is that in the various tests carried out we
can say that most of the values recorded for the French version of the LAMIS-MSHD
dataset are significantly better than those relating to the Arabic version, especially
for the value of the percentage of test patches attributed to the good writer and the
value of the probability of assigning a test patch to its real writer. This may be due to
the complexity of the Arabic language compared to the French language (see Table 2
and Fig. 7).
112 A. Semma et al.
Fig. 7 Percentage of test patches assigned to the right writer
Table 2 Probability of assigning a test patch to the true writer

Patch size Lamis-MSHD Arabic Lamis-MSHD French
32 × 32 29.43 40.11
64 × 64 57.19 70.11
100 × 100 75.11 82
125 × 125 84.75 87.89
150 × 150 89.24 86.95
175 × 175 88.33 73.61
200 × 200 34.18 58.60
Fig. 8 Average training time of CNN during an epoch

Although the very good scores are recorded for patch sizes between 125 × 125
and 150 × 150, but the execution and training time of the CNN seems to be much
higher for these patch sizes with average times going up to 145 min per epoch against
10 min for patches of size 32 × 32 (See Fig. 8). This shows that we certainly gain
in the performance of CNN, but we lose in terms of execution time. So if we have
very large databases like KHATT which contains 1000 writers or QUWI [2] which
contains 1017 writers and to train a Resnet-34 with patch sizes of 32 × 32 we will
have on average a week of training (for 4 million training patches), while for patches
of size 150, for example, the CNN must train for 14 weeks and with more powerful
machines. So, for larger datasets, we must resize the images and train the CNN with
small-sized patches.
5 Conclusion
In this paper, we tried to verify the impact of the choice of the size of the patches
on the performance of convolutional networks in the field of writer identification.
The best scores were recorded for square patches that have dimensions closer to the
average line height of the dataset manuscripts. Certainly, the study cannot give an
absolute answer about the good size of patches to train all CNN, because we did
not test all types of CNN nor all sizes of patches. But, the study offered an answer
among others to the question raised in the abstract: What size of patch to use to train
a CNN model in order to have the best performance?
The study can be improved by investigating the effect of image resizing on the
performance of CNN in the field of writer identification and with the testing of several
types of convolutional networks.
References
1. Abecassis, F.: Opencv-morphological skeleton. Retrieved from Félix Abecassis Projects and
Experiments: http://felix.abecassis.me/2011/09/opencv-morphological-skeleton/geological
mapping at Cuprite Nevada: a rule-based system. Int. J. Remote Sens. 31, 7 (2011)
2. Al Maadeed, S., Ayouby, W., Hassaïne, A., Mohamad Aljaam, J.: Quwi: an arabic and english
handwriting dataset for offline writer identification. In: 2012 International Conference on Fron-
tiers in Handwriting Recognition, pages 746–751. IEEE (2012)
3. Bensefia, A., Paquet, T., Heutte, L.: A writer identification and verification system. Pattern
Recogn. Lett. 26(13), 2080–2092 (2005)
4. Bertolini, D., Oliveira, L.S., Justino, E., Sabourin, R.: Texture-based descriptors for writer
identification and verification. Expert Syst. Appl. 40(6), 2069–2080 (2013)
5. Chawki, D., Labiba, S.M.: A texture based approach for arabic writer identification and veri-
fication. In: 2010 International Conference on Machine and Web Intelligence, pages 115–120.
IEEE (2010)
6. Christlein, V., Bernecker, D., Maier, A., Angelopoulou, E.: Offline writer identification using
convolutional neural network activation features. In: German Conference on Pattern Recogni-
tion, pages 540–552. Springer (2015)
114 A. Semma et al.
7. Christlein, V., Gropp, M., Fiel, S., Maier, A.: Unsupervised feature learning for writer identifi-
cation and writer retrieval. In: 2017 14th IAPR International Conference on Document Analysis
and Recognition (ICDAR), volume 1, pages 991–997. IEEE (2017)
8. Christlein, V., Maier, A.: Encoding cnn activations for writer recognition. In: 2018 13th IAPR
International Workshop on Document Analysis Systems (DAS), pages 169–174. IEEE (2018)
9. Djeddi, C., Gattal, A., Souici-Meslati, L., Siddiqi, I., Chibani, Y., El Abed, H.:. Lamis-mshd: a
multi-script offline handwriting database. In: 2014 14th International Conference on Frontiers
in Handwriting Recognition, pages 93–97. IEEE (2014)
10. Fiel, S., Sablatnig, R.: Writer identification and retrieval using a convolutional neural network.
In: International Conference on Computer Analysis of Images and Patterns, pages 26–37.
Springer (2015)
11. Hannad, Y., Siddiqi, I., Djeddi, C., El-Kettani, M.E.Y.: Improving arabic writer identification
using score-level fusion of textural descriptors. IET Biometr. 8(3), 221–229 (2019)
12. Hannad, Y., Siddiqi, I., El Kettani, M.E.Y.: Writer identification using texture descriptors of
handwritten fragments. Expert Syst. Appl. 47, 14–22 (2016)
13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceed-
ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778
(2016)
14. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: European
Conference on Computer Vision, pages 630–645. Springer (2016)
15. Hinton, G.E., Krizhevsky, A., Sutskever, I.: Imagenet classification with deep convolutional
neural networks. Adv. Neural Inf. Process. Syst. 25, 1106–1114 (2012)
16. Newell, A.J., Griffin, L.D.: Writer identification using oriented basic image features and the
delta encoding. Pattern Recognit. 47(6), 2255–2265 (2014)
17. Rehman, A., Naz, S., Razzak, M.I., Hameed, I.A.: Automatic visual features for writer identi-
fication: a deep learning approach. IEEE Access 7, 17149–17157 (2019)
18. Said, H.E.S., Tan, T.N., Baker, K.D.: Personal identification based on handwriting. Pattern
Recognit. 33(1), 149–160 (2000)
19. Siddiqi, I., Vincent, N.: Writer identification in handwritten documents. In: Ninth International
Conference on Document Analysis and Recognition (ICDAR 2007), volume 1, pages 108–112.
IEEE (2007)
20. Srihari, S.N., Cha, S.-H., Arora, H., Lee, S.: Individuality of handwriting. J. Forensic Sci. 47(4),
856–872 (2002)
21. Xing, L., Qiao, Y.: Deepwriter: a multi-stream deep cnn for text-independent writer identifica-
tion. In: 2016 15th International Conference on Frontiers in Handwriting Recognition (ICFHR),
pages 584–589. IEEE (2016)
22. Yang, W., Jin, L., Liu, M.: Deepwriterid: an end-to-end online text-independent writer identi-
fication system. IEEE Intell. Syst. 31(2), 45–53 (2016)
Network and Cloud Technologies
Optimization of a Multi-criteria
Cognitive Radio User Through
Autonomous Learning
Naouel Seghiri , Mohammed Zakarya Baba-Ahmed ,

Badr Benmammar , and Nadhir Houari
Abstract Dynamic and optimal management of radio spectrum congestion is

becoming a major problem in networking. Various factors can cause damage and
interference between different users of the same radio spectrum. Cognitive radio
provides an ideal and balanced solution to these types of problems (overload and
congestion in the spectrum). The cognitive radio concept is based on the flexible
use of any available frequency band of the radio spectrum that could be detected. In
the world of cognitive radio, we distinguish two categories of networks, namely the
primary ones, which have priority and control over access to the radio spectrum, and
the secondary ones, called cognitive radio networks, which allocate the spectrum
dynamically. In this paper, we focus on the dynamic management of the radio spec-
trum based on a multi-criteria algorithm to ensure the quality of service (QoS) of the
utilization by secondary users. Our approach is to use a multi-agent system based
on autonomous learning and focused on a competitive cognitive environment. In
this paper, we evaluate the secondary user’s performance in an ideal environment of
cognitive radio systems; we use the multi-agent platform called Java Agent Devel-
opment (JADE), in which we implement a program that applies the multi-criteria
TOPSIS algorithm to choose the best primary user (PU) among several PUs detected
in the radio spectrum. Another paper allows scalability over 100 primary users eval-
uating four different types of technologies, namely voice, email, file transfer and
video conferencing, and a comparison at the end of the convergence time for the
latter technology with results from another paper.
N. Seghiri · B. Benmammar
Laboratory of Telecommunication of Tlemcen (LTT), Aboubekr Belkaid University, 13000
Tlemcen, Algeria
M. Z. Baba-Ahmed (B)
Laboratory of Telecommunication of Tlemcen (LTT), Hassiba Ben Bouali University, 02000
Chlef, Algeria
e-mail: m.babaahmed@univ-chlef.dz
N. Houari
Laboratory of Telecommunication of Tlemcen (LTT), ZSoft Consulting, 75010 Paris, France
e-mail: nadhir.houari@protonmail.com
https://doi.org/10.1007/978-981-16-3637-0_9
118 N. Seghiri et al.
1 Introduction
In the last decade, the number of wireless devices has exceeded the world’s popula-
tion. Billions of devices causing a lot of unused spectrum [1]. A big challenge is to
manage and share the allocated spectrum [2]. Conventional radio systems have not
been able to manage these gaps in the radio spectrum. In contrast, intelligent radio
systems, such as cognitive radio systems, manage the spectrum better.
Cognitive radio was officially introduced in 1998 by Joseph Mitola via a seminar
at Royal Inst. of Technology in Stockholm and later published in an article written
by Mitola and Maguire [3]. A cognitive radio is a programmable radio for automatic
detection of available channels and their flexible use in the radio spectrum [4]. By
combining the two systems, traditional and cognitive, we obtain a spectrum with two
types of users, the primary users who have priority and control over the allocation
of their radio spectrum, and the secondary users who dynamically rent a portion of
the spectrum from the primary users. This is referred to as autonomy.
Autonomous computing is not considered as a new technology, but rather a new
holistic, goal-oriented approach to computer system design that holds promise for
the development of large-scale distributed systems [5]. Autonomous computing,
as the name suggests, is a way of designing mechanisms to protect software and
hardware, whether internal or external, in such a way that they can anticipate threats
or automatically restore their function in the event of unforeseen tampering. It was
first introduced by IBM, and research on autonomic agents and multi-agent systems
is heavily inspired by it [6].
A system of multi-agent is a grouping of agents where each has its own capabilities.
It allows us to build complex systems consisting of different interacting intelligent
agents [7]. Each agent can adopt certain behaviors based on local information to
maximize the overall performance of the system [8].
In this paper, we describe a solution for dynamic spectrum allocation in a multi-
agent environment. Here, a multi-criteria decision analysis algorithm is used to find
out the ideal allocation for a secondary user in a radio spectrum with multiple primary
users. We have chosen the Technique for Order of Preference by Similarity to Ideal
Solution (TOPSIS) algorithm, which consists in choosing the alternative with the
shortest geometric distance to the ideal solution and the longest distance to the
anti-ideal solution.
2 Related Work
A cognitive radio terminal can interact with its radio environment to adapt to it, detect
free frequencies and exploit them. It will have the capabilities to efficiently manage
all radio resources. Current research on cognitive radio is mainly focused on the
improvement of detection, analysis and decision techniques [9]. Several approaches
are proposed to optimize it.
Optimization of a Multi-criteria Cognitive Radio … 119
2.1 Bayesian Nonparametric Approaches for CR

Optimization
The Bayesian approach is based on a random model that represents the importance of
the anterior distribution to generate the posterior distribution using Bayes’ theorem.
In [10], the authors proposed the NOnparametric Bayesian channEls cLustering
(NOBEL) scheme. It allows quantifying channels and identifying multi-channel CRN
quality of service levels. In NOBEL, SU observes the channels and extracts the
characteristics of PU’s channels. Then NOBEL exploits these characteristics and
models them using infinite Gaussian mixture model and Gibbs collapsed sampling
method. NOBEL helps SUs find the optimal channel that meets these requirements.
2.2 Reinforcement Learning-Based Approaches for CR

Optimization
Reinforcement learning is one of the most important machine learning techniques in

which desirable behaviors are rewarded and/or undesirable behaviors are sanctioned.
The paper [11] lists the most recent spectrum allocation algorithms which uses rein-
forcement learning techniques in CR networks. There are six algorithms including
Q-learning, improved Q-learning, deep Q-networks, on-policy RL, policy gradient
and actor-critic learning automata. The algorithms’ focal points and impediments
are analyzed in their specific practical application scenarios.
2.3 SVM-Based Approaches for CR Optimization
Support vector machine (SVM) is a very efficient machine learning algorithm for
classification problems. In the paper [12], the authors proposed to use and eval-
uate SVM-based approaches to appropriately classify free channels in the licensed
frequency bands available in a cognitive radio network, i.e., from the best to the least
optimal characteristics for a secondary SU user to choose the best channel.
2.4 ANN-Based Approaches for CR Optimization
An artificial neural network ANN is a computer system based on the way the human
brain works to learn; it consists of a large set of artificially interconnected neurons.
Researchers of cognitive networks have tried to integrate ANN-based techniques to
dynamically access the spectrum. The authors of [13] proposed a spectrum detection
scheme that uses a neural network to determine whether a PU channel is free or busy.
The proposed scheme uses likelihood ratio testing statistics and energy detection to
train the neural network.
2.5 Game Theoretic Approaches for CR Optimization
Game theory is a mathematical model that has gained considerable importance in

scientific research due to its efficiency and accuracy in modeling individual behavior.
The authors in [13] proposed two new approaches based on game theory to model
cooperative spectrum sensing in a cognitive radio environment. The first scenario is
an evolutionary game model, where SUs have the right to choose whether to cooperate
in spectrum detection or not. The second scenario is the Stackelberg game, where
the fusion center (FC) can intervene in the cooperation process to allocate payments
to SUs in return for their participation in the detection.
3 TOPSIS Method
TOPSIS is a multi-criteria analysis method for decision support. It was introduced

in 1981 by Yoon and Hwang [14]. The TOPSIS main idea is about the geometric
distance from both the ideal and the anti-ideal solution, i.e., the most appropriate
solution is the one with the smallest distance from the ideal solution and the larger
distance from the anti-ideal solution [14]. In our work, we implemented this method
to calculate the ideal choice based on the multiple criteria imposed by our secondary
user. The keyword TOPSIS stands for Technique for Order Preference by Similarity
to Ideal Solution [15].
3.1 Ideal and Anti-ideal Solution
Ideal solution: A* = {g1 *, …, gj *, …, gn *}.

With gj * the best value for the jth criterion among all the shares.
Anti-ideal solution: A = {g1 , …, gj , …, gn }.
With gj the worst value for the jth criterion among all the actions.
3.2 Decision Matrix
TOPSIS assumes that we have m options (alternatives) and n attributes/criteria, and

we have the score of each alternative with regard to each criterion.
Let x ij , x ij score of option i in relation to criterion j. We have a matrix D = (x ij )

matrix of n × m. Let J be the set of benefit criteria or attributes (more is better). Let
J be the set of negative criteria or attributes (less is more) [16].
3.3 The Six Steps of the TOPSIS Algorithm
Step1: Development of the normalized decision matrix

xi j
r i j =
m
i=1 xi2j (1)
i = 1...m j = 1...n
Step2: Development of the weighted normalized decision matrix
v i j = w j * ri j
i =1...m j = 1 ...n (2)
⎡ ⎤ ⎡ ⎤
v11 · · · v1n w1 . r11 · · · wn . r1n
⎢ .. . . .. ⎥ == ⎢ .. .. .. ⎥
V =⎣ . . . ⎦ ⎣ . . . ⎦
vm1 · · · vmn w1 . rm1 · · · wn . rmn
Step3: Calculate ideal and negative-ideal solutions
V j+ = max vi j | j ∈ J1 , min vi j | j ∈ J2 (3)
V j− = min vi j | j ∈ J1 , max vi j | j ∈ J2 (4)
J1: set of benefit criteria.

J2: set of cost criteria.
Step4: Determine the separation measure


n +
Si+ = V j − Vi j i = 1 . . . m, j = 1 . . . n (5)
j=1

n −
Si− = V j − Vi j i = 1 . . . m, j = 1 . . . n (6)
j=1
Step5: Determine the relative closeness to the ideal solution
Si−
Pi* = 0 < Pi∗ < 1 (7)
Si− + Si+
Step6: Ranking the preference order.

• Choose the action with the highest similarity index (choice problem).
• Rank the shares in descending order of similarity indexes (ranking
problem) [17].
4 Proposed Approach
In our approach, there are two types of users, primary users (PUs) and secondary
users (SUs). We have defined a negotiation strategy, one-to-many strategy, where a
secondary user (SU) initiates the negotiation with multiple primary users (PUs). In
our case study, there are ten PUs, as shown in Fig. 1.
The SUs have several specific requirements, such as number of channels, band-
width, technology and price. At the beginning of the negotiation, the SU sends the
first hello request to all PUs. The goal of this first request is to find out which PUs are
available; when available we mean that all PUs that have at least a minimum number
of channels and a minimum bandwidth and the required technology or newer and
the price of better are required by the SU. Once a PU acknowledges the request, it
responds with affirmative response if it has at least the minimum requirements of the
SU, or a negative response if it does not have at least one of the requirements of the
SU.
Once the SU has the list of the PUs that respond to its needs, this is where our
work comes in which is to find the best PU among all PUs. We have chosen to
perform this task using the TOPSIS multi-criteria algorithm. We give the list of PUs
that responded with an acknowledgement and their critters (number of channels,
bandwidth, technology…) as input, and as output we expect the best ideal PU, which
best answers our SU needs.
Fig. 1 Proposed scenario
4.1 Flowchart and Objective Functions of the TOPSIS

Algorithm
The flowchart represents the execution steps of our application. First the detection
phase—the SU detects the environment; once it detects a free part of the spectrum, it
broadcasts the minimum number of required channels to all PUs. Second the decision
phase—SU must select a single PU based mainly on the number of channels. The
PUs receive the broadcasted request with the required number of channels, the PUs
that meet this requirement send an acknowledgement that contains their information,
such as the exact number of channels available, the technology used, etc… On the
other hand, the PU, which does not have the required number of channels, rejects the
request. So, it is a matter of which PU is most ideal for the SU. All this is illustrated
in the flowchart (Fig. 2).
Fig. 2 Flowchart and objective functions of the TOPSIS algorithm
5 JADE Simulation
The simulation was performed under Apache NetBeans IDE 12.0 (Integrated Devel-
opment Environment) using the JADE platform, which contains all the components
for controlling control the SMAs, which are explained in more detail below:
In this first part of the simulation, we decided to define a cognitive agent for
the secondary user named SU and ten primary user agents named PU1 to PU10,
recognized by the same SU. This SU agent will communicate with the ten PUs
simultaneously until it finds a PU that is compatible with its requirements (Fig. 3).
Fig. 3 JADE simulation interface of our approach
5.1 Presentation of the Proposed Approach
Our goal in this work is to improve our autonomous learning system by integrating
an algorithm that helps us choose the best primary user based on multiple criteria.
We chose the TOPSIS algorithm because it is simple, flexible and fast to find the
ideal choice. In what follows, we will present our simulation scenario in which we
implement our flowchart for the scenario of a SU communicating with ten PUs.
First, the SU requests three channels to ensure the QoS of the file transfer and
therefore sends requests to all detected PUs. The PUs that have the required number
of channels inform with an ACL message that contains the number of requested
channels and important information about the price of the allocation, the technology,
the allocated time and the bandwidth to be used. Otherwise, the PU that does not
have the required number of channels rejects the request. In this example, we have
eight PUs responding positively with different proposals ranging from the price of the
Fig. 4 Negotiation between SU and the ten PUs
allocation to the technology and bandwidth used and two PUs responding negatively,
PU2 and PU7 (they do not have the required number of channels) (Fig. 4).
5.2 Negotiation Between Secondary and Primary Users
The TOPSIS algorithm and the choice of the best PU

The multiple positive PU responses confuse the SU, and it cannot decide which of
them is optimal. Precisely, the context of the example is to rank the following PUs
that form the choice of the SU (PU1, PU3, PU4, PU5, PU6, PU8, PU9, PU10) using
the TOPSIS algorithm and based on the four criteria listed below:
Data
The first step is to decide a uniform scale of measurement of the levels (scores) to be
assigned to each criterion relative to the corresponding alternative (PU) by defining
numerical values (1–8) generally in ascending order and the linguistic meaning of
each level (from “Not interesting at all” to “Perfectly interesting”). These values are
used to measure both positive (favorable) and negative (unfavorable) criteria.
The Alternatives X Criteria data matrix is determined by assigning each alternative
the level of each of its attributes based on the previously defined scale.
• For positive criteria (time, technology, bandwidth), the higher the score, the more
positive (favorable) the criterion.
• For the negative criterion (price), the higher the score, the more negative
(unfavorable) the criterion.
For each criterion, a weighting is assigned (a weight that reflects the importance
of the criterion in our final choice). The weights must be defined so that their sum is
equal to 1 and are usually defined in %. Even if the weights are not between 0 and 1,
they can always be reduced to the interval [0, 1] by simply dividing each weight by
the sum of all the weights. The following weights are assigned to the four criteria in
order:
• Allocation time: 0.25.
• Technology: 0.25.
• Bandwidth: 0.2.
• Price: 0.3.
Also giving us an interval of these four criteria as follows:
• Allocation time: [between 1 and 24] h.
• Technology: [3G, 3.5G, 3.75G, 4G, 4.5G, 5G].
• Bandwidth: [144 Mbps–10 Gbps].
• Price [from: 120 up to 300] DA per hour.
After the simulation, we found the results as shown in Table 1.
Figure 5 shows the result of the best and worst primary users sharing the spectrum
with the secondary user, among the ten primary users with different criteria.
In conclusion, here is the ranking in descending order of the eight PUs of the most
satisfactory at least in terms of quality of service for file transfer which is given as
follows:
1. PU1 the most favorable 5. PU10

2. PU6 6. PU5
3. PU8 7. PU3
4. PU9 8. PU4 the least favorable.
We notice that PU2 and PU7 do not have required number of channels to share
the spectrum with our secondary user.
Table 1 PU classified after the use of the TOPSIS algorithm

Alternative name Criteria values
PU1 Channel number = 6, Price = 153, Allocated time (h) = 3, Tech = 5G, Bd
= 10,788.112
= 6602.962
PU8 Channel number = 3, Price = 238, Allocated time (h) = 13, Tech = 3.75G,
Bd = 13.385
= 0.7641
Bd = 5.0514
= 0.635
Bd = 11.056
= 66.441
Fig. 5 Simulation results displayed on the console of the Java program
To further strengthen our study, we have opted for a phased study by upscaling for
deeper and better-quality learning and using four QoS for four different technolo-
gies, namely voice, email, file transfer and video conferencing for a secondary user
communicating with multiple PUs.
80%
Number of best suggestion

70%
for each technology

60%
50%
40%
30%
20%
10%
0%
PU1 PU2 PU3 PU4 PU5 PU6 PU7 PU8 PU9 PU10
Video conference 3% 13% 7% 14% 14% 13% 9% 8% 8% 11%
File Transfer 8% 16% 11% 6% 14% 9% 8% 15% 6% 7%
E-mail 7% 13% 13% 11% 13% 11% 9% 10% 7% 6%
voice 6% 10% 11% 9% 11% 9% 12% 11% 8% 13%
TOTAL 6% 13% 11% 10% 13% 11% 10% 11% 7% 9%
TOTAL voice E-mail File Transfer Video conference
Fig. 6 Best suggestion results for SU choosing between ten PUs out of 100 communication attempts
The scaling is done by 100 communication trials of SU with ten PUs requesting
firstly one channel for voice, secondly two channels for email, thirdly three channels
for file transfer and fourthly four channels for video conferencing to find out which
PU is the most optimal. Figure 6 shows the best proposal results between the SU and
the ten PUs for the four technologies.
A comparison between technologies showed that PU4 and PU5 are better for video
conferencing, while PU2 is better for file transfer; PU2, PU3 and PU5 are better for
email; and PU10 is better for voice. For a global view of all technologies, PU2 and
PU5 are the best.
Figure 7 represents a ranking of the PUs for 100 negotiation attempts of a SU
with ten PUs compared to the different technologies.
Now we come to another contribution, namely the convergence time. Figure 8
shows the average convergence time for 100 communication attempts between a SU
and ten PUs.
Figure 8 shows us the average convergence time required between the SU and the
ten PUs to share the spectrum with. One of them for video conferencing, PU1 has the
best time at 55.84 ms, while PU6 is the last with 96 ms; despite this, all users have
a convergence time <150 ms (a time required by the QoS of the technology used for
video conferencing), a value that ensures the QoS required in the literature [18].
And Fig. 9 shows us the comparison of our contribution with the article [19] in
terms of average convergence time, and we find that the best convergence time of our
contribution is much better and optimal compared to their convergence time which
is fixed at 138.9 ms.
The result obtained allows us to say that our cognitive system is optimized,
precisely:
Fig. 7 Best suggestion ranking for SU choosing between 10 PUs out of 100 communication
attempts
120.00
100.00 93.00 96.00

87.76 87.98
Covergence Time in ms
78.66 79.57 81.38

80.00 72.65
61.98
55.84
60.00
40.00
20.00
0.00
PU1 PU2 PU3 PU4 PU5 PU6 PU7 PU8 PU9 PU10
Fig. 8 Average of the convergence time over 100 communication attempts between the SU and the
different PUs in millisecond
1. The cognitive engine has been strengthened with more suggestion data and a
ranking from the best to the worst PUs, which allows the secondary user to
optimize the planning and decision making in the cognitive cycle, also to make
suggestions to other SUs in case of collaboration.
2. A huge win for convergence time optimized for different technologies and espe-
cially for video conferencing, which allows us to extend our work in the future
for real-time technologies.
Last but not least, the results we found show us that the TOPSIS algorithm is
indispensable for negotiation in a multi-agent system to ensure optimal quality of
service in a cognitive radio environment.
Paper [19]
Fig. 9 Comparison of the average convergence time with other work
7 Conclusion
In this paper, we have studied the autonomous behavior of cognitive agents through
multi-agent systems and see their impact on intelligent networks (cognitive radio)
based on the requirements of the secondary user on multiple criteria to choose the
best offer among those of the primary users.
This new contribution allowed us to develop an algorithm based on multiple
criteria to illustrate the communication of the secondary user with the primary users
by selecting the best offer to allocate a part of the spectrum from them with different
requirements. Subsequently, the simulation results using the platform JADE, which
is remarkably close to the ideal case of cognitive radio, proved to be conclusive
with a significant gain in quality of service in cognitive radio systems. The system
can efficiently exploit the spectrum in an opportunistic and reliable manner. Finally,
the results of our new approach are better and optimized than those found in the
literature.
As a perspective, we hope to further strengthen our system based on security
in cognitive radio systems and using our new multi-criteria approach, to test the
scalability of the system and to implement it on a system based on real signals.
References
1. Tomar, G., Bagwari, A., Kanti, J.: Introduction to Cognitive Radio Networks and Applications,
pp. 124–133. CRC Press (2017)
2. Song, M., Xin, C., Zhao, Y., Cheng, X.: Dynamic spectrum access: from cognitive radio to
network radio. IEEE Wirel. Commun. 19(1), 23–29 (2012)
3. Mitola, J., Maguire, G.: Cognitive radio: making software radios more personal. IEEE Pers.
Commun. 6(4), 13–18 (1999)
4. Jaiswal, M., Sharma, A.K., Singh, V.: A survey on spectrum sensing techniques for cognitive
radio. In: Proceedings of the Conference on ACCS, pp. 1–14 (2013)
5. Lin, P., MacArthur, A., Leaney, J.: Defining autonomic computing: a software engineering
perspective. In: Proceedings of the 2005 Australian Software Engineering Conference
(ASWEC’05), 1530-0803/05. IEEE (2005)
6. Baba-Ahmed, M.Z., Benmammar, B., Bendimered, F.T.M: Spectrum allocation for autonomous
cognitive radio networks. IJACT: Int. J. Adv. Comput. Technol. 7(2), 48–59 (2015)
7. Amraoui, A.: Towards a multi-agent architecture for opportunistic cognitive radio. Ph.D. thesis,
University Abou bekr Belkaid Tlemcen (2015)
8. van der Hoek, W., Wooldridge, M.: Multi-agent systems. In: Foundations of Artificial
Intelligence, vol. 3, pp. 887–928. Bradford Books, Cambridge, MA, USA (2008)
9. Kaur, A., Kumar, K.: A comprehensive survey on machine learning approaches for dynamic
spectrum access in cognitive radio networks. J. Exp. Theor. Artif. Intell., 1–40 (2020)
10. Ali, A., Ahmed, M.E., Ali, F., Tran, N.H., Niyato, D., Pack, S.: NOn-parametric Bayesian
channEls cLustering (NOBEL) scheme for wireless multimedia cognitive radio networks. IEEE
J. Sel. Areas Commun. 37(10), 2293–2305 (2019)
11. Wang, Y., Ye, Z., Wan, P., Zhao, J.: A survey of dynamic spectrum allocation based on rein-
forcement learning algorithms in cognitive radio networks. Artif. Intell. Rev. 51(3), 493–506
(2019)
12. Sarmiento, D.A.L., Viveros, L.J.H., Trujillo, E.R.: SVM and ANFIS as channel selection
models for the spectrum decision stage in cognitive radio networks. Contemp. Eng. Sci. 10(10),
475–502 (2017)
13. Patel, D.K., Lopez-Benitez, M., Soni, B., Garcia-Fernandez, A.F.: Artificial neural network
design for improved spectrum sensing in cognitive radio. Wirel. Netw. 26(8), 6155–6174 (2020)
14. Benmammar, B.: Resource allocation in a cognitive radio network using JADE. Research
Report in Telecommunications, Tlemcen University (2015)
15. Loganathan, J., Latchoumi, T.P., Janakiraman, S., Parthiban, L.: A novel multi-criteria channel
decision in co-operative cognitive radio network using E-TOPSIS. In: Proceedings of the
International Conference on Informatics and Analytics, pp. 1–6 (2016)
16. Bhatia, M., Kumar, K.: Network selection in cognitive radio enabled wireless body area
networks. Digit. Commun. Netw. 6(1), 75–85 (2020)
17. Beg, I., Rashid, T.: Multi-criteria trapezoidal valued intuitionistic fuzzy decision making with
Choquet integral based TOPSIS. Opsearch 51(1), 98–129 (2014)
18. Szigeti, T., Hattingh, C., Barton, R., Briley, Jr., K.: End-to-End QoS Network Design: Quality
of Service for Rich-Media & Cloud Networks. Cisco Press (2013)
19. Baba-Ahmed, M.Z., et al.: Self-management of autonomous agents dedicated to cognitive
radio networks. In: International Conference in Artificial Intelligence in Renewable Energetic
Systems. pp. 372–380. Springer, Cham (2019)
MmRPL: QoS Aware Routing for
Internet of Multimedia Things
Hadjer Bouzebiba and Oussama Hadj Abdelkader
Abstract This paper provides an improved version of the routing protocol for low
power and lossy networks (RPL), called multimedia RPL (MmRPL). This protocol is
proposed as a solution for some restrictions in the RPL storing mode. RPL consumes
much energy when the network size increases which may degrade its performance
and cause a degradation of the network performance, resulting in an inconsistency
in the use of the RPL protocol in constrained internet of multimedia things (IoMT)
networks. IoMT applications can be very demanding in terms of quality of service
(QoS) requirements such as minimum delay, reduced rate of control overhead and
low-energy consumption. The proposed extension overcomes the memory overload
challenge and improves the QoS requirements in IoMT networks. The obtained
simulation results proved that the proposed algorithm outperforms the standard RPL
and another extension of RPL, in terms of control-plane overhead, end-to-end delay
and energy consumption.
1 Introduction
The emerging Internet of Things technology (IoT) consists in large networks of

small-sized low-power embedded devices that can support various applications. The
Internet of multimedia things (IoMT) is an extension of the IoT, considered as an
interconnection of multimedia objects in which they are able to acquire, process
and display multimedia contents from and to the real world using specific input and
output devices [1]. With the growth of this type of smart devices, the tendency of
IoMT applications has recently became much more important by getting involved
H. Bouzebiba (B)
STIC Laboratory, University of Tlemcen, 13000 Tlemcen, Algeria
e-mail: hadjer.bouzebiba@univ-tlemcen.dz
O. Hadj Abdelkader
Faculty of Engineering, SYSTEC - Research Center for Systems and Technologies, University of
Porto and Institute for Systems and Robotics, 4200-465 Porto, Portugal
e-mail: hadjabdelkader@fe.up.pt
https://doi.org/10.1007/978-981-16-3637-0_10
134 H. Bouzebiba and O. Hadj Abdelkader
in many fields such as smart homes, smart health, smart vehicles, smart cities ...
etc. This diversity may bring new stringent requirements than the IoT environment,
mostly due to the increase of multimedia content in the network.
For communication purposes, IoT devices use some standardized routing proto-
cols to manage the data transfer within their network. The routing protocol should
be reliable, energy efficient and mainly scalable. As mentioned in [2], a good routing
protocol in this context should extend to satisfy the requirements of large network
size and densities in spite of the resource constraints of wireless sensor networks
(WSNs). The internet engineering task force working group (IETF) Routing over Low
power and Lossy networks (ROLL) have proposed a routing protocol named Rout-
ing Protocol for Low power and Lossy Networks (RPL) [3]. RPL meets the specific
requirements of low power and lossy networks (LLN). For this latter, RPL has rapidly
became the standard routing protocol for IoT. RPL uses a tree approach to organize
the network into a tree structure which allows the data to flow in both upward and
downward directions. In the downward direction, the traffic flow from the root toward
its associated nodes which spoils the scalability of RPL significantly [4]. Moreover, in
storing mode, each RPL router has to store routes for destinations in its sub-DODAG
[5]. This feature generates a limiting factor in RPL, which is the amount of available
memory to store the neighbors in the routing table. Thus, each close node to the root
is obliged to store the routing state for almost the entire destination oriented acyclic
graph (DODAG), which can be challenging for resource-constrained devices [6].
To the best of our knowledge, the standard RPL does not deal with the case where
a parent node cannot accept to add a new downward route when its routing table
is full, this problem can happen in scalable networks such as smart cities. When
the new scalable nodes want to join network, it will consume buffer memory, while
exchanging the control messages besides of the necessary amount of energy taken
during this process. This will negatively impact on the network performance in terms
of delay and energy consumption without the establishment of any proper routing.
In this paper, we propose an RPL extension named MmRPL to tackle the problem
of insufficient storage memory and to optimize the protocol for IoMT networks. The
proposed MmRPL reduces the amount of control-plane overhead by checking the
memory of nodes once having a new connection in the network. The remainder of
this paper is structured as follows. Section 2 provides an overview of RPL protocol.
Section 3 presents a description of the problem statement. Section 4 summarizes
the related work. In Section 5, the proposed solution is discussed. In Section 6, the
performance of MmRPL is evaluated and the results are discussed in comparison with
RPL. The final section concludes the paper and gives some hints about future works.
2 RPL Overview
The RPL [3, 7] represents a distance vector routing protocol, in which a DODAG
based on sets of metrics is constructed. The final destination node in the DODAG
is called the root (as Low Power and Lossy Border Routers (LBR)). The latter acts
MmRPL: QoS Aware Routing for Internet of Multimedia Things 135
like a bridge between the LLN and Internet. RPL supports different types of traffic
: Multipoint to Point (M2P), Point to Multipoint (P2M) and Point to Point (P2P).
Each node in the DODAG is characterized by a rank value, which represents its
distance towards the root node calculated according to an objective function. The
OF [8, 9] determines the rank of each node based on one or more metrics and
selects the optimal route in a DODAG. The DODAG is constructed from the root
by broadcasting ICMPv6 control message (the DODAG Information Object (DIO))
to its neighborhood. This kind of messages contains some configuration parameters
(such as the DODAG roots identity, routing metrics, as well as the rank) needed to
build the topology. Once DIO message is received by neighboring nodes, the node
joining the DODAG will: (1) add the sender prefix to its candidate parent list; (2)
calculate its rank; (3) select the closest node to the root from its candidate parents
which acts as its next hop (preferred parent) toward the root and (4) update the
received DIO message with its own rank and repeat the same procedure until all the
network’s nodes will have an upward route toward the root.
After the construction of the upward route, another type of ICMPv6 control mes-
sages named as destination advertisement object (DAO) is used to build the downward
routes. DAO messages are unicasted to the node’s preferred parent for nodes, which
have already joined the DODAG and want to advertise one or more reachable desti-
nation prefixes including their own. RPL affords two operation modes for downward
routes: the storing and the non-storing mode. The storing mode requires for each
parent that receives a DAO message from one of its children, to store its prefix (DAO
sender address) in its routing table as a next hop prefix. Besides, the DAO receiver
can optionally acknowledges the DAO sender using the DAO-ACK message (DAO
Acknowledgement). The parent by its turn forwards the received DAO to its own
preferred parent repeating the same process until the DAO achieves the DODAG
root. The RPL’s network structure and construction have been illustrated in Fig. 1.
3 Related Work
Since RPL has been considered as an open standard by IETF community, many
enhancements have been proposed to improve its performance [4, 11, 12].
Authors in [11], proposed an improved version of RPL protocol named Enhanced-
RPL that treats the problem of unreachable destination caused by storage limitations
of certain node’s preferred parent. In RPL, when a child node wants to announce its
prefix in the downward routing, it should unicast a DAO message to its preferred
parent, but when the parent’s routing table is full, it will not accept any additional
nodes. This can happen in scalable networks when the number of nodes increases.
As a solution to this problem, in the proposed Enhanced-RPL, a list of candidate
parents was proposed to the child node that has lost the chance to announce itself to
its preferred parent.
Authors in [12] suggested an extension of RPL named MERPL which reduces the
memory consumption by improving the storing mode scalability. In this extension,
Fig. 1 RPL structure [10]
whenever a node reaches a predefined number of N in its routing table, it must

delegate one of its children (which has the highest number of routing table entries)
to play the role of a storing node. Then, the delegating node (parent) must delete all
entries that are related to the child (new storing node) from its routing table. After
that the new reachable destinations selected by the child will be advertised to the
root in a DAO message.
Further, various optimizations of the RPL protocol which take the QoS as routing
metrics have been studied, such as [13–15] which aimed to satisfy a specific QoS.
For example in [13], authors proposed a compliant OF which allows the RPL to
support the multi-instance approach, called OFQS. The OFQS has focused on multi-
objective metrics taking in consideration the following metrics: delay, remaining
energy of the nodes’ battery and quality of link. In [14], the authors measured the
RPL performance in terms of the QoS parameters such as the packet delivery ratio,
the network convergence time, the remaining energy, the latency and the total amount
of overhead. In the same context, in [16], the authors choosed the fuzzy systems in
order to design a new set of contextual OFs for RPL. They specified a set of sensitive
functions “Delivery Quality and Context” (DQCA-OF), by combining alternately the
metrics: ETX, energy consumed and number of hops (HC). Thus, a set of OFs can
be dynamically chosen according to the QoS requirements of the IoT application.
Moreover, in [17], authors proposed an enhanced version of the RPL routing protocol
by providing a new QoS aware OF based on free bandwidth named as FreeBW-RPL.
The proposed OF works on distributing the traffic load in the network by taking, in a
dynamic way, multiple paths for routing the heavy amount of traffic, while creating
an energy balanced routing paths. Another work which aims to perform on a multi-
path parent selection is the on-demand selection (ODeSe) algorithm suggested in
[18]. This algorithm focused on the dynamic conditions at the packet forwarding
time. This work implements the packet automatic repeat request, replication and
elimination and overhearing (PAREO) [19] functions in order to improve both of
reliability and energy efficiency.
A particular RPL improvement has been proposed in [20] called energy effi-
cient optimal parent selection (EEOPS–RPL), in which authors have used the firefly
optimization algorithm in order to extend the lifespan of the IoT network. This algo-
rithm calculates the current location of firefly (each node is considered as a firefly),
attraction of the fireflies, random function, velocity and the global best values in the
network. Thus, during the data transmission, the distance is viewed as the movement
parameter for choosing the optimum parent in the DODAG. The firefly algorithm
offers a fast convergence, while choosing the optimal parent, reduces the packet loss
during the route establishment and expands the lifespan of the entire network.
In any routing protocol, scalability represents an important feature bearing a direct
impact on network’s performance and reliability. In this regard, RPL did not specify
any action to do when a node’s routing table is full and still receiving some solicitation
(DIS or DAO messages) from new nodes or from an already joined node. However,
the aforementioned research did not take in consideration the unnecessary route
establishment and reduction of overhead in IoMT networks. In addition, the routing
protocol should account for the performance of the multimedia communications
within IoMT network in terms of end-to-end delay, overhead and rate of packets
delivered. Moreover, we argue that the problem of scaling with network density
has not been sufficiently analyzed, especially the optimization of the total network
overhead in order to extend the RPL protocol to large data stream routing such as
multimedia data routing. The protocol MmRPL proposed in this paper is dedicated
for these issues especially the overloaded case of an IoMT node.
4 Problem Statement
IoMT’s applications require the satisfaction of many QoS constraints such as mini-
mum amount of overhead, limited delay and energy. However, in the storing mode
of the RPL protocol, nodes may exhaust their memory easily since each node is
required to store the routing information about its sub-DODAG which may lead
to a storage limitation of neighbors and routing tables. In consequences, this lat-
ter represents a more serious problem especially in a scalable network. Generally,
nodes that are around the root are more likely to run out their memory, especially in
large networks such as IoMT; as presented in [12] these nodes may not possess the
required large memory resources. This issue figures as a challenging especially for
the resource constrained devices. Besides, RPL does not provide any action to take
when a parent rejects to install a new downward route (i.e., case of an overloaded
routing table). These issues can impact negatively on the QoS of the IoMT network.
Fig. 2 RPL’s memory saturation case in downward route
Another remaining problem that will obstruct the process, of RPL, is the amount of
overhead exchanged in the network.
4.1 Case Study
In this subsection, we have defined a sample case study of a scalable network with
limited memory storage and resources constrained nodes. In many cases, some of
the node’s preferred parent runs out of its routing table storage such as the node B
depicted in Fig. 2 which has a memory capacity of three routing entries per node.
After that some new nodes (such as node F) want to join the network, at that time
the preferred parent (node B) will try to add its routing entry for a specific target
rendering. Consequently, the announced target will be unreachable and the entire
packet destined for it will be lost. This will also increase the energy consumption.
In order to overcome this issue, we illustrate in Fig. 3 the same network with a pro-
posed solution which aims to stop sending the control messages and the unnecessary
energy loss. Once node F solicits node B (overloaded memory node), this later stops
interacting with any new node till it succeeds to manage its routing table or the node
F finds another reachable parent node.
Fig. 3 The proposed solution by MmRPL
5 The Proposed Solution
Our proposed MmRPL takes into account the scalability of the network in the case
of memory saturation of a new node, while reducing the total overhead. The main
objective focuses on the economy of the number of control messages knowing that
it will also reduce the amount of energy which reflects more benefits for a such
constrained network as IoMT. Our purpose is to avoid the unnecessary messages in
the case of saturation, as depicted in Fig, 4, once the parent node receives a D I S
control message, it will check the memory of its neighboring table. If there is much
space to let the new node join, the RPL’s process will continue. Otherwise, in the
case of overloaded memory, the proposed scheme will prohibit the parent node from
exchanging any control messages which saves its energy until this latter manages to
clear some space in its routing table. Alternatively, the new node will find its new
way to another reachable parent. The use of MmRPL would reflect positively on the
performance of the IoMT network.
6 Performance Evaluation and Discussion
In this section, we explore the performance of MmRPL in scalable network in terms

of control overhead, energy consumption and end to end delay.
We evaluate the MmRPL performance using Cooja emulator [21] which allows us
to test the binary code that could run on real T elos B WSN nodes, besides, it offers
for us a full flexibility to evaluate radio environments and topologies. We have fixed
Fig. 4 Diagram of the proposed solution for the RPL memory saturation case in downward route
the routing table entries N to 3 in order to easily clarify the saturation case. Moreover,
ContikiMAC [22] is employed as the underlying duty cycle MAC layer. Regarding
the radio propagation model, we used the Unit Disk Graph Model: Distance Loss
[23]. Each node sends an application data packet every minute to the root.
All the log files (trace files) of all experiments are analysed by a Perl script in order
to extract the statistical results. The main simulation parameters are summarized in
Table 1.
6.1 Performance Metrics
• Average Power Consumption (APC) In order to compute the energy consumption

of the network, we use a novel technique of the software named Powertrace which
is a tool available in Contiki [24]. This tool tracks the power state by estimating
the energy consumption for CPU processing, packet transmission and listening.
The power trace mechanism provides CPU, low-power mode (lpm), radio listen
and radio transmit values which gives the total time consumed in each power trace
interval. Moreover, we take the average percent radio on time of all the nodes in
the whole network setup. The average power consumption (APC) is calculated
Table 1 Simulation parameters

Parameter Value
Simulation time 500s
Number of nodes [5–20]
Objective function OF0, MRHOF
Routing metric Overhead, Energy consupmtion
DIO, DAO, DIS packet size 30 bytes
Data packet size (50, 100) Bytes
Send interval 1 pckt/min
Ranges of nodes Tx: 50 m , Interference: 100 m
Radio duty cycle ContikiMAC
Radio propagation model Unit disk graph model
according to Eq. 2.
n
i=0 (LPM + CPU + RadioListen + RadioTransmit)
APC = (1)
n
where n represents the number of nodes in the network.
• End-to-end delay The end-to-end delay (Packet Delay) is another evaluation
metric used for measuring the total time taken by packets in order to be successfully
delivered from node to sink.
• Total Overhead The Overhead in RPL is calculated by the Eq. 2.

Overhead = (DIS + DIO + DAO + DAOACK) (2)
6.2 Results Discussion
In Figs. 5a–c and 6a–c, we plot our simulation results after improving the RPL
protocol by reducing the amount of total overhead circulating in the network during
both of the construction and the routing process.
We notice that as long as new nodes join the network (the number of nodes
increases) the performance of MmRPL is showing on by reducing the amount of
overhead around 2000 packets when the network reaches 20 nodes as shows Fig. 5b.
By consequence, the amount of energy consumed by MmRPL network is less than
the one consumed by the conventional RPL as depict all of Fig. 6. At the meantime,
MmRPL results in the same benefits of the total amount of overhead in Fig. 5c when
we tried to expand the simulation time to 14450 seconds in order to observe the
variation of the overhead. As well, we have noticed a stabilization of the average end
to end delay of MmRPL, which takes less delay than the standard RPL, as depict
Fig. 5 Overhead results
Fig. 6. In Fig. 6, we plot the radio transmit power of the saturated node, preferred
parent represented as node 4 in Fig. 3 against time where our proposed MmRPL
outperforms the conventional RPL.
6.3 Comparative Results
This subsection investigates some further analysis about our work by comparing it
with another RPL enhancement based on a combination of multiple sinks support
named MRRD+ (Multiple, RSSI, Rank and dynamic) [25] as illustrated in Table 2.
The first analysis highlighted the impact of the number of nodes on the total
number of control packets. After comparing our simulation parameters (Data packet
size: 50–100 Bytes; Radio propagation model: Unit Disk Graph Model; Simulation
Time: 500 s) with the ones used for the MRRD+ in [25], (Data packet size: 30 Bytes;
Radio propagation model: MRM with random behavior; Simulation Time: 300 s),
respectively, we have noticed that when the network density is equal to 20 nodes,
the MmRPL shows an improvement in terms of the number of control packets. In
another finding, the MmRPL revealed remarkable results better than (MRRD+)-1S,
(MRRD+)-2S, (MRRD+)-3S and (MRRD+)-4S in terms of the average end-to-end
delay. Specifically, in 20 nodes, MRRD+ delay varies from (20–45 ms), while our
Fig. 6 Energy consumption results
Table 2 Comparative results between MmRPL and MRRD+

Routing protocol Performance parameter
Average end to end delay (ms)
MmRPL Network size = 20 12
MRRD+ (1S, 2S, 3S, 4S) 20–45
Routing protocol Performance parameter
Number of control packets
MmRPL Network size = 20 1800
MRRD+ (1S, 2S, 3S, 4S) ≈2200
MmRPL keeps a stable delay consumption for 12 ms, knowing that the DIS sending
interval varies between 20 and 60 DIS message per second.
Even in the presence of existing differences between our simulation parame-
ters and the comparative work’s ones, especially in what concerns the mobility, the
comparison of our work results against the aforementioned work have been very
promising. Thus, the results show that the proposed MmRPL would represent a
good solution for the memory problem, and it reduces the total amount of overhead
in the network and satisfies the QoS requirements of the IoMT networks.
7 Conclusion and Future Work
In this paper, we proposed an extension of the RPL routing protocol named as

MmRPL which satisfies some of the QoS of the IoMT networks. The proposed
MmRPL allows to minimize the total control messages once detecting an event of
node’s memory saturation by avoiding the exchange of unnecessary messages spe-
cially when a new node wants to join the network by soliciting the saturated node
in its vicinity. In this situation, reducing the amount of overhead will help us to
reduce the energy consumption during the exchange process. Besides, the MmRPL
optimizes another IoMT QoS which is the end to end delay taken from the DODAG
construction till the end of the routing process. In future, we will study the validation
of the proposed MmRPL protocol in real embedded IoMT system and test it in large
networks.
References
1. Floris, A., Atzori, L.: Quality of experience in the multimedia internet of things: definition and
practical use-cases. In: 2015 IEEE International Conference on Communication Workshop
(ICCW), pp. 1747–1752. IEEE (2015)
2. Kim, E., Kaspar, D., Gomez, C., Bormann, C.: Problem statement and requirements for 6lowpan
routing. In: Draft-IETF-6LoWPAN-routing-requirements-04, IETF Internet Draft (Work in
Progress) (2009)
3. Winter, T.: RPL: IPv6 routing protocol for low-power and lossy networks (2012)
4. Kiraly, C., Istomin, T., Iova, O., Picco, G.P.:. D-RPL: overcoming memory limitations in RPL
point-to-multipoint routing. In 2015 IEEE 40th Conference on Local Computer Networks
(LCN), pp. 157–160. IEEE (2015)
5. Clausen, T., Herberg, U., Philipp, M.: A critical evaluation of the ipv6 routing protocol for low
power and lossy networks (RPL). In: 2011 IEEE 7th International Conference on Wireless and
Mobile Computing, Networking and Communications (WiMob), pp. 365–372. IEEE (2011)
6. Iova, O., Picco, P., Istomin, T., Kiraly, C.: RPL: the routing standard for the internet of things...
or is it? IEEE Commun. Mag. 54(12), 16–22 (2016)
7. Gaddour, O., Koubâa, A.: RPL in a nutshell: a survey. Comput. Netw. 56(14), 3163–3178
(2012)
8. Thubert, P.: Objective function zero for the routing protocol for low-power and lossy networks
(RPL) (2012)
9. Gnawali, O.: The minimum rank with hysteresis objective function (2012)
10. Safaei, B., Hosseini Monazzah, A.M., Shahroodi, T., Ejlali, A.: Objective function: a key con-
tributor in internet of things primitive properties. In: 2018 Real-Time and Embedded Systems
and Technologies (RTEST), pp. 39–46. IEEE (2018)
11. Ghaleb, B., Al-Dubai, A., Ekonomou, E., Wadhaj, I.: A new enhanced RPL based routing for
internet of things. In: 2017 IEEE International Conference on Communications Workshops
(ICC Workshops), pp. 595–600. IEEE (2017)
12. Gan, W., Shi, Z., Zhang, C., Sun, L., Ionescu, D.: MERPL: a more memory-efficient storing
mode in RPL. In: 2013 19th IEEE International Conference on Networks (ICON), pp. 1–5.
IEEE (2013)
13. Nassar, J., Berthomé, M., Dubrulle, J., Gouvy, N., Mitton, N., Quoitin, B.: Multiple instances
QoS routing in RPL: application to smart grids. Sensors 18(8), 2472 (2018)
14. Joseph Charles , A.S., Kalavathi, P.: QoS measurement of RPL using cooja simulator and
wireshark network analyser (2018)
15. Zier, A., Abouaissa, A., Lorenz, P.: E-RPL: A routing protocol for IoT networks. In: 2018 IEEE
Global Communications Conference (GLOBECOM), pp. 1–6. IEEE (2018)
16. Da Silva Araújo, H., Rodrigues, J.J.P.C., De Al Rabelo, R., De C Sousa, N., Filho, C.C.L.S.J.,
Sobral, J,V.V., et al.: A proposal for IoT dynamic routes selection based on contextual infor-
mation. Sensors (Basel), 18(2), 353 (2018)
17. Bouzebiba, H., Lehsaini, M.; FreeBW-RPL: a new RPL protocol objective function for internet
of multimedia things. Wirel. Pers. Commun. 1–21 (2020)
18. Jenschke, T.L., Koutsiamanis, R.-A., Papadopoulos, G.Z., Montavont, N.: ODeSe: on-demand
selection for multi-path RPL networks. Ad Hoc Netw. 102431 (2021)
19. Koutsiamanis, R.-A., Papadopoulos, G.Z., Jenschke, T.L., Thubert, P., Montavont, N.: Meet
the PAREO functions: towards reliable and available wireless networks. In: ICC 2020-2020
IEEE International Conference on Communications (ICC), pp. 1–7. IEEE (2020)
20. Sennan, S., Somula, R., Luhach, A.K., Deverajan, G.G., Alnumay, W., Jhanjhi, N.Z., Ghosh,
U., Sharma, P.: Energy efficient optimal parent selection based routing protocol for internet
of things using firefly optimization algorithm. Trans. Emerging Telecommun. Technol. e4171
(2020)
21. Osterlind, F., Dunkels, A., Eriksson, J., Finne, N., Voigt, T.: Cross-level sensor network simula-
tion with COOJA. In: Proceedings. 2006 31st IEEE Conference on Local Computer Networks,
pp. 641–648 (2006)
22. Dunkels, A.: The ContikiMAC radio duty cycling protocol (2011)
23. Clark, B.N., Colbourn, C.J., Johnson, D.S.: Unit disk graphs. Discrete Math. 86(1–3), 165–177
(1990)
24. Dunkels, A., Eriksson, J., Finne, N., Tsiftes, N.: Network-level power profiling for low-power
wireless networks. Powertrace (2011)
25. Wang, J., Chalhoub, G.: Mobility support enhancement for RPL with multiple sinks. Ann.
Telecommun. 74(5), 311–324 (2019)
Channel Estimation in Massive MIMO
Systems for Spatially Correlated
Channels with Pilot Contamination
Mohamed Boulouird , Jamal Amadid , Abdelhamid Riadi ,

Abstract This work treats multi-cell (M-C) multi-user (M-U) massive MIMO (M-
MIMO) systems taking into consideration pilot contamination (PC), where Rayleigh
fading channels are correlated in the spatial domain. An appropriate exponential cor-
relation (EC) is using as an approximation model for uniform-linear arrays (Un-LA).
The statistics of minimum mean square error (MMSE), element-wise MMSE (EW-
MMSE), approximate MMSE (Approx.MMSE), and least-squares (LS) estimators
are evaluated and analyzed. The Approx.MMSE estimator uses an imperfect covari-
ance matrix (CM), which relies on sample CM to estimate a true CM provided by the
MMSE. Analytical NMSE formulas for idealistic and realistic CMs are presented
and interpreted. An analytical normalized mean square error (NMSE) formula is also
given for EW-MMSE.
1 Introduction
The M-MIMO technology offers such enhancement in spectral efficiency (S-E) and
energy efficiency (E-E) using spatial multiplexing (S-M) and a high channel gain,
respectively [1, 2]. The literature deals with two families of scenarios for a M-MIMO
systems. The first scenario where the channels are independent [3–5], in which the
M. Boulouird (B)
Smart Systems and Applications (SSA) Group, National School of Applied Sciences
of Marrakesh (ENSA-M), Cadi Ayyad University, Marrakesh, Morocco
e-mail: m.boulouird@uca.ac.ma
J. Amadid · A. Riadi · M. M. Hassani
Instrumentation, Signals and Physical Systems (I2SP) Group, Faculty of Sciences Semlalia,
Cadi Ayyad University, Marrakesh, Morocco
e-mail: jamal.amadid@edu.uca.ac.ma
A. Riadi
e-mail: abdelhamid.riadi@edu.uca.ac.ma
M. M. Hassani
e-mail: hassani@uca.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_11
148 M. Boulouird et al.
expression of channel coefficient is composed of two factors, namely large-scale

fading (LSF) and small-scale fading (SSF), the first is including both path loss and
shadowing, while the second factor is including statistics variations of the channels.
In the second scenario, the channels are spatially correlated (SC) [6–11]. Where the
channel CM describes the correlation between antennas in spatial domain. In reality,
in certain cases channels are dependents (i.e., SC) [6, 12, 13]. That is to say, in certain
measure the element of the channels is correlated. The property of the channel CMs
is to make sense of the spatial correlation among antennas. Where the off-diagonal
elements of CMs usually are nonzero. We consider in our studies SC channels for a
square system M-C M-U M-MIMO under PC.
1.1 Related Works
In [14], the authors have tackled SC channels and their greatness over a system
that performs with M-MIMO technology since SC channels describe more practical
channels and draw a real propagation environment. They present up-to-date results
regarding the core boundaries of M-MIMO, which are not primarily a matter of pilot
contamination rather the capacity to get statistics of the channels. Hence, this result
gives rise to an updated version of M-MIMO, namely M-MIMO 2.0.
The authors in [15] have proposed an approximate model for SC channels for
two arrangements known in the literature as the Un-LA and uniform circular array
employing the Laplacian distribution. A metric for evaluating the performance has
been proposed by which they evaluate the performance of the proposed model. Fur-
thermore, under this metric, the proposed model works best in small-angle spreading
situations.
The authors in [16] discussed the SC channel in the up-link stage, where the
BS has a large number of antennas. Besides, they prolong the LSF concept to SC
channels. For which they have developed a signal-to-interference ratio plus noise
formulation that relies only on SSF factors.
In [6], the authors have dealt with different distributions that describe correlation
among channels to investigate and evaluate the influence of spatial correlation using
either Gaussian, uniform, and Laplacian distribution. The Gaussian distribution is
known as the local scattering model, whereas the uniform distribution is known as
the one-ring model. Besides, they consider that the system operated under a higher
PC level. Furthermore, in each scenario (i.e., for each distribution), they have ana-
lyzed the channel estimation quality using three estimators, namely LS, MMSE,
and EW-MMSE employing the MSE metric. This evaluation is performed using the
effective signal-to-noise ratio, where the best performance is achieved using the local
scattering model (i.e., Gaussian distribution).
Channel Estimation in Massive MIMO Systems . . . 149
1.2 Contributions
The contributions of this work are given in twofold as follows :

1. This work introduced the concept of spatially correlated channels (i.e., that model
a more realistic channel environment ) on ULA using the exponential correlation
model and the influence of this concept on channel estimation.
2. An EW-MMSE channel estimator is proposed, which performs similarly to
MMSE with lower complexity and shown results better than the Approx.MMSE
even if the number of realizations is large.
1.3 Work Scheduling
The rest sections of this work are arranged as follows: In Sect. 2, the system model is
defined including the expression of the received signal and the setup used for CMs. In
Sect. 3, the NMSE expressions are given and analyzed for all used channel estimator.
In Sect. 4, the simulation results are given, in which the efficiency of our proposed
method is evaluated and compared to existing methods. We finished our work by
giving a conclusion in Sect. 5.
2 System Model
Our studies treat a square M-C scheme, where each cell has a BS in the corner and
K user with single antenna in the cell-edge area [17], as depicted in Fig. 1.
All channels are considered as correlated Rayleigh fading (CRF) channels. The
hjlk exemplify vector of the channel from the k-th user in the j-th cell to the N
antennas at l-th BS, which define as hjlk = [hjlk1 , hjlk2 , . . . , hjlkN ]T ∼ CN(0N , Rjlk ).
Where Rjlk ∈ CN×N display a positive semi-defined channel CM. It is remarkable
to underlining that Rijk is not an identity matrix, which characterizes macroscopic
effects including path loss in various directions and channel correlation in the spatial
domain.
The EC model is studied in our work for a Un-LA referring to Bjornson work
[17], with which we modeled the correlation between contiguous antennas. The
inter-antenna correlation can be represented by
[R]m,n = βr |n−m| eiϕ(n−m) 10(νn +νm )//20 (1)
β, ϕ and r are respectively, the LSF coefficient, angle of arrival, correlation co-
efficient/factor. While ν1 , . . . , νM ∼ CN (0, σ 2 ) afford independent LSF variations
through the array. Otherwise, the users in all cells send their Uplink (UP) pilot
sequence (PS) simultaneously. The PSs used in each cell are duplicated in all other
Fig. 1 Square M-C setup
cells (i.e., leading to the PC problem). That is to say, the frequency factor is one.
φk ∈ Cτ is the PS sent by k-th user, where τ is the length of each PS, φk have as
property the following equation φkH φk = 1, ∀k. For all K users in each cell, the global
matrix τ × K of the PS becomes H = IN , ∀k, where = [φ1 , φ2 , . . . , φK ]. The
received UP signal Yj at j-th BS can be defined as
√
L
Yj = q Hjl H + Wj (2)
l=1
where q represents UP transmit power (TP) and WjN ×τ is a noise matrix, each element
of WjN ×τ follow a CN (0, 1).
3 Channel Estimation
3.1 LS Channel Estimation
In general, an estimate for the channel vector hjjk at the j-th BS using LS channel
estimator through adequate statistics [18] is defined as follows
1 L
jjk = θjk = √ Yj φk =
ĥLS hjlk + wjk (3)
q
l=1
Here the vector wjk is the multiplication between the noise matrix and PS of k-th
user divided by square root of pilot power as shown in the following equation
1 1
wjk = √ Wj φk ∼ CN (0N , IN ) (4)
q q
jjk ∼ CN (0N , jk ),
where ĥLS jk represent the covariance of LS channel estimate
L
expressed as jk = l=1 Rjlk + q1 IN . While the error of estimation is represented
jjk = hjjk − ĥjjk which follows CN (0N ,
by h̃LS − Rjjk ). It is dependent of both ĥLS
LS
jk jjk
ĥLS h̃LS jjk , h̃jjk ) = jk − Rjjk .
is determined by Cov(ĥLS LS
and the CM of jjk ;
where jjk
Therefore, the NMSE is computed in accordance with each antenna using LS
estimator, and it is formulated as follows
1 1
εjkLS = jjk − hjjk } =
E{ĥLS 2
Tr[ jk − Rjjk ] (5)
N N
We can notice that the NMSE of LS estimator does not rely on any previously statistics
of the channel (i.e., LSF coefficients). Besides, LS estimator is a linear estimator that
has low complexity and large NMSE compared to MMSE estimator [5]. In the rest
of this subsection, some notes for LS channel estimation are given as follows:
Note 1: If each element of hjjk for all cells, BSs and all users in each cell fol-
lows CN (0, 1). Hence, the CM in this case is a diagonal matrix represented as
follows jk = ψjk IN = ( Ll=1 βjlk + q1 )IN and Rjjk = βjjk IN . Thus, εjkLS = ψjk − βjjk
and βjjk = N1 Tr[ jk ].
Note 2: if hjjk ∼ CN (0, 1), q tends to infinity and under PC, the NMSE of LS

estimator is: εjkLS −→ Ll=1 βjlk when q −→ ∞.
jjk = ĥjlk (i.e., the vectors

Note 3: According to Eq. (3), we can easily remark that ĥLS LS
are parallel). Thus, the BS cannot separate these channels (the channels from the
users that have the same PSs).
3.2 MMSE Channel Estimation
The MMSE channel estimation is an estimator which relies on statistics of the channel
and is among Bayesian estimators category [18]. To estimate the channel hjjk , the
MMSE estimator is determined by
Rjjk −1
MMSE
ĥjjk = θjk = Rjjk jk θjk (6)
jk
For the Gaussian model, MMSE channel estimator has a special property repre-
MMSE
sented by the independence between the vector estimate ĥjjk and the estimation
MMSE
error h̃jjk = hjjk − ĥjjk
MMSE MMSE
, which are randomly distributed vectors as ĥjjk ∼
−1 −1
CN (0N , Rjjk jk Rjjk ) and h̃jjk
MMSE
∼ CN (0N , Rjjk (IN − jk Rjjk )). The estimated
MMSE
vector ĥjjk is not correlated with the received vector (de-spread vector) θjk . Which
MMSE
means Cov(ĥjjk , θjk ) = 0. As a result, θjk is not dependent of both estimated vector
and estimation error vector.
Therefore, we compute the NMSE in accordance with each antenna using MMSE
estimator. Thus, the following result is obtained
1 1 −1
εjkMMSE = E{ĥMMSE
jjk − hjjk 2 } = Tr[Rjjk − Rjjk jk Rjjk ] (7)
N N
We present some notes for MMSE channel estimation as follows.
Note 1: If each element of hjjk for all cells, BSs and all users in each cell follows

CN (0, 1). Therefore, jk = ψjk IN = ( Ll=1 βjlk + q1 )IN and Rjjk = βjjk IN . Thus,
βjjk
2
εjkMMSE = βjjk − ψjk
and βjjk = 1
N
Tr[ jk ].
Note 2: if hjjk ∼ CN (0, 1), q tends to infinity and Under PC, the NMSE of MMSE
β
estimator is εjkMMSE −→ βjjk (1 − L jjkβ ) when q −→ ∞.
l=1 jlk
Note 3: According to Eq. (6), and if Rjjk is invertible, we can conclude that
−1 MMSE
MMSE
ĥjlk = Rjlk Rjjk ĥjjk . In the reference [17], they have been shown that if Rjlk ,
∀l are linearly not dependant, which means Rjjk = αRjlk , ∀l = j where α is a real
number. Consequently, the channels vector are not parallel (i.e., the BS can separate
users sent the same PS). When hjjk ∼ CN (0, 1), according to Eq. (6), we remark that
β
MMSE
ĥjlk = βjlkjjk ĥjjk
MMSE
. The BS cannot separate the users used the same PS because the
βjlk
channel vectors are parallel differed by a factor βjjk
.
3.3 Approximate MMSE Channel Estimator
Generally, to get the CMs Rjlk , ∀l, k is a difficult process that it requires the estimation
of large matrices. In this part, we showed a straightforward and efficient approach to
this issue. According to Eq. (6), we can estimate jk which consists of the sum of Rjlk
plus identity matrix IN times the inverse of q. Thus, we can denote the estimated of
ˆ ˆ
jk by jk , and swapping jk with jk in Eq. (6). A traditional way to this estimation
issue is to approximate the CM including the sample CM [7]. To estimate jk , we
can remark that E[θjk θjkH ] = jk . To approximate jk , we can use the sample CM
below
1
N
ˆ jk = θjk (n)θjkH (n), n = 1, 2, . . . , N (8)
N n=1
where N represents the number of observations (NoO). In addition, E[ ˆ jk ] = jk .

Note that the convergence of the sample CM is proportional to the NoO of the de-
spread vector θjk (n).
1
N
lim ˆ jk = θjk (n)θjkH (n) → jk (9)
N →+∞ N n=1
We show that if N → +∞ ⇒ ˆ jk → jk . During this study, we look to the channels

as ergodic. The covariance of the i-th column of ˆ jk is given by
1
Cov([ ˆ jk ]i ) = [ jk ]i [ jk ]i
T
(10)
N
where [ jk ]i is the i-th column of jk . If hjjk ∼ CN (0, 1), thus [ jk ]i = [ψjk 0]T .
Hence,
1 2
Cov([ ˆ jk ]i ) = ψ IN (11)
N jk
The presence of error in each element of ˆ jk leads to losing its eigenstructure, which
makes each eigenvalue and eigenvector non-aligned with that of jk . While the
MMSE estimator is taking into account the eigenstructure of jk to achieve efficient
channel estimate. Therefore, it has a high effect on the Approx.MMSE performance.
Thus, to get over these problems, a convex combination scheme in [7, 8, 19] is given
to estimate jk (CM)
ˆ jk (η) = η ˆ jk + (1 − η) ˆ diag (12)
jk
This convex combination is a type of regularization that transforms ˆ jk into a full-

rank (F-R) matrix for all η values in the interval [0,1[. Even if when N < N , the
off-diagonal of ˆ jk be under-estimated.
The authors in [6] estimate Rjlk ∀j, l, k for each user in a specific pilot phase. In our
work, we are not considering the individual Rjlk , but we considering the estimation
of jk and performance evaluation under certain parameters. From Eqs. (6), (8), and
(12), ˆ jk is considered as the actual CM. We can give the appro.MMSE formula for
the channel vector ĥjjk as follows
Rjjk
θjk = Rjjk ˆ jk (η)−1 θjk
Approx.MMSE
ĥjjk = (13)
ˆ jk (η)
We suppose that θjk is independent of ˆ jk , that is to say ˆ jk is not estimated using θjk .
In addition, N is high enough to generate a great estimate of jk . The expression
of NMSE for each BS antenna using appro.MMSE is as follows
Approx.MMSE 1 Approx.MMSE
εjk = E{ĥjjk − hjjk 2 }
N (14)
1 −1
= Tr[Rjjk − Rjjk jk (η) Rjjk ]
N
We noticed that when N tends to infinity, the NMSE of Approx.MMSE estimator
and MMSE estimator are the same. The factor η is also chosen so as to the NMSE is
minimized. The authors in [19] have been given an optimization for this problem of
choosing η which minimizes the NMSE with proof.
3.4 EW-MMSE Estimator
The EW-MMSE estimator has low computational compared to the MMSE estimator,
which is also called a diagonalized MMSE estimator. This estimator relies on the
diagonal of the CM, where the off-diagonal of the CM is zero. Therefore, the EW-
MMSE is avoiding the correlation between CM elements. In some work, the EW-
MMSE is used as an alternative estimator [19, 20]. The estimate of hjjk using EW-
MMSE estimator is determined by
Djjk
EW-MMSE
ĥjjk = θjk = Djjk −1
jk θjk (15)
jk
where Djjk ∈ CN ×N and jk ∈ CN ×N are the diagonal matrices of Rjjk and jk ,
EW-MMSE
respectively. The estimated vector ĥjjk EW-MMSE
and the estimation error h̃jjk =
hjjk − ĥjjk
EW-MMSE
are random vectors distributed as ĥjjk EW-MMSE
∼ CN (0N , jjk ) and
EW-MMSE
h̃jjk ∼ CN (0N , ϒ̃jjk ), respectively.
where jjk = qτ Djjk jk ( jk )−1 jk Djjk and ϒ̃jjk = Rjjk − Rjjk jk Dv − Djjk jk Rjjk +
jjk .
EW-MMSE
An important remark is that the vector estimate h̃jjk and the estimation
EW-MMSE
error h̃jjk are correlated (is not the case in the MMSE estimator) which can be
computed as
EW-MMSE
Cov(h̃jjk , h̃jjk
EW-MMSE
) = Djjk −1
jk Rjjk − jk
(16)
= Djjk −1
jk Rjjk − Djjk jk ( jk )
−1
jk Djjk
The expression of NMSE for each BS antenna of the proposed EW-MMSE estimator
is as follows
1
εjkEW-MMSE = E{ĥjjk
EW-MMSE
− hjjk 2 }
N
1
= Tr Rjjk − Rjjk −1 −1
jk Djjk − Djjk jk Rjjk +
N
Djjk −1
jk
−1
jk jk Djjk ,
(17)
where χjk = Rjjk −1 −1
jk Djjk and χjk = Djjk jk Rjjk .
T
The new NMSE expression of the EW-MMSE estimator is as follows
1
εjkEW-MMSE = Tr[Rjjk − χjk − χjkT + Djjk −1
jk
−1
jk jk Djjk ] (18)
N
After computing the NMSE of all methods used in this work, we have defined the
following lemma.
Lemma 1 We consider that the estimated channel vector ĥjjk which is defined by
ĥjjk = jk θjk . In this case, the NMSE for each antenna is defined as follows:
1 1
E{hjjk − jk θjk 2 } = Tr[(IN − jk − Tjk )Rjjk ]
N N (19)
1
+ Tr[(jk jk Tjk )]
N
where jk is a square matrix given by:
⎧
⎪
⎪ IN , LS estimator.
⎪
⎨R −1
jjk jk MMSE estimator.
jk = (20)
⎪
⎪ R ˆ (η) −1
Approx.MMSE estimator.
⎪
⎩
jjk jk
−1
Djjk (jk ) Proposed EW-MMSE estimator.
Demonstration: For the demonstration of NMSE given in Eq. (19), we can easily
get it by direct computation.
4 Simulation Results
During this section, simulation results are presented to prove our theoretical result.
We have used an appropriate model in which the channels are correlated in spatial
domain. Therefore, an EC model is adopted with a factor of correlation r = 0.5 [17].
The standard deviation of LSF σ is fixed to 4. We adopt in our study that L = 4 cells,
K = 2 users in each cell. While τb = 200 is the coherence block and q = 100 mW
is the power of transmission for each device. We can convert q to dBm, which gives
q = 20 dBm.
In all the simulation results, we have considered that N = 100 and η = 0.5 except
the result shown in Figs. 2 and 3, respectively. The NMSE of each antenna can be
written for each estimator as follows E{hjjk − jk θjk 2 }/NTr[Rjjk ]. In addition, all
estimators performances are presented in terms of NMSE.
In Fig. 2, we present the NMSE versus the number of BS antennas. From this
figure, the LS estimator presents the worst performance compared to the others.
NMSE of all used estimators decreases with increasing the number of BS antennas.
It is very important to underline that when the number of BS antennas grows, the
Approx.MMSE and EW-MMSE performances get closer to that of MMSE. However,
the EW-MMSE presents a better performance compared to Approx.MMSE.
Figure 3 shows the NMSE performance against the RFc η. From this figure, the
MMSE, LS, and EW-MMSE estimators have constant NMSE values; meaning that
they do not rely on RFc η. The LS estimator has the largest NMSE compared to
the others, but it is important to underline that the LS estimator does not require
previous knowledge of channel statistics. Unlike LS, the MMSE estimator requires
prior statistics information of the channel. The Approx.MMSE performance is rely-
ing on the value of the factor of η. When η varies from 0 to 0.4, the NMSE of
Approx.MMSE is very close to that of MMSE. We can say that they have similar
NMSE values. In this interval, the effect of off-diagonal elements is not important
in comparison with the diagonal elements. This is resulting from small values of
Fig. 2 NMSE in dependance of the number of BS antennas, N

Fig. 3 NMSE in dependence of the Regularization Factor (RFc), η
the factor η. When η varies from 0.5 to 0.9, the performance of Approx.MMSE is
lower compared with the MMSE estimator, but it is yet surpassing the performance
of LS. When η exceeding 0.9, the Approx.MMSE performance is worse compared
to LS and MMSE estimators. For the reason that the CM is not a F-R matrix when η
exceeding 0.9. On the other side, the EW-MMSE performance is better compared to
the performances of LS and Approx.MMSE. The EW-MMSE performance is very
close to the MMSE estimator for all RFc η values. It is seen as a low-complexity
estimator compared to MMSE.
In Fig. 4, we present the NMSE versus the NoO N , to show the performances
of each estimator under the NoO N . The LS estimator is the worse one, whereas
the MMSE Approx.MMSE and EW-MMSE estimators present higher quality per-
formance than LS. The MMSE estimator supposes that all channel statistics are
perfectly known. From the Fig. 4, we can notice that the NMSE for LS, MMSE, and
EW-MMSE estimators remained the same for all N values; it implies that the per-
formance of LS, MMSE, and EW-MMSE estimators does not rely on N . However,
the Approx.MMSE estimator is approaching the MMSE estimator when the NoO
N increases. It implies that the Approx.MMSE depends on N . In addition, we can
say that the NoO and the error (the errors between the items of the sample CM and
the true CM) are inversely proportional, when N increases. The errors between the
items of the sample CM and the items of the true CM decrease. Nevertheless, the
EW-MMSE estimator is very nearly to MMSE performance, and it presents a better
performance compared to Approx.MMSE and LS estimators.
Fig. 4 NMSE versus the NoO, N
Fig. 5 NMSE versus the transmit pilot power, q

In Fig. 5, we present the NMSE against the UP power q. From this figure, it is
clearly that the LS estimator provides a lowest NMSE values. While, with increasing
the UP Transmit Power, the EW-MMSE and MMSE are nearly the same. On the
other hand, the performance of Approx.MMSE is relying on both the UP Trans-
mit Power q and the NoO N . For small values of q (from 0 to 100), the NMSE
of Approx.MMSE is close to MMSE and EW-MMSE estimators. But regarding q
(q > 100), the MMSE and EW-MMSE estimator performances become better than
Approx.MMSE. Increasing N , the performance of Approx.MMSE approaches to
the MMSE estimator. The Approx.MMSE presents a lower performance compared
to the EW-MMSE and MMSE estimators. Consequently, the EW-MMSE estimator
gives a better performance than Approx.MMSE for all q and N values.
5 Conslusion
This paper has suggested a straightforward and powerful channel estimator in terms
of NMSE performance. The Approx.MMSE estimator has substituted the covariance
matrix of the MMSE estimator through a sample CM. It has presented NMSE results
approaching the MMSE estimator with an increasing number of samples. While, the
worst performance has provided using LS estimator. Nevertheless, the EW-MMSE
has provided better performance than Approx.MMSE. The NMSE results are almost
the same like MMSE estimator with lower complexity.
References
1. Marzetta, T.L.: Noncooperative cellular wireless with unlimited numbers of base station anten-
nas. IEEE Trans. Wirel. Commun. 9(11), 3590–3600 (2010)
2. Larsson, E.G., Edfors, O., Tufvesson, F., Marzetta, T.L.: Massive MIMO for next generation
wireless systems. IEEE Commun. Mag. 52(2), 186–195 (2014)
3. Khansefid, A., Minn, H.: On channel estimation for massive MIMO with pilot contamination.
IEEE Commun. Lett. 19(9), 1660–1663 (2015)
4. De Figueiredo, F.A.P., Cardoso, F.A.C.M., Moerman, I., Fraidenraich, G.: Channel estimation
for massive MIMO TDD systems assuming pilot contamination and frequency selective fading.
IEEE Access 5, 17733–17741 (2017)
5. De Figueiredo, F.A.P., Cardoso, F.A.C.M., Moerman, I., Fraidenraich, G.: Channel estimation
for massive MIMO TDD systems assuming pilot contamination and flat fading. EURASIP J.
Wirel. Commun. Netw. 2018(1), 1–10 (2018)
6. Mandal, B.K., Pramanik, A.: Channel estimation in massive MIMO with spatial channel corre-
lation matrix. In: Intelligent Computing Techniques for Smart Energy Systems, pp. 377–385.
Springer (2020)
7. de Figueiredo, F.A.P., Lemes, D.A.M., Dias, C.F., Fraidenraich, G.: Massive MIMO channel
estimation considering pilot contamination and spatially correlated channels. Electron. Lett.
56(8), 410–413 (2020)
8. Björnson, E., Sanguinetti, L., Debbah, M.: Massive MIMO with imperfect channel covariance
information. In: 2016 50th Asilomar Conference on Signals, Systems and Computers, pp.
974–978. IEEE (2016)
9. Filippou, M., Gesbert, D., Yin, H.: Decontaminating pilots in cognitive massive MIMO net-
works. In: 2012 International Symposium on Wireless Communication Systems (ISWCS), pp.
816–820. IEEE (2012)
10. Adhikary, A., Nam, J., Ahn, J.-Y., Caire, G.: Joint spatial division and multiplexing-the large-
scale array regime. IEEE Trans. Inf. Theory 59(10), 6441–6463 (2013)
11. Yin, H., Gesbert, D., Filippou, M., Liu, Y.: A coordinated approach to channel estimation in
large-scale multiple-antenna systems. IEEE J. Sel. Areas Commun. 31(2), 264–273 (2013)
12. Gao, X., Edfors, O., Rusek, F., Tufvesson, F.: Massive MIMO performance evaluation based
on measured propagation data. IEEE Trans. Wirel. Commun. 14(7), 3899–3911 (2015)
13. Özdogan, Ö., Björnson, E., Larsson, E.G.: Massive MIMO with spatially correlated Rician
fading channels. IEEE Trans. Commun. 67(5), 3234–3250 (2019)
14. Sanguinetti, L., Björnson, E., Hoydis, J.: Toward massive MIMO 2.0: understanding spatial
correlation, interference suppression, and pilot contamination. IEEE Trans. Commun. 68(1),
232–257 (2019)
15. Forenza, A., Love, D.J., Heath, R.W.: Simplified spatial correlation models for clustered MIMO
channels with different array configurations. IEEE Trans. Veh. Technol. 56(4), 1924–1934
(2007)
16. Adhikary, A., Ashikhmin, A.: Uplink massive MIMOfor channels with spatial correlation. In
2018 IEEE Global Communications Conference (GLOBECOM), pp. 1–6. IEEE (2018)
17. Björnson, E., Hoydis, Jakob, Sanguinetti, L.: Massive MIMO has unlimited capacity. IEEE
Trans. Wirel. Commun. 17(1), 574–590 (2017)
18. Sengijpta, S.K.: Fundamentals of Statistical Signal Processing: Estimation Theory (1995)
19. Shariati, N., Björnson, E., Bengtsson, M., Debbah, M.: Low-complexity polynomial channel
estimation in large-scale MIMO with arbitrary statistics. IEEE J. Sel. Topics Signal Process.
8(5), 815–830 (2014)
20. Björnson, E., Hoydis, J., Sanguinetti, L.: Massive MIMO networks: Spectral, energy, and
hardware efficiency. Found. Trends Signal Process. 11(3–4), 154–655 (2017)
On Channel Estimation of Uplink TDD
Massive MIMO Systems Through
Different Pilot Structures
Jamal Amadid , Mohamed Boulouird , Abdelhamid Riadi ,

Abstract This work is considered as a comparative study in which the quality of

channel estimation (CE) in massive multiple-input multiple-output (M-MIMO) sys-
tems is studied by operating at Uplink (UL) phase according to a time division duplex
(TDD) scheme using commonly known channel estimators existing in the literature.
The least squares (LS) and minimum mean square error (MMSE) channel estimators
are investigated with three categories of pilots, namely regular pilots (RPs), time-
superimposed (or superimposed ) pilots and staggered pilots (StP). Two patterns of
frequency reuse (FR) per category are used. The simulation results showed that by
increasing the number of BS antennas with a fixed number of symbols dedicated to
the UL phase and vice versa, the normalized mean square error (NMSE) of the LS
and MMSE estimators using the superimposed pilot (SuP) or StP is asymptotically
approaches the NMSE of the LS and MMSE estimators using the RP, respectively.
An asymptotic behavior is studied for two different FR scenarios.
1 Introduction
M-MIMO cellular networks rely on a large number of antennas (NoA) at the base
stations (BS) to serve a large number of users. M-MIMO technology has attracted
considerable interest as a candidate for future cellular systems [1, 2]. Respecting
J. Amadid · A. Riadi · M. M. Hassani

Instrumentation, Signals and Physical Systems (I2SP) Group, Faculty of Sciences Semlalia,
Cadi Ayyad University, Marrakesh, Morocco
e-mail: jamal.amadid@edu.uca.ac.ma
A. Riadi
e-mail: abdelhamid.riadi@edu.uca.ac.ma
M. M. Hassani
e-mail: hassani@uca.ac.ma
M. Boulouird (B)
Smart Systems and Applications (SSA) Group, National School of Applied Sciences
of Marrakesh (ENSA-M), Cadi Ayyad University, Marrakesh, Morocco
e-mail: m.boulouird@uca.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_12
162 J. Amadid et al.
the NoA at the BS, these systems offer a major enhancement in the UL stage, in
the same way improving the energy efficiency (EE) and spectral efficiency (SE),
when the accurate channel state information (CSI) is convinced to be obtainable at
reception [3–5]. By using linear processing at the BS [6, 7], the throughput has been
increased under advantageous spreading conditions [8]. The previously mentioned
advantages of a large MIMO cellular network depend on the presumption that such
BS has access to reliable CSI. The CE process in both multiplexing modes (i.e.,
TDD and frequency division duplex (FDD) is performed by involving orthogonal
training sequences. Whereas for the CE phase in M-MIMO systems, the FDD mode
is considered as an impractical use [4, 9]. While TDD mode is largely applied and
became the most promising for M-MIMO. Since we want to build a network that
can perform successfully under any form of the propagation environment, in TDD
mode, the CSI is available at the BS when the pilot sequence and data are received at
the BS. While the channel reciprocity scenario is always used [10], and the CE can
be made with more accuracy.
Despite the advantage provided by TDD mode for M-MIMO, the M-MIMO sys-
tem has constraint of pilot contamination (PC) resulting from the duplication of the
same pilot sequences (PS) in contiguous cells, which cannot disappear even if the
NoA at the BS reaches infinity. Therefore, PC keeps till now as a choke-point for the
large MIMO TDD systems [3, 6, 11].
1.1 Related Works
In the literature, CE regarding the PC problem has been addressed in several works.
In [12], the authors rely on the hypothesis that adjacent cells coordinate their trans-
mission signals based on second-order statistics to facilitate the CE process. In the
case of non-coordination among the cells, a lot of works focused on mitigating PC
in the CE phase. In [13, 14], authors deal with CE concerning the PC problem using
singular-value decomposition (SVD) and semi-blind channel estimate to avoid PC.
For practical use, the BS has no information regarding channel statistics of the con-
tiguous cells. The authors in [15–17] suggested an estimator based on maximum
likelihood (ML) which can afford similar accuracy to that of MMSE without know-
ing any information about channel statistics of the contiguous cells. To summarize,
the previously mentioned literature dealt with the send of pilots followed by payload
data (PD) symbols (herein indicated to as RPs or time-multiplexed (TMu) pilots [18–
20]). In mobile communication scenarios, the channel coherent time is restricted by
the users’ mobility. In this case, the RP intended to present the worst performance.
Alternatively, to RP, current studies have centered on the SuP channel estimate in
the UL M-MIMO [19–22], where SuP is regarded as a supported pilot scheme. By
comparing SuP and RP, no added time for services is needed for SuPs. Thus, it can
effectively contribute better SE compared to RP [23] which demand added time to
accomplish that service. The power allocation strategy across SuPs and PD was stud-
On Channel Estimation of Uplink TDD Massive MIMO Systems . . . 163
ied and evaluated in [24]. In [25], the authors have been introduced the SuP channel
estimate in traditional MIMO systems.
In recent years, numerous studies [26–30] have been conducted on M-MIMO
systems with SuPs and have concluded that they are efficient in avoiding PC problems.
However, the assorted pilot style is subject to co-interference of PD symbols, which
frequently restricts its effectiveness especially in the situation of low signal-to-noise
ratio (SNR).
1.2 Organization of This Work
The main parts of our study are outlined as follows: First, the system model is
presented in Sect. 2. Next, LS performance is assessed for three categories of pilots
using NMSE in Sect. 3. Then, the MMSE estimator is discussed for three pilot
categories in Sect. 4. After that, the results of the simulation are presented in Sect. 5
in which we affirm our theoretical study. Finally, our final remarks are summarized
in Sect. 6.
1.3 Contributions of This work
The main concern of this paper is to study the UL channel estimation for M-MIMO
cellular networks. The two major contributions of this work are as follows:
1. Investigate and evaluate the performances of LS and MMSE estimators using
either Rer pilots and Sup pilots with different frequency reuse schemes.
2. Introduce an Stg pilot, which considered as a particular case of Sup pilots, and
analyze this pilot type under different frequency reuse schemes.
2 System Model
Our model deals with a multi-cell (MC) multi-user (MU) scenario in the UL phase.
The TDD mode is used with L cells and K users with a single antenna in each cell
with M K (M is the BS antennas). Generally, in communication networks, a band
of symbols is assumed in which the channel coherent is considered. In our work,
this symbol band is symbolized by C and presumed to be split into two sub-band
Cup and Cdl defined the number of time slots in UL and Downlink, successively. The
spreading matrix received at the jth BS symbolized by Yj ∈ CM ×Cup , which can be
expressed as
√
L−1 K−1
Yj = T
qlk gjlk slk + Nj (1)
l=0 k=0
Here gjlk ∈ CM ×1 represents the channel from user k in the lth cell to the jth BS.
While slk represents the vector of symbols dispatched by user k in the lth cell,
where qlk represents the power by which the symbols slk are dispatched. In addition,
Nj ∈ CM ×Cup represents the noise matrix, where each column of Nj is distributed as
CN (0; σ 2 ). In this paper, we adopt the assumption that the columns are not dependent
on each other.
Generally, the channel gjlk ∼ CN (0M ; βjlk 11M ) expressed in function of two coef-
ficients, namely small-scale fading (SSF) and large-scale fading (LSF) coefficients,
where the SSF is defined by the quick change over the phase and amplitude of a sig-
nal. Its SSF coefficient is regarded as complex normal distribution CN (0; 1). While
the LSF coefficient includes path loss or attenuation of the path as well as log-normal
shadowing. Furthermore, for the coherence duration time, the channel gjlk is assumed
to be static, which means that the channel is supposed to be constant over C sym-
bols. While βjlk is supposed to be consistent for a considerably longer duration. The
symbol slk in Eq. (1) depends on the type of pilot dispatched on the UL. Whenever
RP is employed, the pilots are dedicated to some of the components of slk , and the
rest is considered for PD. On the other hand, when the SuPs is employed, all the
pilots and PD dispatched alongside each other.
We supposed that there is a synchronization of all pilots. This hypothesis is typi-
cally used in the large MIMO research [3, 12, 31]. However, such a device is simple
to examine numerically using that hypothesis. In reality, the synchronization of a
wide-area network may not be feasible. In this work, the CE quality gained from RP,
SuP, and StP on LS and MMSE estimators is evaluated using NMSE.
3 Least Square Channel Estimation
In this section, the performances are studied, evaluated and discussed for the LS
channel estimation for three pilot schemes.
3.1 Regular Pilot
The RP has been used in many works in the literature [15–17]. In this category of
pilots, each user dispatches a pilot/training sequence of length τ for channel estimates
which is followed by PD. The PS used in this subsection are taken/extracted from a
unitary matrix ∈ Cτ ×τ such that H = τ Iτ , where every PS is represented by a
column of this matrix. These PS are orthogonal and shared over r RP cells. Meaning
that, at every r RP th cell, the PS ψlk that is dispatched by user k is re-used in all r RP
cells, where r RP is defined as r RP = τ/K and K symbolizes the number of user per
cell. Hence, the LS channel estimate using RP is formulated as [8, 20, 32]

ls qlk
RP
ĝjjk = gjjk + gjlk + njk (2)
qjk
l∈Pj (r RP )\j
√
Here njk = Nj ψjk /(τ qlk ) and the cells employed the same PS as cell j are referred
to as subgroup Pj (r ). When employing the RP scheme together with LS channel
RP
estimate, the NMSE is formulated as
ls E{ĝjjk
RP
− gjjk 2 }
N MSE RP =
jk
E{gjjk 2 }
⎛ ⎞ (3)
1 ⎝ qlk σ2 ⎠
= βjlk +
βjjk qjk τ qjk
l∈Pj (r
RP )\j
The NMSE expression in (3) depends on interference from contiguous cells. In other
words, from the cells that employ the same PS as cell j (i.e., PC) which occurs when
using the same pilot in the previously mentioned subgroup Pj (r RP ).
3.2 Superimposed Pilots
The SuPs are the second category introduced in our work in which the users dis-
patched pilots accompanying PD with reduced power (i.e., slk = ρdlk + λφlk ). The
two parameters λ2 , ρ 2 > 0 are the UL transmit power assigned for pilot and PD
successively. Under the constraint, ρ 2 + λ2 = 1. The LS channel estimate using SuP
is formulated as [20]
L−1 K−1
SuPls qnk ρ qnp T ∗ Nj φlk∗
ĝjlk = gjnk + gjnp dnp φlk + √ (4)
qlk Cu λ n=0 p=0 qjk λCup qlk
n∈Pj (r SuP )
Here φlk ∈ CCup and dlk ∈ CCup are, successively, the pilot and PD symbols dis-
patched by the user k in the ith cell. In this case, the Cup orthogonal SuPs are reused
in all r SuP cells. Here r SuP = Cup /K, K symbolizes the number of users in each cell.
Besides, the cells employed the same PS as cell j are referred to as subgroup Pj (r SuP ).
Furthermore, The PS used in this subsection are extracted from a unitary matrix
∈ CCup ×Cup such that H
= Cup ICup . Hence, φlkH φnp = δlk δnp . When employing
the SuP scheme together with LS channel estimate, the NMSE is formulated as
ls E{ĝjjk
SuP
− gjjk 2 }
N MSE SuP =
jk
E{gjjk 2 }
(5)
ρ 2 qnp
L−1 K−1
1 qlk σ2
= βjlk + βjnp +
βjjk qjk Cu λ2 n=0 p=0 qjk λ2 Cu qjk
l∈Pj (r
SuP )\j
The NMSE expression in (5) depends on interference from contiguous cells as in the
previous scheme, added to an additional interference term comes from sending pilot
alongside PD.
3.3 Staggered Pilots
The StPs are the third category of pilots studied in our work, where users in each cell
are staggering their pilot communications. Guaranteed that if the users of a specific
cell send UL pilots, users in the rest of r StP − 1 cells send PD [33, 34]. This pilot
category is considered as a particular case of the SuP, where the pilot power pp for
this category depends on the length of coherence time Cup as well as the length of
the PS τ used in the RP cases. The PD power pd for this category of pilot depends
on PD power in the former discussed category as exemplified the equation below
⎧
⎪
⎨pp = qλ Cup /τ
2
⎪
pd = qρ 2 (6)
⎪
⎪
⎩P = Cup blkdiag{
τ 0, . . . , L−1 }
Considered Yn ∈ CM ×τ as the spread matrix attained the jth BS when the users in
the nth cell ( where 0 n r StP ) sent UL pilots. Remark that the index j has been
removed from Yn for simplicity.
√ √
Yn = qlk pp gjlk φnk
T
+ qlk pd gjlk (dlkn )T + Nn (7)
l∈Pn (r SuP ) k l ∈P
/ n (r SuP ) k
(dlkn ) represents the vector of the data symbols dispatched at the nth block by the user
k in cell l. The LS channel estimate using StP is formulated as
qlk
StPls 1 pd qlp n T ∗
ĝjnk = gjlk + gjlp (dlp ) φnk
qnk Cu pp qnk
l∈Pn (r SuP ) l ∈P
/ n (r SuP ) p
(8)
∗
Nn φnk
+ √
Cup pp qnk
When employing the StP scheme together with LS channel estimate, the NMSE is
formulated as
ls E{ĝjjk
SuP
− gjjk 2 }
N MSE StP =
jk
E{gjjk 2 }
qlk
1 qlk pd
= βjlk + βjlk (9)
βjjk qjk Cup pp qjk
l∈Pj (r
SuP )\j l ∈P
/ j (r SuP ) k
σ2
+
pp Cup qjk
As in the case of SuPs, the NMSE expression in (9) depends on interference from
contiguous cells, which belongs to the same Pj (r SuP ) subgroup (As mentioned early
the cells employed the same PS as cell j are referred to as subgroup Pj (r SuP )).
An additional interference term comes from dispatching UL data over other cells
simultaneously with the sent pilots on Pj (r SuP ) subgroup.
4 MMSE Channel Estimation
In this section, the performances of the MMSE channel estimate for three pilot
schemes are studied, evaluated and discussed . As in the previous section (Sec. 3),
in which we discussed the LS channel estimate for three categories of pilots. In the
same manner, this section discusses the MMSE channel estimate for the three pilot
categories discussed in the previous section. We assumed that the same symbols and
properties elaborated in the previous section are valid in this section.
4.1 Regular Pilot
As in Sect. 3.1, this subsection evaluates and studies the MMSE channel estimate for
a system employing RPs. Assuming that the same symbols and properties elaborated
in Sect. 3.1 are valid in this subsection. The RPs have been addressed in several
works in the literature[15–17]. Therefore, the MMSE channel estimate using RP is
written as follows [8, 20, 32]

qlk
θjkRP = ĝjjk
RP
= gjjk + gjlk + njk (10)
qjk
l∈Pj (r RP )\j
√
Here, njk = Nj ψjk /(τ qlk ) and the cells employed the same PS as cell j are referred to
as subgroup Pj (r ), where the RP scheme is used with the MMSE channel estimate.
RP
In this case, the MMSE channel coefficient is written as follows

mmse βjjk βjjk

RP
ĝjjk = θjk = θ
RP jk
(11)
qlk σ2
l∈Pj (r RP ) qjk βjlk + jk
τ qjk

qlk σ2
where RP
jk = l∈Pj (r RP ) qjk βjlk + τ qjk
. The metric NMSE of the MMSE estimator
using the RP is formulated as follows
mmse
mmse E{ĝjjk
RP
− gjjk 2 }
N MSE RP =
jk
E{gjjk 2 }
(12)
1 qlk σ2
= βjlk +
RP
jk qjk τ qjk
l∈Pj (r
RP )\j
The NMSE formula in (12) relies on interference from neighboring cells meaning
that from a cell that uses the same PS as cell j. This happens in our scenario when
the same pilot is used in the previously described subgroup Pj (r RP ).
4.2 Superimposed Pilots
As in Sect. 3.2, this subsection investigates and discusses the MMSE channel estimate
for a system working under SuPs. Considering that the same symbols and properties
elaborated in Sect. 3.2 are valid in this subsection. The SuPs have a large benefit for
M-MIMO systems [20, 21] where the pilot and PD are dispatched simultaneously.
The MMSE channel estimate using SuP is formulated as [20]
L−1 K−1
ls qnk ρ qnp T ∗ Nj φ ∗
θjkSuP = ĝjlk
SuP
= gjnk + gjnp dnp φlk + √ lk (13)
qlk Cu λ n=0 p=0 qjk τ qlk
n∈Pj (r SuP )
where the SuP scheme is used with the MMSE channel estimate. In this case, the
MMSE channel coefficient is written as follows
mmse βjjk
SuP
ĝjjk = θjkSuP
qlk ρ2 L−1 K−1 qnp σ2
l∈Pj (r SuP ) qjk βjlk + Cu λ2 n=0 p=0 qjk βjnp + λ2 Cu qjk
(14)
βjjk
= θ SuP
SuP jk
jk

qlk ρ2 L−1 K−1 qnp σ2
Here SuP
jk = l∈Pj (r SuP ) qjk βjlk + Cu λ2 n=0 p=0 qjk βjnp + λ2 Cu qjk
. When
employing the SuP scheme together with MMSE channel estimate, the MSE is for-
mulated as
mmse
mmse E{ĝjjk
SuP
− gjjk 2 }
N MSE SuP =
jk
E{gjjk 2 }

ρ 2 qnp
L−1 K−1
1 qlk σ2
= βjlk + βjnp +
SuP
jk qjk Cu λ2 n=0 p=0 qjk λ2 Cu qjk
l∈Pj (r SuP )\j
(15)
The NMSE expression in (15) depends on interference from contiguous cells as in
the previous scheme, added to an additional interference term comes from sending
pilot alongside PD.
4.3 Staggered Pilots
As in Sect. 3.3, this subsection introduces StP as a particular case of SuPs. As stated
previously in Sect. 3.3, the users in each cell are staggering their pilot communica-
tions. Guaranteed that if the users of a specific cell send UL pilots, users in the rest
of r StP − 1 cells send PD [33, 34]. We assume that the same symbols and proper-
ties elaborated in Sect. 3.3 are valid in this subsection. Hence, the MMSE channel
estimate for a system working under StP is expressed in the following form
ls
θjkStP = ĝjnk
StP

qlk 1

pd qlp
n T ∗
= gjlk + gjlp (dlp ) φnk
qnk Cu pp qnk (16)
l∈Pn (r SuP ) l ∈P
/ n (r SuP ) p
∗
Nn φnk
+ √
Cup pp qnk
The expression of the MMSE channel coefficient when using StP is written as follows
mmse βjjk
StP
ĝjjk = θjkStP
qlk pd qlk σ2
l∈Pj (r SuP )\j qjk βjlk + Cup pp l ∈P
/ j (r SuP ) k qjk βjlk + pp Cup qjk
βjjk
= θ StP
StP jk
jk
(17)
qlk pd qlk σ 2
where StP
jk = l∈Pj (r SuP )\j qjk βjlk + Cup pp l ∈P
/ j (r SuP ) k qjk βjlk + pp Cup qjk
When using the StP scheme with MMSE channel estimate, the NMSE is formu-
lated as follows
ls E{ĝjjk
SuP
− gjjk 2 }
N MSE StP =
jk
E{gjjk 2 }
qlk
1 qlk pd
= StP
βjlk + βjlk (18)
jk qjk Cup pp qjk
l∈Pj (r SuP )\j l ∈P
/ j (r SuP ) k
σ2
+
pp Cup qjk
As in the case of SuPs, the NMSE expression in (18) depends on interference from
contiguous cells which belongs to the same Pj (r SuP ) subgroup (As mentioned early
the cells employed the same PS as cell j are referred to as subgroup Pj (r SuP )).
An additional interference term comes from dispatching UL data over other cells
simultaneously with the sent pilots on Pj (r SuP ) subgroup.
5 Simulation Results
Simulation results are provided in this section to validate our theoretical analysis
given in the previous sections. This section aims to evaluate and compare the perfor-
mances of LS and MMSE channel estimates using the NMSE metric. For a system
using L = 91 cells (five tiers of cells) and K = 10 users per cell for all pilot categories
aforementioned. Users are distributed across the cells. With the aim of studying the
PC effect, we assume that the users are at a distance greater than 100 m from the BS,
where the shadowing effect is taken into consideration, which is usually assumed to
obtain from tall buildings. We analyze the performance of LS and MMSE for pilot
categories discussed in previous sections under two FR schemes (r = 3, r = 7). The
SNR value for the UL phase is fixed to 10 dB.
Figure 1 shows the NMSE in dependence on Cup , which presents the number of
symbols used in the UL phase. The number of BS M antennas is fixed at 100 in all
simulation except where M is varied. As Cup increases, the performance provided
through the LS estimator using SuPs and StPs in both FR cases (r SuP = 3, r StP = 3;
r SuP = 7, r StP = 7) is asymptotically closed to the performance provided from RPs
in both FR cases (r RP = 3, r RP = 7), respectively. In addition, system performance
is improved by using FR equal to 7 (which is visualized in the NMSE values for all
pilot categories). Noted that the effect of FR is a major factor in the performance
of SuP and StPs (i.e., overcome the NMSE gap between SuP and StPs) since as FR
increases. The performance obtained with SuPs is close to that obtained with StPs
(similar behavior).
Fig. 1 NMSE in dependence on the number of symbols in the UL Cup for the LS estimator using
three different pilot categories and considering two cases of FR
Figure 2 shows the NMSE in dependence on Cup for the MMSE estimator under
different FR. As Cup increases, the performance afforded by the MMSE estimator
using SuP and StPs in the two FR cases (r SuP = 3, r StP = 3; r SuP = 7, r StP = 7)
is asymptotically closed to the performance afforded by the RPs in the two FR
cases (r RP = 3, r RP = 7) respectively. It is worth noting that the performance of the
MMSE estimator is better than that of the LS estimator. Furthermore, the system
performance is improved with a FR of 7 compared to the case of 3. Besides, note
that the impact of FR is crucial for the performance of the SuP and StPs, where the
difference between the NMSE of the SuP and StPs using the MMSE estimator is
relatively small compared to the case of the NMSE of the SuP and StPs using LS.
Figure 3 shows the NMSE versus M . The performances of the LS estimator are
presented for three categories of pilots under two FR values. The number of symbols
Cup in the UL phase is fixed at 35 in all simulations except where Cup is varied. For
the case where r SuP = r StP = 3, a large gap is given between the NMSE of the SuP
and StPs for small values of M. While in the case of r SuP = r StP = 7, this gap is
relatively narrow. As M increases, this gap becomes quite narrow and the NMSE of
the SuP and StPs asymptotically approaches to the NMSE of the RPs for both FR
scenarios.
Fig. 2 NMSE in dependence on the number of symbols in the UL Cup for the MMSE estimator
using three different pilot categories and considering two cases of FR
Figure 4 shows the NMSE in dependence on M . The performances of the MMSE

estimator are presented for three categories of pilots under two FR values. It is
obvious that the performance of the MMSE estimator is better than that of the LS
estimator (by comparing the results provided in Figs. 3 and 4). For the case where
r SuP = r StP = 3, a large gap is given between the NMSE of the SuP and StPs for small
values of M emphasizing. This difference is less than that provided by LS under the
same conditions. Whereas in the case of r SuP = r StP = 7, this gap is relatively narrow.
As M increases, this gap became rather narrow and the NMSE of the SuP and StPs
is asymptotically approaches the NMSE of the RPs for both FR scenarios.
Fig. 3 NMSE in dependence on the NoA M at the BS for the LS estimator using three different
pilot categories and considering two cases of FR
6 Conclusion
In this work, we have studied and analyzed the quality of CE for the M-MIMO system
in the UL phase. The TDD scheme is operated for three categories of pilots. We have
assessed CE quality employing the LS and MMSE channel estimators for regular,
SuP, and StPs for two different FR scenarios. We have shown that when the number
of symbols dedicated to the UL phase increases, an asymptotic behavior using LS
and MMSE estimators with staggered and SuPs is observed. Wherein their NMSE
approaches the NMSE of LS and MMSE estimators that employ RP pilots as the
number of symbols Cup dedicated to the UL phase increases. Furthermore, we also
studied the performance of our system under the NoA at the BS, where an identical
asymptotic behavior or curve shape is obtained. We have also studied the impact of
FR, where we have concluded that the performance is improving by using FR of 7.
While a very small gap in terms of the NMSE is obtained between staggered and
SuPs by using FR of 7 where this gap is very narrow using the MMSE estimator in
comparison to LS estimator.
Fig. 4 NMSE in dependence on the NoA M at the BS for the MMSE estimator using three different
pilot categories and considering two cases of FR
References
1. Boccardi, F., Heath, R.W., Lozano, A., Marzetta, T.L., Popovski, P.: Five disruptive technology
directions for 5g. IEEE Commun. Mag. 52(2), 74–80 (2014)
2. Osseiran, A., Boccardi, F., Braun, V., Kusume, K., Marsch, P., Maternia, M., Queseth, O.,
Schellmann, M., Schotten, H., Taoka, H., et al.: Scenarios for 5g mobile and wireless commu-
nications: the vision of the metis project. IEEE Commun. Mag. 52(5), 26–35 (2014)
3. Ngo, H.Q., Larsson, E.G., Marzetta, T.L.: Energy and spectral efficiency of very large multiuser
MIMO systems. IEEE Trans. Commun. 61(4), 1436–1449 (2013)
4. Lu, L., Li, G.Y., Swindlehurst, A.L., Ashikhmin, A., Zhang, R.: An overview of massive MIMO:
benefits and challenges. IEEE J. Sel. Topics Signal Process. 8(5), 742–758 (2014)
5. Rusek, F., Persson, D., Lau, B.K., Larsson, E.G., Marzetta, T.L., Edfors, O., Tufvesson, F.:
Scaling up MIMO: Opportunities and challenges with very large arrays. IEEE Signal Process.
Mag. 30(1), 40–60 (2012)
6. Hoydis, J., Ten Brink, S., Debbah, M.: Massive MIMO in the UL/DL of cellular networks:
How many antennas do we need? IEEE J. Sel. Areas Commun. 31(2), 160–171 (2013)
7. Yang, H., Marzetta, T.L.: Performance of conjugate and zero-forcing beamforming in large-
scale antenna systems. IEEE J. Sel. Areas Commun. 31(2), 172–179 (2013)
8. Marzetta, T.L.: Noncooperative cellular wireless with unlimited numbers of base station anten-
nas. IEEE Trans. Wirel. Commun. 9(11), 3590–3600 (2010)
9. Björnson, E., Larsson, E.G., Marzetta, T.L.: Massive MIMO: ten myths and one critical ques-
tion. IEEE Commun. Mag. 54(2), 114–123 (2016)
10. Paulraj, A.J., Ng, B.C.: Space-time modems for wireless personal communications. IEEE Pers.
Commun. 5(1), 36–48 (1998)
11. Larsson, E.G., Edfors, O., Tufvesson, F., Marzetta, T.L.: Massive MIMO for next generation
wireless systems. IEEE Commun. Mag. 52(2), 186–195 (2014)
12. Yin, H., Gesbert, D., Filippou, M., Liu, Y.: A coordinated approach to channel estimation in
large-scale multiple-antenna systems. IEEE J. Sel. Areas Commun. 31(2), 264–273 (2013)
13. Ngo, H.Q., Larsson, E.G., EVD-based channel estimation in multicell multiuser MIMO systems
with very large antenna arrays. In: 2012 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), pp. 3249–3252. IEEE (2012)
14. Guo, K., Guo, Y., Ascheid, G.: On the performance of EVD-based channel estimations in mu-
massive-MIMO systems. In: 2013 IEEE 24th Annual International Symposium on Personal,
Indoor, and Mobile Radio Communications (PIMRC), pp. 1376–1380. IEEE (2013)
15. Khansefid, A., Minn, H.: On channel estimation for massive MIMO with pilot contamination.
IEEE Commun. Lett. 19(9), 1660–1663 (2015)
16. de Figueiredo, F.A.P., Cardoso, F.A.C.M., Moerman, I., Fraidenraich, G.: Channel estimation
for massive MIMO TDD systems assuming pilot contamination and flat fading. EURASIP J.
Wirel. Commun. Netw. 2018(1), 1–10 (2018)
17. de Figueiredo, F.A.P., Cardoso, F.A.C.M., Moerman, I., Fraidenraich, G.: Channel estimation
for massive MIMO TDD systems assuming pilot contamination and frequency selective fading.
IEEE Access 5, 17733–17741 (2017)
18. Guo, C., Li, J., Zhang, H.: On superimposed pilot for channel estimation in massive MIMO
uplink. Phys. Commun. 25, 483–491 (2017)
19. Upadhya, K., Vorobyov, S.A., Vehkapera, M.: Downlink performance of superimposed pilots
in massive MIMO systems in the presence of pilot contamination. In: 2016 IEEE Global
Conference on Signal and Information Processing (GlobalSIP), pp. 665–669. IEEE (2016)
20. Upadhya, K., Vorobyov, S.A., Vehkapera, M.: Superimposed pilots are superior for mitigating
pilot contamination in massive MIMO. IEEE Trans. Signal Process. 65(11), 2917–2932 (2017)
21. Zhang, H., Pan, D., Cui, H., Gao, F.: Superimposed training for channel estimation of OFDM
modulated amplify-and-forward relay networks. Science China Inf. Sci. 56(10), 1–12 (2013)
22. Li, J., Zhang, H., Li, D., Chen, H.: On the performance of wireless-energy-transfer-enabled
massive MIMO systems with superimposed pilot-aided channel estimation. IEEE Access 3,
2014–2027 (2015)
23. Zhou, G.T., Viberg, M., McKelvey, T.: A first-order statistical method for channel estimation.
IEEE Signal Process. Lett. 10(3), 57–60 (2003)
24. Huang, W.-C., Li, C.-P., Li, H.-J.: On the power allocation and system capacity of OFDM
systems using superimposed training schemes. IEEE Trans. Veh. Technol. 58(4), 1731–1740
(2008)
25. Dai, X., Zhang, H., Li, D.: Linearly time-varying channel estimation for MIMO/OFDM systems
using superimposed training. IEEE Trans. Commun. 58(2), 681–693 (2010)
26. Zhang, H., Gao, S., Li, D., Chen, H., Yang, L.: On superimposed pilot for channel estimation
in multicell multiuser MIMO uplink: large system analysis. IEEE Trans. Veh. Technol. 65(3),
1492–1505 (2015)
27. Upadhya, K., Vorobyov, S.A., Vehkapera, M.: Superimposed pilots: an alternative pilot structure
to mitigate pilot contamination in massive MIMO. In: 2016 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), pp. 3366–3370. IEEE (2016)
28. Li, F., Wang, H., Ying, M., Zhang, W., Lu, J.: Channel estimations based on superimposed
pilots for massive MIMO uplink systems. In: 2016 8th International Conference on Wireless
Communications & Signal Processing (WCSP), pp. 1–5. IEEE (2016)
29. Die, H., He, L., Wang, X.: Semi-blind pilot decontamination for massive MIMO systems. IEEE
Trans. Wirel. Commun. 15(1), 525–536 (2015)
30. Wen, C.-K., Jin, S., Wong, K.-K., Chen, J.-C., Ting, P.: Channel estimation for massive MIMO
using Gaussian-mixture Bayesian learning. IEEE Trans. Wirel. Commun. 14(3), 1356–1368
(2014)
31. Björnson, E., Hoydis, J., Kountouris, M., Debbah, M.: Massive MIMO systems with non-ideal
hardware: Energy efficiency, estimation, and capacity limits. IEEE Trans. Inf. Theory 60(11),
7112–7139 (2014)
32. Fisher, R.A.: On the mathematical foundations of theoretical statistics. Philos. Trans. R. Soci.
Lond. Ser. Containing Pap. Math. Phys. Char. 222(594–604), 309–368 (1922)
33. Kong, D., Daiming, Q., Luo, K., Jiang, T.: Channel estimation under staggered frame structure
for massive MIMO system. IEEE Trans. Wirel. Commun. 15(2), 1469–1479 (2015)
34. Mahyiddin, W.A.W.M., Martin, P.A., Smith, P.J.: Performance of synchronized and unsynchro-
nized pilots in finite massive MIMO systems. IEEE Trans. Wirel. Commun. 14(12), 6763–6776
(2015)
NarrowBand-IoT and eMTC Towards
Massive MTC: Performance Evaluation
and Comparison for 5G mMTC
Adil Abou El Hassan , Abdelmalek El Mehdi , and Mohammed Saber
Abstract Nowadays, the design of 5G wireless network should consider the Internet
of Things (IoT) among the main orientations. The emerging IoT applications need
new requirements other than throughput to support a massive deployment of devices
for massive machine-type communication (mMTC). Therefore, more importance is
accorded to coverage, latency, power consumption and connection density. To this
purpose, the third generation partnership project (3GPP) has introduced two novel
cellular IoT technologies enabling mMTC, known as NarrowBand IoT (NB-IoT)
and enhanced MTC (eMTC). This paper provides an overview of NB-IoT and eMTC
technologies and a complete performance evaluation of these technologies against
the 5G mMTC requirements is presented. The performance evaluation results show
that these requirements can be met but under certain conditions regarding the system
configuration and deployment. At the end, a comparative analysis of the performance
of both technologies is conducted mainly to determine the limits and suitable use
cases of each technology.
1 Introduction
Internet of Things (IoT) is seen as a driving force behind recent improvements in

wireless communication technologies such as third generation partnership project
(3GPP) long-term evolution advanced (LTE-A) and 5G New Radio (NR) to meet the
expected requirements of various massive machine-type communication (mMTC)
applications. The mMTC introduce a new communication era where billions of
A. Abou El Hassan (B) · A. El Mehdi · M. Saber

SmartICT Lab, National School of Applied Sciences,
Mohammed First University Oujda, Oujda, Morocco
e-mail: a.abouelhassan@ump.ac.ma
A. El Mehdi
e-mail: a.elmehdi@ump.ac.ma
M. Saber
e-mail: m.saber@ump.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_13
178 A. Abou El Hassan et al.
devices, such as remote indoor or outdoor sensors, will need to communicate with
each other, while connected to the cloud-based system.
The purpose of 5G system design is to cover three categories of use cases: enhanced
mobile broadband (eMBB), massive machine-type communication (mMTC), as well
as ultra reliable low-latency communication (uRLLC) [1].
The benefit of 5G system is the flexibility of its structure, which allows the use
of a common integrated system to cover many use cases, by using a new concept
which is network slicing based on SDN (Software-Defined Networking) and NFV
(Network Function Virtualization) technologies [2].
3GPP has introduced two low-power wide area (LPWA) technologies for IoT in
Release 13 (Rel-13): NarrowBand IoT (NB-IoT) and enhanced machine-type com-
munication (eMTC) which were designed to coexist seamlessly with existing LTE
systems. The 3GPP Rel-13 core specifications for NB-IoT and eMTC were final-
ized in June 2016 [3, 4], whereas Rel-14 and Rel-15 enhancements were completed,
respectively, in June 2017 and June 2018 [3, 4]. About the Rel-16 enhancements,
they are underway and scheduled for completion in 2020 [1]. In Rel-15, 3GPP has
defined in its work five requirements of 5G mMTC in terms of coverage, throughput,
latency, battery life and connection density [5].
The aim of this paper is to determine the system configuration and deployment
required for NB-IoT and eMTC technologies in order to fully meet the 5G mMTC
requirements. In addition, a comparative analysis is performed of the performances
of NB-IoT and eMTC technologies against the 5G mMTC requirements, in order to
determine the limits and suitable use cases of each technology.
The remainder of the paper is organized as follows. Section 2 presents the related
works. In Sect. 3, overviews of both NB-IoT and eMTC technologies are provided.
This is followed, in Sect. 4, by a complete performance evaluation of NB-IoT and
eMTC technologies against 5G mMTC requirements in terms of coverage, through-
put, latency, battery lifetime and connection density. In addition, the enhancements
provided by the recent 3GPP releases are also discussed. A comparative analysis
of the performances evaluated of NB-IoT and eMTC technologies is presented in
Sect. 5 in order to specify the limits and suitable use cases of each technology. Finally,
Sect. 6 concludes the paper.
2 Related Works
Many papers address 3GPP LPWA technologies including NB-IoT and eMTC and
non-3GPP LPWA technologies such as LoRa and Sigfox. El Soussi et al. [6] propose
an analytical model and implement NB-IoT and eMTC modules in discrete-event
network simulator NS-3, in order to evaluate only battery life, latency and connection
density. Whereas Jörke et al. [7] present typical IoT smart city use cases such as waste
management and water metering to evaluate only throughput, latency and battery life
of NB-IoT and eMTC. Pennacchioni et al. [8] analyze the performance of NB-IoT in
a massive MTC scenario focusing on only the evaluation of coverage and connection
NarrowBand-IoT and eMTC Towards Massive MTC: Performance Evaluation . . . 179
density, by choosing a smart metering system placed in a dense urban scenario as a

case study. However, Liberg et al. [9] focus on NB-IoT technology only but provide
a performance evaluation against 5G mMTC requirements. On the other hand, Krug
and O’Nils [10] compare the delay and energy consumption of data transfer covering a
various IoT communication technologies such as Bluetooth, WiFi, LoRa, Sigfox and
NB-IoT. However to our knowledge, there is no paper covering the evaluation of the
performances of NB-IoT and eMTC technologies against 5G mMTC requirements,
as well as the comparative analysis of these performances. This motivated us to
perform a comparative analysis of the evaluated performances of NB-IoT and eMTC
technologies against 5G mMTC requirements, in order to highlight the use cases of
each technology.
3 Overview of Cellular IoT Technologies: NB-IoT and

eMTC
3.1 Narrowband IoT: NB-IoT
The 3GPP design aims for Rel-13 were low cost and low-complexity devices, long
battery life and coverage enhancement. For this purpose, two power saving techniques
have been implemented to reduce power consumption of device: Power saving mode
(PSM) and extended discontinuous reception (eDRX) introduced in Rel-12 and Rel-
13, respectively, [7, 11]. The bandwidth occupied by the NB-IoT carrier is 180 kHz
corresponding to an one physical resource block (PRB) of 12 subcarriers in an LTE
system [11]. There are three operation modes to deploy NB-IoT: as a stand-alone
carrier, in guard-band of an LTE carrier and in-band within an LTE carrier [11, 12].
In order to coexist with LTE system, NB-IoT uses orthogonal frequency division
multiple access (OFDMA) in downlink with the identical subcarrier spacing of 15
kHz and frame structure as LTE [11]. Whereas NB-IoT uses in uplink single-carrier
frequency division multiple access (SC-FDMA) and two numerologies which use
15 kHz and 3.75 kHz subcarrier spacings with 0.5 ms and 2 ms slot durations,
respectively, [11]. The restricted QPSK and BPSK modulation schemes are used in
downlink and uplink by NB-IoT device with a single antenna [3, 11]. Also, NB-IoT
defines three coverage enhancement (CE) levels in a cell: CE-0, CE-1 and CE-2
corresponding to the maximum coupling loss (MCL) of 144 dB, 154 dB and 164 dB,
respectively, [8].
Two device categories Cat-NB1 and Cat-NB2 are defined by NB-IoT which cor-
respond to the device categories introduced in Rel-13 and Rel-14, respectively. The
maximum transport block size (TBS) supported in uplink by Cat-NB1 is only 1000
bits compared to 2536 bits for Cat-NB2. For downlink, the maximum TBS supported
by Cat-NB1 is only 680 bits compared to 2536 bits for Cat-NB2 [3].
The signals and channels used in downlink (DL) are as follows: Narrowband pri-
mary synchronization signal (NPSS), narrowband secondary synchronization signal
(NSSS), narrowband reference signal (NRS), narrowband physical broadcast channel

(NPBCH), narrowband physical downlink shared channel (NPDSCH) and narrow-
band physical downlink control channel (NPDCCH). NPDCCH is used to transmit
downlink control information (DCI) for uplink, downlink and paging scheduling [3,
11].
Whereas only one signal and two channels are used in uplink (UL): Demodulation
reference signal (DMRS), narrowband physical uplink shared channel (NPUSCH)
and narrowband physical random access channel (NPRACH). Two formats are used
for NPUSCH which are: Format 1 (F1) and Format 2 (F2). NPUSCH F1 is used by
the user equipment (UE) to carry uplink user’s data to the evolved Node B (eNB),
whereas NPUSCH F2 is used to carry uplink control information (UCI) which are the
DL hybrid automated repeat request acknowledgement (HARQ-ACK) and negative
ACK (HARQ-NACK) [11].
For cell access, the UE must first synchronize with the eNB using NPSS and
NSSS signals to achieve time and frequency synchronization with the network and
cell identification. Then, it receives narrowband master information block (MIB-
NB) and system information block 1 (SIB1-NB) carried by NPBCH and NPDSCH,
respectively, from eNB to access the system [11].
3.2 Enhanced Machine-Type Communication: eMTC
The overall time structure of the eMTC frame is also identical to that of the LTE frame
described in Sect. 3.1. eMTC reuses an identical numerology as LTE, OFDMA and
SC-FDMA are used in downlink and uplink, respectively, with subcarrier spacing of
15 kHz [12]. The eMTC transmissions are limited to a narrowband size of 6 PRBs
corresponding to 1.4 MHz including guardbands. As the LTE system has a bandwidth
from 1.4 to 20 MHz, a number of non-overlapping narrowbands (NBs) can be used
if the LTE bandwidth exceeds 1.4 MHz [4]. Up to Rel-14, eMTC device uses QPSK
and 16-QAM modulation schemes with a single antenna for downlink and uplink.
Whereas support for 64-QAM in downlink has been introduced in Rel-15 [4].
Two device categories are defined by eMTC: Cat-M1 and Cat-M2 corresponding
to device categories introduced in Rel-13 and Rel-14, respectively. Cat-M1 has only
an maximum channel bandwidth of 1.4 MHz compared to 5 MHz for Cat-M2 [4].
In addition, Cat-M2 supports a larger TBS of 6968 bits and 4008 bits in uplink
and downlink, respectively, compared to 2984 bits in both downlink and uplink for
Cat-M1 [4].
The following channels and signals are reused by eMTC in downlink: Physical
downlink shared channel (PDSCH), physical broadcast channel (PBCH), primary
synchronization signal (PSS), secondary synchronization signal (SSS), positioning
reference signal (PRS) and cell-specific reference signal (CRS). MTC physical down-
link control channel (MPDCCH) is the new control channel which has the role of
carrying DCI for uplink, downlink and paging scheduling [4, 12].
Whereas for uplink, the following signals and channels are reused: Demodulation
reference signal (DMRS), sounding reference signal (SRS), physical uplink shared
channel (PUSCH), physical random access channel (PRACH) and physical uplink
control channel (PUCCH) which conveys UCI [4, 12].
For cell access, the UE uses the PSS/SSS signals to synchronize with the eNB, and
PBCH which carries the master information block (MIB). After decoding the MIB
and then the new system information block for reduced bandwidth UEs (SIB1-BR)
carried by PDSCH, the UE initiates the random access procedure using PRACH to
access the system [12].
4 NB-IoT and eMTC Performance Evaluation
4.1 Coverage
The MCL is a common measure to define the level of coverage a system can support.
It is depending on the maximum transmitter power (PTX ), the required signal-to-
interference-and-noise ratio (SINR), the receiver noise figure (NF) and the signal
bandwidth (BW) [13]:
MCL = PTX − (SINR + NF + N0 + 10log10 (BW)) (1)
where N0 is the thermal noise density which is a constant equal −174 dBm/Hz.
Based on the simulation assumptions given in Table 1 according to [14] and using
(1) to calculate MCL, Tables 2 and 3 show the NB-IoT and eMTC channel coverage,
respectively, to achieve the MCL of 164 dB which corresponds to the 5G mMTC
coverage requirement to be supported [5].
Tables 2 and 3 also indicate the required acquisition time and block error rate
(BLER) associated with each channel to achieve the targeted MCL of 164 dB. From
the acquisition times shown in Tables 2 and 3, we note that to achieve the MCL of
164 dB at the appropriate BLER, it is necessary to use the time repetition technique
for the simulated channels.
Table 1 Simulation and system model parameters

Parameter Value
System bandwidth 10 MHz
Channel model Tapped delay line (TDL-iii/NLOS)
Doppler spread 2 Hz
NB-IoT mode of operation Guard-band
eNB Rx/Tx 4/2 and 4/4 only for NPSS/NSSS transmissions
Device Rx/Tx 1/1
Table 2 Downlink and uplink coverage of NB-IoT

Assumptions Downlink physical channel Uplink physical channel
for simulation
NPBCH NPDCCH NPDSCH NPRACH NPUSCH F1 NPUSCH F2
TBS (Bits) 24 23 680 – 1000 1
Acquisition 1280 512 1280 205 2048 32
time (ms)
BLER (%) 10 1 10 1 10 1
Max transmit 46 46 46 23 23 23
power (dBm)
Transmit 35 35 35 23 23 23
power/Carrier
(dBm)
Noise figure 7 7 7 5 5 5
NF (dB)
Channel 180 180 180 3.75 15 15
bandwidth
(kHz)
Required −14.5 −16.7 −14.7 −8.5 −13.8 −13.8
SINR (dB)
MCL (dB) 163.95 166.15 164.15 164.76 164 164
Table 3 Downlink and uplink coverage of eMTC

Assumptions Downlink physical channel Uplink physical channel
for simulation
PBCH MPDCCH PDSCH PRACH PUSCH PUCCH
TBS (Bits) 24 18 328 – 712 1
Acquisition 800 256 768 64 1536 64
time (ms)
BLER (%) 10 1 2 1 2 1
Max transmit 46 46 46 23 23 23
power (dBm)
Transmit 39.2 36.8 36.8 23 23 23
power/Carrier
(dBm)
Noise figure 7 7 7 5 5 5
NF (dB)
Channel 945 1080 1080 1048.75 30 180
bandwidth
(kHz)
Required −17.5 −20.8 −20.5 −32.9 −16.8 −26
SINR (dB)
MCL (dB) 163.95 164.27 163.97 164.7 164 165.45
Fig. 1 NPDSCH scheduling cycle (Rmax = 512; G = 4) at the MCL
Fig. 2 NPUSCH F1 scheduling cycle (Rmax = 512; G = 1.5) at the MCL
4.2 Throughput
The downlink and uplink throughputs of NB-IoT are obtained according to the
NPDSCH and NPUSCH F1 transmission time intervals issued from NPDSCH and
NPUSCH F1 scheduling cycles, respectively, and using the simulation assumptions
shown in Tables 1 and 2. While the downlink and uplink throughputs of eMTC are
determined based on the PDSCH and PUSCH transmission time intervals issued from
PDSCH and PUSCH scheduling cycles respectively and the simulation assumptions
given in Tables 1 and 3. The MAC-layer throughput (THP) is calculated with the
following formula:
(1 − BLER)(TBS − OH)
THP = (2)
PDCCH Period
where PDCCH period is the period of physical downlink control channel of NB-IoT
and eMTC that are NPDCCH and MPDCCH, respectively, and OH is the overhead
size in bits corresponding to the radio protocol stack.
Figure 1 depicts NPDSCH scheduling cycle of NB-IoT according to [14], where
the NPDCCH user-specific search space is configured with a maximum repetition
factor Rmax of 512 and a relative starting subframe periodicity G of 4.
Based on BLER and TBS given in Table 2 and using an overhead (OH) of 5 bytes,
a MAC-layer THP in downlink of 281 bps is achieved according to the formula (2).
The NPUSCH F1 scheduling cycle depicted in Fig. 2 corresponds to scheduling
of NPUSCH F1 transmission once every fourth scheduling cycle according to [14],
which ensures a MAC-layer THP in uplink of 281 bps according to the formula (2)
and based on BLER and TBS given in Table 2 and an overhead (OH) of 5 bytes.
Figure 3 depicts the PDSCH scheduling cycle of eMTC which corresponds to
scheduling of PDSCH transmission once every third scheduling cycle, where the
Fig. 3 PDSCH scheduling cycle (Rmax = 256; G = 1.5) at the MCL
Fig. 4 PUSCH scheduling cycle (Rmax = 256; G = 1.5) at the MCL
MPDCCH user-specific search space is configured with Rmax of 256 and a rela-
tive starting subframe periodicity G of 1.5 according to [14]. Whereas the PUSCH
scheduling cycle depicted in Fig. 4 corresponds to scheduling of PUSCH transmis-
sion once every fifth scheduling cycle according to [14].
From BLER and TBS indicated in Table 3 and the use of an overhead (OH) of 5
bytes, the MAC-layer throughputs obtained in downlink and uplink are 245 bps and
343 bps respectively according to the formula (2).
As part of 3GPP Rel-15, 5G mMTC requires that downlink and uplink troughputs
supported at the MCL of 164 dB must be at least 160 bps [5]. As can be seen,
the MAC-layer throughputs of both NB-IoT and eMTC technologies meet the 5G
mMTC requirement.
It should be noted that the BLER targets associated with each NB-IoT and eMTC
channel require the acquisition times shown in Tables 2 and 3, respectively. Therefore,
the throughput levels of NB-IoT and eMTC can be further improved by using the
new Cat-NB2 and Cat-M2 device categories, respectively, which support a larger
TBS in downlink and uplink and also enhanced HARQ processes.
4.3 Latency
The latency should be evaluated for the following procedures: Radio resource con-
trol (RRC) resume procedure and early data transmission (EDT) procedure that has
been introduced in Rel-15 and allowing the device to terminate the transmission of
small data packets earlier in RRC-idle mode. Figures 5 and 6 depict the data and
signaling flows corresponding to the RRC Resume and EDT procedures respectively
that are used by NB-IoT. Whereas the data and signalling flows corresponding to the
Fig. 5 NB-IoT RRC resume

procedure
Fig. 6 NB-IoT EDT

procedure
RRC Resume and EDT procedures used by eMTC are illustrated in Figs. 7 and 8,
respectively.
The latency evaluation is based on the same radio related assumptions and the
system model given in Table 1, whereas the packet sizes used and the latency evalu-
ation results of NB-IoT and eMTC at the MCL of 164 dB are shown in Tables 4 and
5, respectively, according to [14].
As can be seen from Tables 4 and 5, the 5G mMTC target of 10 s latency at the MCL
of 164 dB defined in 3GPP Rel-15 [5] is met by NB-IoT and eMTC technologies,
for both RRC Resume and EDT procedure. However, the best latencies of 5.8 and
5 seconds obtained by NB-IoT and eMTC, respectively, using the EDT procedure
are mainly due to the multiplexing of the user data with Message 3 on the dedicated
traffic channel, as shown in Figs. 6 and 8, respectively.
Fig. 7 eMTC RRC resume

procedure
Fig. 8 eMTC EDT

procedure
4.4 Battery Life
The RRC resume procedure is used in battery life evaluation instead of the EDT
procedure since EDT procedure does not support uplink TBS larger than 1000 bits
which requires long transmission times. The packet flows used to evaluate battery
life of NB-IoT and eMTC are the same as shown in Figs. 5 and 7, respectively,
where DL data corresponds to the application acknowledgment regarding receipt
of UL report by the eNB. Four levels of device power consumption are defined,
including transmission (PTX ), reception (PRx ), Idle-Light sleep (PILS ) corresponding
to device in RRC-Idle mode or RRC-Connected mode but not actively transmitting
or receiving, whereas Idle-Deep sleep (PIDS ) corresponds to power saving mode.
Table 4 Packet sizes and results of NB-IoT latency evaluation

RRC Resume procedure EDT procedure
Random access 7 bytes Random access 7 bytes
response: Msg2 response: Msg2
RRC Conn. Resume 11 bytes RRC Conn. Resume 11 + 105 bytes
request: Msg3 request: Msg3 + UL
report
RRC Conn. Resume: 19 bytes RRC Conn. Release: 24 bytes
Msg4 Msg4
RRC Conn. Resume 22 + 200 bytes
complete: Msg5 +
RLC Ack Msg4 + UL
report
RRC Conn. Release 17 bytes
Latency 9s Latency 5.8 s
Table 5 Packet sizes and evaluation results of eMTC latency

RRC resume procedure EDT procedure
Random access 7 bytes Random access 7 bytes
response: Msg2 response: Msg2
RRC Conn. Resume 7 bytes RRC Conn. Resume 11 + 105 bytes
request: Msg3 request: Msg3 + UL
report
RRC Conn. Resume: 19 bytes RRC Conn. Release: 25 bytes
Msg4 Msg4
RRC Conn. Resume 22 + 200 bytes
complete: Msg5 +
RLC Ack Msg4 + UL
report
RRC Conn. Release 18 bytes
Latency 7.7 s Latency 5s
The battery life in years is calculated using the following formula according to
[13]:
Battery energy capacity
Battery life [years] = (3)
E day
365 ×
3600
Where E day is the device energy consumed per day in Joule and calculated as follows
:
E day = [(PTX × TTX + PRx × TRx + PILS × TILS ) × Nrep ] + (PIDS × 3600 × 24)
(4)
Table 6 Simulation and system model parameters for battery life evaluation
Parameter Value
LTE system bandwidth 10 MHz
Channel model and Doppler spread Rayleigh fading ETU—1 Hz
eNB power and antennas configuration NB-IoT: 46 dBm (Guard-band,
In-band)—2Tx/2Rx 43 dBm
(Stand-alone)—1Tx/2Rx
eMTC: 46 dBm—2Tx/2Rx
Device power and antennas configuration 23 dBm—1Tx/1Rx
Table 7 Traffic model and device power consumption

Message format
UL report 200 bytes
DL application acknowledgment 20 bytes
Report periodicity Once every 24 h
Device power consumption levels
Transmission and reception power PTx : 500 mW—PRx : 80 mW
consumption
Idle mode power consumption PILS : 3 mW—PIDS : 0.015 mW
As for TTX , TRx and TILS , they correspond to overall times in seconds for transmission,
reception and Idle-Light sleep, respectively, according to packet flows of NB-IoT and
eMTC shown in Figs. 5 and 7, respectively. While Nrep corresponds to the number
of uplink reports per day.
The simulation and system model parameters used to evaluate the battery life of
NB-IoT and eMTC are given in Table 6 according to [15, 16]. While the assumed
traffic model according to Rel-14 scenario and device power consumption levels used
are given in Table 7.
Based on the transmission times of the signals and downlink and uplink channels
given in [15] and using the formulas (3) and (4) with the simulation assumptions
given in Table 7 and a 5 Wh battery, the evaluated battery lifes of NB-IoT to achieve
the MCL of 164 dB in in-band, guard-band and stand-alone operation modes are
11.4, 11.6 and 11.8 years, respectively. Whereas the evaluated battery life of eMTC
to achieve the MCL of 164 dB is 8.8 years according to the assumed transmission
times given in [16].
The 5G mMTC requires battery life beyond 10 years at the MCL of 164 dB,
supposing an energy storage capacity of 5Wh [5]. Therefore, NB-IoT achieves the
targeted battery life in all operations modes. However, eMTC does not fulfill the
5G mMTC targeted battery life. In order to significantly increase eMTC battery life,
the number of base station receiving antennas should be increased to reduce UE
transmission time. Therefore, if the number of base station receiving antennas is 4
instead of only 2, the evaluated battery life is 11.9 years which fulfills the 5G mMTC
target according to [14].
To further increase the battery life of NB-IoT and eMTC, the narrowband wake-
up signal (NWUS) and MTC WUS signal (MWUS) introduced in 3GPP Rel-15 can
be implemented, respectively. Since these signals allow the UE to remain in idle
mode until informed to decode NPDCCH/MPDCCH channel for a paging occasion,
thereby achieving energy saving.
4.5 Connection Density
The 5G mMTC target on connection density which is also part of the international
mobile telecommunication targets for 2020 and beyond (IMT-2020), requires the
support of one million devices per square kilometer in four different urban macro
scenarios [5]. These scenarios are based on two channel models (UMA A) and (UMA
B) and two distances of 500 and 1732 m between adjacent cell sites denoted by ISD
(inter-site distance) [17].
Based on the simulation assumptions given in Table 8 and the non-full buffer sys-
tem level simulation to evaluate connection density of NB-IoT and eMTC according
to [18], Fig. 9 shows the latency required at 99% reliability to deliver 32 bytes
payload as a function of the connection requests intensity (CRI) to be supported,
corresponding to the number of connection requests per second, cell and PRB.
It should be noted that the latency shown in Fig. 9 includes the idle mode time to
synchronize to the cell and read the MIB-NB/MIB and SIB1-NB/SIB1-BR. Know-
ing that each UE must submit a connection request to the system periodically, we
can calculate the connection density to be supported (CDS) per cell area using the
following formula:
CRI · CRP
CDS = (5)
A
Table 8 System level simulation assumptions of urban macro scenarios

Parameter Value
Frequency band 700 MHz
LTE and eMTC system bandwidths 10–1.4 MHz
Operation mode of NB-IoT In-band
Cell structure Hexagonal grid with 3 sectors per size
Pathloss model UMA A, UMA B
eNB power and antennas configuration 46 dBm—2Tx/2Rx
UE power and antennas configuration 23 dBm—1Tx/1Rx
Fig. 9 Intensity of connection requests in relation to latency
where CRP is the periodicity of connection requests

√ given in seconds and the hexag-
onal cell area A is calculated by: A = ISD2 · 3/6.
For NB-IoT, to evaluate the connection density per PRB and square kilometer
depicted in Fig. 10 and which corresponds to the overall number of devices that
successfully transmit a payload of 32 bytes accumulated over two hours, the CDS
is obtained from (5) using the CRI values of Fig. 9 and a periodicity of connection
requests of two hours. While for eMTC, to evaluate the connection density per
narrowband and square kilometer shown in Fig. 11, the CDS is determined from (5)
using the CRI values of Fig. 9, a periodicity of connection requests of two hours and
a scaling of a factor 6 corresponding to the eMTC narrowband of 6 PRBs.
As can be seen from Fig. 10, in the two scenarios corresponding to the 500 m
ISD, more than 1.2 million devices per PRB and square kilometer can be supported
by an NB-IoT carrier with a maximum 10 s latency. However, only 94000 and 68000
devices per PRB and square kilometer can be supported using the (UMA B) and
(UMA A) channel models, respectively, with an ISD of 1732 m within the 10-s
latency limit. Indeed, in the 1732 m ISD scenario, the density of base stations is 12
times lower than with a 500 m ISD. Therefore, this difference in base station density
results in differences of up to 18 times between the connection densities relating to
scenarios of 500 and 1732 m ISD.
For the 500-m ISD scenario shown in Fig. 11, a single eMTC narrowband can
support up to 5.68 million devices within the 10-s latency limit, by addition of 2
Fig. 10 Connection density

of NB-IoT
in relation to latency
Fig. 11 Connection density

of eMTC
in relation to latency
further PRBs to transmit PUCCH. For the 1732 m ISD and (UMA B) scenario, cell
size is a 12 times larger which explains a eMTC carrier can only support 445,000
devices within a limit of latency of 10 s.
Also, to further improve connection density of eMTC, sub-PRB resource alloca-
tion for uplink that has been introduced in 3GPP Rel-15 can be used in the case of a
scenario with a low-base station density.
5 Comparative Analysis of the Performance of NB-IoT

and eMTC Technologies
Figure 12 depicts the diagram comparing the performance of NB-IoT and eMTC
technologies evaluated in Sect. 4 in terms of coverage, throughput, latency, battery
life and connection density. The latencies shown in Fig. 12 are that obtained with
EDT procedure, while the connection densities are represented by the best values
obtained of the supported intensity of connection requests (CRI) from Fig. 9 within
the 10-s latency limit, and that correspond to the same urban macro scenario using
500 m ISD and (UMA B) channel model. The 5G mMTC requirement of CRI shown
in Fig. 12 corresponds to the targeted CRI obtained from (5) to achieve one million
devices per square kilometer for 500 m ISD scenario.
From Tables 2 and 3, it can be seen that NPUSCH F1 and PUSCH channels need
the maximum transmission times to reach the coverage target of 164 dB. Thus, for
NB-IoT, NPDCCH must be configured with 512 repetitions to achieve the targeted
BLER of 1%, while the maximum configurable repetition number for NPDCCH is
2048. Whereas for eMTC, MPDCCH needs to be configured with the maximum
configurable repetition number, i.e., 256 repetitions to reach the targeted BLER of
Fig. 12 Performance comparison diagram of NB-IoT and eMTC technologies

1%. Therefore, to support operations in extreme coverage, NB-IoT technology can

be considered more efficient than eMTC technology.
As shown in Fig. 12, eMTC can offer significantly higher uplink throughput due
to the larger device bandwidth and reduced processing time. In addition, Fig. 12
shows that eMTC performs slightly better than NB-IoT in terms of latency using the
EDT procedure. The justification is that NPDCCH and MPDCCH achieve an MCL
of 164 dB for a transmission times of 512 ms and 256 ms, respectively, according to
Tables 2 and 3. Therefore, eMTC technology is capable of serving IoT applications
requiring relatively short response times such as End-Device positioning and voice
over LTE (VoLTE).
Figure 12 shows that eMTC is slightly more efficient than NB-IoT in terms of
battery life to achieve the MCL of 164 dB, but only if the number of base station
receiving antennas is 4 instead of only 2 according to [14]. In fact, the increase of the
number of base station receiving antennas improves the uplink throughput, thereby
reducing UE transmission time to achieve energy savings.
Figure 12 also indicates that NB-IoT offers a higher connection density than
eMTC, which is due to the efficient use of sub-carrier NPUSCH F1 transmissions
with a large number of repetitions. Therefore, NB-IoT technology is likely to meet
IoT applications requiring a massive number of connected devices, such as smart
metering system for gas, electricity and water consumption.
6 Conclusion
To conclude, this paper shows that the five 5G mMTC targets are achieved by both
NB-IoT and eMTC technologies. However, the results of performance evaluation
show that the performances are achieved except under certain conditions regarding
system configuration and deployment, such as the number of repetitions configured
for channels transmission, the number of antennas used by the base station and the
density of base stations. Regarding the coverage and connection density, NB-IoT
offers a better performances than eMTC and precisely for the scenario of a high-base
station density with 500 m inter-site distance. While eMTC performs more efficiently
than NB-IoT in terms of throughput, latency and battery life. Therefore, NB-IoT can
be claimed to be the best performing technology for IoT applications supporting
operations in extreme coverage and requiring a massive number of devices. On the
other hand, to meet the requirements of IoT applications that need relatively shorter
response times, eMTC is the most efficient technology to choose.
References
1. Ghosh, A., Maeder, A., Baker, M., Chandramouli, D.: 5G evolution: a view on 5G cellular
technology beyond 3GPP release 15. IEEE Access 7, 127639–127651 (2019). https://doi.org/
10.1109/ACCESS.2019.2939938
2. Barakabitze, A.A., Ahmad, A., Mijumbi, R., Hines, A.: 5G network slicing using SDN and
NFV: a survey of taxonomy, architectures and future challenges. Comput. Netw. 167, 106984
(2020). https://doi.org/10.1016/j.comnet.2019.106984
3. Ratasuk, R., Mangalvedhe, N., Xiong, Z., Robert, M., Bhatoolaul, D.: Enhancements of nar-
rowband IoT in 3GPP Rel-14 and Rel-15. In: 2017 IEEE Conference on Standards for Commu-
nications and Networking (CSCN), pp. 60–65. IEEE (2017). https://doi.org/10.1109/CSCN.
2017.8088599
4. Ratasuk, R., Mangalvedhe, N., Bhatoolaul, D., Ghosh, A.: LTE-M evolution towards 5G mas-
sive MTC. In: 2017 IEEE Globecom Workshops (GC Wkshps), pp. 1–6. IEEE (2018), https://
doi.org/10.1109/GLOCOMW.2017.8269112
5. 3GPP: TR 38.913, 5G: Study on scenarios and requirements for next generation access tech-
nologies Release 15, version 15.0.0. Technical Report, ETSI. https://www.etsi.org/deliver/etsi_
tr/138900_138999/138913/15.00.00_60/tr_138913v150000p.pdf (2018)
6. El Soussi, M., Zand, P., Pasveer, F., Dolmans, G.: Evaluating the performance of eMTC and NB-
IoT for smart city applications. In: 2018 IEEE International Conference on Communications
(ICC), pp. 1–7. IEEE (2018). https://doi.org/10.1109/ICC.2018.8422799
7. Jörke, P., Falkenberg, R., Wietfeld, C.: Power consumption analysis of NB-IoT and
eMTC in challenging smart city environments. In: 2018 IEEE Globecom Workshops, GC
Wkshps 2018—Proceedings, pp. 1–6. IEEE (2019). https://doi.org/10.1109/GLOCOMW.
2018.8644481
8. Pennacchioni, M., Di Benedette, M., Pecorella, T., Carlini, C., Obino, P.: NB-IoT system
deployment for smart metering: evaluation of coverage and capacity performances. In: 2017
AEIT International Annual Conference, pp. 1–6 (2017). https://doi.org/10.23919/AEIT.2017.
8240561
9. Liberg, O., Tirronen, T., Wang, Y.P., Bergman, J., Hoglund, A., Khan, T., Medina-Acosta,
G.A., Ryden, H., Ratilainen, A., Sandberg, D., Sui, Y.: Narrowband internet of things 5G
performance. In: IEEE Vehicular Technology Conference, pp. 1–5. IEEE (2019). https://doi.
org/10.1109/VTCFall.2019.8891588
10. Krug, S., O’Nils, M.: Modeling and comparison of delay and energy cost of IoT data transfers.
IEEE Access 7, 58654–58675 (2019). https://doi.org/10.1109/ACCESS.2019.2913703
11. Feltrin, L., Tsoukaneri, G., Condoluci, M., Buratti, C., Mahmoodi, T., Dohler, M., Verdone,
R.: Narrowband IoT: a survey on downlink and uplink perspectives. IEEE Wirel. Commun.
26(1), 78–86 (2019). https://doi.org/10.1109/MWC.2019.1800020
12. Rico-Alvarino, A., Vajapeyam, M., Xu, H., Wang, X., Blankenship, Y., Bergman, J., Tirronen,
T., Yavuz, E.: An overview of 3GPP enhancements on machine to machine communications.
IEEE Commun. Mag. 54(6), 14–21 (2016). https://doi.org/10.1109/MCOM.2016.7497761
13. 3GPP: TR 45.820 v13.1.0: Cellular system support for ultra-low complexity and low throughput
Internet of Things (CIoT) Release 13. Technical Report, 3GPP. https://www.3gpp.org/ftp/
Specs/archive/45_series/45.820/45820-d10.zip (2015).
14. Ericsson: R1-1907398, IMT-2020 self evaluation: mMTC coverage, data rate, latency & battery
life. Technical Report, 3GPP TSG-RAN WG1 Meeting #97. https://www.3gpp.org/ftp/TSG_
RAN/WG1_RL1/TSGR1_97/Docs/R1-1907398.zip (2019)
15. Ericsson: R1-1705189, Early data transmission for NB-IoT. Technical Report, 3GPP TSG
RAN1 Meeting #88bis. https://www.3gpp.org/ftp/TSG_RAN/WG1_RL1/TSGR1_88b/Docs/
R1-1705189.zip (2017)
16. Ericsson: R1-1706161, Early data transmission for MTC. Technical Report, 3GPP TSG
RAN1 Meeting #88bis. https://www.3gpp.org/ftp/TSG_RAN/WG1_RL1/TSGR1_88b/Docs/
R1-1706161.zip (2017)
17. ITU-R: M.2412-0, Guidelines for evaluation of radio interface technologies for IMT-2020.
Technical Report, International Telecommunication Union (ITU) (2017). https://www.itu.int/
dms_pub/itu-r/opb/rep/R-REP-M.2412-2017-PDF-E.pdf
18. Ericsson: R1-1907399, IMT-2020 self evaluation: mMTC non-full buffer connection density
for LTE-MTC and NB-IoT. Technical Report, 3GPP TSG-RAN WG1 Meeting #97. https://
www.3gpp.org/ftp/TSG_RAN/WG1_RL1/TSGR1_97/Docs/R1-1907399.zip (2019)
Integrating Business Intelligence
with Cloud Computing: State of the Art
and Fundamental Concepts
Hind El Ghalbzouri and Jaber El Bouhdidi
Abstract The majority of the problems that organizations are currently facing is the
lack of use of cloud computing as shared resources, and it is always looking to become
smarter and more flexible by trying to use recent and powerful technologies, such
as business intelligence solutions. The business intelligence solution is considered a
quick investment and an easy-to-deploy solution. It has become very popular with
organizations that process a huge amount of data. Thus, to make this solution more
accessible, we thought to use cloud-computing technology to migrate the business
intelligence system using frameworks and adapted models to process the data in a
way that is efficient. The most important goal is to satisfy users, in terms of security
and availability of information. There are many benefits to using cloud BI solution,
especially in terms of cost reduction. This paper will address an important definition
regarding cloud computing and business intelligence, the importance of each one, and
the combination of both, evoking a cloud BI, we will present the management risks
to take into account before proceeding to solutions, and the benefits and challenges
of cloud computing will also discussed by comparing existing scenarios and their
approach. The perspective of our future research will be based on this state of the art
that remains an important opening for future contributions.
1 Introduction
Over recent years, augmentation of business application has become very huge,
the data and information stored in different business systems are also increasing,
the business intelligence has become a trendy technology used in lot of company,
and especially the organizations specialized in digital transformation. The business
intelligence has evolved rapidly to the level of recent technology like new software
H. El Ghalbzouri (B) · J. El Bouhdidi

SIGL Laboratory, National School of Applied Sciences Tetuan, Tetuan, Morocco
e-mail: h.elghalbzouri@uae.ac.ma
J. El Bouhdidi
e-mail: jaber.elbouhdidi@uae.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_14
198 H. El Ghalbzouri and J. El Bouhdidi
and hardware solutions, using BI process, the organizations become more scalable,
intelligent and flexible at the data management level, business intelligence has been
historically one of the most resource intensive applications, and it helps the decisions
makers to have a clear visibility and make a better decision to take a good strategy
to improve their business.
However, at a decline of economic, business or other activities, the majority of
organizations find it difficult to make a huge funding investment in technology and
human resources, so for this, they always look for implementing a software solutions
and options to improve theirs business with a lower costs because the costs can be
immensely expensive using recent technologies.
Cloud computing is a model for managing, storing and processing data online via
the Internet, it is an economic solution to resolve the problem of high costs, because
they are subscription-based witches means that we are paying a monthly rental fee,
which is inclusive of all underlying hardware infrastructure and software technology,
and this technical advantages attract organizations to migrate theirs BI system to the
cloud computing. The latter conceptualized three models services: SAAS—software
as service, PAAS platforms as a service, IAAS infrastructure as services, each one
of those has advantages and inconvenient, but the most used service in organization
currently is the SAAS service. Because of its performance and accessibility via any
Web browser, there is no need to install software or to buy any hardware.
Business intelligence contains a set of theories and methodologies that allow to
transform a unstructured data into significant and useful information for business and
decisions makers, BI offers to users an elastic utilization of storage and networking
resources that cloud computing gives a resilient pay-as-you-go manner. The inte-
gration of BI system into cloud environment needs to respect a lot of characteristic
of each BI component and look for the interdependencies between those, and the
deployment of cloud BI presents a lot of challenges at the technical side, conceptual
and organizational.
There are many challenges of cloud BI, including ownership, data security and
cloud providers confidence or host used for deployment. The cloud BI is the hosted
buzzword talked in every industry, and it is the combined power of cloud computing
and bi technology. A data center powered by BI technology can provide access to data
and monitor business performance. This enables the acquisition of a massive data
warehouse with 10 to 100 terabytes of relational databases (RDBMS). This elastic
data evolution is what makes cloud BI so powerful, so for that we can provide a
framework to help organizations to move their data or system to cloud environment,
but like each technology, there is many risks to take into account in terms of security,
financial factor, response time and how to choose the best service cloud to insure a
good migration with minimizing risks. These issues will be very important to take in
account before proceeding to any solution. In this paper, we will discuss and define
the existing approach and their future orientations. We will present also the existing
frameworks and their implementation aspect related to cloud BI system, afterward
we will compare existing scenarios and discussing them in terms of the effectiveness,
of each one, in order to optimize the quality of requirements.
Integrating Business Intelligence with Cloud Computing … 199
2 Overview of Fundamental Concepts
Cloud computing systems deal with large volumes of data using almost limitless
computing resources, while data warehouses are multidimensional databases that
store huge volumes. Combining cloud computing with business intelligence systems
is among the new solution addressed by most of the organization, and in the next
section, we will clarify the basic terms and discuss the associated concepts and their
adaptation to business intelligence domain.
2.1 Cloud Computing
Cloud computing is described as a new model represented by a pool of systems

in which computing infrastructure resources are connected into a network over the
Internet, and it offers a scalable infrastructure for management of data resources.
Adapting a cloud computing as a solution, the costs may be reduced significantly.
The old concept of cloud computing is grid computing, previously grid computing
was used of free resources, which mean that all computers connected to a network
to solve just a single problem at the same time, so if one system fails, there is a high
risk of others to fail. Currently, the extended version of grid computing is named
as cloud computing or technology vendors like IBM, Google, Microsoft [1], so this
new concept of cloud computing tries to resolve this issue by using all the systems
in the network so that if one system fails, another will automatically replace it.
The National Institute of Standards and Technology (NIST) defines cloud
computing as “a model for enabling ubiquitous, convenient, on-demand network
access to a shared pool of configurable computing resources (e.g., networks, servers,
storage, applications and services) that can be rapidly provisioned and released with
minimal management effort or service provider interaction” [2].
Cloud computing offers mobility features known as mobile cloud computing.
Mobile cloud computing is defined as a “new worldview for portable applications,
where the information handling and storage are relocated from the nearby clients to
intense and centralized computing platforms situated in the clouds” [3].
Cloud resources can be reassigned according to customer needs. For example,
some countries do not allow their user data to be stored outside their borders. In
order to achieve this, cloud providers create an infrastructure that can reside in each
country that this offers flexibility in terms of the use of several time zones to work
with.
2.2 Characteristics’ and Deployment Model
According to NIST, cloud computing has four characteristics: on-demand self-

service, broad network access, resource pooling, rapid elasticity and measured
service.
Regarding the “on-demand self-service” and “broad network access”, a consumer
can forecast computer capabilities, such as server time and network storage, and can
access to functionality over the network using multiples platforms client (e.g., mobile
phones, tablets, laptops and workstations).
Cloud infrastructure is characterized by rapid elasticity and resource pooling, for
example the IT resource provider is pooled to serve multiple consumers using the
multi-tenant model, it is rapidly elastic, the capacities are available, and it is often
unlimited and can be scaled to any amount at any time. In addition, cloud infras-
tructure is a very measured service because the cloud system automatically monitors
and optimizes shared resources. It can be monitored and reported on, providing
transparency for both the service provider and the consumer of the service being
used.
Cloud computing has proven adequate for hosting multiple databases, processing
analytic workloads and providing a database as a service. Cloud computing is a new
model that resources of its infrastructures are provided as a service on the Internet.
All data owners can use all services provided by the cloud computing and outsource
their data to him, enjoying a good quality of services.
Cloud computing architecture has three types of services such as software as a
service (SaaS), platform as a service (PaaS) and infrastructure as a service (IaaS).
- Infrastructure as a service (IaaS): The IaaS providers offer the physicals or virtual
machines to request needs of the costumers and help them to deploy theirs logical
solutions, this service is very benefic for the client who needs flexible and security
infrastructure with reduced costs, and the costumer with this service pays only for
the allocated resource, without worrying for hardware or software maintenance.
With IaaS service, the costumer does not manage or control the underlying cloud
infrastructure, and he has just a limited control of certain network component (e.g.,
host firewalls).
• Platform as a service (PaaS): The PaaS service is most service used for cloud
because its facilitates the implementations and testing software solutions, it
provides also a necessary resources needed to run an application, and they are
automatically allocated so that user does not need to do it manually, also in
PaaS service. The consumer does not manage or control the underlying cloud
infrastructure, including the network, servers, operating systems or storage, but
controls the applications and possibly configuration settings for the application
hosting environment.
• Software as a service (SaaS): The SaaS service is described as pay-per-use service,
where cloud providers offer to theirs users a complete configured solution (hard-
ware and software), the costumer or organization has just to pay a monthly or
annually subscription fee that will depend on the customization of the BI solution
and resources allocated to this company. SaaS can offer full access to a business to
implement their BI solution, with a benefits that concern maintenance of software
and hardware solutions, the whole implementation is backed by the provider, and
costumer does not have to worry about this.
Likewise, the customer does not manage or control the underlying cloud infras-
tructure, including network, servers, operating systems, storage or even individual
application features, but there is an exception that may be possible of the user-specific
application configuration [4, 5].
3 Business Intelligence
Business intelligence is a generic term used for combining multiple technologies and
processes, it is a set of software tools and hardware solutions, it stores and analyzes
data in order to help in decisions making, and business intelligence is both a process
and a product. The process consists of methods that organizations use to develop
their useful information that can help organizations to survive and predict certainly
the behavior of their competitors.
Business intelligence system includes specific components. For example, data
warehouses, ETL tools, tools for multidimensional analysis and visualization.
3.1 Data Warehouse
Data warehouse is a database used for reporting and processing data, and it is a central
repository of data created for integration of the data from multiple heterogeneous
sources that support analytical reporting.
Data warehouses are used to historize data and create reports for management,
such as annual and quarterly comparisons.
3.2 ETL
ETL process is responsible for extracting data from multiple heterogeneous sources,
its necessary role is transformation of the data from many different formats into a
common format, and after that, we load it in a data warehouse.
3.3 Multidimensional Analysis
Online analytical processing (OLAP) can be taken as an example of multidimen-

sional analysis. It allows the analysis of user information from different databases of
multiple systems at the same time.
Relational databases are considered like two-dimensional, and OLAP process
is multidimensional, which means that users can analyze multidimensional data
interactively from multiple perspectives.
3.4 Restitutions
This component is very important, the main objective of the restitution of the data
is to communicate a good information for a decision-makers and to show them the
results in a clear and effective way, through graphical presentation like dashboards,
more and more the data collected well, the decision-maker takes the right decision of
their business, and this helps them to communicate good hypotheses and prediction
for the future.
3.5 BI Architecture
To process this huge data in a BI system and integrate it through a cloud environment,
we need to run a basic architecture based on the BI cloud solution, this architecture
contains cloud-compatible components that facilitate interaction between them, and
in order to achieve this migration, cloud computing provides an environment based
on the use of multiple services separated by layers forming the hardware and software
system [6] (Fig. 1).
Data integration: This is related to ETL tools that are needed to transform the data
purifying process.
Database: This is related to multidimensional or relational databases.
Data warehousing tools: It is related to a package of applications and tools that
allow the maintenance of data warehouse.
Bi tools: This analyzes the data that are stored in data warehouse.
Hardware: It is related to storage and networks on which data will be physically
stored.
Software: This refers to everything related to the operating systems and drivers
necessary to handle the hardware.
Fig. 1 BI on the cloud architecture
4 Integrating Business Intelligence into Cloud Computing
4.1 Cloud BI
The business intelligence solution based on a cloud computing platform is called

“Cloud Business Intelligence". Cloud business intelligence is a revolutionary concept
that makes it possible to deploy the BI system to the cloud environment, which is
easy to deploy and flexible, also with reduced cost, cloud BI is based on various
services offered by the cloud, for example, software as a service business intelligence
(SAAS BI), and it is a software of delivery model for business intelligence in which
applications are typically deployed outside the corporate firewall hosted in Web
site and accessed by a secure Internet connection for the end users [7]. BI cloud
technology is sold by providers, with a subscription or pay-per-view basis, instead
of using a traditional licensing model with annual maintenance fees.
4.2 Migration of BI to the Cloud
Integrating BI into a cloud environment will solve the problem of technology obso-
lescence and is an advantage for organizations in term of scalability that will be
achieved, no matter how company’s data are complex. This integration is not related
only for scalability, but also for elasticity and ease of use. By Elasticity that’s refer to
ability of a BI system to absorb continually information’s from new added software
[8].
Technologies considered by Dresney Advisory include cloud technologies,

including query and reporting tools, OLAP, data mining, analytics, ad hoc anal-
ysis and query tools and dashboards. The most researched features of cloud BI in
2018 are dashboards, advanced visualization, data integration and self-service [9].
Some of the features of BI migration to the cloud include the ease of deployment
described in the following example: A company that decides at one time to build a
new online solution would say helpdesk for its customers. This can be implemented in
the cloud and integrated into BI processes in a short time, without having to purchase
additional hardware, such as servers, so helpdesk software will run as well.
Implementing a Web solution based on BI system to help decision-making process
is also determined by portability and eases of access on any Web browser, that is,
the user will visualize the information using various devices; however, it is inside or
outside a company and will be constantly informed [10].
Pros and cons of integration of a BI solution into cloud environment are as follows.
Pros:
• Scalability and elasticity;
• Reduced costs;
• Ease of use and access;
• Availability;
• Hardware and software maintenance.
Cons:
• Privacy
• Government regulations (where applied).
As stated before in the cloud computing section of this article, privacy remains an
issue. BI solutions are not an exception. The security provided by BI solutions is only
at an user interface (UI) level. The data stored on cloud database is exposed to the
provider. Government regulations are, in some cases, a barrier in the migration of BI
solutions of companies to a cloud infrastructure outside the border. This represents
a downside in terms of cloud computing expenses. The cloud providers that are
located in the same country with an organization might have higher costs than foreign
providers.
4.3 Comparison Between Private and Public Cloud BI
To make a good decision for business in order to migrate their business intelligence
system to cloud environment, we must choose best type of cloud computing such us
the private, public or hybrid one. The hybrid deployment combines IT services from
private and public deployments [11]. In terms to ensure a high security of the data
migrated in public cloud, organizations have little control of their resources, because
the data are open for public use and accessed via the Internet. It does not require
a maintenance or need time changing because the cloud provider is responsible for
public cloud: It is a shared cloud that means that we are paying only for what we
need, it is used especially in testing Web sites or pay-per-use applications. In public
cloud, organizations have little control of their resources, because the data are open
for public use and accessed via the Internet. It does not require a maintenance or
need time changing because the cloud provider is responsible for it.
The BI infrastructure software platforms available for cloud providers hosting are
SAP, RightScale, blue insight (IBM), Web-Sphere, Infromatica and Salesforce.com.
For these platforms, we find those who are public and private one, for example,
• RightScale: It is a public cloud BI, and it is open source and publicly available,
for all and not just for organizations. It is a full BI cloud that all business intelli-
gence functions are processed. For example, report generation, online analytical
processing, comparative and predictive analysis.
• Blue Insight (IBM): It is a private BI cloud and is not open source. The data
management technique that is used by it has more than a petabyte of data storage.
It not only supports forecasting, but it is also scalable and flexible.
• Salesforce.com: It is a public BI cloud and is not open source. The data manage-
ment technique it uses is automated data management. It supports forecasting, but
it is not flexible, and it has low scalability.
• Informatica: It is a public BI cloud and is not open source. The data management
technique that is used by it is data migration, replication and archiving. It not only
supports forecasting, but it is also not flexible, and it has low scalability [12, 13]
This comparison concludes that the best public solution for cloud BI is RightScale,
while for private solutions, we can take only those implemented by IBM (Table 1).
Table 1 Characteristics of public, private, hybrid cloud computing

Features Public cloud Private cloud Hybrid cloud
Scalability Very high Limited Very high
Reliability Moderate Very high Medium to high
Security Totally depends on High-level security Secure
service providers
Performance Low to medium Good Good
Cost Free Depends on resource Lower than private cloud
located
Flexibility High Very high Very high
Maintenance No maintenance There is a maintenance
Time saving Yes No Yes
Pricing Pay per use Fixed Fixed
Examples Amazon EC2, VMWARe, Microsoft, IBM, VMWARe, vCloud,
Google appEngine KVM, XEN, IBM Eucalyptus
5 Inspected Scenarios
With the rapid evolution of business intelligence technology and its new concept of
cloud integration, many research and studies have been done with implementation of
various solutions. Now, it has become difficult for societies to choose the best one.
In our case of study, several scenarios based on business intelligence cloud will be
analyzed. In this section, we will discuss two of these scenarios used to illustrate the
issues discussed previously.
5.1 Scenario 1: OLAP Cloud BI Framework Based

on OPNET Model
This scenario deals with the hosting of BI systems in the cloud based on OLAP cubes
by integrating them on the Web.
The data structures used in the OLAP cube must be converted to XML files based
on DTD structures to be compatible with the Web object component (Web cubes),
and this solution provides better performance for exploring data in the cloud.
For this, we integrate an OLAP framework comprising the dashboards and the data
analytics layer as SAAS model, for the integration of data warehouse and OLTP/DSS
databases as PAAS model and for the underlying servers, databases we integrate it
as IAAS model.
In our case, we used the OPNET model, and this network can be used to integrate
BI and OLAP applications that has been designed in such a way that the load can
be evenly distributed to all the relational database management systems (RDBMS)
servers in such a way that all RDBMS servers are evenly involved in receiving and
processing the OLAP query load. The main architecture of OPNET model is that it
comprises two large domains—the BI on the cloud domain and the extranet domain
comprising six corporates having 500 OLAP users in each as shown in Figs. 2 and 3
The application clouds are IP network cloud objects comprising application server
arrays and database server arrays, connected to a cloud network [14].
In the following figure, the BI framework contains four numbers of Cisco 7609
series layer 3 high and routing switches connecting in such a way that the load can
be evenly distributed. The cloud switch 4 is routing all inbound traffic to the servers
and sends their responses back to the clients.
The cloud switches 1 and 3 are serving four RDBMS servers, and the cloud switch
2 is serving all the OLAP application servers. An array of five numbers of OLAP
application servers and an array of eight numbers of RDBMS servers. The blue dotted
lines from each OLAP server are drawn to all the RDBMS servers indicating that each
OLAP server will use the services of all the RDBMS servers available in the array to
process a database query. The customer’s charge is routed to the OLAP application
servers using destination preference settings on the client objects configured in the
extranet domain [14] (Fig. 4).
Fig. 2 Architecture of OPNET model [14]
Fig. 3 BI on the cloud architecture [14]

Fig. 4 Extranet domain comprising six corporate shaving 500 OLAP users in each corporate [14]
OLAP queries are 10 to 12 times heavier than normal database queries. This
explains that each query extracts multidimensional data from several schemas, so
the query load in OLAP transactions is very high. For example, if the OLAP service
on a cloud can be used by hundreds of thousands of users, the back-end databases
must be partitioned in parallel to manage the OLAP query load. The centralized
schema object must be maintained with all tenant details, such as—identification,
user IDs, passwords, access privileges, users per tenant, service level agreements and
tenant schema details [14, 15].
A centralized schema object can be designed to contain the details and privileges
of all tenants on the cloud. The IAAS provider should ensure a privacy and control,
both of load distribution and response time pattern, the OLAP application hosted on
the cloud may be not compatible with the services, so for this, we can use the SAAS
provider that can allow the creation of an intermediate layer to host a dependency
graph that helps in dropping the attributes not needed in the finalized XML data cube.
BI and OLAP must have a high level of resources as a multilayer architecture
composed of multidimensional OLAP cubes with multiplexed matrices representing
the relationships between various business variables. All the cubes send OLAP
queries to data warehouses stored in RDBMS servers. The response time size of
an OLAP query is typically 10–12 times greater than an ordinary database query.
5.2 Scenario 2: Model of Cloud BI Framework by Using

Multi-cloud Providers
The approach of this scenario is about the factors, which affect the migration of BI to
cloud, so that we adapt organizational requirements and different deployment model
to alternative cloud service models, such as the example system shown in Figs. 5 and
6.
This model of framework helps decision-makers to take into account of cloud BI
system as well as security, cost and performance.
(1) Bi-user represents the organization’s premises where BI system is running
before the cloud migration.
In this deployment, the BI-user pushes and extracts the data to cloud environ-
ment. Push and pull communications are secured by encrypting the dataflow with
the transport layer security/secure sockets layer cryptographic protocols (TLS/SSL).
Fig. 5 BI system in a cloud provider [16]

Fig. 6 BI system in a cloud provider [16]
Once the Bi-user transfers its data to cloud premises, the BI tools start to run in
cloud environment and analyze all this data which are stored in data warehouse to
generate data analysis and report, in order to be accessed by different devices such
as the workstation, tablet or mobile phone shown at rounded circle (3) in Fig. 6.
Regarding the trust of the organization to the data transferred, our approach takes
a partial migration strategy using more than one cloud provider, to insure security
and opting for partial migration that sensitive data stay locally and other components
move to the cloud providers, while others stay locally. And for Bi-user pushes, the
anonymized data needed to use BI tools on IaaS, PaaS, or SaaS platforms to leverage
the additional scalable resources for available BI tools.
This partial migration is done with using more than one cloud provider—namely
cloud provider and cloud provider 2—see rounded circle (2) to insure a portability,
synchronization module and high security using end-to-end SSL/TLS encryption to
secure the communication between cloud premises as shown in Fig. 3 [16] (Fig. 7).
This approach gives users an updated data, however, of the number of cloud
providers used to explore it.
For this, we address a data synchronization with a globally unique identifier
(GUID) to enforce consistency among data transferred from source to target data
storage and harmonize data over time.
By deploying a copy of a BI system to different cloud environments with harmo-
nized data between them, we avoid a problem of vendor lock in and that ameliorates
a resilience of Bi systems, this solution gives an isolation of a system so all failures
that can have do not attack our components, and we can manage the BI system from
a safe computing environment outside, if we observe a failure to control it. Also, in
case of failure, BI system tolerates it as the framework ensures it with availability, by
letting the overall system transparently use the BI system running in another cloud
provider with model of synchronization.
The mechanisms of this framework work as interactions between the data in local
premises and cloud environment, and this interaction can be affected by several risks
for example:
• The loss of data can happen during the migration of the system BI to cloud
environment, because the size of the data to be transferred to cloud environments
has implications in terms of the cost of large-scale communications and overall
system performance, so this cloud migration framework can recover and save the
data to avoid this case, so for this, our framework re-computes the data transferred
and compares it with stored one [17].
• Security, we supported by granting access to these data only to users with a user
role related to them and the necessary level of authorization.
For the sensitive data we do tokenization method to replace the original data, for
example we use the Social Security Number with randomly generated values. The
Integration of Business Intelligence into the cloud keep the original format of the
data and preserve the functionality running on the cloud premises, and we translate
from token to real data in the cloud provider side [18].
6 Comparison of the Scenarios
Discussing about these two scenarios for the integration of BI systems in the cloud,
we can see that each of them has its strengths and weaknesses at the same time.
The first scenario is based on the use of OLAP framework based on OPNET which
is a structured network model with three different cloud services: SAAS, PAAS and
IAAS, which are offered by providers. The use of network modeling OPNET and
OLAP framework, OLAP queries are 10 to 12 times heavier than normal database
queries. Because each query makes an effort to extract multidimensional data from
multiple schemas and the load in OLAP is very high, this implies that when multiple
users need to use this service the load on the back-end databases must be balanced
and partitioned with schemas in parallel to handle OLAP queries. BI and OLAP
must have a high level of resources as a multilayer architecture composed of multi-
dimensional OLAP cubes with multiplexed matrices representing the relationships
between different variables in the enterprise.
Concerning the second scenario, we discussed a partial migration of the BI system
to the cloud, using more than one cloud provider, in this migration, we benefit with
a high level of security of the data so that sensitive data stays locally, also using
more than one cloud provider that insures performance at the data transfer level, so
Bi-user is always informed with updated data over time, however, of the number of
cloud providers used to explore the data. But with this solution, some data can be lost
while the migrating, so that is because of the implications in terms of costs induced
by large-scale communications and overall system performance. That is why our
cloud migration framework backs up and recovers data in the event of a disaster to
protect against this eventuality, so lost data are still a problem in cloud BI migration
at different levels, moreover, that every organization has a level of security that it
wants to implement for its solution.
7 Conclusion and Perspective
Cloud computing in recent years became a trend of the majority of organization who
uses business intelligence process, and it has a very important role to facilitate the
integration and access to the information with a level of performance. The cloud BI
solution has been improved with his flexibility of implementation, scalability and
high performance of software and hardware business intelligence tools.
In this paper, we discuss the importance of business intelligence for decision-
making and the importance to integrate it into the cloud environment, in order to
make it flexible to access into the data.
We discussed also about some considerations that we should take into account,
and in order to choose a best service for cloud, we define some components and
architecture of BI. In additionally, the benefits and inconvenience of cloud BI have
been discussed; finally, we compared public, private and hybrid cloud with the char-
acteristics of each, we made a case study of existing solutions, and we compare them
with taking into account two important scenarios.
The cloud BI has many benefits in terms of data processing performance, but some
challenges still need more researches; for example, security challenges, performance
and response time of requests in the OLAP process will be different and much more
complex. In the next step of our research, we will develop other application scenario
to verify it in the practice. So, this state of the art has significant openings for future
contributions, and it is only the beginning of studies on future challenges.
References
1. Kumar, V., Laghari, A.A., Karim, S., Shakir, M., Brohi, A.A.: Comparison of Fog computing
& cloud computing. Int. J. Math. Sci. Comput. (2019)
2. Mell, P., Grance, T.: The Nist Definition of Cloud Computing, pp. 800–145. National Institute
of Standards and Technology Special Publication (2011)
3. Laghari, A.A., He, H., Shafiq, M., Khan, A.: Assessing effect of Cloud distance on end user’s
Quality of Experience (QoE). In: 2016 2nd IEEE International Conference on Computer and
Communications (ICCC), pp. 500–505. IEEE (2016)
4. http://faculty.winthrop.edu/domanm/csci411/Handouts/NIST.pdf
5. Cloud Computing: An Overview. http://www.jatit.org/volumes/researchpapers/Vol9No1/10V
ol9No1.pdf
6. Mohbey, K.K.: The role of big data, cloud computing and IoT to make cities smarte, Jan 2017
7. https://searchbusinessanalytics.techtarget.com/definition/Software-as-a-Service-BI-SaaS-BI
8. Patil, S., Dr. Chavan, R.: Cloud business intelligence: an empirical study. J. Xi’an Univ.
Architect. Technol. (2020) (KBC North Maharashtra University, Jalgaon, Maharashtra, India)
9. Bastien, L.: Cloud Business Intelligence 2018: état et tendances du marché Cloud BI’, 9 april
2018
10. Tole, A.A.: Cloud computing and business intelligence. Database Syst. J. V4 (2014)
11. Westner, M., Strahringer, S.: Cloud Computing Adoption. OTH Regensburg, TU Dresden
12. Kasem, M., Hassanein, E.E.: Cloud Business Intelligence Survey. Faculty of Computers and
Information, Information Systems Department, Cairo University, Egypt
13. Rao, S., Rao, N., Kumari, K.: Cloud Computing : An Overview. Associate Professor in
Computer Science, Nova College of Engineering, Jangareddygudem, India
14. Al-Aqrabi, H., Liu∗, L., Hill, R., Antonopoulos, N.: Cloud BI: Future of business intelligence
in the Cloud
15. https://onlinelibrary.wiley.com/doi/abs/https://doi.org/10.1002/cpe.5590
16. Juan-Verdejo, A., Surajbali1, B., Baars2, H., Kemper, H.-G.: Moving Business Intelligence to
Cloud Environments. CAS Software A.G, Karlsruhe, Germany
17. https://www.comparethecloud.net/opinions/data-loss-in-the-cloud/
18. https://link.springer.com/chapter/10.1007/978-3-319-12012-6_1
Distributed Architecture for
Interoperable Signaling Interlocking
Ikram Abourahim, Mustapha Amghar, and Mohsine Eleuldj
Abstract The interoperability in railway systems and especially in railway sig-

naling interlocking is an issue for mobility need and to master the evolution of
technology. Today, traffic of trains need a continuous communication, and the diver-
sity of technologies makes the interoperability difficult. In Europe, some projects
are in development to solve the interoperability problem. The European Rail Traf-
fic Management System (ERTMS) is the first project deployed: It aims to establish
an exchange of signaling information between interlocking and train. EULYNX is
another project interested in standardization of interfaces between field equipment
and computer interlocking. In this paper, we propose an architecture of computer
interlocking that deal with the interoperability between adjacent calculator through a
combination between functional blocks of IEC 61499 standard and service-oriented
architecture (SOA). Moreover, the combination is executed on a distributed mode of
sub-calculators that compose the calculator of the computer interlocking.
1 Introduction
Railway signaling is the system allowing a fluent mobility of trains and, at the same
time, ensuring its security.
The main roles of railway signaling are: the spacing between successive trains and
traffic in two opposite directions on the same track between stations; the management
of internal movement at the station and the protection of trains from speeding and
the risk of derailment.
The research work for this paper is the result of a collaboration between EMI and ONCF.
I. Abourahim (B) · M. Amghar · M. Eleuldj

Mohammed V University in Rabat, Rabat, Morocco
e-mail: ikramabourahim@research.emi.ac.ma
M. Amghar
e-mail: amghar@emi.ac.ma
M. Eleuldj
e-mail: eleuldj@emi.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_15
216 I. Abourahim et al.
To meet those requirements, a set of principles, rules, and logic process put
together in the railway signaling system to ensure the safety of trains’ mobility by
acting on field equipment (points and signals . . .). The command of those equipment
and the verification of their statutes are usually done from a signaling machine called
interlocking.
Actually, most of infrastructure managers migrate to computer interlocking that
allows an easy management of field equipment and gives new functions and services.
But this new technology is faced to a lack of homogeneity, and then, a difficulty of
communication between interlocking proposed by the various suppliers especially
on borders. Also, each modification or upgrade of the infrastructure requires a partial
change that may cost more than the total change of the interlocking.
The first project initiated in Europe to deal with interoperability issue is ERTMS
[1]. This system aims to facilitate trains’ mobility from a country to an other without a
big investment through a direct transmission of signaling information to the onboard
system of the train.
A group of railway infrastructure managers in Europe has carried out a project enti-
tled EULYNX [2], since 2014, to standardize the communication protocol between
computer interlocking and field equipment independently of their industrial supplier.
An other need of interoperablity is the communication between interlockings on
border. Actually, most solutions deployed in different countries are using electrical
logic even if they rely on computer interlockings in each side of border. This solution
cannot be general but is realized specifically for each case differently.
Unfortunately, there are not enough articles in the literature that study the inter-
operability between computer interlocking in the signaling railway field due the
lack of knowledge synergies between manufacturer competitors since R&D is their
respective competitive advantage.
Our paper presents in the first part a review about signaling systems and some
existing architecture of computer interlocking. In the second part, we introduce our
approach for interoperability of computer interlocking that unifies the architecture
and facilitates the communication on borders between stations through SOA [3]
and IEC 61499 standard [4]. Moreover, in third part, we explain our proposition of
distributed architecture model to combine the interoperability with better processing
of the computer interlocking . After that, we analyze results of the execution of our
proposed architecture. And finally, we conclude with summary and perspectives of
the project.
2 Railway Signaling System: Existing
2.1 Railways Control System
Railway systems, like all transport systems, are in continuous evolution and develop-
ment. They take advantage of new technologies to improve operations and services.
Distributed Architecture for Interoperable Signaling Interlocking 217
Fig. 1 Railway control system architecture
Safety is a major advantage of rail, particularly for signaling system. Rail traffic
safety depends on the reliability of signaling systems especially when it is about an
automatic command and control system.
The railway control system allows to continuously supervise, control, and adjust
the train operations, ensuring a safe mobility of trains at all times through a continuous
communication between interlocking and field equipment.
Being primarily electrical or mechanical, field equipment needs intermediate ele-
ments to communicate with computer interlocking called objects controllers. Then,
we get a global network ensuring a safe and reliable interaction as shown in Fig. 1.
For high-level process supervisory management, many architectures of signaling
system are deployed to allow an interconnection and a continuous exchange between
object controllers of all field equipment and the calculator of computer interlocking.
This calculator is also operating in an interconnection with the control station.
2.1.1 Control Station
The control station receives information from the interlocking and makes it available
to the operator by connecting it to the control device of the railway system. It provides
the following features: Visualization of the state of signaling equipment; ordering
the desired routes; route control; the positions of the trains; the state of the areas;
remote control of remote stations.
2.1.2 Interlocking
The interlocking being the intermediary between the control station and field equip-
ment, it receives its inputs from both systems and also from the neighboring inter-
locking. It does the necessary processing and calculation, then returns the orders to
the field equipment and the updates to the control station, and it sends data back to
the neighboring interlocking. Its main function is to ensure operational safety and
adequate protection against the various risks that may arise and affect the safety of
persons and property.
To ensure safe operation, the interlocking must respect IEC 61508 standard. The
latter applies mainly in cases where it is the programmable automaton that is respon-
sible for performing security functions for programmable electrical, electronic or
electromechanic systems. IEC 61508 defines analytical methods and development
methods for achieving functional safety based on risk analysis. It also determines
the levels of security integrity (SIL) to be achieved for a given risk.
The SIL can be defined as an operational safety measure that determines recom-
mendations for the integrity of the security functions to be assigned to safety systems.
There are four levels of SIL and each represents an average probability of failure
over a 10-year period.
• SIL 4: Very significant impact on the community resulting in a reduction of the
danger from 10,000 to 100,000.
• SIL 3: Very important impact on the community and employees reducing the
danger from 1000 to 10,000
• SIL 2: Significant protection of installation, production, and employees reducing
the danger from 100 to 1000.
• SIL 1: Low protection of installation and production resulting in a reduction in
danger of 10–100.
For any type of interlocking, the SIL 4 level is required.
2.1.3 Object Controller
The object controller subsystem consists of several types of hardware cards, its role
is to connect the computer interlocking to field equipment: signals, switch engines,
etc., and return their state to the interlocking.
The object controller system has two main functions:

• Control communications with the computer switch station.
• Provide the electrical interface to track equipment.
The object controller (OC) receives commands from the central interlocking system
(via the transmission system and transmission controller unit) and executes them,
converting the software order into the appropriate electrical signal for the field object.
It handles trackside objects and returns state information to the central calculator.
2.1.4 Field Equipment
Field equipment refers to different object of trackside which acts locally for the safety
of train movement: like signals, ERTMS balises, track circuits, switches, track pedals,
etc.
• Signals: The signals are essentially used to perform the following functions: stop
signals, speed-limiting signals, and directions signals. Each of these functions
usually includes an announcement signal and an execution or recall signal.
• ERTMS balises Point-to-track transmission transmitter, using magnetic transpon-
der technology. Its main function is to transmit and/or receive signals. The
Eurobalise transmits track data to trains in circulation. For that, the Eurobalise
is mounted on the track, in the center or on a crossbar between two rails. The
data transmitted to the train comes either from the local memory contained in the
Eurobalise or from the lateral electronic unit (LEU), which receives the input sig-
nals from the lights or the interlocking and selects the appropriate coded telegram
to transmit.
• Track circuits (CdV): allows an automatic and continuous detection of the presence
or absence of vehicles at all points in a specific section of lane. By therefore, it
provides information on the state of occupation of an area that will be used to ensure
train spacing, crossing announcements, and electrical immobilization of switches.
Its detection principle effectively ensures that the signal is closed entrance to the
area not only as soon as an axle enters it, but also when an incident intervenes on
the track (broken rail, track shunted by a bar metal closing the track circuit, etc.).
• Switches: are a constituent of the railway that allows support and guidance of a
train in the a given route during a crossing. The motors are used to move and main-
tain railway switches in the appropriate position. Switches can be automatically
controlled from a station or on the ground by an authorized person.
• Track pedals: These devices, also called pedal repeaters (RPds), are located near
the track, and they are intended to indicate the presence of a train in a part of the
track where they are located. When a train passes, its axles press a pedal and close
an electrical circuit. This pedal remains supported until the passage of the last axle.
Fig. 2 Decentralized architecture
2.2 Computer Interlocking Architectures
As the computer interlocking enables to take advantage of new technologies and

the fluidity of communication that emanates from them, many architectures are used
actually to ensure a continuous communication between elements of signaling control
systems.
There are two types of architecture that are globally implemented: Decentralized
and Centralized Architectures.
2.2.1 Decentralized Architecture
In this architecture (Fig. 2), interlocking exchanges data with objects controllers and
the station control related to its area of control, and there is a direct link for exchange
between adjacent interlocking. This direct link, when it is about different suppliers
of computer interlocking, in an electromechanic interface because the protocol of
communication is most of the time different and needs the acceptance of suppliers to
collaborate. But when we have the same supplier, the serial or Ethernet link is chosen,
and the communication is adequate with the context of computer interlocking.
2.2.2 Centralized Architecture
In this architecture (Fig. 3), data from OCs is send to the interlocking managing the
area where those objects are located, and all interlocking of a region exchange with
the same control station that called central command station.
Fig. 3 Centralized architecture
When the command and control are centralized, the communication between
adjacent interlocking does not need a direct and specific link, also more functionalities
become possible like:
• Tracking Trains: Each train has an identifier that is monitored along the route.
• Programmable List: allows to prepare a route order list for one or many days.
• Automatic Routing: regarding to a number of train and its position a route or
itinerary is commanded automatically.
Some of those operations need an interaction with external system through the exter-
nal server.
3 Interoperability of Computer Interlocking
3.1 The European Initiatives for Railway Signaling

Interoperability
Rail transport accompanies openness and free movement between European coun-
tries. So, to ensure a security of trains traffic, European infrastructure managers
needed to unify their signaling systems in order to establish interoperability through
new projects like European Rail Traffic Management System (ERTMS) [1] and EUL-
YNX project [2]
3.1.1 ERTMS
Before the implementation of ERTMS [1], each country had its own traffic man-
agement system used for transmission of signaling information between train and
track-side.
Then, the European Union (EU), conducted by the European Union Agency for
Railway (ERA), leaded the development of ERTMS with principal suppliers of sig-
naling systems in Europe.
The main target of ERTMS is to promote the interoperability of trains in Europe.
It aims to greatly enhance safety, increase efficiency of train transports, and enhance
cross-border interoperability of rail transport in EU.
3.1.2 EULYNX Project
The implementation of computer interlocking result in a difficulty of communication

with field equipment when suppliers are different. So the European community,
especially 12 European Infrastructure Managers, leaded an initiative called EULYNX
project [2, 5]. This project aims to standardize interfaces and communication protocol
between field equipment and computer interlocking.
3.2 Software Architecture Proposition for Interoperability
Until now, the problem of interoperability between computer interoperability is not

approached specifically in research and literature articles and R&D for industrials is
their competitive advantage.
In our approach, we choose to deal with software architecture to meet the challenge
of interoperability and homogeneity of signaling interlocking. Then, we rely on two
principles of software architecture from the fields of computer science and industrial
computing which are the SOA [3] and the functional blocks according to IEC 61499
[4]. This combination has yielded evidence of success regarding global industrial
interoperability and the ease of hot upgrading without interrupting production [6].
3.2.1 IEC 61499 Standard
The international standard IEC 61499 [4], dealing with the topic of function blocks
for industrial process measurement and control systems, was originally published in
2005 and revised in 2012.
The IEC 61499 standard [4] relies on an execution model driven by the event. This
execution model allows a rationalization of the execution of all functions according
to a justified order and need.
3.2.2 SOA: Service-Oriented Architecture
Service-oriented architecture [3] (SOA) is a form of mediation architecture that is an

application for services (Software components) implemented in interaction model
with :
• strong internal consistency using a pivot exchange format, usually XML or JSON.
• loose external couplings using an interoperable interface layer, usually a WS (Web
Service).
The service-oriented architecture is a very effective response to the problems that
companies face in terms of re-usability, interoperability, and reduction of coupling
between the different systems implemented on their information systems. Thus, we
distinguish three types of services in the SOA [3] architecture:
• service provider: which provides the information or the service/function.
• service requester: which requests the information or the service/function.
• service repository: the directory of all available services.
4 Distributed Architecture for Interoperable Signaling

Interlocking
Some work in railway signaling domain deal with the approach of distributed archi-
tecture [7, 8] in different ways independently of interoperability issue.
To perform the proposition of interoperability through functional blocs, we con-
sider a new proposal of distributed architecture for signaling system.
Indeed, functional blocks regarding IEC 61499 standard allow to decompose a
system on elementary functions that can be executed in a distributed environment
respecting a synchronous logic. So, we choose to keep a central control and super-
vision and distribute calculation related to interlocking.
Previously, in central architecture, the central calculator is connected directly to
object controllers (OC). Each OC is an intermediary between the calculator and the
field equipment only for data exchange (Fig. 4).
So, as a distributed configuration, we propose for each station a network of sub-
systems as shown in Fig. 5:
• Principal functions are executed in the central calculator.
• Functions related to field equipment in borders of station are executed on auxiliary
station calculator and only needed information is sent to central calculator. As an
example, if we consider a plan station like Fig. 6, we will cut the station into two
parts, left and right; then, the equipment of each part is linked to the auxiliary
station calculator right or left.
• Functions related to outside area between stations are executed on auxiliary block
calculator and only needed information is sent to central calculator.
Fig. 4 Central process deployment
Fig. 5 Distributed process deployment

Fig. 6 Station plan
Fig. 7 Functional block diagram Interlocking—SysML
As a result of the choice of software architecture explains in previous part, we

made the functional model shown in Fig. 7 and explained in [5].
This model allows to categorize functions of interlocking in families of functional
blocks and then their distribution in different calculators of our proposed distributed
architecture (Fig. 5).
As centralized architecture is the most architecture used for computer interlock-
ing commissioning around the world, we choose to make a comparison between
centralized and distributed architectures.
4.1 Quantitative Calculation Comparison
Regarding to Moroccan principals of railway signaling, we made a new distribution

of functions of interlocking. Instead of calculating all the functions in the central
calculator, we chose to distribute them between the central and the auxiliary station
or the auxiliary block calculators.
Each function is related to an execution time. That result on execution time of

cycle modeled in (1):

N
texe = ti , (1)
i=1
texe : cycle execution time.

ti : execution time of each function depending on the number of variables and oper-
ations.
N: total number of functions.
So reducing the number of functions executed by a calculator allows reducing
time of execution in a cycle by each calculator and then ensure .
4.1.1 Distribution on Auxiliary Station
At each station, we separated the functions of field equipment on either side of the
station into an auxiliary station calculator that results on having two auxiliary station
calculator as mentioned in Fig. 11.
In Table 1, we give a quantitative overview about functions that we keep executing
in central calculator and those which we migrate to auxiliary station calculators. This
distribution respect categories are mentioned in Fig. 7.
We notice that the numbers mentioned in Table 1 are related to all the functions
that we use not all in the same station or in the same auxiliary station because they
relate to different possible configurations in the station.
4.1.2 Distribution Auxiliary Block
In auxiliary block, the functions executed are related to signals and areas. Indeed
between stations, train traffic management is done automatically through the logical
Table 1 Computing interlocking functions

Categories Total functions Central execution Auxiliary station
execution
Itinerary 103 98 5
Point 11 9 2
Area 18 5 13
Signal 195 0 195
Protection 12 9 3
Authorization 56 10 38
Field switch 10 0 10
Total 405 159 246
link between signals and the occupation of the areas that frame them. So all functions
are executed in the auxiliary block, and only results are sent to the central calculator
for supervision need.
4.2 Flow Exchange Variables
For a centralized or distributed architecture, the exchange of the state of variables

between functions is essential to enable a global execution of the system with coher-
ence and logical synchronization. Each category of functions has a flow of data
exchanged as shown in Fig. 8.
In the case of centralized architecture, all information collected by object con-
trollers is sent to the central calculator. But in the case of distributed architecture,
only the results of the auxiliary calculators are sent to central calculator.
As an example, we choose signal category to made a comparison of data flow between
central and distributed architectures due to the fact that the functions of this category
are calculated entirely in auxiliary calculator.
For internal variables of signal’s functions, we can reduce the flow of data from
central to object controller when calculation is made in the auxiliary calculator, so
the exchange of 56 variables in centralized architecture (Fig. 9) is reduced to 0
variables in distributed architecture (Fig. 10). Also, it allows to reduce data flow to
central calculator from 45 variables that are sent from object controllers (Fig. 9) to
23 variables sent from auxiliary calculator (Fig. 10). If we consider a linear model
(2) for communication time:
tcom = a + b.K v (2)
tcom : communication time.

a: latency.
b: debit.
K v : nombre of variables.
So, reducing the number of variables exchanged allows reducing time of commu-
nication.
5 Analysis and Discussion of Results
To perform our choice of distributed architecture, we made a simulation through

ISaGRAF simulator regarding to the deployment network shown in Fig. 11.
This simulation combined the use of functional blocks respecting the IEC 61499
standard and their execution in a distributed architecture for computer interlocking.
We can distinguish in Fig. 3 in the center of the simulator of the central computer
or calculator; the simulators that surround it relate to the auxiliary station computers
and finally at borders the auxiliary block computers. For the simulation, we were
Fig. 8 Exchange diagram
Fig. 9 Flow signal’s data—Centralized architecture
Fig. 10 Flow signal’s data—Distributed architecture
content with two auxiliary block, but in reality we can find four or more depending
on the extent of the distance between the stations.
The results of simulation ensure, on one side, the equivalent between central
and distributed architecture regarding synchronous logic of execution of functions
related to the interlocking. On the other side, the distribution of functions’ execution
allows a reduction at the level of the charge on central calculator and the time of
execution cycle (texe ) as well as the decrease in the flow of exchange (tcom ) between
the interlocking and the field equipment.
Fig. 11 Distributed architecture-deployment network
6 Conclusion and Perspectives
Technological change and the need for more speed for train mobility requires support
for infrastructure and especially rail signaling systems. Then, the use of computer
interlocking for signaling management and the interoperability between interlocking
and related equipment becomes an evidence.
Moreover, our proposal takes into account interoperability through elementary
functions by the use of functional blocks regarding IEC 61499 standard and also
through a distributed architecture of calculators that facilitate the exchange on borders
between stations in the same country or between different countries.
A simulation confirmed the relevance of the model using some functions respect-
ing signaling principals of Morocco. An other simulation will be considered in the
upcoming work steps having for objective the parameter test of real stations. Through
the analysis of different parameters, mainly the respect of the process scheduled in
the distributed system architecture model and the exchange security expectations,
we can then consider the deployment phase.
References
1. The ERTMS/ETCS signaling system an overview on the standard European interoperable sig-
naling and train control system. http://www.railwaysignalling.eu
2. EULYNX JR EAST Seminar 2016. International Railway Signaling Engineers (IRSE)
3. Newcomer, E., Lomow, G: Understanding SOA with Web Services, 14 décembre 2004
4. Christensen J.H. (Holobloc Inc, Cleveland Heights, OH USA), Strasser, T. (AIT Austrian Insti-
tute of Technology, Vienna AT), Valentini, A. (O3 neida Europe, Padova IT), Vyatkin, V. (Univer-
sity of Auckland, NZ), Zoitl, A. (Technical University of Vienna, AT): The IEC 61499 Function
Block Standard: Overview of the Second Edition. Presented at ISA Automation Week (2012)
5. Abourahim, I., Amghar, M., Eleuldj, M.: Interoperability of signaling interlocking and its cyber-
security requirements. In: 2020 1st International Conference on Innovative Research in Applied
Science, Engineering and Technology (IRASET)
6. Dai, W. Member IEEE, Vyatkin, V., Senior Member IEEE, Christensen, J.H., Dubinin, V.N.:
Bridging service-oriented architecture and IEC 61499 for flexibility and interoperability. IEEE
Trans. Ind. Inform. 11(3) (2015)
7. Pascale, A., Varanese, N., Maier, G., Spagnolini, U.: A wireless sensor network architecture for
railway signaling, Dip. Elettronica e Informazione, Politecnico di Milano, Italy. In: Proceedings
of the 9th Italian Networking Workshop, Courmayeur, 11–13 2012
8. Hassanabadi, H., Moaveni, B., Karimi, M., Moaveni, B.: A comprehensive distributed archi-
tecture for railway traffic control using multi-agent systems. Proc. Ins. Mech. Eng. Part F: J.
Rail Rapid Trans. 229(2), 109–124 (2015) (School of Railway Engineering, Iran University of
Science and Technology, Narmak, Tehran, Islamic Republic of Iran)
A New Design of an Ant Colony
Optimization (ACO) Algorithm
for Optimization of Ad Hoc Network
Hala Khankhour, Otman Abdoun, and Jâafar Abouchabaka
Abstract In this paper we have used a new approach of the ACO algorithm to solve
the problem of routing data between two nodes, the source to the destination in the AD
HOC network, specifically, we have improved a new variable GlobalACO to decrease
the cost between the ants (cities), and to better manage the memory management
where the ants stored the pheromones. Indeed, we used the BENCHMARK instances
to evaluate our new approach and compared them with the other article after we
applied this new approach to an AD HOC Network topology. The simulation results
of our new approach show convergence and speed with a smaller error rate.
1 Introduction
Since their inception, Mobile wireless sensor networks have enjoyed ever-increasing
success within industrial and scientific communities [1], AD Hoc wireless communi-
cation networks consist of a large number of mobile sensor nodes that can reposition
themselves, get organized in the network, and move to another node to increase the
coverage area and reach the destination, and of course, interact with the physical
environment [2]. And each node powered by a battery, then the lifespan of a wireless
sensor network depends on the lifespan of the energy resources of the sensor nodes,
the size of the network; therefore the challenge is to have reliable and fast communi-
cation with all these constraints on the Ad Hoc sensor network. In addition to these
constraints, the researchers have shown that routing in vehicular networks is an NP-
hard problem with several conflicting goals [3, 4]. Therefore, the time taken by an
exact method to find an optimal solution is exponential and sometimes inapplicable.
For this reason, this challenge can be reduced to an optimization problem to be solved
H. Khankhour (B) · J. Abouchabaka

Computer Science Department, Science Faculty, University IBN Tofail, Kenitra, Morocco
e-mail: Hala.khankhour@uit.ac.ma
O. Abdoun
Computer Science Department, Faculty Polydisciplinary, University Abdelmalek Essaadi,
Larache, Morocco
https://doi.org/10.1007/978-981-16-3637-0_16
232 H. Khankhour et al.
with approximate methods called metaheuristics in polynomial time, a metaheuristic

can be adapted to a specific problem to find high-quality solutions [5] like routing in
Ad Hoc big map network, some algorithms are inspired by nature, like Ant Colony
Optimization, Genetic Algorithm, and Simulated Annealing [6–8], and others that
are not inspired by nature, such as iterative local search and tabu search [9]. In a real
Ad Hoc network, it is possible to have many obstacles between two or more nodes.
These obstacles can attenuate the transmitted signal and disrupt communication, in
addition to the constraint of the battery for each node. It is therefore necessary to find
a brief reliable path between a source node and a destination node. To achieve this
goal, a new design is presented in this article, about the heuristic routing algorithm;
it is based on Ant Colony Optimization (ACO) for Ad Hoc Networks.
Several specific cases of ACO meta-heuristics have been proposed in Dorigo’s
literature. According to history, the strongest and most efficient systems are the ant
system (1991, 1992, 1996), ant colony system (ACS 1997) [10, 11] for more details,
see the difference between them [12].
The first ACO routing algorithm was dedicated to wired networks, they have used
the proactive routing protocol, relying mainly on ants to find the path, but it does not
adapt to changing topologies, so does not apply to ad hoc networks [13]. In 2008,
Yu Wan-Jun proposed an ACO-AHR algorithm using the reactive routing protocol
[14]. Another algorithm proposed by Abdel-Moniem, The MRAA algorithm based
on-demand Distance Vector (AODV) for Ad Hoc Network, the goal is to find the best
path in short delay [15]. The MAR-DYMO algorithm proposed by Correia, based on
two ACO procedures to be applied to take advantage of routing in Ad Hoc Network
[16]. Xueyang Wang proposed a new algorithm called ACO-EG for finding the best
path between two nodes based on evolutionary graph theory [17].
In this paper, our strategy is to create a new design algorithm based on the ACO
algorithm, we have applied this algorithm on ad hoc networks to find the best path in
a short time while avoiding several constraints such as loss of node energy, collisions,
loss of packets.
The paper is presented as follows, Sect. 2 will explain the algorithm ACO and
the related works, in Sect. 3 will describe our new design algorithm of ACO (Glob-
alACO), the efficiency of our new design algorithm on BENCHMARK instances,
and the Ad hoc network, and finally the conclusion in Sect. 5.
2 Presentation of the Algorithm ACO
Ants are small insects, weigh 1–150 mg, and measure from 0.01 to 3 cm, These
social insects form colonies that contain millions of ants (Fig. 1).
the body of the ant is divided into three major parts:
• The head is the support of the antennae (extremely developed sensory receptors)
and of the mandibles (members located at the level of the mouth which are in the
form of toothed and powerful pincers).
A New Design of an Ant Colony Optimization … 233
Fig. 1 Description of the ant
• The thorax allows communication between the head and the abdomen, supported
by three pairs of very long and very thin legs that allow ants to move in all
directions and all possible positions.
• The abdomen contains the entire digestive system and the motor of the blood
system [10].
Ant colony optimization is an iterative population-based algorithm where all indi-
viduals share a common knowledge that allows them to guide their future choices
and to indicate to other individuals directions to follow or on the contrary to avoid.
Strongly inspired by the movement of groups of ants, this method aims to build
the best solutions from the elements that have been explored by other individuals.
Each time an individual discovers a solution to the problem, good or bad, he enriches
the collective knowledge of the colony. So, each time a new individual will have to
make choices, he can rely on collective knowledge to weigh his choices [5] (Fig. 2).
To use the natural name, individuals are ants who will move around in search of
solutions and who will secrete pheromones to indicate to their fellows whether a path
is interesting or not. If a path is found to be heavily pheromized, it will mean that
many ants have judged it as part of an interesting solution and that subsequent ants
should consider it with interest.
In the literature, the first ACO algorithm to be proposed by Dorigo was the Ant
(AS) [10, 11]. After each turn between source and destination, the ants are updated
with all the pheromone values traveled. The edges of the graph are the components
of the solution, then the update of the pheromones between the cities r and s is as
follows (1) [18]:
Fig. 2 How the ant finds a path

m

τ (r, s) ← (1 − ρ)τ (r, s) + τ (r, s)k (1)
k=1
where 0 < ρ < 1 is the evaporation rate, m is the number of ants and τ (r, s)k is
the quantity of pheromone put on edge (r, s) by the k-th oven (2):
⎧
⎨ 1 {if ant k uses edges (. . .) in its tour}
τ (r, s) = L k (2)
⎩
0 otherwise
where L k is the tour length of the k-th ant.

Origo mentioned in his article [19], that ants are initially randomly distributed in
cities, so an ant colony algorithm is an iterative population-based algorithm where all
individuals share a common knowledge that allows them to guide their future choices
and indicate to other individuals directions to follow or on the contrary to avoid. Once
the city tour has been completed, an ant k deposits a quantity of pheromone on each
edge of its route; then a pheromone update is necessary [20].
In general, the ants are used not to meet on a common path, that’s why we used
the update of the local pheromones, the rule (2), to encourage the ants to visit edges
not yet visited [17].
This helps to mix the cities so that the cities visited at the start of ant visits are
visited later in another ant tour.
The pheromone level is updated by applying the local equation update rule (3)
τ (r, s) = (1 − ρ).τ (r, s) + ρ(r, s)(r, s) (3)
where τ (r, s) is the quality of pheromone on the edge (r, s) at time t.

ρ: is a parameter governing the decrease of pheromones such that 0 < ρ < 1.
3 Proposed Approach: New Design of ACO in AD HOC

(GlobalACO)
In this part, we will apply the ACO method on Ad Hoc Network to approach the
optimal solution of the problem big map network AD Hoc. An ant k placed on the
city i at instant t will choose the next city j according to the visibility n of this city and
the quantity of pheromones t deposited on the arc connecting these two cities, other
algorithmic variants drop the pheromone on the nodes of the network Ad Hoc. The
choice of the next city will be made stochastically, with a probability of choosing
the city j given by the following algorithm:
Initialization of pheromone tracks;

Initialization pheromone trails;
Place the ants in the source;
Loop as long as the stop criterion has not been reached:
– Build the solutions component by component, crossing the disjunctive
graph.
– Use of a heuristic.
– Updating pheromone tracks, the update done globally.
End of the loop.
4 Proposed Simulations and Results
To evaluate our work, as a first step we used the instances of BENCHMARK, to

evaluate our algorithm by comparing with another work of Darren M. Chitty [19]
just for the best solution, the latter proposed a new PartialACO variant, whose aim is
to reduce memory constraints, that is, the PartcialACO variant does not update partial
for the best tower. In our article, we added a new variable is GlobalACO, this last
makes the global update the best trick; after different tests and configuration runs.
The approximate solutions found are In Table 1 we compared the results obtained
with Darren M. Chitty (PartialACO) and our result (GlobalACO); knowing that the
stopping criterion is 100 iterations; In Table 1 we also note the error of the value of
the solution obtained, according to Eq. (4):
the solution obtained − the solution optimal

Er % = × 100 (4)
the solution optimal
According to Table 1 we notice that for the pcb442 instance, the error rate for the
GlobalACO variant (0.51%) is very small compared to the rate of the PartialACO
variant (1.14%), and this is also the case for the d657 instance, the error rate of
Table 1 Comparison between PartialACO and GlobalACO by using TSP instances

Problem ACO
Name Optimal Result PartialACO Er% Result GlobalACO Er%
pcb442 50,779 51357.8806 1.14 51,038 0.51
d657 48,912 50320.6656 2.88 49,394 0.98
rat783 8806 8997.9708 2.18 8904 1.12
pr1002 259,045 265806.0745 2.61 262,930 1.49
pr2392 378,032 396971.4032 5.01 388,616 2.75
GlobalACO (0.98%) is very small compared to the rate of the PartialACO variant
(2.88%), and the same is the case for large cities like for example the instance pr2392,
the error rate of the variant GlobalACO (2.79%) is small compared to the rate of the
PartialACO variant (5.01%).
The illustrated Fig. 3 shows that the GlobalACO algorithm is closer to the optimal
than the PartialACO algorithm.
In Fig. 4, we notice that there is a large distance between the PartialACOEr
et GlabalACOEr, this means that our algorithm GlobalACO gave better results
compared to the work of Darren M. Chitty PartialACO for the five instances, and
it seems that the GlobalACO algorithm converges quickly. After testing our Glob-
alACO algorithm on TSP instances, we will now apply our algorithm to the AD HOC
network.
We suggest more comparative studies between the simulation used (GlobalACO)
and other approaches using the genetic algorithm (AG), for example, in the article by
Fig. 3 MTSP comparison with PartialACO and GLOBALACO

Fig. 4 Comparison error rate between PartialACO and GLOBALACO
Esra’a Alkafaween and Ahmad B. A. Hassanat [21], they proposed a genetic algo-
rithm to produce the offspring using a new mutation operator named “IRGIBNNM”,
subsequently, they created a new SBM method using three mutation operators, to
solve the traveling salesman problem (TSP). This approach is designed on the combi-
nation of two mutation operators; random mutation and knowledge-based mutation,
the goal is to accelerate the convergence time of the genetic algorithm proposed.
Table 2 compares the results obtained by Esra’a Alkafaween and Ahmad B. A.
Hassanat and our results (GlobalACO) for the 12 instances.
From Table 2 and Fig. 5, we notice that the error rate is very small for our result of
GlobalACO algorithm compared to the result of the New SBM algorithm, especially
for the number of cities greater than 40,000, as well as the results obtained by our
program are almost the same as those of the literature, this means that the size of the
cities will have a great effect on the outcome of the problem.
From Fig. 6, the error rates obtained by our program (GlobalACO) are almost zero,
and close to the results of the literature, so our algorithm GlobalACO is powerful for
Large Scale TSP Instances.
4.1 Solve the Sensor Network Big Map Using GlobalACO
In this section we applied the GlobalACo algorithm to a sensor array, first, we consid-
ered the starting location as the source sensor and the food location as the destination
sensor, the ant antennas as the sensor antennas, and the tour as the circuit on the AD
Table 2 Comparison between New SBM and GlobalACO by using 12 TSP instances
Problem AG ACO
Name Optimal New SBM Er% Result GlobalACO Er%
eil51 426 428 0.47 426 0
a280 2579 2898 12.37 2582 0.12
bier127 118,282 121,644 2.84 118,285 0.002
kroA100 21,282 21,344 0.29 21,286 0.018
berlin52 7542 7544 0.02 7542 0
kroA200 29,368 30,344 3.32 29,370 0.007
pr152 73,682 74,777 1.49 73,686 0.005
lin318 42,029 470,06 11.84 42,033 0.016
pr226 80,369 82,579 2.75 80,370 0.0012
ch150 6528 6737 3.2 6528 0
st70 675 677 0.29 675 0
rat195 2323 2404 3.49 2325 0.08
Fig. 5 MTSP comparison with NewSBM and GlobalACO
HOC network, from the source node to the destination node. After several searches,
unfortunately almost there is no AD HOC topology to work with, I found a Chang
Wook Ahn topology on this article [22].
As shown in Fig. 7 we generated a network topology with 20 nodes and we
displayed the results found in Table 2, after several runs we found the total path cost
equals 142 just in 0.015 s.
Fig. 6 Comparison error rate between NewSBM and GlobalACO
Fig. 7 The topology used in AD HOC by using ACO

5 Conclusion
This article presents the optimization of the AD HOC network by using the ACO
metaheuristic, the execution of the results show that the use of the GlobalACO
variant gave better results, which means, that the data flow from the source node to
the destination node will be done in a faster way while keeping up the energy of each
node before the termination of the AD HOC network communication.
References
1. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: Wireless sensor networks: a survey,
Comp. Net. 38(4), 393–422 (2002)
2. Sagar, S., Javaid, N., Khan, Z. A., Saqib. J., Bibi, A., Bouk, S. H.: Analysis and modeling
experiment performance parameters of routing protocols in manets and vanets, IEEE 1lth
International Conference, 1867–1871 (2012)
3. Cai Zheng, M., Zhang, D.F., Luo, l.: Minimum hop routing wireless sensor networks based
on ensuring of data link reliability. IEEE 5th International Conference on Mobile Ad-hoc and
Sensor Networks, pp. 212–217 (2009)
4. Eiza, M.H., Owens, T., Ni, Q., Shi, Q.: Situation-aware QoS routing algorithm for vehicular
Ad Hoc networks. IEEE Trans. Veh. Technol. 64(12) (2015)
5. Hajlaoui, R., Guyennet, H., Moulahi, T.: A Survey on Heuristic-Based Routing Methods in
Vehicular Ad-Hoc Network: Technical Challenges and Future Trends. IEEE Sens.S J., 16(17),
September (2016)
6. Alander, J.T.: An indexed bibliography of genetic algorithms in economics, Technical Report
Report (2001)
7. Okdem, S., Karaboga, D.: Routing in Wireless Sensor Networks Using an Ant Colony Op-
timization (ACO) Router Chip. 9(2), 909–921 (2009)
8. Kumar, S., Mehfuz, S.: Intelligent probabilistic broadcasting in mobile ad hoc network: a PSO
approach”. J. Reliab. Intell. Environ. 2, 107–115 (2016)
9. Prajapati, V. K., Jain, M., Chouhan, L.: Tabu Search Algorithm (TSA): A Comprehensive
Survey “, Conference 3rd International Conference on Emerging Technologies in Computer
Engineering Machine Learning and Internet of Things (ICETCE) (2020)
10. Voss, S.: Book Review: Morco Dorigo and Thomas Stützle: Ant colony optimization (2004)
ISBN 0-262-04219-3, MIT Press. Cambridge. Math. Meth. Oper. Res. 63, 191–192 (2006)
11. Stalling, W.: High-Speed networks: TCP/IP and ATM design principles. Prentice-Hall,
Englewood Cliffs, NJ (1998)
12. Sharkey. P.: Ant Colony Optimisation: Algorithms and Applications March 6 (2014)
13. Xiang-quan, Z., Wei, G., Li-jia, G., Ren-ting, L.: A Cross-Layer Design and Ant-Colony
Optimization Based Load-Balancing Routing Protocol for Ad Hoc Network (CALRA). Chin.
J. Electron.7(7), 1199–1208 (2006)
14. Yu, W.J., Zuo, G.M., Li, Q.Q.: Ant colony optimization for routing in mobile ad hoc networks.
7th International Conference on Machine Learning and Cybernetics, pp. 1147–1151 (2008)
15. Abdel-Moniem, A. M., Mohamed, M. H., Hedar, A.R.: An ant colony optimization algorithm
for the mobile ad hoc network routing problem based on AODV protocol. In Proceedings of
10th International Conference on Intelligent Systems Design and Applications, pp. 1332–1337
(2010]
16. Correia, S.L.O.B., Celestino, J., Cherkaoui, O.: Mobility-aware ant colony optimization routing
for vehicular ad hoc networks. IEEE Wireless Communications and Networking Conference,
pp. 1125–1130 (2011)
17. Wang, X., Liu, C., Wang, Y., Huang, C.: Application of Ant Colony Optimized Routing Algo-
rithm Based on Evolving Graph Model In VANETs, 17th International Symposium on Wireless
Personal Multimedia Communications (WPMC2014)
18. Chitty, M.D: Applying ACO to large scale TSP instances. Adv. Comput. Intell. Syst. 350,
104–118 (2017)
19. Rana, H., Thulasiraman, P., Thulasiram, R.K.: MAZACORNET: Mobility Aware Zone based
Ant Colony Optimization Routing for VANET, IEEE Congress on Evolutionary Computation
June 20–23, pp. 2948-2955, Cancún, México (2013)
20. Tuani A.F., Keedwell E., Collett M.: H-ACO A Heterogeneous Ant Colony Optimisation
Approach with Application to the Travelling Salesman Problem. In: Lutton E., Legrand P.,
Parrend P., Monmarché N., Schoenauer M. (eds.) Artificial Evolution. EA 2017. Lecture Notes
in Computer Science, vol 10764. Springer (2018)
21. Alkafaween. E., Hassanat. A.: Improving TSP solutions using GA with a new hybrid mutation
based on knowledge and randomness, Computer Science, Neural and Evolutionary Computing
(2018)
22. Ahn, C.W., Ramakrishna, R. S.: A Genetic Algorithm for Shortest Path Routing Problem and
the Sizing of Populations, IEEE Trans. Evol. Comput. 6(6) (2002)
Real-Time Distributed Pipeline
Architecture for Pedestrians’
Trajectories
Kaoutar Bella and Azedine Boulmakoul
Abstract Cities are suffering from traffic accidents. Every one results in significant
material or human injuries. According to WHO (World Health Organization), 1.35
million people perish each year as a consequence of road accidents and more end up
with serious injuries. One of the most recurrent factors is distracted driving. 16% of
pedestrian injuries were triggered by distraction due to phone use, and the amount
of pedestrian accidents caused by mobile distraction continues to increase, some
writers call Smombie a smartphone zombie. Developing a system to eliminate these
incidents, particularly those caused by Smombie, has become a priority for the growth
of smart cities. A system that can turn smartphones from being a cause of death to
a key player for pedestrians’ safety. Therefore, the aim of this paper is to develop
a real-time distributed pipeline architecture to capture pedestrians’ trajectories. We
collect pedestrians’ positions in real-time using a GPS tracker mounted in their smart
phones. The collected data will be displayed to monitor trajectories and stored for
analytical use. To achieve real-time distribution, we are using delta architecture.
To enforce this pipeline architecture, we are using open-source technologies such
Traccar as GPS tracking Server and Apache Kafka to consume the collected data such
as messages, Neo4j to store the increasing data collected for analytical purposes, as
we use Spring boot for API development, and finally.
1 Introduction
Road incidents can only lead to tragedies. Whether due to speeding, poor road struc-
ture, or due to human error, it must be handled. The implementation of a safety
This work was partially funded by Ministry of Equipment, Transport, Logistics and Water-Kingdom
of Morocco, The National Road Safety Agency (NARSA) and National Center for Scientific and
Technical Research (CNRST). Road Safety Research Program. An intelligent reactive abductive
system and intuitionist fuzzy logical reasoning for dangerousness of driver-pedestrians interactions
analysis.
K. Bella · A. Boulmakoul (B)

LIM/IOS, FSTM, Hassan II University of Casablanca, Casablanca, Morocco
https://doi.org/10.1007/978-981-16-3637-0_17
244 K. Bella and A. Boulmakoul
Fig. 1 Estimated
intersection collision
system for pedestrians has become a must in presence of the mortality rates that
are climbing every year due to injuries [1–3]. As mentioned before, smartphone can
be a negative player in this scenario, so in a time driven by technology when it is
used for social use case improvement, it is obvious to take advantage of this negative
player for our benefit. In this paper, we are using smartphone as GPS trackers to
collection pedestrians’ locations in order to assemble their trajectory. By knowing
each pedestrians and drivers’ location we can estimate their next future position and
alert if a collision is about to happen (Fig. 1).
Collecting positions for multiple users in real-time, results in big amount of data
[4, 5]. This data must be processed for real-time monitoring and stored for analytical
uses. However, collecting and handling these massive data presents challenges in
how to perform optimized online data analysis. Since speed is very critical in this use
case. In order to implement such a complex architecture and to achieve the segments
of the reactive manifesto (responsive, resilient, elastic, message-driven) we need a
robust architecture with highly scalable frameworks. To collect the locations, we are
using Traccar as a GPS tracking server. The collected positions are consumed by
a messaging queue. Based on previous work, traditional messaging queues such as
ActiveMQ can manage small amounts of information and retain the distribution state
of each one, resulting in a lower throughput and no horizontal scale because of the
lake of replication concept.
Therefore, we used Apache Kafka. Kafka is a stream-processing platform built
by LinkedIn and currently developed by the Apache Software Foundation. Kafka
aims to provide low-latency ingestion of large amounts of events. It’s highly scalable
due to partition replications, providing higher availability too. Now we collect the
information, we need to store them for future use. If the right database meets the
usage requirement, it can ease and speed up the exploitation of these data which is
a key player for the system responsiveness. For our use case where we value data
connection for semantics, we used a graph database Neo4j. A graph database is
significantly simpler and more expressive than relational one and we won’t have to
worry about out-of-band processing, such as MapReduce.
Real-Time Distributed Pipeline Architecture … 245
Recent developments in information and communication technologies are seen

as an essential design vector for smart cities [4, 6–12]. One of these areas is that of
mobility in a city. Currently, various cities are considering innovative ways to reduce
emissions by increasing active mobility. Considerations related to pedestrian safety
are a major challenge for cities. Technologies such as spatial databases, geographic
information systems, internet of things, mobile computing, spatial data mining, and.
define the fundamental basis for research in the field of urban computing. This work
is part of this area and plans to put these technologies at the service of road safety.
The remainder of this article is organized as follows; the next section describes
the architecture to connect different components of this pipeline. Section 3 details
the implementation of the proposed architecture. A case study, describing how the
software solution is tested, is presented in Sect. 4. The article concludes with a
discussion and proposes future work in Sect. 5.
2 Architecture
Defining the architecture of the pipeline is a challenging action. The architecture

describes how the various elements interact with each other. The goal of this paper is
to provide an efficient data flow in order to incorporate real-time data processing. We
initially implemented lambda architecture, but due to higher costs when running our
jobs, we switched to delta architecture, which is an upgrade of Lambda architecture.
It unifies the batch and streaming layers to avoid the gap between the two layers and
we won’t have to treat data differently [4, 13, 14] (Fig. 2).
In this paper the batch layer is not 100% used, only for formatting and conversion
purposes, even though we chose it for future objectives (Fig. 3).
This architecture allows us first to have an exact view of data in real time and
a view on analyzed and batched data (Table 1). Traccar Server is the source of our
real-time data (locations). Apache Kafka is used for data processing in real time as a
message broker. Spring boot for Restful APIs development. And to display real time
and batched data for final user, we are using ReactJs as a frontend framework.
Fig. 2 Delta architecture layers

Fig. 3 Pipeline architecture components and data flow
Table 1 Data flow details

Ref Description
(1) Http request to Session API to establish connection (session token)
(2) Reply with cookies
(3) Position {longitude, latitude}
(4) Position {longitude, latitude}: Kafka producer
(5) Sends data to NEO4J after conversion (batch Layer)
2.1 Traccar GPS Tracker
Traccar Server Traccar is an open-source GPS tracking system [15]. It is built

of Jetty Java HHTP, Natty network pipeline framework, and MySQL database. For
every connection, it creates a pipeline of event handlers. The received messages from
GPS devices are formatted and stored in the database (SQL database). In this paper,
pedestrians and drivers’ locations are recorded to Traccar server continuously from
Traccar client application installed in their smartphones. Our web service can access
Traccar server, to retrieve collected data from Traccar client; longitude, latitude,
and speed. Although our web server must be assigned an access token, in order to
communicate with the server by Http API (Figs. 4 and 5).
Based on research, there are various sets of Traccar server API. In our use case, we
only need session and position APIs. Our web server sends access token parameters
in Session API request, in order to initiate the connection. Traccar server sends in
response cookie string to establish trusted connection. The cookie is essential use
position and devices APIs. Position API is used to read users locations (longitude
and latitude) and speed. We can get users locations in real time with a time difference
Fig. 4 Communication
between Traccar client and
our spring boot application
Fig. 5 Android GPS

tracking application
configuration
of three seconds or less. Our webserver establishes the connection using the access
token, and whenever a new client is connected it collects their locations and speed
at all time. Using these locations, we can set pedestrians trajectories using Traccar
client (Table 2).
Pedestrians In order to keep track of pedestrian’s locations, a GPS tracker is needed.

In this paper, we are using and Android GPS tracker application developed by Traccar.
In order to establish the connection, we must first set some configurations.
Table 2 Parameters pf Traccar HTTP API

HTTP Request/Response
API
Session Request:
http://{ServerIP/Domain}/api/session?token=Cy4J8amzLZ1DpzYAw76TpEDfcRWPi5yU
Response:
HTTP/1.1 200 OK
Date: Wed, 18 Nov 2020 06:55:10 GMT
Server: Jetty(9.4.20.v20190813)
Expires: Thu, 19 Nov 2020 00:00:00 GMT
Content-Type: application/json
Content-Length: 532
Set-Cookie: JSESSIONID = node09ys9frpwk3i51458s79g37dgc5.node0; Path =/
Keep-Alive: timeout = 5, max = 100
Connection: Keep-Alive
Position Request:
http://{ServerIP/Domain}/api/positions
Reply:
accuracy: 1500
device Id: 1
fix Time: “2020-11-18T06:21:27.000+0000”
id: 19
latitude: 33.584807
longitude: −7.584743
speed: 0
2.2 Kafka
Apache Kafka is an open-source distributed streaming platform, licensed under the

Apache license [16]. Kafka is written in java and Scala; it implements a publish-
subscribe messaging system that is designed to be fast and scalable. Message brokers
are middleware that allow applications to exchange messages (events) in cross plat-
forms. In order to provide reliable message storage and guaranteed delivery, message
brokers rely on a message queue that stores and orders the messages until the
consumer can process them. This strategy prevents the loss of valuable data and
enables systems to continue functioning even in the face of the intermittent connec-
tivity or latency issues. Messages are sent by producers, a given message concerns a
subject, in Kafka this is called a Topic. The consumer subscribes to one or more topics.
He will therefore receive all the messages concerning these subscribed topics. This
architecture is an abstraction that provides us developers with a standard of handling
the flow of data between an application’s services so that we can focus on the core
logic.
Configuration Kafka cluster can be composed of multiple brokers. Each broker

is identified with an ID and can contain certain topic partitions. In this paper, we
are using a single node with 2 brokers as Kafka configuration (replication-factor
= 2), so that when a broker is down, another one can serve the data of the topic.
This architecture is able to handle more producers. However, this is still a basic
Fig. 6 Kafka cluster consisting of one node and two brokers
configuration and, since Kafka is distributed in nature, a cluster typically consists of

multiple nodes with several brokers. The results shown in this paper are based on
a two-brokers architecture in order to reveal the first step towards distributing the
system. Zookeeper is responsible for managing the load over the nodes (Fig. 6).
2.3 Spring Boot
Spring boot is a Java-based framework for building web and enterprise applications.
This framework provides a flexible way to configure Java beans and database trans-
actions. It provides also powerful for managing Rest APIs as well as it contains an
embedded Servlet Container. We chose It, to abstract the Api service configuration
to focus on writing our logic instead of spending time configuring the project and
server. It provides a template as a high-level abstraction for sending messages, as
well as support for Message-driven POJOs with @KafkaListener annotations and a
listener container. With this service, we can create Topics and different instances of
producer and consumers very easily.
2.4 Neo4j Instantiation
When dealing with a huge amount of data, storing and retrieving these data become
a real challenge. In this paper, not only, we are dealing with a lot of data but also
our data is highly interconnected. According to previous researches, Cypher is a
promising candidate for a standard graph query language. This supports our choice
of using a graph database. A graph database saves the data in an object format
represented as a node and binds it together with edges (association). It uses Cypher
as a query language that allows us to store and retrieve data from the graph database.
The syntax of Cypher offers a visual and logical way of matching node patterns and
relationships in the graph. Also, we can use the sink connector with Kafka to move
data from Kafka topics to Neo4j using Cypher templates.
2.5 React Js
We are using ReactJs for handling view layer for mobile and web application. ReactJs
is a JavaScript frontend framework; it is component-based Single Page Applications
(SPA) framework. It suits well for data presentation in real time since we can update
data without refreshing the page. Its main purpose is being fast and simple. It uses
the conception of Virtual Dom for performance optimization. Each DOM object
has a corresponding virtual DOM object. When an object is updated, all the DOM
changes, which sound incredibly inefficient, but the cost is negligible because the
virtual DOM can update so fast. React compares the virtual DOM with its snapshot,
which was taken just before the update. By comparing the latest virtual DOM with
the pre-update version, ReactJs only updates the one changed. Each semantic data is
represented in a higher component. Which will help us structure the project; keep it
maintainable and easy to read. In this case, we are using two HOCs: Historical data
and real-time data.
3 Architecture Components Implementation
3.1 Processing Real-Time Data Operation
The architecture of this paper consists of several components which are GPS Tracking
device system, Message broker, database, and a view platform. Traccar client sends
pedestrians locations with a 3 s delay to Traccar server. Kafka producer collects
from Traccar server these locations using restful APIs in our Spring boot service and
publishes it to the convenient topic. Kafka consumer listens to this topic and each
record is sent to Neo4j as nodes. The core concept is to get data from Kafka and
dispatch it to Neo4j and controllers. The diagram below explains the flow of Data
(Fig. 7).
Fig. 7 Traccar sequence diagram for the real-time locations monitoring and historical data
Fig. 8 Graph of pedestrian’s

locations
3.2 Neo4J Instance
In graph databases, data are presented under graph format. The figure below repre-
sents the data saved with hibernate instance, each message is a node under position
format {longitude, latitude, speed, userId} related by “move to” connection (Fig. 8).
3.3 Testing
We used kafka-*-perf-test library to measure read and write throughput and to stress
test the cluster. First testing our producer with 1000 messages. We chose not to
Table 3 Producer test results

Parameter Result
start.time 2020-11-27 21:38:28:413
end.time 2020-11-27 21:38:29:002
compression 0
message.size 100
batch.size 400
total.data.sent.in.MB 50
MB.sec 0.269
total.data.sent.in.nMsg 1000
nMsg.sec 239.61
Table 4 Consumer test

Parameter Result
results
start.time 2020-11-27 22:40:27:403
end.time 2020-11-27 22:40:28:002
fetch.size 1951429
data.consumed.in.MB 2.2653
MB.sec 3.8700
data.consumed.in.nMs 1001
nMsg.sec 40931.2567
specify message size during tests since our messages are not very large. We’ve set
initial-message-id to generate test data. The results of our producer are as follow
(Table 3).
The result of our consumer tests given in the following (Table 4).
According the results, we can conclude that configuration is resilient enough
for the number of data set. Traccar client provides us with pedestrians’ positions
and their speed each three seconds. Although, while testing we realized that some
positions are not 100% accurate. And after testing two different trajectories each
with approximately 50 positions, 87% of the positions are received: bases on each
three seconds we receive a location.
4 Results
In this paper, we have implemented a real-time pipeline for pedestrians’ trajectories,

using open-source technologies. From our view layer implemented with ReactJs,
we retrieve data using REST API from our Spring boot application. Pedestrian’s
trajectories are displayed on a Google Map using our ReactJs application, as shown
in the figure (Fig. 9).
Fig. 9 Pedestrian trajectory
We can see pedestrians’ trajectory, but at a certain point where positions are not
recorded properly either because the position is not sent properly to Traccar server
or delay time. We record accident locations, intersections where accidents happen
more often; with this simple peace of information, we can categorize intersections
as red zone, orange zone, or green zone. According to the zone type we are going to
rise up the collision percentage as shown in Fig. 10. Accidents are yellow dots, and
according to the number of accidents we categorize the intersection as dangerous or
normal.
Fig. 10 Intersection
categories
5 Conclusion and Future Work
Nowadays, traffic accidents are one of the most serious problems of the transportation
system. Pedestrians’ safety is treated as a priority to solve in order to upgrade to a
smart and safe city ecosystem. In this paper, we present real-time distributed pipeline
architecture for pedestrians’ trajectories. Our primer concern is pedestrians because
it’s the weakest component on roads accidents. The main challenge was to define
an optimized architecture to provide real-time processed data. Using Traccar GPS
tracking server, we are collecting pedestrians’ and drivers’ positions from Android
application installed in their smart phones. Using this data, we can set trajectories
and visualize them. For future work we aim, to estimate intersection collisions in
order to alert the pedestrian. For more accuracy, we want to record more information
besides positions and speed.
References
1. Hussian, R., Sharma, S., Sharma, V.: WSN applications: automated intelligent traffic control
system using sensors. Int. J. Soft Comput. Eng. 3, 77–81 (2013)
2. MarketWatch: Inattention is leading cause of deadly pedestrian accidents in el
paso. https://kfoxtv.com/news/local/inattention-leading-cause-ofdeadly-pedestrian-accidents-
in-el-paso-say-police, (2019)
3. Tribune, C.: Look up from your phone: Pedestrian deaths have spiked (2019) https://www.chi
cagotribune.com/news/opinion/editorials/ct-editpedestrian-deaths-rise-20190301-story.html
4. Maguerra, S., Boulmakoul, A., Karim, L., et al.: Towards a reactive system for managing big
trajectory data. J. Ambient Intell. Human Comput. 11, 3895–3906 (2020). https://doi.org/10.
1007/s12652-019-01625-3
5. Bull, A., Thomson, I., Pardo, V., Thomas, A., Labarthe, G., Mery, D., Diez, J.P., Cifuentes, L.:
Traffic congestion the problem AND how to deal with it. United Nations Publication, Santiago
(2004)
6. Atluri, G., Karpatne, A., Kumar, V.: Spatio-temporal data mining: A survey of problems and
methods. ACM Comp. Surveys 51(4), Article No. 83 (2018)
7. Boulmakoul, A., Bouziri, A.E.: Mobile object framework and fuzzy graph modelling to boost
HazMat telegeomonitoring. In: Garbolino, E., Tkiouat, M., Yankevich, N., Lachtar, D. (eds.)
Transport of dangerous goods. NATO Science for Peace and Security Series C: Environmental
Security. Springer, Dordrecht (2012)
8. Das, M., Ghosh, S.K.: Data-Driven Approaches for Spatio-Temporal Analysis: A Survey of
the State-of-the-Arts. J. Comput. Sci. Technol. 35, 665–696 (2020). https://doi.org/10.1007/
s11390-020-9349-0
9. D’silva, G.M., Khan, A., Gaurav, J., Bari, S.: Real-time processing of IoT events with historic
data using Apache Kafka and Apache Spark with dashing framework, 2017 2nd IEEE Interna-
tional Conference on Recent Trends in Electronics, Information & Communication Technology
(RTEICT), Bangalore, pp. 1804–1809 (2017), https://doi.org/10.1109/rteict.2017.8256910
10. Duan, P., Mao, G., Liang, W., Zhang, D.: A unified spatiotemporal model for short-term traffic
flow prediction. IEEE Transactions on Intelligent Transportation Systems (2018)
11. Goodchild, M.F.: Citizens as sensors: the world of volunteered geography. GeoJournal, 211–
221 (2007)
12. Chen, L., Roy, A.: Event detection from Flickr data through wavelet-based spatial analysis. In
CIKM’09, 523–532 (2009)
13. Marz, N., Warren, J.: Big data: principles and best practices of scalable real time data system,
ISBN 9781617290343, Manning Publications (2015)
14. Psomakelis, E., Tserpes, K., Zissis, D., Anagnostopoulos, D., Varvarigou, T.: Context agnostic
trajectory prediction based on λ-architecture, Future Generation Computer Systems 110, 531–
539 (2020). ISSN 0167-739X, https://doi.org/10.1016/j.future.2019.09.046
15. Watanabe, K., Ochi, M., Okabe, M., Onai, R.: Jasmine: a real-time local-event detection system
based on geolocation information propagated to microblogs. In CIKM ’11, 2541–2544 (2011)
16. Abdelhaq, H., Sengstock, C., Gertz, M.: Eventweet: Online localized event detection from
twitter. VLDB 6, 12 (2013)
Reconfiguration of the Radial
Distribution for Multiple DGs by Using
an Improved PSO
Meriem M’dioud, Rachid Bannari, and Ismail Elkafazi
Abstract PSO is one of the famous algorithms that help to find the global solution;
in this study, our main objective is to improve the result found by the PSO algorithm
to find the optimal reconfiguration by adjusting the inertia weight parameter. In this
paper, I select the chaotic inertia weight parameter and the hybrid strategy using the
combination between the chaotic inertia weight and the success rate, these kinds of
parameters are chosen due to their accuracy, and they give the best solution compared
with other types of parameter. To test the performance of this study, I used the IEEE
33 bus in the case of the presence of the DGs, and a comparative study is done to
check the reliability and the quality of these two suggested strategies. In the end, it
is noticed that the reconfiguration by using the chaotic inertia weight gives a better
result than the hybrid strategy and the other studies: reduce losses, improve the
voltage profile at each node, and give the solution at a significant time.
1 Introduction
Distributed generation (DG) is an old idea that appears for a long time. But it is still
a good technology that provides the electricity at or near where it will be consumed
[1]. When it is integrated into the electrical network, the distributed generation can
allow reducing the losses at the level of the transmission and distribution network [2].
It is an emerging strategy used to help the electrical production plants to follow client
consumption and used to minimize the quantity of electricity that should be produced
at power generation plants, besides, help to reduce the environmental impacts that
leftover by the electrical production plants [3]. The authors of [4] have used a method
to determine the forecasts of the electrical consumption, and based on the electrical
quantity requested, in the same direction, the authors of [5] have focused on the
M. M’dioud (B) · R. Bannari

Laboratory Sciences Engineering ENSA, Ibn Tofail University, Kenitra, Morocco
I. Elkafazi
Laboratory SMARTILAB, Moroccan School of Engineering Sciences, EMSI, Rabat, Morocco
https://doi.org/10.1007/978-981-16-3637-0_18
258 M. M’dioud et al.
treatment of the injection of intermittent energy sources. Thus, they have chosen to
use some techniques of energy management and check the response to demand, and
they have concluded their paper by studying the importance of the smart grids to
check a balance between the supply and the demand inter electricity producers and
consumers.
However, the authors of [6] have discussed the injection of the intermittent energy
generators will give rise to impacts on the reliability of the electrical system (maybe
voltage exceed the voltage plane limits), and they have proposed some solution that
helps us to insert the intermittent power in electrical systems with a high injection
rate. And the authors of [7] have introduced a supervisor algorithm (the predictive
control model) to control the energy produced aiming to minimize the total cost of
production.
During these last years, the electrical enterprises have been oriented toward new
techniques, to enhance, and minimize energy exploitation by searching for suitable
reconfigurations, so that these new strategies encourage the reduction of losses. To
find the solutions to this type of problem, various studies have been done, such as
the authors of [8] have based on the modified shark smell optimization to search for
the new reconfiguration of the radial distribution with and without DG aiming to
minimize the total power system losses. On the other hand, the authors of [9] have
suggested using a simple algorithm for distribution system load with DG, where
they have considered the DGs as negative loads. In the article [10], the author gives
a comparison between two types of DG (PQ bus, it is mean that the integrated power
is considered as a load with a negative sign, the second type is the PV bus is mean
that the reactive power injected in the network counts on the expected tension on the
bus, but the active power injected is considered constant).
In this vision of the issue, the authors of [11] have used the genetic algorithm to
reduce the losses and optimize the voltage at each node after they used the forward–
backward method to apply the load flow analysis, aiming to predict the optimum plan
for a power distribution network with the presence of multiple DGs. The authors of
[12] have discussed in their study the influence resulting from the presence of DG
in the network on the losses and the voltage profile of the electrical network. In
this vision, [13] have studied the performance analysis of the electrical distribution
system with and without DGs. And the authors of [14] have used the prim particle
swarm optimization algorithm to search the new reconfiguration and reduce losses
of the electrical network with the presence of the DGs.
In this paper, I will use an enhanced and adaptive PSO algorithm based on the
adjusting strategy of the inertia weight parameter, this algorithm is chosen thanks
to its high speed to meet an optimal and best solution, thereby it easy to implement
and it bases on simple equations. And the main goal of this research is to find a
significant reduction concerning the value of the losses and a significant amelioration
in the voltage profile. For testing the performance and the quality of this proposed
method, this paper focused on the IEEE 33 bus with the presence of the DGs, and a
comparative study is done to compare the results with other recent studies described
in the following paragraph.
Reconfiguration of the Radial Distribution for Multiple DGs … 259
2 Related Works
On this side, the authors of [15] have tried to solve this problem by using the particle
swarm optimization algorithm focused on the decreasing inertia weight parameter
“w” to update the new position and velocity of the particle then have applied the
backward forward sweep to define the power flow analysis. The authors of [16] have
chosen also the linear decreasing inertia weight by using the sigmoid transformation
to restrict the value of the velocities. On another study, the authors of [17] have used
the linear decreasing weight by eliminating the wend in the second term. Regarding
the authors of [18] have done the comparative analysis to determine the best equation
for inertia weight that helps to enhance the quality of the PSO algorithm and to check
the reliability and the performance of their paper they focused on five mathematic
equations, and they have concluded that the chaotic inertia weight is the best tech-
nique to define a result with higher accuracy; however, the random inertia weight
technique is best for better efficiency. On the other hand, the authors of [19] have
thought to combine the swarm success rate parameter with the chaotic mapping to
define the new inertia weight parameter and to validate the performance and the
quality of their proposed tool, they examined this new parameter by solving the
five functions (Griewank, Rastrigin, Rosenbrock, Schaffer f6, and Sphere), and they
concluded that the swarm success rate is a useful tool to improve the performance
of any swarm focused on the optimization algorithms. The authors of [8] have used
the modified shark smell optimization to reduce the losses and improve the voltage
profile this algorithm have the same idea as the particle swarm optimization, and
they have concluded that this method helps to find the solution in a significant time
and enhance the profile voltage at each node. Therefore, in this study, the main goal
is to adjust the inertia weight parameter based on these two best strategies already
described in the previous paragraph (the chaotic inertia weight and the swarm success
rate combined with the chaotic mapping), aiming to find the optimal reconfiguration
of the radial distribution network in the case of the presence of multiple DGs with
losses minimized and voltage profile improved. To show the performance of these
techniques, it is important to test these suggested techniques on IEEE 33 bus with
DGs to compare the solution of this paper with the other studies focused on the
particle swarm optimization algorithm.
To study this issue, this paper is divided into five parts. Section 3 introduces the
main objective of this study, gives the objective function, and defines the constraints
and describes the main steps of the chosen method, and presents the case studies
with the presence of DGs. Section 4 discusses and gives an analysis study about the
found results and makes the difference between this studies with the other recent
works. And in the fifth and the final section, I conclude the research and I present an
idea about the future research.
3 Suggested Algorithm
3.1 Problematic
The main cause that encourages several companies to search for some strategies
to reduce losses comes from the peak demand it is means when all resources are
operating at maximum, this last one gives rise to unnecessary expenses for the electric
companies.
When the load increases, these losses increase. This paper has studied the problem
using the famous reconfiguration strategy of the electrical network, by the implemen-
tation of data using MATLAB software to find the optimal solution using the PSO
algorithm to find a new structure of the network with minimum total losses.
Objective Function. As described above, in this paper it is important to reduce
the losses, I use the following expression to calculate this last one:

Losses = 1 ∈ S R1 ∗ I12 (1)
With S is the set of the system edges and I l is the current of line l, Rl is the current
of line l.
This optimization problem is solved under the following constraints [20].
Constraints.
Kirchhoff’s law:
I ∗ A = 0 (2)
where I: row vector of current of each line of graph and X: incidence matrix of
graph.
(Aij = 0 if there are no arcs between i and j; Aij = 1 else);
Tolerance limit:
|(V jn − V j)| V jn |≤ εj max (3)
Vjn nominal voltage at node j, Vj is voltage in node j and ε j max is tolerance limit
at node j [4] (±5% for HTA and + 6%/−10%BT).
Admissible current constraint:
Il ≤ Il,maxadm (4)
Il current of line l and Il,maxadm .: current maximum admissible of line l.

Radial topology constraint. To have a simple and to keep the security of the
electrical network, it is better to choose the radial configuration. It means that on
each loop exists an open line. To have this topology, the system should follow these
constraints.
Total number of main loops:

Nmain loops = Nedge − Nnode + 1 (5)
where N edge is the total edges of the network, N node is the total number of nodes,
and N main loops are the total number of loops in the system.
The total number of sectionalizing switches
Nedge = Nnode − 1 (6)
The total number of tie lines should be equal to the number of main loops in the
electrical system.
To study this issue, I break up this problem into two important parts, the first one
regarding the radial electrical system with the presence of the DGs. The second one
is focused on the Newton–Raphson methods to do the power flow module. This load
flow method is chosen to its advantage to a fast convergence rate [21].
3.2 Network with DGs
For this study of reconfiguration with the presence of DGs, I take the case of IEEE
33 bus with tie line is 33–37 as shown in Fig. 1.
Table 6 in appendices gives the line and load data of this network [22]. In this
paper, we assume that the DGs data as shown in Table 1:
In this vision, according to the insertion of DG, the power integrated into a node
linked to a DG will be modified. And, to update the new value of active and reactive
power, I based on the following formulas [24].
P = Pload − PDG (7)
Q = Q load − PDG (8)
PDG = a ∗ Q DG (9)
To comprehend how the electrical system with DG works, Fig. 2 makes things
easier.
So, the losses in this case become:

Ploss = R ∗ (Pload − PDG )2 + (Q load − (±Q DG ))2 V 2 (10)
where
R is the line resistance.
Ploss is the line losses.
Fig. 1 Network of IEEE 33 bus with DGs
Table 1 DGs data [23]

Location(bus) Size (MW) Power factor
28 0.1 0.95
17 0.2 0.95
2 0.14 0.98
32 0.25 0.85
Fig. 2 DG injected as a negative load to the bus

Pload active power consumption of load.

Qload reactive power consumption of load.
(PDG , QDG ) is the active and reactive power output of distributed generation.
a is the power factor of DG.
3.3 PSO Algorithm
PSO is a metaheuristic using to search for the optimal solution, invented by [25]. This
optimization method is based on the collaboration of individuals among themselves.
In this study, due to the feature of the simple travel rules in the solution space, the
particles can gradually find the global minimum. This algorithm follows these steps:
Step 1: In the first step, initialize the number of particles and the number of tie
lines by respecting the condition of the system is in radial nature (Table 4).
Step 2: Initialize iteration number (maxiter), inertia coefficient (w1 and w2 ), and
acceleration coefficients (C 1 and C 2 ), the initial velocity of each particle is randomly
generated (Table 5 in appendices).
Step 3: Identify the search space for each D dimension (all possible reconfigura-
tion).
Step 4: Apply the Newton–Raphson method [21] to load flow analysis.
Step 5: Define the best value among all pbest values.
Step 6: Find the global best and identify the new tie switches.
Step 7: Update the velocity and new position for each D dimension of the ith
particle using the following equation:
Select a random number z in the interval [0, 1] and use a chaotic mapping by
using the logistic mapping to set inertia weight coefficient [18]:
z(iter + 1) = 4 ∗ z(iter) ∗ (1 − z(iter)) (11)
or use the success rate [19]:

1 f Pbestit < f Pbestit−1
Successit = (12)
0 f Pbestit ≥ f Pbestit−1

n

Succrate = Successit n (13)
i=1
z(iter + 1) = 4 ∗ Succrate ∗ (1 − Succrate ) (14)
According to the choice of the inertia weight calculation strategy, adjust and
calculate the inertia coefficient by using [25].
W = (w1 − w2 ) ∗ (maxiter − iter)/maxiter + w2 ∗ z(iter + 1) (15)

Now, I use this value of the inertia weight to update the new velocity and of the
new position.
Update velocity by using [25]:
Vi (iter + 1) = W ∗ Vi (iter) + C1 ∗ r1 ∗ (Pi (iter) − X i (iter))

+ C2 ∗ r2 ∗ (G(t) − X i (iter)) (16)
Update position by this equation [25]:
X i (iter + 1) = X i (iter) + Vi (iter + 1) (17)
Define the new fitness values for the new position [25]

Pbestti f xit+1 > Pbestit
Pbestt+1 = (18)
t
xit+1 f xit+1 ≤ Pbestit
Define the global best by using [25]

Gbest = min Pbestit+1 (19)
Step 8: Until iter = maxiter, go to step 4. Else print the optimal results.
Step 9_: Display results.
4 Test and Results
4.1 Chaotic Inertia Weight Results
Using the chaotic inertia weight, the set of the tie line becomes 2–13–26–29–33
instead of 33–34-35–36––37 in the base case. Figure 3 shows the new reconfiguration.
In Table 2, a comparative study is done; in this paper, I compare the result of
the PSO algorithm by using the chaotic inertia parameter with the study of [24] and
the [26] study where the authors solve this same problem by using the minimum
spanning tree method.
In Fig. 4, it is clear that the profile voltage is improved compared with the initial
case, and the minimum voltage equals to 0.9359 p.u at node 30 instead of 0.8950 p.u
at node 18 in the base case. Table 7 in the appendices gives the value of the voltage
at each node.
As already described in the previous table, the reconfiguration using the chaotic
inertia weight gives the best results, where the losses in this case equal to 0.1005 MW
and this value is better than the other study [24] where the losses equal to 0.1241 MW
and better than the case of the reconfiguration by using Prim’s algorithm [26].
Fig. 3 Network after reconfiguration using the chaotic inertia weight
Table 2 Comparative study

Base Case with GA with DG [19] Prim’s algorithm Proposed PSO
DG with DG [18] with DG using
the chaotic
inertia weight
Tie line 33–34–35–36–37 12–15–18–21–22 12–27–33–34-35 2–13–26–29–33
Node of the 18 – 25 30
minimum
voltage
Minimum 0.8950 0.9124 0.95025 0.9359
voltage profile
(p.u)
Total Losses 0.2465 0.1241 0.1331 0.1005
(MW)
Concerning the value of the voltage profile at each node, it is noticed that the
minimum voltage profile is improved as compared with the base case (0.8950p.u at
node 18) and the reconfiguration using GA [24] (0.9124p.u), so the voltage profile
at each node is enhanced. In addition, we shouldn’t forget to point out that with this
strategy we need (46.68 s) to have the solution.
Fig. 4 Voltage profile in the case of the chaotic inertia weight
4.2 Combination Method Results
In this case, after reconfiguration of the network by using the combination of success
rate and the chaotic inertia weight to adjust the inertia weight parameter aiming to
enhance the PSO algorithm, we found the new tie line is 2–22–30–33-34 as shown
in Fig. 5.
On other hand and focused on Fig. 6 that gives the value of the voltage profile
at each node in two cases (before and after reconfiguration). This figure shows that
the minimum voltage profile in this case equals to 0.9115p.u at node 31. Table 7 in
appendices gives the value of the voltage at each node for this case.
Table 3 gives the results found by using these two strategies suggested to improve
the inertia weight parameter, the other recent studies, and the base case.
Focused on the result given in the previous table, it is clear that the value of the
losses in the case when the combination strategy equals to 0.115 MW, this value is
lesser than the base case 0.265 MW and the [24] 0.1241 MW and 0.1331 in the case
of the reconfiguration by using Prim’s algorithm [26]. But compared with the case of
the PSO algorithm using the chaotic inertia weight, it is noticed that this last strategy
gives a better result than the combination strategies.
Concerning the voltage profile, it is clear that the minimum value of the voltage
profile; in this case, it is equal to 0.9115p.u, and this value is better than the base case
and is almost similar to the [24], but the other studies give more improved values
than this case (0.95025p.u [26] and 0.9359p.u for the case of the PSO using the
Fig. 5 Network after reconfiguration using the combination strategy
Fig. 6 Voltage profile in the case of the combination strategy

268
Table 3 Comparative analysis

Base Case with DG GA with DG [24] Prim’s algorithm with DG Proposed PSO with DG Proposed PSO with DG
[26] using the chaotic inertia using the combination
weight strategy
Tie line 33–34-35–36–37 12–15–18–21–22 12–27–33–34–35 2–13–26–29–33 2–22–30–33-34
Node of the minimum 18 – 25 30 31
voltage
Minimum voltage profile 0.8950 0.9124 0.95025 0.9359 0.9115
(p.u)
Total losses (MW) 0.2465 0.1241 0.1331 0.1005 0.115
M. M’dioud et al.
chaotic inertia weight). It is interesting to point out that this strategy takes (55.178 s)
to execute and give the result.
5 Conclusion
Aiming to check the condition of the generation following the consumption, several
studies are interested in using the reconfiguration of the radial distribution system to
reduce the losses in the electrical system. So, the main objective of this study is to
find the new reconfiguration of the network by using two kinds of inertia weight (the
chaotic inertia weight and the hybrid of the chaotic inertia weight and the success
rate) to improve the results found by the PSO algorithm used in other recent studies.
To perform the reliability of these strategies, I select to test these strategies in
the case of IEEE 33 bus with the presence of the DGs. And a comparative study is
done to compare the results found by this paper with other recent studies. In the end,
using these strategies helps to enhance the network; also, the reconfiguration using
the chaotic inertia weight to adjust the PSO algorithm gives an improved result than
the combination strategy.
In the next research, it seems interesting to encourage the authors to do the next
search about using the PSO algorithm to find the optimal allocation and sizing of
DGs to improve the reconfiguration and reduce losses in the network.
Appendix
See Tables 4, 5, 6, and 7
Table 4 Set of the loops for

Loops Dimensions Switches
IEEE 33 bus [8]
1 Sd1 8–9–10–11–21–33–35
2 Sd2 2–3–4–5–6–7–18–19–20
3 Sd3 12–13–14–34
4 Sd4 15–16–17–29–30–31–36–32
5 Sd5 22–23–24–25–26–27–28–37
Table 5 Parameters of the

Parameter Value
proposed PSO
C1 2 * rand(1)
C2 2 * rand(1)
Wmax 0.9
Wmin 0.4
Population size 20
Dimension of search space 5
Maximum iteration 100
Table 6 Line and load data of IEEE 33 bus [22]

Line data Load data
Branch N° From bus To bus R (ohm) X(ohm) Pl(Kw) Ql(kvar)
1 1 2 0.0922 0.047 100 60
2 2 3 0.493 0.2511 90 40
3 3 4 0.366 0.1864 120 80
4 4 5 0.3811 0.1941 60 30
5 5 6 0.819 0.707 60 20
6 6 7 0.1872 0.6188 200 100
7 7 8 0.7114 0.2351 200 100
8 8 9 1.03 0.74 60 20
9 9 10 1.04 0.74 60 20
10 10 11 0.1966 0.065 45 30
11 11 12 0.3744 0.1238 60 35
12 12 13 1.468 1.155 60 35
13 13 14 0.5416 0.7129 120 80
14 14 15 0.591 0.526 60 10
15 15 16 0.7463 0.545 60 20
16 16 17 1.289 1.721 60 20
17 17 18 0.732 0.574 90 40
18 2 19 0.164 0.1565 90 40
19 19 20 1.5042 1.3554 90 40
20 20 21 0.4095 0.4784 90 40
21 21 22 0.7089 0.9373 90 40
22 3 23 0.4512 0.3083 90 50
23 23 24 0.898 0.7091 420 200
24 24 25 0.896 0.7011 420 200
25 6 26 0.203 0.1034 60 25
(continued)
Table 6 (continued)
Line data Load data
26 26 27 0.2842 0.1447 60 25
27 27 28 1.059 0.9337 60 20
28 28 29 0.8042 0.7006 120 70
29 29 30 0.5075 0.2585 200 600
30 30 31 0.9744 0.963 150 70
31 31 32 0.3105 0.3619 210 100
32 32 33 0.341 0.5302 60 40
Tie lines
33 8 21 2 2
34 9 15 2 2
35 12 22 2 2
36 18 33 0.5 0.5
37 25 29 0.5 0.5
Table 7 Voltage Profile in

Bus The chaotic inertia weight The combination strategy
the case of using the chaotic
parameter
inertia weights and the
combination strategy 1 1 1
2 0.999172639148097 0.998909397340444
3 0.950119615573435 0.941816046915740
4 0.950197864595776 0.941915894499951
5 0.950384769075342 0.942172719114258
6 0.950736541598505 0.942745626162071
7 0.950801915260580 0.942272190455390
8 0.952178374115133 0.942044120286597
9 0.953986728146658 0.943153382841413
10 0.959346721345702 0.944666257465305
11 0.960258850341073 0.945036272631173
12 0.961996158279058 0.945788205077869
13 0.961283469701170 0.935454716873337
14 0.945915236387610 0.931030577352812
15 0.946340607670596 0.927628810209586
16 0.944554248961896 0.923934940402118
17 0.940230512256466 0.916188864414032
18 0.939000617939255 0.913919720659541
19 0.997624030527326 0.996806758504366
20 0.984803200597631 0.979011354882791
21 0.981438432428948 0.974175549465857
(continued)
Table 7 (continued)
Bus The chaotic inertia weight The combination strategy
parameter
22 0.975847552100016 0.965979448836100
23 0.950113458716196 0.948086801989005
24 0.950239293567017 0.947921749672686
25 0.950513319395836 0.947084596452350
26 0.950698911078949 0.942977540694717
27 0.950729767133923 0.943325296224972
28 0.950681330399383 0.945007617974172
29 0.950602669016140 0.946348853096167
30 0.935947175640918 0.946457584854114
31 0.936910083625147 0.911254236300635
32 0.937366905217155 0.911662243051541
33 0.938123317946990 0.912578877413074
References
1. Peng, F.Z.: Editorial Special Issue on Distributed Power Generation. IEEE Trans. Power
Electron. 19(5), 2 (2004)
2. Carreno, E.M., Romero, R., Padilha-Feltrin, A.: An efcient codifcation to solve distribution
network reconfguration for loss reduction problem. IEEE Trans. Power Syst. 23(4), 1542–1551
(2008)
3. Salama, M.M.A., El-Khattam, W.: Distributed generation technologies, defnitions and benefts.
Electric Power Systems Research 71(2), 119–128 (2004)
4. Multon, B.: L’énergie électrique: analyse des resources et de la production.Journées de la
section électrotechnique du club EEA (1999)
5. Strasser, T., Andrén, F., Kathan, J., Cecati, C., Buccella, C., Siano, P., Leitão, P., Zhabelova,
G., Vyatkin, V., Vrba, P., Mařík, V.: A Review of Architectures and Concepts for Intelligence
in Future Electric Energy Systems. IEEE Trans. Industr. Electron. 62(4), 2424–2438 (2014)
6. Caire, R.: Gestion de la production décentralisée dans les réseaux de distribution.Institut
National Polytechnique de Grenoble, tel-00007677 (2004)
7. Xie, L., Ilic, M.D.: Model predictive dispatch in electric energy systems with intermittent
resources. In IEEE International Conference on Systems, Man and Cybernetics (2008)
8. JUMA, S.A.: Optimal radial distribution network reconfiguration using modified shark smell
optimization. (2018) http://hdl.handle.net/123456789/4854
9. Sivkumar, M.: A Simple Algorithm for Distribution System Load Flow with Distributed Gener-
ation, In IEEE International Conference on Recent Advances and Innovations in Engineering,
Jaipur, India (2014)
10. Gallego, LA, Carreno E., Padilha-Feltrin, A.: Distributed generation modeling for unbal-
anced three-phase power flow calculations in smart grids. In Transmission and Distribution
Conference and Exposition: Latin America (T&D-LA) (2010)
11. Chidanandappa, R., Ananthapadmanabha, T.:Genetic algorithm based network reconfiguration
in distribution systems with Multiple DGs for time varying loads. SMART GRID Techno. 21,
460–467 (2015)
12. Ogunjuyigbe, A., Ayodele, T., Akinola, O.: Impact of distributed generators on the power loss
and voltage profile of sub-transmission network. J. Electr. Syst. Inf. Technol. 3, 94–107 (2016)
13. Ahmad, S., Asar, A.U., Sardar, S., Noor, B.: Impact of distributed generation on the reliability
of local distribution system. IJACSA 8(6), 375–382 (2017)
14. Ma, C., Li, C., Zhang, X., Li, G., Han, Y.: Reconfiguration of distribution networks with
distributed generation using a dual hybrid particle swarm optimization algorithm.Hindawi
Math. Probl. Eng. 2017, 11 (2017)
15. Sudhakara Reddy, A.V., Damodar Reddy, M.: “Optimization of network reconfiguration by
using particle swarm optimization. In 1st IEEE International Conference on Power Electronics,
Intelligent Control and Energy Systems (ICPEICES) (2016)
16. Tandon, A., Saxena, D.: Optimal reconfiguration of electrical distribution network using selec-
tive particle swarm optimization algorithm. In International Conference on Power, Control and
Embedded Systems (2014)
17. Inji Ibrahim Atteya: Hamdy Ashour, Nagi Fahmi and Danielle Strickland, Radial distribu-
tion network reconfiguration for power losses reduction using a modified particle swarm
optimisation, . Open Access Proceedings J. 2017(1), 2505–2508 (2017)
18. Bansal, J.C., Singh, P.K., Saraswat, M., Verma, A., Jadon, S.S., Abraham, A.: Inertia weight
strategies in particle swarm optimization. In Third World Congress on Nature and Biologically
Inspired Computing (2011).
19. Arasomwan, A.M., ADEWUMI, A.O.: On adaptive chaotic inertia weights in particle swarm
optimization. In IEEE Swarm Intelligence Symposium (2013)
20. Enacheanu, F.: outils d’aide à la conduite pour les opérateurs des réseaux de distribution
(2008).https://tel.archives-ouvertes.fr/tel-00245652
21. Sharma, A., Saini, M., Ahmed, M.: Power flow analysis using NR method. In International
Conference on Innovative Research in Science, Technoloy and Management, Kota, Rajasthan,
India (2017)
22. Baran, M.E., Wu, F.F.: Network reconfiguration in distribution systems for loss reduction and
load balancing. IEEE Trans. Power Delivery 4(2), 1401–1407 (1989)
23. Jangjoo, M.A., Seifi, A.R.: Optimal voltage control and loss reduction in microgrid by active
and reactive power generation.J. Intell. & Fuzzy Syst., 27, 1649–1658 (2014)
24. Moarrefi, H., Namatollahi, M., Tadayon, M.: Reconfiguration and distributed Generation(DG)
placement considering critical system condition. In 22nd International Conference and
Exhibition on Electricity Distribution (2013)
25. Kennedy, J., Eberhart, R.: Particle swarm optimization. In International Conference on Neural
Networks (1995)
26. M’dioud, M., ELkafazi, I., Bannari, R.: An improved reconfiguration of a radial distribution
network by using the minimum spanning tree algorithm. Solid State Technol. 63(6), 9178–9193
(2020)
On the Performance of 5G Narrow-Band
Internet of Things for Industrial
Applications
Abdellah Chehri, Hasna Chaibi, Rachid Saadane, El Mehdi Ouafiq,

and Ahmed Slalmi
Abstract Manufacturing industry is continuously evolving since the very beginning

of the industrial era. This modernization is undoubtedly the outcome of continuous
new technology development in this field, which has kept the industries on the verge,
looking for new methods for improving productivity enhancement and better oper-
ational efficiency. The advent of 5G will provide the world of industry to connect
its infrastructures to digitize people and machines to optimize production flows.
Narrow-band-IoT addresses “Massive IoT” type use cases, which involve deploying
a large energy-efficient quantity. These low-complex objects do not need to commu-
nicate very frequently. 5G will provide the ability to develop new uses previously
impossible or complex to implement. Consequently, it will complement the range of
network solutions already in place in the company, giving it the keys to accelerating
its transformation. This paper evaluates the 5G-NR-based IoT air interface with the
FEC with industrial channel models. Low-density parity-check (LDPC), polar, turbo
code, and TBCC are assumed.
1 Introduction
At this time, the 4G cellular networks have existed for several years. It is time to look
forward and see what the future will bring regarding the next generation of cellular
networks: the fifth generation, most often referred to as 5G [1, 2].
A. Chehri
University of Quebec in Chicoutimi, 555, Boul. de L’Université, G7H 2B1 Saguenay, QC, Canada
e-mail: achehri@uqac.ca
H. Chaibi
GENIUS Laboratory, SUP MTI, 98, Avenue Allal Ben Abdellah, Hassan-Rabat, Morocco
R. Saadane (B) · E. M. Ouafiq
SIRC/LaGeS-EHTP, EHTP, Km 7 Route, El Jadida 20230, Morocco
A. Slalmi
Ibn Tofail University, Kenitra, Morocco
https://doi.org/10.1007/978-981-16-3637-0_19
276 A. Chehri et al.
Today, “everyone” is already online, and with the emerging Internet of Things,
everything will also be online—everywhere, and always. There is a demand for
ubiquitous access to mobile services.
The “Internet of Things,” which must also be counted among the most signif-
icant technological trends, is often mentioned in conjunction with M2M commu-
nication. The risk of confusion is high here: Even if the two approaches overlap,
they are still two different things. What they have in common is the goal of auto-
mated data exchange between devices. IIoT, or the “Industrial Internet of Things,”
is mainly aimed at private users. There are various carrier networks in the area of
M2M communication, i.e., cellular radio or GPRS, which are an option [3].
Furthermore, classic M2M communication involves point-to-point applications.
On the other hand, in the context of the IoT, a standardized, open approach is
practiced. Ultimately, however, it is already foreseeable that both technologies will
converge, complement each other, and one day may no longer be separable. For
example, many M2M providers have already started to integrate cloud functions into
their offerings [4, 5].
3G and 4Gs important goal were to achieve constant coverage for the same
services in both outdoor and indoor scenarios. According to Chen and Zhao, 5G
will be a heterogeneous framework, and backward compatibility will not be manda-
tory indoors and outdoors [6]. The improvement of the user equipment is expected to
provide the ability to support simultaneous connections, both indoors and outdoors.
The advent of 5G will provide the world of industry with the means to connect
its infrastructures to digitize people and machines to optimize production flows.
5G must provide a unified vision of connectivity within the enterprise regardless
of the network. The core network is designed to encompass all types of access
to natively. However, the arrival of technology does not mean the “end” of other
systems; on the contrary. Each has its characteristics and specific uses, such as NB-
IoT, a global industry standard, open, sustainable, and scalable. It complements the
other technologies already defined, such as Sigfox and LoRA; NB-IoT addresses
“Massive IoT” type use cases, which involve deploying a large quantity of energy-
efficient. These low-complex objects do not need to communicate very frequently.
5G will provide the ability to develop new uses previously impossible or complex to
implement. Consequently, it will complement the range of network solutions already
in place in the company, giving it the keys to accelerating its transformation.
IoT requires massive connectivity where several low-cost devices and sensors
communicate.
The deployment of wireless technology in wider and more critical industrial appli-
cations requires deterministic behavior, reliability, and predictable latencies to inte-
grate the industrial processes more effectively. Real-time data communication and
information reliability in the wireless channels are some of the major concerns of the
control society regarding NB-IoT, and hence, suitable improvements in NB-IoT are
required to ensure desired reliability and time sensitivity in emergency, regulatory,
and supervisory control systems.
On the Performance of 5G … 277
This is being labeled as the fourth industrial revolution or Industry 4.0. There are
many advantages brought by 5G cutting-edge technologies for industrial automation
scenarios in the drive for industry 4.0
This paper is organized as follows. Section 2 describes the prominent 5G use cases.
Section 3 introduces terminology and description of the Industrial Internet of Things,
the main pillar of Industry 4.0. Section 4 describes the 5G NR (New Radio) interface.
The performance of 5G narrow-band Internet of Things for industrial applications is
given in Sect. 5. Finally, Sect. 6 presents our conclusions.
2 5G Use Cases
To give the reader an idea about which communication challenges 5G is expected

to solve, the most important use cases will be presented. Use cases are descriptions
of interactions between an actor, typically the user, and a system, to achieve a goal.
They also make it clearer to understand the background for the 5G requirements, and
they are likely to be the driver for the 5G technology that will be developed [7].
Some of the 5G applications will be old and familiar while introducing new
and more diverse services is also expected. The existing dominating human-centric
communication scenarios will be complemented by an enormous increase in commu-
nication directly between machines or fully automated devices. Hence, the use cases
for 5G may be split into two following main categories: mobile broadband and the
Internet of Things (IoT).
According to forecasts, the growth in data traffic volumes will be exponential
for the next years. In the next decade, the total volume of mobile traffic is expected
to increase a thousand times the traffic volume today, mostly due to the increasing
number of connected devices.
Compared to the mobile broadband use case, the IoT category covers the use cases
where devices and sensors communicate directly with each other, without a user being
present at all times. This is referred to as machine-to-machine (M2M) communication
[8], and this type of communication will be an essential part of the emerging IoT.
Here, devices and items embedded with sensors, electronics, software, and network
connectivity collect and exchange data. Sensors or devices for M2M communications
will be integrated daily using objects such as cars, household appliances, textiles,
and health-critical appliances [9].
Various standardization organizations have technical working groups responsible
for machine-to-machine communication (M2M) and the Internet of the future. The
third-generation partnership project (3GPP) deals with M2M communication under
the term machine-type communication (MTC) and has started to standardize it in
the 3GPP specification Release 10. In Release 11, some of the proposed suggested
functions address and control the devices [10].
The Internet Engineering Task Force (IETF) oversees many of the necessary
protocols in the Internet of Things. These include tried and tested protocols such as
IP, TLS, and HTTP, which are widespread and ensure interoperability.
However, newer protocols have also been developed to consider the changed
conditions in M2M communication, such as the constrained application protocol
(CoAP) and the communication protocol IPv6 over low-power wireless personal
area network (6LoWPAN). Other necessary protocols in M2M communication are
message protocols such as Extensible Messaging and Presence Protocol (XMPP),
MQ Telemetry Transport (MQTT), and Advanced Message Queuing Protocol
(AMQP). Other protocols that enable the management of the devices, such as device
management (DM) and lightweight (LW) M2M from Open Mobile Alliance (OMA)
and TR-069 from Broadband Forum (BBF), were also proposed [11].
To ensure and develop M2M standards, the European Telecommunications Stan-
dards Institute (ETSI) founded a technical committee in 2009. The requirements were
defined, which in addition to security and communication management, also the func-
tional requirements a horizontal platform for M2M communication. This platform
should ensure that communication with a wide variety of sensors and actuators is
possible in a consistent manner for different applications.
3 Industrial Internet of Things (IIoT)
IIoT is a variant of the IoT that is used in the industrial sector. The Industrial Internet
of Things can be used in many industries, in manufacturing, in agriculture, in hospi-
tals, in institutions, in the field of health care, or the generation of energy and
resources. One of the most critical aspects is improving operational effectiveness
through intelligent systems and more flexible production techniques [12–15].
With IIoT, industrial plants or gateways connected to them send data to a cloud.
Gateways are hardware components that connect the devices of the industrial plants
and the sensors to the network. The data is processed and prepared in the cloud.
This allows employees to monitor and control the machines remotely. Besides, an
employee is informed if, for example, maintenance is necessary [16].
The Industrial Internet of Things also uses object-oriented systems connected
to the network and can interact with each other. These devices are equipped with
sensors whose role is to collect data by monitoring the production context in which
they operate. The information stored in this way is then analyzed and processed,
helping to optimize business processes.
From predictive maintenance to the assembly line, the Industrial Internet of
Objects (IIoT) offers a significant competitive advantage regardless of the industry.
Sensors play a central role in IIoT. They collect vast amounts of data from different
machines in one system. To be able to meet this challenge, the sensors in the IIoT
area must be significantly more sensitive and precise than in the IoT environment.
Even minor inaccuracies in the acquisition of the measurement data can have fatal
consequences such as financial losses. The IIoT offers numerous advantages for
industry:
1. Production processes can be automated.
2. Processes can be adapted flexibly and in real-time to changing requirements.

3. Machines recognize automatically when they need to be serviced.
4. Some of the maintenance can be carried out by the machines themselves.
5. Disturbances and production interruptions are minimized.
6. Throughput and production capacity increase.
The IIoT also brings some challenges that need to be addressed in the future
intensive maintenance of the machines in terms of software and firmware. Both have
to be kept up to date to close security gaps or prevent them from arising in the first
place.
Secure transmission of the encrypted data to the cloud is a prerequisite. Otherwise,
it will be easy for hackers to get hold of sensitive data.
So far, IIoT devices from different manufacturers are not compatible with
each other. The reason is the lack of uniform standards and manufacturer-specific
protocols.
High effort for the processing, protection, and storage of the data. In order to cope
with the substantial amounts of data, applications, and databases from the big data
area have to be used.
NB-IoT (for narrow-band IoT) is a serious asset when the goal is energy use.
This infrastructure technology could provide ten years of autonomy to a 5Wh
battery through energy optimization and ambient energy recovery. NB-IoT oper-
ates a 200 kHz frequency band and is used for fixed sensors that do not need a small
volume of data like water or electricity meters.
LTE-M technology is widely used in mobile telephony since it is compatible with
existing networks and does not require new modems. Its transfer rate is high, LTE-M
will also evolve in environments such as remote monitoring or autonomous vehicles.
Another advantage: LTE-M supports voice exchanges and the mobility of objects.
4 5G NR Radio Interface
With the advent of the IoT, the issues related to Industry 4.0, and experts’ prediction
to have more than 75 billion objects connected using a wireless network by 2025, it
is necessary to create technologies adapted to these new needs. This standard allows
connected objects to communicate large volumes of data over very large distances
with very high latency.
NB-IoT or narrow-band IoT, or LTE-M2M is low consumption and long-range
technology (LPWAN) validated in June 2016, operating differently.
Like LoRa and Sigfox, this standard allows low-power objects to communicate
with external applications through the cellular network.
The communication of these objects via NB-IoT is certainly not real-time but
must be reliable over time. By relying on existing and licensed networks, operators
are already in charge of their quality of service. They will thus be able to guarantee
a quality of service (QoS) sufficient for this operation type.
NB-IoT builds on existing 4G networks from which several features and mech-
anisms are inherited. It is therefore compatible with international mobility thanks
to roaming, also called roaming. This also means that these networks are acces-
sible under license and are managed by operators specialized in the field. Experts,
therefore, manage the quality of the network in the area.
NB-IoT is considered 5G ready, which means that it will be compatible with this
new transmission standard when it is released.
For NR, the relevant releases are Release 14, 15. In Release 14, a number of
preliminary activities were done to prepare for the specification of 5G. For instance,
one study was carried out to develop propagation models for spectrum above 6 GHz.
Another study was done on scenarios and requirements for 5G and concluded at
the end of 2016. Besides, a feasibility study was done of the NR air interface itself,
generating several reports covering all aspects of the new air interface. Rel’15 will
contain the specifications for the first phase of 5G.
NR DownLink (DL) and UpLink (UL) transmissions are organized into frames.
Each frame lasts 10 ms and consists of 10 subframes, each of 1 ms. Since multiple
OFDM numerologies are supported, each subframe can contain one or more slots.
There are two types of cyclic prefix (CP): normal CP, each slot conveys 14 OFDM
symbols. Extended CP, each slot shares 12 OFDM symbols. Besides, each symbol
can be assigned for DL or UL transmission, according to the slot format indicator
(SFI), which allows flexible assignment for TDD or FDD operation modes. In the
frequency domain, each OFDM symbol contains a fixed number of sub-carriers.
One sub-carrier allocated in one OFDM symbols is defined as one resource
element (RE). A group of 12 RE is defined as one resource block (RB). The total
number of RBs transmitted in one OFDM symbol depends on the system bandwidth
and the numerology. NR supports scalable numerology for more flexible deploy-
ments covering a wide range of services and carrier frequencies. It defines a positive
integer factor m that affects the sub-carrier spacing (SCS), the OFDM symbol, and
cyclic prefix length.
A small sub-carrier spacing has the benefit of providing a relatively long cyclic
prefix in absolute time at a reasonable overhead. In contrast, higher sub-carrier spac-
ings are needed to handle, for example, the increased phase noise at higher carrier
frequencies [17]. Note that the sub-carrier spacing of 15, 30, and 60 kHz wide apply
to carrier frequencies of 6 GHz or lower (sub-6), while the sub-carrier spacing of 60,
120, and 240 kHz apply to above 6 GHz carrier frequencies [18].
An NB-IoT channel is only 180 kHz wide, which is very small compared to
mobile broadband channel bandwidths of 20 MHz. So, an NB-IoT device only needs
to support the NB-IoT part of the specification. Further information about the specifi-
cation of this category can be found in the 3GPP technical report TR 45.820: Cellular
system supports for ultralow complexity and low throughput Internet of Things [19]
(Fig. 1).
The temporal and frequency resources which carry the information coming from
the upper layers (layers above the physical layer) are called physical channels [20].
There are several physical channels to specify for the uplink and downlink:
Fig. 1 NR framing structure
1. Physical downlink shared channel (PDSCH): used for downlink data transmis-
sion.
2. Physical downlink control channel (PDCCH): used as a downlink for informa-
tion control, which includes the scheduling decisions required for the reception
of downlink data (PDSCH) and for scheduling granting authorization to transmit
data uplink (PUSCH) by a UE.
3. Physical broadcast channel (PBCH): used for broadcasting system information
required by a UE to access the network;
4. Physical uplink shared channel (PUSCH): used for uplink data transmission (by
a UE).
5. Physical uplink control channel (PUCCH): used for uplink control informa-
tion, which includes: HARQ acknowledgment (indicating whether a downlink
transmission was successful or not), schedule request (request network time-
frequency resources for uplink transmissions), and downlink channel status
information for link adaptation.
6. Physical random-access channel (PRACH), used by a UE to request the
establishment of a connection called random access.
When a symbol is sent through a physical channel, a delay created by the propa-
gation signal taking different paths may cause the reception of several copies of the
same frame. A cyclic prefix is added to each symbol to solve this problem, consisting
of samples taken from its end and tied at its beginning. The goal is to add a guard
time between two successive symbols. If a CP’s length is longer than the maximum
channel propagation, there will be no inter-symbol interference (ISI), which means
that two successive symbols will not interfere. It also avoids inter-carrier interference
(ICI), which causes the loss of orthogonality between the sub-carriers. It uses a copy
of the last part of the symbol that plays a guard interval [21].
5G NR can use a spectrum from 6 GHz to 100 GHz. The 5G system’s bandwidth
is increased by ten times (from 100 MHz in LTE-A to 1 GHz +) compared to LTE-A
technology. Bands for NR are basically classified as low, middle, and high bands,
and these bands can be used depending on the application described below:
1. Low bands below 1 GHz: most extended range, e.g., mobile broadband and
massive IoT, e.g., 600, 700, 850/900 MHz
2. Medium bands 1 GHz to 6 GHz: wider bandwidths, for example, eMBB and
critical, for example, 3.4–3.8 GHz, 3.8–4.2 GHz, 4.4–4.9 GHz
3. High bands above 24 GHz (mm-Wave): extreme bandwidths, for example,
24.25–27.5 GHz, 27.5–29.5, 37–40, 64–71 GHz.
OFDM has 15 kHz sub-carrier spacing with a 7% (4.69 µs) cyclic prefix. The
numerology for LTE has been specified after an extensive investigation in 3GPP.
For NR, it was easy for 3GPP to aim for OFDM numerology similar to LTE-
like frequencies and deployments. 3GPP, therefore, considered different sub-carrier
spacing options near 15 kHz as basic numerology for NR. There are two important
reasons to keep LTE numerology as the base numerology:
1. Narrow-band-IoT (NB-IoT) is a new radio access technology (already
deployed since 2017) supporting massive machine-type communications. NB-
IoT provides different deployments, including in-band deployment within an
LTE operator enabled by the selected LTE numerology. NB-IoT devices are
designed to operate for ten years or more on a single battery charge. Once such
an NB-IoT device is deployed, the incorporating carrier will likely be rearmed
to NR during the device’s life.
2. NR deployments can take place in the same band as LTE. With an adjacent LTE
TDD carrier, the network controller should adopt the same uplink/downlink
switching model as the LTE TDD protocol. Each numerology where (an integer
multiple of) a subframe is 1 ms can be aligned to regular subframes in LTE. In
LTE, duplex switching occurs in special subframes. To match the direction of
transmission in special subframes, the same numerology as in LTE is required.
This implies the same sub-carrier spacing (15 kHz), the same OFDM symbol
duration (66.67 µs), and the same cyclic prefix (4.69 µs)
5 Performance of 5G Narrow-Band Internet of Things

for Industrial Applications
A particular emphasis has been done on the design of forwarding error correction
(FEC) solutions to support the underlying constraints efficiently. In this regard, it has
been taken into account the codes used in 5G-NR (Rel’ 15) channels, low-density
parity-check (LDPC) for data and polar code for control, and those code (TBCC) for
control. This scenario’s potential requirements are the use of lower-order modulation
schemes with shorter block size information to satisfy low-power requirements.
The advanced channel coding schemes with robust error protection with low-
complexity encoding and decoding is preferred. The candidate coding scheme for
the next 5G based IoT system is: polar code, low-density parity-check (LDPC), turbo
code, and tail-biting convolutional code (TBCC) [22].
In the time domain, physical layer transmissions are organized into radio frames.
A radio frame has a duration of 10 ms. Each radio frame is divided into ten subframes
of 1 ms duration. Each subframe is then divided into locations [23].
This section evaluates the 5G-NR-based IoT air interface with the FEC schemes
previously and with industrial channel models.
The large-scale wireless channel characteristics were evaluated from 5 to 40 GHz
frequency for industrial scenario (Fig. 2).
Since the data traffic generated by IoT applications requires a small volume, it has
been considered a data bit range from 12 up to 132 with a step of 12. The segmentation
(and de-segmentation) block has not been considered due to the small data packet to
transmit. Furthermore, from upper layers is randomly generated and does not refer
to a specific channel. LDPC, polar, turbo code, and TBCC are assumed.
3GPP agreed to adopt polar codes for the enhanced mobile broadband (eMBB)
control channels for the 5G NR (new radio) interface. At the same meeting, 3GPP
agreed to use LDPC for the corresponding data channel [24].
The polar code keeps performing better than other codes, achieving 3 dB, with
respect to the turbo code (Fig. 3).
6 Conclusion
Modularity, flexibility, and adaptability of production tools will be the rule in Industry
4.0. 5G will allow the integration of applications that leverage automation and robo-
tization. The optimized flow rates and the ability to integrate numerous sensors
to ensure preventive and predictive maintenance of production tools constitute a
prospect of increasing the reliability of Industry 4.0. The consolidation of indus-
trial wireless communication into standards leads to an increase in deployments
throughout various industries today. Despite the technology being considered mature,
plant operators are reluctant to introduce mesh networks into their process, despite
their very low energy profiles. While very promising, 5G will not take hold quickly,
with its high costs slowing mass distribution. Between its business model, the price
of the connection, and the cost of electronics, it will take a few years to see it flourish
everywhere.
Fig. 2 Path loss versus frequency for V-V and V-H polarization for indoor channels of mm-wave
bands [25]
Fig. 3 BER vs. SNR for different FEC techniques
References
1. GPP: Release 15. Retrieved March 2, 2017. http://www.3gpp.org/release-15

2. Carlton, A.: 5g reality check: Where is 3gpp on standardization? Retrieved March 18th, 2017
3. Slalmi, A., Chaibi, H., Saadane, R., Chehri, A., Jeon, G.: 5G NB-IoT: Efficient Network
Call Admission Control in Cellular Networks. In: concurrency and computation: practice and
experience. Wiley, e6047. https://doi.org/10.1002/cpe.6047
4. Slalmi, A., Chaibi, H., Chehri, A., Saadane, R., Jeon, G., Hakem, N.: On the ultra-reliable and
low-latency communications for tactile internet in 5G era. In: 24th International Conference on
Knowledge-Based and Intelligent Information & Engineering Systems, Verona, Italy, 16–18
September 2020
5. Slalmi A., Saadane R., Chehri A., Kharraz H.: How Will 5G Transform Industrial IoT: Latency
and Reliability Analysis. In: Zimmermann A., Howlett R., Jain L. (eds) Human Centred Intel-
ligent Systems. Smart Innovation, Systems and Technologies, vol 189. Springer, Singapore
(2020)
6. Chen, S., Zhao, J.: The requirements, challenges, and technologies for 5 g of terrestrial mobile
telecommunication. IEEE Comm. Magazine 52(5), 36–43 (2014)
7. Nokia: White paper: 5 g use cases and requirements (2014)
8. Sabella, A., Wbben, D.: Cloud technologies for flexible 5 g radio access networks. IEEE
Communications Magazine 52(5), 68–76
9. Tehrani, M.N., Uysal, M., Yanikomeroglu, H.: Device-to-device communication in5g cellular
networks: Challenges, solutions and future directions. IEEE Communications Magazine 52(5),
86–92 (2014)
10. Kunz, A., Kim, H., Kim, L., Husain, S.S.: Machine type communications in 3GPP: From
release 10 to release 12. 2012 IEEE Globecom Workshops, Anaheim, CA, pp. 1747–1752
(2012). https://doi.org/10.1109/GLOCOMW.2012.6477852
11. Wahle, S., Magedanz T., Schulze, F.: The OpenMTC framework—M2M solutions for smart
cities and the internet of things. IEEE International Symposium on a World of Wireless, Mobile
and Multimedia Networks (WoWMoM), San Francisco, CA (2012)
12. Slalmi, A., Saadane, R., Chehri, A.: Energy Efficiency Proposal for IoT Call Admission Control
in 5G Network. In: 15th International Conference on Signal Image Technology & Internet Based
Systems, Sorrento (NA), Italy, November 2019
13. Chehri, A., Mouftah, H.: An empirical link-quality analysis for wireless sensor networks. Proc.
Int. Conf. Comput. Netw. Commun. (ICNC), 164–169 (2012)
14. Chehri, A., Chaibi, H., Saadane, R., Hakem, N., Wahbi, M.: A framework of optimizing the
deployment of IoT for precision agriculture industry, vol 176, 2414–2422 (2020). ISSN 1877-
0509, KES 2020
15. Chehri, A.: The industrial internet of things: examining how the IIoT will improve the predictive
maintenance. Ad Hoc Networks, Lecture Notes of the Institute for Computer Sciences, Smart
Innovation Systems and Technologies, Springer (2019)
16. Chehri, A.: Routing protocol in the industrial internet of things for smart factory monitoring:
Ad Hoc networks, Lecture Notes of the Institute for Computer Sciences, Smart Innovation
Systems and Technologies, Springer (2019)
17. GPP, 5G NR; Overall Description; Stage-2, 3GPP TS 38.300 version 15.3.1 Release 15, October
2018
18. GPP TS 38.331 v.15.1.0: NR. Radio Resource control (RRC), Protocol Specification, 2015
19. GPP. TS 45.820 v2.1.0: Cellular System Support for Ultra Low Complexity and Low
Throughput Internet of Things, 2015
20. Furuskär A., Parkvall, S., Dahlman, E., Frenne, M.: NR: The new 5G radio access technology.
IEEE Communications Standards Magazine (2017)
21. Chehri, A., Mouftah, H.T.: New MMSE downlink channel estimation for Sub-6 GHz non-line-
of-sight backhaul. In: 2018 IEEE Globecom Workshops (GC Workshops), Abu Dhabi, United
Arab Emirates, pp. 1–7 (2018). https://doi.org/10.1109/GLOCOMW.2018.8644436
22. GPP. TS 38.213 v15.1.0: Physical Layer Procedures for Control, 2018
23. Vardy, T.: List decoding of polar codes. IEEE Trans. Inf. Theory 61(5), 2213–2226 (2015)
24. Tahir, B., Schwarz, S., Rupp, M.: BER comparison between Convolutional, Turbo, LDPC, and
Polar codes. I: 2017 24th International Conference on Telecommunications (ICT), Limassol,
pp. 1–7 (2017)
25. Al-Samman, A.M., Rahman, T.A., Azmi, M.H., Hindia, M.N., Khan, I., Hanafi, E.: Statistical
Modelling and Characterization of Experimental mm-Wave Indoor Channels for Future 5G
Wireless Communication Networks. PLoS ONE 11(9), (2016)
A Novel Design of Frequency
Reconfigurable Antenna for 5G Mobile
Phones
Sanaa Errahili, Asma Khabba, Saida Ibnyaich, and Abdelouhab Zeroual
Abstract The purpose of this paper is to design a new frequency reconfigurable

patch antenna that operates in two different frequency bands. The planned antenna
is designed on dielectric substrate of Rogers RT5880 with 2.2 relative permittivity.
The total size of the antenna is 6 × 5.5 × 1.02 mm3 . The proposed antenna is
formed of a positive intrinsic negative diode (PIN diode), placed at the radiating
patch to achieve frequency reconfigurability based on the switching state of the PIN
diode. The simulation of the proposed antenna is implemented using CST microwave
studio. The performance of antenna is analyzed from the reflection coefficient, the
surface current distribution, and the radiation pattern. The antenna has two resonant
frequencies for 5G applications: 26.15 GHz and 46.1 GHz.
1 Introduction
With the emergence of new standards, telecommunication systems must be able

to combine several standards on the same antenna. Reconfigurable antennas [1–
3] are an important part of applications in wireless communication because their
operation can be modified dynamically [4], which can be very advantageous for
several applications. In addition, reconfigurability allows the antenna to offer more
functionality.
Reconfigurable antennas must be able to adapt to their environment by changing
their operating frequency [5], and/or their polarization [6] and/or radiation pattern
[7, 8].
There are several reconfigurability techniques on designing the reconfig-
urable antenna like employing electronic, mechanical, or optical switching [9–
11]. However, the electronic switching is more frequently used compared to other
approaches, that is due to her efficiency and reliability. The technics of electronic
switching comprise the PIN diodes, varactor diodes, field effect transistor (FET),
S. Errahili (B) · A. Khabba · S. Ibnyaich · A. Zeroual

I2SP Laboratory, Faculty of Science, Cadi Ayyad University Marrakech, Marrakech, Morocco
e-mail: sanaa.errahili@ced.uca.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_20
288 S. Errahili et al.
and radio-frequency microelectromechanical system(RF MEMS) switches. Other

approaches based on substrate agility. Many reconfigurable antenna is proposed in
the recent years.
In this paper, we propose a noval reconfigurable patch antenna that can be tuned in
two frequency bands by changing the geometry of the radiation patch by using a PIN
diode. The proposed reconfigurable antenna is based on the PIN diode ON and OFF
status between radiating elements, and it is able to select a very separate frequency
band. The proposed antenna covers two bands of the fifth generation: 26.15 and
46.1 GHz.
2 Design of the Proposed Antenna
The design of the proposed reconfigurable antenna is presented in Fig. 1, which

shows the side view of the antenna.
The proposed patch [12, 13] antenna has designed on 1.6 mm thickness of Rogers
RT5880 dielectric with relative permittivity 2.2, and the dimension of the substrate
is 6 × 5.5 × 1.02 mm2 . The metallization of the patch and the ground plane is
considered as copper, which has a uniform thickness t. The excitation is launched
through microstrip feed line, which has a width Wf .
Parameters value of the proposed antenna is listed in Table 1.
As shown in Fig. 1, a PIN diode is located on the radiating patch. The PIN diode
Fig. 1 Geometry of the proposed antenna

A Novel Design of Frequency Reconfigurable Antenna … 289
Table 1 Parameters of the

Parameter Value (mm)
proposed antenna
W 60
L 100
Ws 5.5
Ls 6
Wp 3
Lp 3
h 0.95
t 0.035
Wf 0.18
Lf 2
is used to connect the two parts of the patch. When the PIN diode is on state OFF,
the antenna has only the main patch and the antenna operates at 26.15 GHz. The
second configuration, when the PIN diode is on state ON, so the proposed antenna
includes the two parts of patch, the proposed antenna operates at 46.1 GHz.
Figure 2 represents the equivalent circuit model of a PIN diode. The model used
is that proposed in [14, 15], it is a simplified RLC equivalent circuit of the PIN diode
that does not take account of the “surface mounting” effect. It consists of a parasitic
inductor (L) in series with an intrinsic capacity (C) and an intrinsic resistance (R),
which are connected in parallel (Fig. 2b). When the PIN diode is in the OFF state,
the values of R, L, and C are, respectively, equal to R2 , L 1 and C 1 . Conversely, when
the PIN diode is in the ON state, the capacitance does not intervene, and the values
of R and L are, respectively, equal to R1 and L1 (Fig. 2a).
In this work, the PIN diode MPP4203 is used as a switch. The circuit parameters
are L 1 = 0.45 nH, R1 = 3.5, R2 = 3 k, and C 1 = 0.08 pF.
Fig. 2 PIN diode equivalent

circuit [16]: a ON state,
b OFF state
In this section, the simulated results of the proposed reconfigurable antenna: reflection
coefficient S11, surface current distribution, and radiation pattern are presented.
The proposed antenna is designed, optimized, and simulated using CST studio
suite.
Figures 3 and 4 show the simulated reflection coefficient of the proposed
reconfigurable antenna.
There are two operating frequencies that can be varied by varying the diode PIN
states (which is inserted in the antenna): OFF and ON states.
The resonant frequencies are:
Fig. 3 Reflection coefficient of the proposed antenna when the PIN diode is in OFF state
Fig. 4 Reflection coefficient of the proposed antenna when the PIN diode is in ON state
• The first resonant frequency f 1 = 26.15 GHz with the reflection coefficient of −
21.29 dB, the value of the bandwidth at −10 dB is 24.91–27.62 GHz, if the diode
PIN is OFF.
• The second resonant frequency f 2 = 46.1 GHz with the reflection coefficient of −
22.21 dB, the value of the bandwidth at −10 dB is 43.22–47.34 GHz,if the diode
PIN is ON.
Figures 5 and 6 represent the surface current distribution of the proposed
reconfigurable antenna for the resonant frequencies.
In the first configuration, the PIN diode is ON, the two radiators are connected as
seen in Fig. 6, the strong current distribution appears from the feed position to the
Fig. 5 Surface current distribution of the proposed antenna if the PIN diode in OFF state
Fig. 6 Surface current distribution of the proposed antenna if the PIN diode in ON state
left side of the main radiator, and it passes from the diode to the second radiator on
the left.
In the second configuration, the proposed antenna operates with only the main
radiator when the PIN diode is OFF. As shown in Fig. 5, the strong current distributes
from the feed position to the top sides of the main triangle radiator.
Figures 7 and 8 show the simulated radiation patterns of the proposed antenna
with different switching states of the PIN diode, and it is plotted at 26.15 GHz and
46.1 GHz in 3D view. As observed, the antenna presents a good radiation performance
with a max gain value of 5.34 dB for the ON state of the PIN diode and 6.08 dB for
the OFF state.
Fig. 7 Radiation pattern of the proposed antenna if the PIN diode in OFF state
Fig. 8 Radiation pattern of the proposed antenna if the PIN diode in ON state
Fig. 9 Reflection coefficient of the proposed antenna simulated with CST and HFSS
4 Validation of the Results
For checking the previous results obtained by the CST microwave studio, we use
another simulator program named Ansys HFSS software. The reflection coefficient
of the proposed antenna for ON and OFF states of the PIN diode connected in the
antenna is shown in Fig. 9. Firstly for the simulation with HFSS software, if the PIN
diode in ON state, we obtained three resonant frequencies between them we have
the principal frequency at 46.6 GHz.
Secondly, if the PIN diode in OFF state, we obtained two resonant frequencies so
that the important frequency to us is 26.4 GHz.
So the resonant frequencies obtained by HFSS are shifted a little compared to those
obtained by CST. Also, we notice that some resonant frequencies become more or
Less significative. But the resonant frequencies still in the band of 24.25–27.5 GHz
if the PIN diode is on OFF state and 45.5–50.2 GHz if the PIN diode is on ON state.
These changes because the simulators are not same, due to different computational
techniques involved, HFSS is based on finite element method (FEM) which is more
accurate for designing antennas, while CST is based upon finite integration technique
(FIT) [17] and is also popular among antenna designers due to ease in simulations.
5 Proposed Millimeter Wave Antenna Array for 5G
The proposed array contains eight reconfigurable antenna placed on the top of a
mobile phone PCB like it shows Fig. 10 [18–20]. The overall size of the mobile
phone PCB is 60 × 100 mm2 . Simulations have been done using CST software to
Fig. 10 Configuration of the proposed MIMO Antenna for 5G: a back view, b front view, and
c zoom view of the antenna array
validate the feasibility of the proposed frequency reconfigurable array antenna for
Millimeter Wave 5G handset applications [21, 22].
It can be seen that the proposed 5G array is compact in size with dimensions L a ×
W a = 25 × 3.2 mm2 (Fig. 1c). Furthermore, there is enough space in the proposed
mobile phone antenna to include 3G and 4G MIMO antennas [23, 24]. The antenna
is designed on a Rogers RT5880 substrate with thickness h and relative permittivity
2.2.
Figures 11 and 12 show the S-parameters (S1,1–S8,1) of the array for the two
conditions of the PIN diodes (ON/OFF conditions). As illustrated, the mutual
coupling between antennas elements of array is good. Furthermore, it can be seen
that the array has good impedance adaptation at 26.15 GHz (all diodes in OFF state)
and at 46.1 GHz (all diodes in ON state).
The 3D radiation paterns of the proposed antenna at 26.15 and 46.1 GHz are
illustrated in Figs. 13 and 14 showing that the proposed reconfigurable antenna array
a good beam steering property with the max gain value of 6.39 and 6.3, respectively.
6 Conclusion
In this paper, a new frequency reconfigurable patch antenna is designed, optimized,

and simulated with CST studio suite. The proposed antenna can reconfigure on two
different frequency bands of the fifth generation using PIN diodes, with the reflection
Fig. 11 Simulated S-parameters of the proposed 5G mobile phone antenna, if all diodes are in OFF
state
Fig. 12 Simulated S-Parameters of the proposed 5G mobile phone antenna, if all diodes are in ON
state
coefficient less than −10 dB those are 24.91–27.62 GHz and 43.22–47.34 GHz. The
overall structure size of the designed antenna 6 × 5.5 × 1.02 mm3 .
This antenna is useful for 5G applications [25].
Fig. 13 Simulated of the proposed 5G mobile phone antenna, if all diodes are in OFF state
Fig. 14 Simulated of the proposed 5G mobile phone antenna, if all diodes are in ON state
References
1. Lim, E.H., Leung, K.: Reconfigurable Antennas. In: Compact multifunctional antennas for
wireless systems, pp. 85–116. Wiley (2012). https://doi.org/10.1002/9781118243244.ch3
2. Loizeau, S.: Conception et optimisation d’antennes reconfigurables multifonctionnelles et ultra

large bande. Ph.D. Dissertation (2009)
3. Guo, Y.J., Qin, P.Y.: Reconfigurable antennas for wireless communications. In: Chen, Z.,
Liu, D., Nakano, H., Qing, X., Zwick, T. (eds) Handbook of antenna technologies. Springer,
Singapore (2016). https://doi.org/10.1007/978-981-4560-44-3_119
4. Bernhard, J.T.: Reconfigurable antennas. Morgan & Claypool Publishers (2007). ISNB
1598290266, 9781598290264
5. Ismail, M.F., Rahim, M.K.A., Zubir, F., Ayop, O.: Log-periodic patch antenna with tunable
frequency (2011)
6. Hsu, Shih-Hsun, Chang, Kai: A novel reconfigurable microstrip antenna with switchable
circular polarization. IEEE Antennas Wirel. Propag. Lett. 6(2007), 160–162 (2007)
7. Dandekar, K.R., Daryoush, A.S., Piazza, D., Patron, D.: Design and harmonic balance analysis
of a wideband planar antenna having reconfigurable omnidirectional and directional patterns.
5 (2013)
8. Nikolaou, S., Bairavasubramanian, R., Lugo, C., Car-rasquillo, I., Thompson, D.C., Ponchak,
G.E., Papapolymerou, J., Tentzeris, M.: Pattern and frequency reconfigurable annular slot
antenna using PIN diodes. IEEE Trans. Antennas Propag. 54,(2), 439–448 (2006)
9. Sarah El Kadri. 2011. Contribution à l’étude d’antennes miniatures reconfigurables en
fréquence par association d’éléments actifs. Ph.D. Dissertation
10. Kumar, D., Siddiqui, A.S., Singh, H.P., Tripathy, M.R., Sharma, A.: A Review: Techniques and
Methodologies Adopted for Reconfigurable Antennas. In: 2018 International Conference on
Sustainable Energy, Electronics, and Computing Systems (SEEMS). 1–6 (2018). https://doi.
org/10.1109/SEEMS.2018.8687361
11. Salleh, S.M., Jusoh, M., Seng, L.Y., Husna, C.: A review of reconfigurable frequency switching
technique on micostrip antenna. J. Phys.: Conf. Ser. 1019, 012042 (2018). https://doi.org/10.
1088/1742-6596/1019/1/012042
12. Fang, D.G.: Antenna theory and microstrip antennas. CRC Press, Taylor & Francis Group,
New York (2015)
13. Zhang, Z.: Antenna design for mobile devices. Wiley (2011). Print ISBN:9780470824467,
Online ISBN:9780470824481, https://doi.org/10.1002/9780470824481
14. Ismail, M.F., Rahim, M.K.A., Majid, H.A.: The Investigation of PIN diode switch on reconfig-
urable antenna. In: 2011 IEEE International RF & Microwave Conference. IEEE, pp. 234–237
(2011)
15. Lim, J., Back, G., Ko, Y., Song, C., Yun, T.: A reconfigurable PIFA using a switchable pin-diode
and a fine-tuning varactor for USPCS/WCDMA/m-WiMAX/WLAN. IEEE Trans. Antennas
Propag. 58(7), 2404–2411 (2020). https://doi.org/10.1109/TAP.2010.2048849
16. Balanis, C.A.: Modern antenna handbook. Wiley, Hoboken (2008)
17. Balanis, C.A.: Antenna Theory: Analysis and Design, 3rd ed. John Wiley, Hoboken, NJ (2005)
MLA (8th ed.)
18. Sanayei, S., Nosratinia, A.: Antenna selection in MIMO systems. IEEE Commun. Mag. 42(10),
68–73 (2004). https://doi.org/10.1109/MCOM.2004.1341263
19. Li, Y., Sim, C., Luo, Y., Yang, G.: 12-Port 5G Massive MIMO Antenna Array in Sub-6 GHz
Mobile Handset for LTE Bands 42/43/46 Applications. IEEE Acc. 6, 344–354 (2018). https://
doi.org/10.1109/ACCESS.2017.2763161
20. Sanayei S., Nosratinia, A.: University of Texas at Dallas, Antenna Selection in MIMO Systems,
IEEE Communications Magazine, October (2004)
21. Rappaport, T.S., Sun, S., Mayzus, R., Zhao, H., Azar, Y., Wang, K., Wong, G. N., Schulz, J.
K., Samimi, M., Gutierrez, F.: Millimeter wave mobile communications for 5G cellular: It will
work! IEEE Access 1, 335–349. [6515173] (2013). https://doi.org/10.1109/ACCESS.2013.226
0813
22. Liu, D., Hong, W., Rappaport, T.S., Luxey, C., Hong, W.: What will 5G Antennas and Propa-
gation Be? IEEE Trans. Antennas Propag. 65(12), 6205–6212 (2017). https://doi.org/10.1109/
TAP.2017.2774707
23. Sharawi, M.S.: Printed MIMO antenna engineering. Electronic version: https://books.google.
co.ma/books?id=7INTBAAAQBAJ&lpg=PR1&ots=aHyFM1I5Wi&dq=mimo%20antenna&
lr&hl=fr&pg=PR7#v=onepage&q=mimo%20antenna&f=false
24. Li, Y., Desmond Sim, C.-Y., Luo, Y., Yang, G.: 12-Port 5G Massive MIMO Antenna Array
in Sub-6 GHz Mobile Handset for LTE Bands 42/43/46 Applications, 2169–3536 (c) (2017)
IEEE, https://doi.org/10.1109/access.2017.2763161
25. Hong, W., Baek, K.-H., Lee, Y., Kim, Y., Ko, S.-T.: Study and prototyping of practically large-
scale mm wave antenna systems for 5G cellular devices, IEEE Communications Magazine,
September (2014)
Smart Security
A Real-Time Smart Agent for Network
Traffic Profiling and Intrusion Detection
Based on Combined Machine Learning
Algorithms
Nadiya El Kamel , Mohamed Eddabbah, Youssef Lmoumen,

and Raja Touahni
Abstract Cyber-intrusions are constantly growing due to the ineffectiveness of the

traditional cyber security tools and filtering systems-based attacks detection. In the
last decade, significant techniques of machine and deep learning were employed to
resolve the cyber security issues. Unfortunately, the results are still imprecise with a
lot of shortcomings. In this paper, we present a real-time cyber security agent based
on honeypots technology for real-time data collection and a combination of machine
learning algorithms for data modeling that enhances modeling accuracy.
1 Introduction
On the Internet, each connected node represents a target for a black hat, while
the reasons behind cyber-attacks are various [1–4]. By exploiting vulnerabilities,
intruders gain access into private networks, they may spy, steal, or sabotage the data,
etc. Hence, in order to protect their sensitive data, companies deploy more and more
security solutions. On the other hand, attackers develop their tools too, by adopting
new techniques to avoid detection and filtering systems. In December 2020s, many
leader companies compromised over the SUNBURST hack campaign including even
security tools providers [5]. Intruders exploited the Solarwinds Orion platform update
N. El Kamel (B) · R. Touahni

Laboratory of Electronic Systems, Information Processing, Mechanics and Energetics, Faculty of
Sciences, Ibn Tofail University, Kenitra, Morocco
e-mail: nadiya.elkamel@uit.ac.ma
R. Touahni
e-mail: touahni.raja@uit.ac.ma
M. Eddabbah
LABTIC Laboratory ENSA, Abdelmalek Essaadi University Tangier, Tangier, Morocco
Y. Lmoumen
CIAD UMR 7533, Univ. Bourgogne Franche-Comté, UTBM, 90010 Belfort, France
https://doi.org/10.1007/978-981-16-3637-0_21
302 N. El Kamel et al.
file to add vulnerabilities and backdoors. They had employed a combination of tech-
niques to infiltrate into 17 K of the Solarwinds customers’ networks [6]. In order to
detect such sophisticated attacks, we design a smart agent-based attacks detection
using a combination of machine learning algorithms for flow modeling and honeypot
techniques for an updatable database conception. The honeypot is a security resource
implemented for being probed, attacked, or compromised [7–9]. It was proposed to
automatically consider any interaction detected as a malicious activity. The gener-
ated log files data will be aggregated and modeled using a combination of machine
learning classifiers, to enhance precision and future attacks detection. Next sections
are devoted firstly to discussing some related works and secondly to explain the smart
agent functions, advantages, and use cases.
2 Related Works
Many cyber security solutions proposed in the last decade, but results still present
some limitations and shortcomings [10] while all Internet-providers seek to protect
themselves against fraudulent use of their data, stealing, sabotage and against all
malicious activities on computer systems. The most recent works in cyber security
focus on machine and deep learning algorithms for attacks data modeling [11, 12].
Pa, Y et al. [13] suggest a honeypot-based approach, for malware detection based
on corresponding signatures generated from a honeypot system. This method is still
limited and unable to detect new signatures or new kinds of malwares. Moreover, a
machine learning-based solution represents a promising candidate to deal with such
a problem, due to its ability to learn and teach over time.
R. Vishwakarma et al. [14] present an IoT combating method against DDoS
attacks, based on IoT honeypot-generated data for dynamic training of a machine
learning model. The proposed method allows detecting zero-day DDoS attacks,
which has emerged as an open challenge in defending IoT against DDoS attacks.
P. Owezarski et al. [10] studied an unsupervised anomalies learning-based attacks
characterization, using honeypot system for data construction. This study is based on
clustering techniques as subspace clustering, density-based clustering, subspace clus-
tering, and evidence accumulation for classifying flow ensembles in traffic classes.
The proposed study in this work does not require a training phase.
K. Lee et al. [15] used machine learning algorithms (SVM) to automatically
classify social spam for network communities such as Facebook and MySpace based
on a social honeypot information collection.
T. Chou et al. [16] suggest a three-layer hierarchy structure-based intrusion detec-
tions, consisting of an ensemble of classifiers groups, each consisting of a combina-
tion feature selecting classifiers. They applied different machine learning algorithms
and feature subset to solve uncertainty problems and maximize the diversity. In
terms of detection rate (DR), false-positive rate (FPR) and classification rate (CR),
the results demonstrate that the hierarchy structure performs better than a single
classifier-based intrusion detection.
A Real-Time Smart Agent for Network Traffic Profiling … 303
G. Feng et al. suggest in [17] a linkage defense system for improving private
network security by linking honeypots with the network security. The defense
network centroid honeypot treats suspicious flows arriving from the traditional tools,
while blocking network access depends on the honeypot state; if the honeypot is
damaged, then the correspondent intruder will be blocked by the firewall.
3 A Real-Time Security Agent Based on Machine Learning

and Honeypot Technologies
Manual cyber security tools reconfiguration represents some shortcomings in terms

of time and money. While it is an illusion to think that a lock and a key represent
perfect security defense, intruders develop their strategies constantly to add back-
doors, to break the peer username/password, to hide their command traffic, and even
to hide data in media (Steganography technique). IDS, Firewalls, and IPS protect
systems from traditional hack tools and tactics but are still ineffective to detect
hidden command and data traffic. There is no contact between them to block intru-
sion detected in an IDS by firewalls [17], and they represent a passive solution when
it is about zero-day and future attacks [18]. The real cyber security challenge is
accepting the probability of an imminent attack and understanding what is really
going on within information systems.
The main objective of this work is to design a real-time cyber security agent, based
on machine learning algorithms combination and honeypots technology. Machine
learning algorithms-based intrusion modeling allows automatic attacks detection,
while the honeypot system is a deceptive technology that allows intruders control of
fack machines [19], traffic capturing, data collection and ensure the logging of newly
coming malware features [20, 21].
Based on these technologies, the smart agent takes functions of packet inter-
ception, information collection, suspicious profile creation (database construction),
comparison of suspicious profiles with attacker’s database profiles for making
decisions, database update, and firewall alarm if an attacker profile is detected (Fig. 1).
3.1 Profiles Creation (Phase 1)
Information security policy focuses on mechanisms that insure data integrity, avail-
ability, and confidentiality, which consists of traffic monitoring and filtering. While
unknown profiles detection’s time is the most critical point for intrusion detection
systems. For this reason, we develop a smart agent for detecting attacks and for
constructing a shared and an updatable database for protecting Internet content-
providers from current and future attacks. In the first stage, the agent takes func-
tions of packet interception and hackers profile creation based on the transport and
Fig. 1 Security agent-based

attacks detection
application information gathered within a honeypot that emulates fake services, and
machine learning for data modeling using an hierarchical structure of algorithms that
maximize the detection precision.
The originality of the honeypot lies in the fact that the system is voluntarily
presented as a weak source able to hold the attention of attackers [8]. The general
purpose of honeypots is to make the intruder believe that he can take control of a real
production machine, which will allow the smart agent to model the compromising
data gathered as a profile, and making decision to send alarms when it is about
an intruder profile. The honeypot classification depends on its interactions level.
Low-interaction honeypot offered a limited services emulation set; for example,
it cannot emulate an file transfer protocol (FTP) service on port 21, but emulate
only the login command or just one other command, and it records a limited set
of information, monitor only known activities. The advantage of low-interaction
honeypots lies in their simplicity of implementing and managing and pose little
risk since the attacker is limited. Medium-interaction honeypots give little more
access than low-interaction honeypots [22] and offer better emulation services. So
a medium-interaction honeypots enable logging more advanced attacks, but they
require more time to implement and a certain level of expertise. For high-interaction
honeypots, we provide the attacker with real operating systems and real applications
[23]. This type of honeypots allows gathering a lot of information about the intruder
as it interacts with a real system, examining all behaviors and techniques employed,
and checking if it is about a new attack.
In this work, we employ a high-level interaction honeypot for extracting a huge
amount of information about intruders, and we exploit it again after the decision
phase, and if an attacker is detected, we configure the firewall in the way that it
redirect him into the honeypot system one more time, for limiting his capacity of
developing tools and tactics. In a network company, the productive network consists
of, for example, HTTP, database, monitoring, and management servers. A network of
honeypot must be deployed and configured to run the fake services of the productive
Fig. 2 Hacker profile creation architecture
network (S1, S2, S3, etc.), which suspicious flows are redirected to by the network
firewall system (Fig. 2).
At the profile creation phase (Fig. 2), suspicious flows will be redirected to a
network of honeypots servers, allowing intruders the control of fack servers [24];
the transport and application layer collected information will be aggregated into
vectors Vuser . Qualitative data such as IP address is stocked directly in the profile,
while quantitative data is classified into homogeneous classes (time inter packet,
number of packets per flow… etc.) using an hierarchical structure of machine learning
algorithms combine classification algorithm and linear regression (Algorithm 1).
Machine learning techniques combination consists of mixing classification and
regression algorithms, whose purpose is to maximize the precision level of the fitting
models.
In this paper, we propose a combination composed of two-layers hierarchy
structure, consisting of a classification algorithm, that divides flows quantitative
data into homogenous subsets at the first layer, and linear regression for each
subset modeling and classifying at the second layer (Fig. 3). The advantages of
the hierarchical classification method lie in increasing modeling precision, reliable
detection, and false alarms avoiding. For the smart agent development, this technique
represents a keystone for suspicious and attacker profiles creation and update.
Fig. 3 Flow modeling process
Algorithm 1: Learning
INPUT
K //number of clusters
j j j
Vi j = (V1 , V2 , …, Vi ) //Hacker j data array (vector of vectors)
START
Akl = Q1A (V ij ) //qualitative adaptation
Ak’l’ = Q2A (V ij ) //quantitative adaptation
f = c-1//linear regression order = space dimension -1
for j=1; i<c; j++ //For each row of Ak’l’
(Cj [K], Rj [K]) = K-means (Ail’ , K) //creation of clusters center et radius
for j=1; i<c; j++
C L j [ f ] = Linear Regression (C j [K], R j [K], Ail’ //Regression of every clusters
OUTPUT//Hacker profile
Akl //qualitative adaptation
CLc[f] //linear regression coefficients
Fig. 4 Decision stage
At the profile creation phase, the initialization of these three parameters is crucial:
the number of clusters, centroids initialization, and the linear function weights. The
higher initialization precision of these parameters increases the precision of the fitting
model.
3.2 Attacks Detection (Phase 2)
In the decision phase (Algorithm 2), the smart agent functions are packet interception,
information collection, suspicious profiles creation, comparison of suspicious
profiles with database profiles based on the distance metric (Fig. 4) [20], and
administrator alarm if an attacker profile is detected. Hence, it represents a traffic
cop that detects malicious profiles and alarms firewall systems to cut the route. The
advantages of this contribution lie in the automatic detection of malicious profiles,
construction of new dataset that reflect current network situations and the latest attack
trends, and limiting the capacity of developing intruder’s tools by redirecting him
into the honeypot system one more time. Hence, by making the intruder interact
again with the honeypot, the profiles will be enriched with the gathered information
and then reinforcing the learning.
Algorithm 2: Decision
INPUT
K //number of clusters
j j j
V = (V1 , V2 , …, Vi ) //user j information array (vector of vectors)
(HAkl , HCLc[f]) //Hackers profiles
START
UAkl = Q1A (V ij ) //qualitative adaptation
Ak’l’ = Q2A (V ij ) //quantitative adaptation
f = c−1//linear regression order = space dimension -1
for j=1; i<c; j++ //For each row of Ak’l’
(Cj [K], Rj [K]) = K-means (Ail’ , K) //creation of clusters center et radius
for j=1; i<c; j++
U C L j [ f ] = Linear Regression (C j [K], R j [K], Ail’ ) //Linear Regression of every clusters
OUTPUT
Distance(HCLc[f], UCLj[f])
Isequal(HAkl , UAkl )
In the Euclidean sense, H C L c [ f ] is the closest vector to H C L j [ f ].
4 Conclusion
In this paper, we have presented a smart cyber security agent based on machine
learning algorithms combination for automatic decision instead of manual treatment
and the deceptive honeypot for data collection. The proposed agent is not just an auto-
matic security tool for new and zero-day attacks detection, but it allows constructing
a new updatable database that reflects current network situations, latest attack trends,
and future attack techniques. Next works will be devoted to implementing the smart
agent in a real environment (cloud infrastructure) in order to test its performances
and compare it with other cyber security solutions.
References
1. Matin, I.M.M., Rahardjo, B.: The Use of Honeypot in Machine Learning Based on Malware
Detection: A Review. 1–6 (2020)
2. Matin, I.M.M., Rahardjo, B.: Malware detection using honeypot and machine learning. 7th
International Conference on Cyber and IT Service Management 7, 1–4 (2020)
3. Singh, J., Singh, J.: A survey on machine learning-based malware detection in executable files.
Journal of Systems Architecture 101861 (2020)
4. Iwendi, C., Jalil, Z., Javed, A.R., Reddy, T., Kaluri, R., Srivastava, G., Jo, O.: Keysplitwater-
mark: Zero watermarking algorithm for software protection against cyber-attacks. IEEE Access
8, 72650–72660 (2020)
5. Bowman, J.: How the United States is Losing the Fight to Secure Cyberspace. (2021)
6. Oxford, A.: SolarWinds hack will alter US cyber strategy. Emerald Expert Briefings (2021)
7. Jiang, K., Zheng, H.: Design and Implementation of A Machine Learning Enhanced Web
Honeypot System. pp. 957–961. IEEE, (Year)
8. Spitzner, L.: Honeypots: tracking hackers. Addison-Wesley Reading (2003)
9. Karthikeyan, R., Geetha, D.T., Vijayalakshmi, S., Sumitha, R.J.I.j.f.R., Technology, D.i.:
Honeypots for network security. 7, 62–66 (2017)
10. Owezarski, P.: Unsupervised classification and characterization of honeypot attacks. pp. 10–18.
IEEE, (2014)
11. Berman, D.S., Buczak, A.L., Chavis, J.S., Corbett, C.L.: A survey of deep learning methods
for cyber security. Information 10, 122 (2019)
12. Liu, H., Lang, B.: Machine learning and deep learning methods for intrusion detection systems:
A survey. applied sciences 9, 4396 (2019)
13. Pa, Y.M.P., Suzuki, S., Yoshioka, K., Matsumoto, T., Kasama, T., Rossow, C.: IoTPOT:
Analysing the rise of IoT compromises. (2015)
14. Vishwakarma, R., Jain, A.K.: A honeypot with machine learning based detection framework
for defending IoT based botnet DDoS attacks. pp. 1019–1024. IEEE, (Year)
15. Lee, K., Caverlee, J., Webb, S.: Uncovering social spammers: social honeypots + machine
learning. pp. 435–442. (2010)
16. Chou, T.-S., Fan, J., Fan, S., Makki, K.: Ensemble of machine learning algorithms for intrusion
detection. pp. 3976–3980. IEEE, (2019)
17. Feng, G., Zhang, C., Zhang, Q.: A design of linkage security defense system based on honeypot.
pp. 70–77. Springer, (2013)
18. Matin, I.M.M., Rahardjo, B.: Malware detection using honeypot and machine learning. pp. 1–4.
IEEE, (2020)
19. Seungjin, L., Abdullah, A., Jhanjhi, N.Z.: A Review on Honeypot-based Botnet Detec-
tion Models for Smart Factory. International Journal of Advanced Computer Science and
Applications 11, (2020)
20. El Kamel, N., Eddabbah, M., Lmoumen, Y., Touahni, R.: A Smart Agent Design for Cyber
Security Based on Honeypot and Machine Learning. Security and Communication Networks
2020, (2020)
21. Ng, C.K., Pan, L., Xiang, Y.: Honeypot frameworks and their applications: a new framework.
Springer (2018)
22. Negi, P.S., Garg, A., Lal, R.: Intrusion detection and prevention using honeypot network for
cloud security. pp. 129–132. IEEE, (2020)
23. Wang, H., Wu, B.: SDN-based hybrid honeypot for attack capture. pp. 1602–1606. IEEE,
(2019)
24. Naik, N., Jenkins, P., Savage, N., Yang, L.: A computational intelligence enabled honeypot for
chasing ghosts in the wires. Complex & Intelligent Systems 1–18 (2020)
Privacy Threat Modeling in Personalized
Search Systems
Anas El-Ansari, Marouane Birjali, Mustapha Hankar,

and Abderrahim Beni-Hssane
Abstract Personalized search systems simplify information access on the Web by

providing adapted results for each user’s query according to his preferences and
needs. Such systems depend mainly on collecting user data to create profiles that
represent user interests. While collecting user data is essential for improving the
quality of the returned results, it has also raised serious concerns regarding user
privacy. Giving the user a personalized browsing experience usually comes at the
cost of his privacy. Thus, most people are afraid of using such applications. Current
search engines lack dedicated privacy-preserving features and do not fulfill people’s
expectations in terms of privacy. Researchers have been investigating solutions to
overcome this issue and create privacy-aware personalized systems. In this paper,
we present an overview of the current privacy threats in personalized search systems
with a threat modeling approach. Furthermore, we examine techniques and solutions
for user data security and privacy protection.
1 Introduction
The information amount available on the Web grows continuously preventing people
from easily obtaining desired items or information [1]. This dilemma highlights a
pressing demand for effective personalized search systems capable of simplifying
information access and item discovery considering the user’s interests and prefer-
ences. Most studies usually focus on improving personalization quality and disre-
garding user privacy issues. One of the challenges facing personalized systems is
the privacy protection problem [2] since giving the user a personalized browsing
experience comes at the cost of his privacy. Thus, people are afraid of using such
applications.
As the current search engines lack dedicated privacy-preserving features and
do not fulfill people’s expectations in terms of privacy, alternative search engines
A. El-Ansari (B) · M. Birjali · M. Hankar · A. Beni-Hssane

LAROSERI Laboratory, Computer Science Department, Sciences Faculty, Chouaib Doukkali
University, El-Jadida, Morocco
https://doi.org/10.1007/978-981-16-3637-0_22
312 A. El-Ansari et al.
have emerged: metasearch engines (e.g., DuckDuckGo [3]) and search engines (e.g.,
Qwant [4, 5]). The former enhances existing search engines by focusing on the
privacy protection of their users, while the latter develops a search engine that does
not exploit users’ information. Nevertheless, these alternatives do not implement
any specific privacy-preserving mechanisms. Instead, they claim, in their terms of
service, that they do not collect any personal information of their users. For instance,
DuckDuckGo affirms that they also store searches, but not in an identifiable form,
as they do not collect IP addresses or any identifiable user information.
The lack of protection for such data raises privacy issues. For instance, the AOL
query logs scandal [6] when AOL research released a file on its Web site with over
twenty million search queries for about 650,000 users. New York Times identified a
user from the published file by cross-referencing it with some phone-book listings.
AOL admitted it was an error and removed the file, yet others redistributed the file
on mirror sites. This example not only raises panic among users but also dampens
the data publishers’ enthusiasm in offering improved personalized services.
Besides, as these systems’ implementations are not publicly available and as they
do not explicitly provide the data they log, users cannot be confident in the privacy
protection obtained by these solutions. Users can only trust these services and hope
that their data are in a safe and privacy-preserving storage space. Researchers have
been investigating solutions to overcome this issue and create search engines that
ensure a privacy protection by design as in [7].
For organizations that collect or manage user data, security and privacy should
be mandatory as it is for individuals who own it. It is the primary concern when
undertaking the process of protecting fundamentally sensitive information such as
identities, finances, and health records.
Without privacy and security measures, other malicious third parties would gain
access to large amounts of possibly damaging data [8]. However, the distinction
between user privacy and user data security is not clear for everyone, and the terms
are often misplaced or confused as the same thing.
Data security and user privacy are two fundamental components for a successful
data protection strategy [9], so safeguarding data is not limited to one single concept
of the two.
The difference between both topics is not in the execution, implementation, or
results but the philosophy supporting them. Specifically, it is a matter of which data
require protection, what protection mechanism is employed, from whom, and who
is responsible for this protection.
Data security aims at preventing unapproved access to user data via leaks or
breaches, despite who the unapproved party is. To achieve this, companies use tech-
nologies and tools such as user authentication, firewalls, network limitations, and
even internal security measures to prevent such access. Also, security technologies
like encryption and tokenization further protect user data by making it unreadable
at the moment a breach occurs and can stop attackers from potentially revealing
massive amounts of the user’s sensitive data.
Privacy, however, focuses on ensuring that the sensitive data an organization
stores, processes, or transmits are ingested compliantly and with consent from the
Privacy Threat Modeling in Personalized Search Systems 313
owner of that sensitive data. It means informing users upfront of which types of data
the system collects, for what purpose, and who has access to it. The individual must
then agree to the terms of use, allowing the organization that ingests data to use it in
line with its stated purposes.
So, privacy is more about using data responsibly and following the wishes of users
to prevent using it by unauthorized parties. However, it can also include security-type
measures to ensure privacy protection. For instance, efforts to prevent the linking of
sensitive data to its data subject or natural person such as de-identifying personal
data, obfuscating it, or storing it in different places to reduce the possibility of re-
identification are other privacy provisions.
A personalized system can apply security controls without satisfying privacy
requirements, yet privacy issues are hard to control without employing efficient
security practices.
In this work, we study the privacy threat in personalized search systems, and we
propose a threat model for privacy-preserving personalized search systems. More-
over, we discuss privacy protection and cryptographic solutions that can be used in
personalized search systems.
The paper is organized as follows. The next section addresses the privacy threat
modeling in personalized search. In Sect. 3, we present privacy protection solutions.
While the fourth section addresses the cryptographic solutions. The paper concludes
and points to future work ideas in the conclusion section.
2 Privacy in Personalized Search
To protect privacy in personalized search systems, we must consider two contra-

dicting effects. First, such systems improve search quality by collecting more user
data. Second, it must hide the sensitive data collected and stored in a user profile to
place the privacy risk under control.
2.1 Personalized Search Systems
We can classify the personalized search systems into three distinct structures, based
on the storage place of the user profile (server side or client side) and its use for
personalization [10].
Server-side personalization: The system stores user profiles on the server side.
This structure requires the user to have an identifying account. The server then creates
and updates the profile either explicitly from the user’s input (requesting the user to
list his interests) or implicitly by collecting the user’s browsing history (e.g., query
and click-through history). The latter method needs no additional work from users
and contains a better description of his interests. Some search engines, like Google
Personalized, adopted this architecture. Most systems with such a structure ask users
to provide consent before collecting and using their data. If the user accords his
permission, the search system will hold all the personally identifiable data possibly
available on the server. Thus, from the user’s perspective, this architecture provides
a low level of privacy protection.
Client-side personalization: Storing the profile on the user device, the client
sends queries to a search engine and receives results, as in an ordinary Web search
scenario. The client agent also performs a query expansion to generate a new person-
alized query before sending it to the search engine. Furthermore, as in [11], the client
agent ranks the search results to match user preferences.
Client–Server collaborative personalization: This structure is a balance of the
previous two. The profile is still on the client side, but the server also participates in
search personalization. At query time, the client agent extracts a sub-profile from the
user profile to send it to the search engine along with the query. The search engine
then uses the received context to personalize the results.
2.2 Privacy Threats
Personalized search systems pose several risks to users’ privacy. This section
discusses potential privacy threats based on the location of the breach on the user’s
sensitive data.
Threats on the client side: Exploiting a user device to store the collected sensitive
data (user profile) by personalized search systems introduces risks. Some of these
threats may lead to critical problems for both users and service providers.
Cross-site scripting (XSS) is among the prevalent vulnerabilities in recent Web
applications. An attacker can execute scripts within the context of the Web site under
attack. Different types of XSS exist with the same result allowing for the execution of
malicious codes in the browser of the user, allowing the attacker to access sensitive
user data.
Client-side SQL injection (csSQLi) is a new form of well-known SQL injection
attacks that have emerged recently due to the introduction of database support on
clients (Google Gears and HTML 5 SQL databases). A popular mechanism used
in conjunction with SQL injections is a mechanism called stacked queries (SQ),
allowing an attacker to execute his query irrespective of the original one. The attacker
adds the SQ to an original one through the use of a semicolon. SQL injection with
SQ can be a powerful combination, as it allows executing arbitrary SQL commands
on the database, especially on old browsers’ versions.
Client-side data corruption or leakage is when the user, or an attacker controlling
his device, changes or corrupts the stored data or retrieves sensitive information.
Numerous attacks can result in a data leakage/corruption, including malware, XSS,
csSQLi, and threats exploiting vulnerabilities in the user’s browser or device. To
ensure data security and lower the risk of exploiting client-side data storage vulner-
abilities, both service providers and users need to implement preventive measures
(e.g., encryption, digital signatures, and access control mechanisms). Furthermore,
output encoding mechanisms and parameterized queries prevent XSS and csSQLi,
respectively.
Threats on the server side: Reports of privacy breaches on the server side
affecting personalized systems dominate the news with increasing frequency. Most
personalization systems store the collected user data on their servers (the first target
for attackers). We can classify server-side privacy threats into two categories: insider
threats and outsider threats.
Insider privacy threats come from inside the service-providing organization (mali-
cious or negligent insiders, infiltrators, or even the organization’s intentions). This
category involves data leakage (as in the AOL scandal), data misuse (as in the
Facebook–Cambridge Analytica scandal) when using data for other purposes. Data
brokerage is also an insider threat (when aggregating data and reselling the valuable
categories of customers to third parties). The theft of private or commercially relevant
information can come from inside the organization.
Outsider privacy threats are ever-present and pose a real danger. Any system
connected to the Internet is at risk. Historically, the most famous data breaches were
typically outsider type. In 2016, the email giant and search engine Yahoo had their
systems compromised in a data breach, causing an information loss of more than
500 million users [12]. eBay also reported that an attacker exposed its entire list of
145 million clients’ accounts in May 2014 [13]. Many more companies on the web
have suffered from data breaches (Adobe, Canva, LinkedIn, Zynga, etc.). While these
breaches cost millions of dollars, outsider threats are usually the ones targeted with
the traditional security measures (firewall, passwords, encryption, etc.) to prevent
potential attacks (malware attacks, phishing, XSS, SQL injections, password attacks,
etc.). However, securing the server is a challenging and complicated task that is hard
to accomplish.
Failure to implement efficient security controls like patches/updates or secure
configurations, replacing default accounts, or disabling unnecessary back-end
services can compromise data confidentiality and integrity. Moreover, introducing
an additional measure to enhance security may increase vulnerability and expose
the system to further threats. The answer to this problem is to understand system
vulnerabilities and implement a risk-mitigation approach taking into consideration
insider and outsider threats.
Communication channel threats: Internet serves as an electronic chain
connecting a client to a server. Messages on this network travel an arbitrary path
from the client device to a destination point. The message passes through a number
of intermediate nodes on the network before arriving at the final destination. It is
difficult to ensure that each computer on the Internet, through which messages pass,
is secure and non-hostile.
Web applications often use the HTTP protocol for client–server communication,
which communicates all information in plain text. Even when they provide transport-
layer security through the use of the HTTPS protocol, if they ignore certificate vali-
dation errors or revert to plain text communication after a failure, they can jeopardize
security by revealing data or facilitating data tampering. Compression side-channel
attacks such as CRIME and BREACH presented concrete and real-world examples
of HTTPS vulnerabilities in 2012 and 2013. Since then, security experts confront
new attacks on TLS/SSL every year, especially with servers using version 1.2 of
the TLS communication protocol, which supports encryption and compression. The
combination of encryption and compression algorithms presented security flaws that
allowed the attackers to open the content of the encrypted HTTP header and use the
authentication token within the cookie to impersonate a user.
These attacks, among others, can lead to a privacy threat called eavesdropping
(a sniffing or man-in-the-middle attack), which consists of collecting information
as it is sent over a channel by a computer or other connected devices. The attacker
takes advantage of unprotected network communications to obtain information in-
transit or once received by the user. Moreover, the advancement toward the future 5G
networks is rapid and expected to offer high data speed, which will eventually result
in increasing data flows in the communication channels between users and servers.
This fact will raise the user’s concerns about privacy protection in these networks.
2.3 Privacy Threat Model
Since implementing effective privacy-preserving systems requires understanding the

range of potential privacy threats in this field, it is important to study and define a
privacy threat model. The threat model described in this section is based on the
aforementioned privacy threats in personalized search systems.
The idea is to model how an attacker could threaten the user’s privacy and conduct
an attack. As described in Fig. 1, we investigate three different threat scenarios:
• The attacker controls the user’s device and has access to the whole user profile.
• The communication channel is insecure, the attacker can eavesdrop and collect
the user’s queries with portions of his profile, and then using auxiliary information
(online ontology) and social engineering skills, he can guess the user’s profile.
Fig. 1 Threat model for personalized search systems

• The server is compromised, and the attacker can collect data sent by the user and
then guess the original user profile as in the second scenario.
In Sect. 2.2, we summarized potential methods that an attacker may use to access
the client’s device, the server, or the communication channel. Since the user profile
is generally stored on the client side, encryption is mandatory to secure the user’s
data and reduce the risk in the first scenario.
During the browsing session in the personalized search system, a user sends many
queries to the server, with short portions of his profile. As we mentioned in the second
and third scenarios, an attacker can obtain a significant amount of the original profile,
by collecting the sub-profiles and using the online ontology to figure the rest.
Considering the user profile P, each time a user enters a query q the system sends
a part of P. If the attacker captures each generalized profile Gi , it is possible after
n query to guess a significant portion of the profile P using the online taxonomy.
And, even if the generalized profile Gi contains no private data, the attacker can still
obtain the profile by comparing the Gn to the ontology.

n
Gi = Gn → P (1)
i=1
where n is a number of queries (depends on the user activity and time).

To illustrate how an attacker can breach user privacy, Fig. 2 shows an example of
a user profile (a) with two generalized profiles (Ga and Gb ).
Fig. 2 Ontology-based user profile

The gray concepts in this figure reflect the user’s private data. And the general-
ized profiles contain no sensitive data because the system stops at the parent nodes.
However, in Ga for example, the attacker can retrieve the sub-tree of security relying
on the taxonomy (b) in the same Fig. 2, where security is the parent of two nodes
including a private one (privacy). Therefore, if the probability of touching both
branches is equal, the attacker has 50% confidence in privacy leading to high privacy
risk.
3 Privacy Protection Solutions
To reduce the aforementioned risks, a number of regulations and measures can be

adopted in personalized search systems.
3.1 Privacy Laws and Regulations
Data are considered as the new oil, enabling new opportunities for advanced analytics
(e.g., personalization) like never before, but this does not come without cost. By
collecting user information, companies have to guarantee the use of data according
to the latest privacy regulations. And various types of data require different measures;
those generated in the healthcare sector, for example, are sensitive from a privacy
point of view.
General Data Protection Regulations (GDPR) [14] was the first regulative initia-
tive to mention location data explicitly in the privacy-sensitive data context, as this
kind of information can infer some sensitive user information. The study in refer-
ence [15] proved that using a dataset where the system collected user location hourly,
four spatio-temporal points (e.g., GPS location) were sufficient to recognize 95% of
the individuals. Such conclusions increased the need for privacy-related regulations
leading organizations to adopt new approaches such as anonymizing or removing the
collected data on the customer’s request. Anonymizing user data provides limited
privacy protection. In the EU AI Guidelines, service providers and organizations
should evaluate the potential adverse consequences of user data used in AI systems
on human rights and, with these consequences in consideration, choose a careful
strategy based on suitable risk prevention and reduction measures.
California Consumer Privacy Act (CCPA) is the most comprehensive state data
privacy legislation to date is the California Consumer Privacy Act (CCPA), signed
into law in June 2018 and was effective in January 2020. The CCPA is cross-sector
legislation that introduces important definitions and broad individual consumer rights
and imposes substantial duties on entities or persons that collect personal data about
or from a California resident.
State Data Privacy Laws: In addition to the general regulations and laws, the
USA has several data security and privacy laws among its states, territories, and
localities. Currently, 25 state attorney generals in the USA oversee data privacy
laws governing the collection, use, storage, safeguarding, and disposal of personal
information collected from their residents, especially regarding the data breach noti-
fications or the security of the social security numbers. Some apply to governmental
entities only, some apply to private entities, and some apply to both.
Many organizations have already begun addressing this issue by implementing
privacy-enhancing technologies (PET), which are not obligatory as per current
privacy regulations yet represent the next measure toward a more ethical and secure
data usage.
3.2 Privacy-Enhancing Technologies
Researchers try to address the privacy protection problems in personalized systems

by protecting user identification using techniques such as the pseudo-identity, no
identity, the group identity, and no personal information. Most efforts focus on the
second level. For example, [16] provided online anonymity for users by creating a
group profile of a number of users. The de-identification techniques are usually used
in personalized systems that collect user’s personally identifiable information (PII)
and works on aggregate data. However, these techniques are vulnerable to attacks
such as the unsorted matching attacks, temporal attacks, and complementary release
attacks. Authors in [17] provided a detailed survey on de-identification techniques
for privacy protection.
Others focus on techniques protecting the sensitive user data (user profile). In this
type of privacy-preserving system, three main technique lines are available, namely
differential privacy, randomized perturbation, and individual privacy.
Differential privacy (DP) is becoming widely accepted as a model for privacy
protection during the past years. This solution preserves privacy by making it difficult
for the attacker to assume the presence/absence of an individual in the dataset. This
technique is designed to work on aggregate data and is most suitable for big data.
For instance, authors in [18] proposed another DP model for neighborhood-based
collaborative filtering capable of selecting neighbor privately but failed to main-
tain a reliable trade-off between privacy and accuracy. Authors in [19] designed a
privacy-built-in client agent that perturbs user data on the client device. However,
the perturbed data utility decreased due to an inhered process volatility. Another
work in [20] proposes a probabilistic model for mobility dataset releasing based on
differential privacy to give users control over privacy level in location-based services.
The main limitations of differential privacy are first, the noise amount added to
the data (more noise for higher privacy risks), which reduces the data utility. It is also
vulnerable against insider threats that come from inside the personalization server.
Authors in [21] presented a survey of DP techniques for more details on the subject.
Randomized perturbation (RP) is another noise-based technique proposed in
[22]. Authors claim that they can obtain accurate recommendations while adding
randomness from a specific distribution to the original user data to counter informa-
tion exposure. The chosen range of randomness is based only on experience, and this
method does not have a provable privacy protection guarantee. Another work in [23]
proposed a multi-leveled privacy-preserving approach for CF systems by perturbing
ratings before submitting them to the server. Yet the results showed a decrease in
utility. Authors in [24] presented a hybrid method for a privacy-aware recommender
system by combining DP with RP to offer more privacy protection. However, the
recommendation accuracy loss with this approach is significant. Most noise-based
techniques share the problem of utility loss [25].
Both DP and RP techniques are designed for aggregate data privacy, ignoring
each user’s privacy requirements, and they both decrease the personalization accu-
racy. Lately, users’ privacy concerns increased due to unethical data aggregation
practices in many recommendation systems. For this reason, in our work, we focus
on individual data privacy.
Individual privacy (IP): As an IP solution, authors in [26] claim that they can
achieve better results with a privacy guarantee if the personalization is only performed
based on less sensitive user data. The idea is to expose only the insensitive part of the
profile to the search engine. Yet, even when using this approach, an attacker or the
server can still collect a significant portion of the user profile. Authors in reference
[27] proposed a privacy-protecting framework for book search based on the idea of
constructing a group of likely fake queries associated with each user query to hide
the sensitive topics in users’ queries. This approach focused on limited query types
with no support for general ones and no practical implementation.
4 Cryptographic Solutions
Cryptographic techniques such as homomorphic encryption (HE), searchable

symmetric encryption (SSE), or garbled circuits used as a mechanism to protect
private data confidentiality. Authors in [28] present a privacy-preserving solution
to generate recommendations using HE and data packing. Also, authors in [29]
used partially HE to design two privacy-preserving protocols for trust-oriented POI
recommendation based on the off-line encryption and parallel computing.
A recent work [30] proposed a fully HE scheme for private IR in the cloud; the
paper contains theoretical analysis but with no proof-of-concept implementation. A
survey by [31] on keyword search with public key encryption can help understand
the use of encryption in private information retrieval.
4.1 Secure k-NN Algorithm
The secure k-nearest neighbor (SkNN) algorithm [32] is used in the searchable
encryption area to encrypt documents and queries presented in a vector space model
of size m. SkNN algorithm allows calculating the similarity [33] (the dot product)
between an encrypted document vector and an encrypted query vector without any
need for decryption. The SkNN algorithm is composed of three functions: KeyGen,
Enc, and Eval.
4.2 Attribute-Based Encryption
The attribute-based encryption (ABE) is an encryption method used to apply an

access control policy to control access to a data collection [34]. It consists of adding
some characteristics to the encrypted ciphertext and the user’s private key. During the
decryption process, ciphertext decryption is possible only if the number of matching
attributes between the ciphertext and the private key exceeds a certain threshold.
Authors in [35] proposed an encryption method called key-policy attribute-based
encryption (KP-ABE) that consists of storing an access structure in the user’s secret
key and attributes in the encrypted data. This method can achieve fine-grained access
control and brings more flexibility in the management of the users than the previous
technique. Nevertheless, KP-ABE has the disadvantage of not being intuitive.
Authors in [36] proposed an alternative method called CP-ABE (ciphertext-policy
attribute-based encryption) that works in the same way as the KP-ABE method except
that the access structure was stored in the ciphertext, while the attributes are in the
user private key. During the decryption process, if some private key attributes satisfy
the data access policy, then ciphertext decryption is possible. Otherwise, the user
has no right to access this data, and his private key cannot decrypt the ciphertext.
For example, (Pediatrics∧(Doctor1∨Doctor2)) is the access policy in the data. If the
user secret key contains the attributes “Pediatrics” and “Doctor1”, then the user has
the right to access the data and can decrypt the ciphertext using his private key. After
that, several approaches based on the CP-ABE method appeared in the literature.
4.3 Homomorphic Encryption (HE)
Homomorphic encryption is a cryptosystem for performing mathematical operations

on encrypted data generating encrypted results which, after decryption, is the same as
the result of these operations on unencrypted data [37]. Several HE systems appeared
in the literature.
Paillier cryptosystem is an asymmetric encryption probabilistic algorithm. It
is an additive homomorphic encryption method that can calculate two ciphertexts
addition without any need for decryption. The Paillier cryptosystem applications can
be in many fields that need to protect data privacy, such as the searchable encryption
area.
Ideal lattice-based fully homomorphic encryption: Gentry and Boneh proposed
the first fully homomorphic encryption (FHE) scheme based on ideal lattices [38].
The encryption process consists of hiding the message by adding noise. Decryp-
tion consists of using the secret key to remove the noise from the ciphertext. This
scheme can perform homomorphic operations on arbitrary depth circuits. For that,
the authors start by constructing a somewhat homomorphic encryption (SWHE)
scheme performing limited operation number. However, the noise size grows after
each arithmetic operation, mostly when it consists of multiplication until it becomes
not possible to decrypt. To avoid this problem, the authors proposed a technique
called “bootstrapping” that consists of refreshing the ciphertext by reducing the noise
size. This technique allows us to transform a SWHE scheme into a FHE scheme.
Nevertheless, this scheme is still a theoretical model because of its inefficiency.
Fully homomorphic encryption over the integers: Authors in [39] proposed
a HE scheme that is similar to the one proposed by Gentry and Boneh [38] except
that it is simpler and less efficient since it works with integers instead of ideals. This
scheme has semantic security based on the hardness hypothesis of the approximate
great common divisor (GCD) problem.
Homomorphic encryption from learning with errors: Brakerski and Vaikun-
tanathan proposed an asymmetric fully homomorphic encryption (FHE) scheme
that operates over bits [40]. This scheme uses the ring learning with error (RLWE)
assumption proposed in [41]. It manipulates polynomials using integer coefficients.
Leveled fully homomorphic encryption: The term (leveled) describes this
approach because of the private key updating performed at each level of a circuit.
The latter represents an arithmetic function and is composed of a set of gates. Each
gate refers to the operation of addition or multiplication. The leveled homomor-
phic encryption [40] introduces two main techniques, key switching and modulus
switching.
A personalized search system must ensure privacy protection to earn the user’s trust.
Otherwise, only a minority of users will use it, to whom the personalized experience
is more important than their privacy.
In this paper, we reviewed different risks and threats on the user privacy in a threat
modeling approach. Moreover, we presented several techniques to preserve the user’s
privacy in personalized search along with cryptographic solutions to protect the user
sensitive data.
In the future, we plan to implement those solutions on a search system presented
in [42] personalized using an ontology-based profiling method described in [43].
References
1. El-Ansari, A., Beni-Hssane, A., Saadi, M.: An improved modeling method for profile-based
personalized search. In Proceedings of the 3rd International Conference on Networking,
Information Systems & Security (pp. 1–6). (2020). https://doi.org/10.1145/3386723.3387874
2. El Makkaoui, K., Ezzati, A., Beni-Hssane, A., Motamed, C.: Cloud security and privacy model
for providing secure cloud services. In 2016 2nd international conference on cloud computing
technologies and applications (CloudTech) (pp. 81–86) (2016). https://doi.org/10.1109/CloudT
ech.2016.7847682
3. Parsania, V. S., Kalyani, F., Kamani, K.: A comparative analysis: DuckDuckGo vs. Google
search engine. GRD J.S-Glob. Res. Dev. J. Eng.-Ing 2(1), 12–17 (2016)
4. Tisserand-Barthole, C.: Qwant. com, un nouveau moteur français. Netsources 104, 14 (2013)
5. El-Ansari, A., Beni-Hssane, A., Saadi, M.: An ontology based social search system. In:
Networked Systems: 4th International Conference, NETYS 2016. https://doi.org/10.13140/
RG.2.2.28161.79200
6. Roelofs, W.: The AOL scandal: an information retrieval view (2007)
7. El-Ansari, A., Beni-Hssane, A., Saadi, M.: An enhanced privacy protection scheme for Profile-
based personalized search. International J. Adv. Trends Comput. Sci. Eng. 9(3), (2020). https://
doi.org/10.30534/ijatcse/2020/241932020
8. El Makkaoui, K., Ezzati, A., Hssane, A. B.: Challenges of using homomorphic encryption
to secure cloud computing. In 2015 International Conference on Cloud Technologies and
Applications (CloudTech) (pp. 1–7). IEEE. (2015). https://doi.org/10.1109/CloudTech.2015.
7337011
9. Chen, D., Zhao, H.: Data security and privacy protection issues in cloud computing. In: 2012
International Conference on Computer Science and Electronics Engineering, Vol. 1, pp. 647–
651 (2012)
10. El-Ansari, A., Beni-Hssane, A., Saadi, M., El Fissaoui, M.: PAPIR: privacy-aware personalized
information retrieval. J. Ambient. Intell. Humaniz. Comput. (2021). https://doi.org/10.1007/
s12652-020-02736-y
11. Hawalah, A., Fasli, M.: Dynamic user profiles for web personalisation. Expert Syst. Appl.
42(5), 2547–2569 (2015)
12. Trautman, L.J., Ormerod, P.C.: Corporate Directors’ and Officers’ Cybersecurity Standard of
Care: The Yahoo Data Breach. Am. UL Rev. 66, 1231 (2016)
13. Minkus, T., Ross, K. W.: I know what you’re buying: Privacy breaches on ebay. In: International
Symposium on Privacy Enhancing Technologies Symposium, pp. 164–183. Springer, Cham
(2014)
14. Voigt, P., Von dem Bussche, A.: The eu general data protection regulation (gdpr). A practical
guide, 1st ed.. Springer International Publishing, Cham (2017)
15. De Montjoye, Y.A., Hidalgo, C.A., Verleysen, M., Blondel, V.D.: Unique in the crowd: The
privacy bounds of human mobility. Sci. Rep. 3, 1376 (2013)
16. Zhu, Y., Xiong, L., Verdery, C.: Anonymizing user profiles for personalized web search. In:
Proceedings of the 19th international conference on World wide web, pp. 1225–1226 (2010)
17. Tomashchuk, O., Van Landuyt, D., Pletea, D., Wuyts, K., Joosen, W.: A Data Utility-Driven
Benchmark for De-identification Methods. In: International Conference on Trust and Privacy
in Digital Business, pp. 63–77. Springer, Cham (2019)
18. Zhu, T., Li, G., Ren, Y., Zhou, W., Xiong, P.: Differential privacy for neighborhood-based
collaborative filtering. In: Proceedings of the 2013 IEEE/ACM International Conference on
Advances in Social Networks Analysis and Mining, pp. 752–759 (2013)
19. Shen, Y., Jin, H.: Epicrec: Towards practical differentially private framework for personalized
recommendation. In: Proceedings of the 2016 ACM SIGSAC conference on computer and
communications security, pp. 180–191 (2016)
20. Zhang, J., Yang, Q., Shen, Y., Wang, Y., Yang, X., Wei, B.: A differential privacy based proba-
bilistic mechanism for mobility datasets releasing. J. Ambient. Intell. Hum.Ized Comput. 1–12
(2020)
21. Desfontaines, D., Pejó, B.: Sok: Differential privacies. Proceedings on Privacy Enhancing
Technologies 2020(2), 288–313 (2020)
22. Zhu, J., He, P., Zheng, Z., Lyu, M.R.: A privacy-preserving QoS prediction framework for web
service recommendation. In 2015 IEEE International Conference on Web Services, pp. 241–
248. IEEE (2015)
23. Polatidis, N., Georgiadis, C.K., Pimenidis, E., Mouratidis, H.: Privacy-preserving collaborative
recommendations based on random perturbations. Expert Syst. Appl. 71, 18–25 (2017)
24. Liu, X., Liu, A., Zhang, X., Li, Z., Liu, G., Zhao, L., Zhou, X.: When differential privacy meets
randomized perturbation: a hybrid approach for privacy-preserving recommender system.
In International Conference on database systems for advanced applications, pp. 576–591.
Springer, Cham (2017)
25. Siraj, M.M., Rahmat, N.A., Din, M.M.: A survey on privacy preserving data mining approaches
and techniques. In Proceedings of the 2019 8th International Conference on Software and
Computer Applications, pp. 65–69 (2019)
26. Shou, L., Bai, H., Chen, K., Chen, G.: Supporting privacy protection in personalized web
search. IEEE Trans. Knowl. Data Eng. 26(2), 453–467 (2012)
27. Wu, Z., Li, R., Zhou, Z., Guo, J., Jiang, J., Su, X.: A user sensitive subject protection approach
for book search service. J. Am. Soc. Inf. Sci. 71(2), 183–195 (2020)
28. Erkin, Z., Veugen, T., Toft, T., Lagendijk, R.L.: Generating private recommendations efficiently
using homomorphic encryption and data packing. IEEE Trans. Inf. Forensics Secur. 7(3),
1053–1066 (2012)
29. Liu, A., Wang, W., Li, Z., Liu, G., Li, Q., Zhou, X., Zhang, X.: A privacy-preserving framework
for trust-oriented point-of-interest recommendation. IEEE Access 6, 393–404 (2017)
30. Wang, X., Luo, T., Li, J.: An Efficient Fully Homomorphic Encryption Scheme for Private
Information Retrieval in the Cloud. Int. J. Pattern Recognit Artif Intell. 34(04), 2055008 (2020)
31. Zhou, Y., Li, N., Tian, Y., An, D., Wang, L.: Public Key Encryption with Keyword Search in
Cloud: A Survey. Entropy 22(4), 421 (2020)
32. Lei, X., Liu, A. X., Li, R.: Secure knn queries over encrypted data: Dimensionality is not
always a curse. In 2017 IEEE 33rd International Conference on Data Engineering (ICDE),
pp. 231–234. IEEE (2017)
33. Erritali, M., Beni-Hssane, A., Birjali, M., Madani, Y.: An approach of semantic similarity
measure between documents based on big data. Int. J. Electr. Comput. Eng. 6(5), 2454 (2016).
https://doi.org/10.11591/ijece.v6i5.10853
34. Li, J., Zhang, Y., Ning, J., Huang, X., Poh, G. S., Wang, D.: Attribute based encryption with
privacy protection and accountability for CloudIoT. IEEE Transactions on Cloud Computing
(2020)
35. Li, J., Yu, Q., Zhang, Y., Shen, J.: Key-policy attribute-based encryption against continual
auxiliary input leakage. Inf. Sci. 470, 175–188 (2019)
36. Qiu, S., Liu, J., Shi, Y., Zhang, R. (2017). Hidden policy ciphertext-policy attribute-based
encryption with keyword search against keyword guessing attack. Sci. China Inf. Sci. 60(5),
052105.
37. El Makkaoui, K., Beni-Hssane, A., Ezzati, A.: Cloud-ElGamal: An efficient homomorphic
encryption scheme. In: 2016 International Conference on Wireless Networks and Mobile
Communications (WINCOM), pp. 63–66. IEEE. (2016). https://doi.org/10.1109/WINCOM.
2016.7777192
38. Gentry, C., Boneh, D.: A fully homomorphic encryption scheme, vol. 20, No. 9, pp. 1–209.
Stanford University, Stanford (2009)
39. Van Dijk, M., Gentry, C., Halevi, S., Vaikuntanathan, V.: Fully homomorphic encryption
over the integers. In Annual International Conference on the Theory and Applications of
Cryptographic Techniques, pp. 24–43. Springer, Berlin, Heidelberg (2010)
40. Brakerski, Z., Gentry, C., Vaikuntanathan, V.: (Leveled) fully homomorphic encryption without
bootstrapping. ACM Transactions on Computation Theory (TOCT) 6(3), 1–36 (2014)
41. Lyubashevsky, V., Peikert, C., Regev, O.: On ideal lattices and learning with errors over rings.
Journal of the ACM (JACM) 60(6), 1–35 (2013)
42. El-Ansari, A., Beni-Hssane, A., Saadi, M.: A multiple ontologies based system for answering
natural language questions. In: Europe and MENA cooperation advances in information and
communication technologies, pp. 177–186. Springer, Cham. (2017). https://doi.org/10.1007/
978-3-319-46568-5_18
43. El-Ansari, A., Beni-Hssane, A., Saadi, M.: An ontology-based profiling method for accurate
web personalization systems. J. Theor. Appl. Inf. Technol. 98(14), 2817–2827 (2020)
Enhanced Intrusion Detection System
Based on AutoEncoder Network and
Support Vector Machine
Sihem Dadi and Mohamed Abid
Abstract In the recent years, Internet of Vehicles (IoV) became the subject of
many searches. IoV relies on Vehicular Ad-hoc NETwork (VANET) which is based
on Vehicle-to-Vehicle (V2V), Vehicle-to-Infrastructure (V2I) communication, and
Vehicle-to-Everything (V2X). Due to heterogeneous communications in VANET,
vehicles are increasingly becoming vulnerable to intruders or attackers. Several
research works proposed solutions to detect intrusion, some of them used deep learn-
ing which uses a set of algorithms called neural networks in different fields, such
as health, economic, transport and many others domains. In this paper, an enhanced
intrusion detection system (IDS) based on AutoEncoder (AE) network and support
vector machine (SVM) is proposed. Our goals are to detect five main types of attacks
(DoS, DDoS, Wormhole, Black hole and Gray hole attack) that VANET may face
by combining the ability of support vector machine (SVM) to exploit large amounts
of data with the strength of features extraction provided by AutoEncoder (AE). The
experimental results show that the enhanced intrusion detection system (IDS) is capa-
ble to reach a high level of accuracy, also we prove by security analysis that our new
solution detects successfully these attacks.
1 Introduction
Smart and connected vehicles are inspiring researchers to develop new uses cases
to exploit their benefits. That why, Vehicular Ad-hoc NETwork (VANET) is the
backbone to intelligent transportation system (ITS) research. VANET is a mobile
network allowing vehicles to communicate with each other, with the aim of improving
road safety through the exchange of alerts between vehicles.
S. Dadi (B) · M. Abid

Faculty of sciences of Gabes, Laboratory Hatem Bettaher Irescomtah, University of Gabes,
Gabes, Tunisia
e-mail: sihem.dadi@fsg.u-gabes.tn
M. Abid
e-mail: mohamed.abid@enig.rnu.tn
https://doi.org/10.1007/978-981-16-3637-0_23
328 S. Dadi and M. Abid
In VANET, there are many types of communications, while the most popular ones
are: Vehicle to Vehicle V2V and Vehicle to Infrastructure (V2I) communication which
means a communication between the Road Side Units (RSU), a fixed infrastructure
and vehicle. Also, vehicles can communicate with any objects and this case is known
as Vehicle to Everything (V2X).
Through this network, vehicles can exchange control, alert or "other" messages,
depending on the application and the environmental context. Due to the communi-
cation of vehicles within VANET and with others entities in the Internet, they can
be easily a target for many attacks. Detecting successful or failed attack attempts is
important to secure networks (servers, end-hosts and other assets). Some attackers
want to change the content of the information exchanged in the network, owing to
dynamic topology which may result loss of connectivity as well as unreliable con-
nections. To secure VANET, many tools are used such as ciphering exchanging data,
Intrusion Detection System (IDS) and Intrusion Prevention System (IPS). IDS are
defined as the last line of defense during an attack. They allow the detection of attacks
targeting a vehicle or a network. However, they do not offer mechanisms in response
to attacks. IDS informs administrator that an attack has been detected; however, this
information is only relevant if used. On the other hand, intrusion prevention system
(IPS) has the capabilities of detecting, isolating and blocking malicious attacks in
real time.
In parallel, a new technology has emerged and succeeded in a short time to pen-
etrate several domains for different purposes, which is deep learning. The latter is a
subfield of machine learning: a set of algorithms that try to imitate the functioning
of the human brain. Among the areas in which deep learning has been introduced,
we cite: Military, health, biology, transport,... and many others. In the transport field,
deep learning is used for many objectives, but the most important among them are:
object detection, vehicle trajectory prediction, traffic flow prediction and intrusion
detection. Also, deep learning provides intelligent solutions since it is based on neural
network.
The idea of our solution is to design an enhanced IDS based on deep learning,
AutoEncoder (AE) network and Support Vector Machine (SVM).
This paper is organized as follows: Sect. 2 discusses the theoretical background
that mean the technologies and tools to be used in our system. Section 3 describes
related works. Section 4 is devoted to the proposed solution using neural networks.
We utilized the UNSW-NB15 data set to evaluate the performance metrics of the
proposed enhanced IDS, and we analysed its Security Analysis . Finally, we conclude
the paper and give some future works.
2 Theoretical Background
This section presents IDS, VANET and deep learning.

Enhanced Intrusion Detection System Based on AutoEncoder Network . . . 329
2.1 Intrusion Detection System
An intrusion detection system is a software or hardware device that automates the

intrusion detection process which consists of supervising events coming from net-
work in order to detect signs of intrusions [1]. Therefore, an IDS is a monitor-only
application designed to identify and report on anomalies but after a hackers already
damages network infrastructure.
Hereafter, we present the different classes of IDS and the working process of each
model.
• IDS classes: Mainly as shown in Fig. 1, an IDS can be categorized into two main
classes: based on its location in the network or by its detection method. The latter
type is subdivided into two subclasses:
Misuse/Signature-based detection: It is based on signatures which are stored in
knowledge base; this model has a high accuracy rate of detection but only for
known attacks.
Anomaly-based detection: This model is smarter than the previous one, because
it does not require signatures to detect intrusion and also it can identify unknown
attacks depending on similar behavior of other intrusions. It can be accomplished
via the use of artificial intelligence and deep learning algorithms. This method is
prone to high false positive rate of detection.
• IDS working process: An IDS working process depends on the detection model. In
the case of signature-based detection, it monitors all the network packets and based
on analyzing signatures stored on the knowledge base, it detects potential malware
and matches the suspicious activities in the network. In contrast, an anomaly-based
detection follows these steps to successfully detect intrusions: It starts monitoring
the network traffic and analyses its pattern against predefined norms or knowledge
base. Then, if an abnormal traffic is identified, then it launches alerts to report
unusual behavior.
Fig. 1 Intrusion detection

system classification
2.2 Vehicular Ad Hoc Network VANET
Vehicular Ad hoc NETwork called “VANET”, a derivation of Mobile Ad hoc NET-

work (MANET), was firstly introduced in 2001 to guarantee vehicle and driver
security.
• VANET Architecture: Typically, there are four main components that make up the
VANET networks: OBU, RSU, TA and AU [2] (As shown in Fig. 2).
On Board Unit (OBU): As its name suggests, is a unit attached to each vehicle
supporting Intelligent Transport System ITS, with the aim of exchanging informa-
tion with the approximate OBUs or with RSUs.
Road Side Unit (RSU): Standing for the Base Station and acting as the gateway
between vehicles and road services provided by VANET.
Trusted Authority (TA): Responsible for assigning digital certifications (unique
identifier) to RSUs and OBUs in the network.
Application Unit (AU): A device equipped within the vehicle that uses the appli-
cations provided by the provider using the communication capabilities of the OBU.
• Communication types in VANET: Fig. 3 shows the most popular communication
types in VANET: Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I)
[3]:
Vehicle-to-Vehicle (V2V): Which allows vehicles to communicate with each other
using IEEE WAVE [3]. Vehicle-to-Infrastructure (V2I): Which allows the vehi-
cle to communicate with the infrastructure such as traffic light or road side units
(RSU), using Wi-Fi and 4G/LTE [3].
A vehicle in VANET can communicate with other entities outside its original
network. This leads to tremendous growth in the number of its communications.
Hence, many new types of communication emerged such as V2N (Vehicle-to-
Fig. 2 VANET architecture

Fig. 3 Communication types in VANET
Network), V2P (Vehicle-to-Pedestrian), V2S (Vehicle-to-Sensor), V2G (Vehicle-

to-Grid) and V2D (Vehicle-to-Device) and more others in the near future ....
• Attacks in VANET: The number of attacks that a car can face, in a VANET, is
increasingly countless. These attacks can either shut down or degrade network
performance [4]. Among these threats, we focus on five attacks that appear to be
the most dangerous: Denial of Service, Distributed Denial of Service, Wormhole,
Black and Gray hole attacks.
Denial of Service Attack (DoS): is the act of blocking or preventing access to
authorized users. This attack is the most dangerous one, because it threatens the
network availability.
Distributed Denial of Service Attack (DDoS): In DDOS attack, multiple mali-
cious vehicles launch attack on a legitimate vehicle from different locations, and
they may use different time slots to send those messages. So, it is very difficult to
prevent or track this attack.
Wormhole Attack: It is an attack in which two attackers locate themselves strate-
gically in the network. Then, the attackers keep on listening to the network and
record the wireless information.
Black hole Attack: It happens when one or more nodes drop communications
between other nodes. It is mainly a kind of denial of service, where the black hole
is a node that always responds positively with an Route Reply (RREP) to every
Route Request (RREQ) message , even if it does not have a legitimate route to the
destination.
Gray Hole Attack: This attack is a category of a black hole attack in that the
malicious node deletes packets but, in this case, this is done in a partial or a
selective way. In fact, the malicious node can switch between two modes: Either
it remains in normal mode as a harmless node or it switches to the attack mode.
2.3 Deep Learning
Deep learning is a subfield of machine learning, which is essentially a multi-layer

neural network. The latter is a set of algorithms which mimic the human brain operates
in order to recognize underlying relationships in a set of data [5]. A neural network
has the following architecture in Fig. 4.
In general, a neural network is essentially composed of:
Input Layer: It is composed of one or more nodes; each node receives input from
an external source. Every node is connected with another node from the next layer,
with a particular weight.
Hidden Layers: As their name indicates, these layers are hidden from the external
world, a neural network has at least 5 to 10 layers or even more. Calculations are
executed in these layers, on data received from input layer in order to generate results.
Output Layer: Receives the results generated from the hidden layers.
Nowadays, there are many types of neural networks in deep learning which are
used for different purposes. Among these types, we cite: Convolutional Neural Net-
works (CNN) and its derivations ( DCN, DN, . . .), Recurrent Neural Networks (RNN)
and its improvements ( LSTM and GRU), also Generative Adversarial Networks
(GAN) and its subtypes ( DCGAN, WGAN, . . .) and finally AutoEncoder network
(AE) and its subcategories ( SEA, SSEA, . . .).
Fig. 4 Neural network structure

There are as many neural network classes as there are classifiers; among these
classification algorithms, we cite: Logistic Regression (LR), Naïve Bayes, Stochastic
Gradient Descent, K-Nearest Neighbours (K-NN), Decision Tree (DT), Random
Forest (RF), Support Vector Machine (SVM), . . . In this paper, we chosen to use of
AE and SVM which are detailed in the following items.
• AutoEncoder (AE) network: AutoEncoder is a neural network approach that
teaches a network to ignore noise, in order to learn efficient data representations
(encoding) [6]. it is an unsupervised learning technique.
AutoEncoders are widely used for: dimensionality reduction, image compression,
denoising and generation and for features extraction . . . and many others. As shown
in Fig. 5, an AutoEncoder network is composed of two main gates which are:
Encoder: receives the input and transforms it to a new representation, which is
usually called a code or latent variable.
Decoder: receives the generated code at the encoder and transforms it to a recon-
struction of the original input.
AE hyperparameters: There are four hyperparameters that we need to set before
training an AE:
– Code size: number of nodes in the middle layer.
– Number of layers: the autoencoder can be as deep as we like.
– Number of nodes per layer: the AE architecture we are working on is called a
stacked AE since the layers are stacked one after another.
– Loss function: we either use mean squared error (mse) or binary crossentropy.
If the input values are in the range [0, 1], then we typically use crossentropy,
otherwise we use the mean squared error.
Similarly for the feature extraction phase, it can be achieved via many algorithms
such as Kernel PCA, neural networks, . . .. In this paper, we select AE to realize
this task because according to [7], this neural network is the best choice to extract
features.
• Support vector machine (SVM): SVMs are a family of machine learning algorithms
that solve problems of classification, regression or anomaly detection [8]. They
are known for their solid theoretical guarantees, their great flexibility and their
ease of use even without great knowledge of data mining. Figure 6 illustrates
SVM principle which is simple: separate the data into classes using a border as
“simple” as possible, so that the distance between the different groups of data
and the border which separates them is maximum . This distance is also called
“margin” and SVMs are thus referred to as “wide margin separators”, the “support
vectors” being the data closest to the border. There are many methods to classify
a set of data such as: artificial and deep neural networks, decision tree, random
forest, . . .. But and based on many comparative studies such as [9], it shows that
support vector machine (SVM) is more powerful in the classification phase and
that will be our choice to classify the selected features.
Fig. 5 AutoEncoder
architecture
Fig. 6 SVM principle
3 Literature Survey
Thanks to its ability to detect attacks with high accuracy, intrusion detection is the
most authentic technique to protect VANET. This is why it is necessary to have
a strong detection mechanism such as approaches based on deep learning. In this
context, several approaches have been proposed:
Multilayer Perceptron-Based Distributed Intrusion Detection System for Internet
of Vehicles is an intrusion detection approach proposed by Anzer and Elhadef [10].
DoS, User to Root attack (U2R), Root to Local attack (R2L) and probe attacks are
successfully identified, in this approach. Results are in form of prediction, classifi-
cation and confusion matrix.
Sangve et al. suggested an algorithm for detecting rogue nodes in VANET [11].
Rogue nodes broadcast false information report that’s why, in the actual approach,
authors use anomaly-based detection approach to detect rogue nodes which are intro-
duced in the system and are successfully detected.
Zhao et al. deployed an intrusion detection system for detecting intrusion in
VANET [12]. In this work, they used two neural networks: Deep Belief Network
(DBN) to extract features and Probabilistic Neural Network (PNN) in the classifica-
tion phase.
In [13], Maglaras combined the dynamic agents and static detection to design
intrusion detection system in VANET. By this approach, only DoS attack is success-
fully identified and detected.
Zeng et al. implemented DeepVCM which is a deep learning-based method for
intrusion detection in VANET [14]. DeepVCM consists of two models Convolu-
tional Neural Network (CNN) algorithm for features extraction and Long Short-Term
Memory (LSTM) algorithm for classification. In this paper, authors identify the Dos,
DDoS, Black hole, Wormhole and Sybil attack.
In order to detect distributed denial-of-service attack in VANET, Gao et al. devel-
oped a distributed network intrusion detection system [15]. They used Spark-ML-
RF-based algorithm to train the system, and both Random Forest algorithm for
classification and T Decision Tree to perform the classification process.
In the same context, Vatilcar et al. introduced an intrusion detection system in
VANET using deep learning method [16]. This work was achieved by the use of Deep
Belief Network (DBN) to extract the meaningful features and Restricted Boltzmann
Machine (RBM) in classification phase. In the next section, we present the design of
our proposed solution.
4 Enhanced IDS Based on AE and SVM
In this section, we present the design of an enhanced IDS which aims to detect attacks
in the VANET, using AutoEncoder network and SVM algorithm. This solution is
better comparing with previous ones since none of the latter can detect all the five
attacks mentioned earlier, and they are cumbersome and the tools used are difficult
to train.
Also, we use of a special dataset to train and test our system. We manage to study
its performances. Finally, we analyse the security of our solution.
Our solution is based on the same architecture (see Fig. 7) presented in paper [17]
which is the common architecture for all the previous solutions. It is composed of
two modules:
Profiling module: contains the features trained off-line.
Monitoring module: detects a type of an incoming packet after feature extraction. If
the monitoring module identifies a new attack type, the profiling module may update
the database of the profiling module for upcoming packets.
Fig. 7 IDS based on deep learning algorithms common architecture
4.1 The Proposed Architecture
The enhanced IDS aims to successfully identify: DoS, DDoS, Wormhole, Black
and Gray hole attacks. Our method provides an enhanced IDS that combines the
advantages of AE to extract features and SVM algorithm to classify features already
extracted to solve intruders detection problem.
We implemented the new IDS and make performance analysis using UNSW NB-
15 dataset which was developed in 2015 [18], and it is a dataset for network intrusion
detection systems containing two million and 540,044 records. This data set involves
nine attack categories and 49 features.
The partitioned dataset contains ten categories, one normal and nine attacks,
namely generic, exploits, fuzzers, DoS, reconnaissance, analysis, backdoor, shell-
code and worms. Figure 8 shows in detail the class distribution of the UNSW-NB15
dataset.
Fig. 8 Class distribution of the UNSW-NB15 dataset

Fig. 9 Overall of the proposed system
Based on the common architecture already mentioned in the previous works, our
IDS is composed of three modules, which are respectively: Profiling, Monitoring
and Detection module as shown on Fig. 9.
Profiling Module: is divided into three main phases which are :
• Data preprocessing: In this phase, both encoding and normalization stages take
place: the encoding process consists of converting symbol features to numerical
values, and the normalization one refers to range these encoded values between 0
and 1.
• Features selection/extraction: Aims to reduce the number of features in a dataset
by creating new features from the existing ones (and then discarding the original
features). For our IDS, this phase is reached via the use of AutoEncoder (AE)
network, the encoder receives preprocessed data from the previous phase and
compresses it to a latent space representation, and then the encoder sends it to the
decoder which had to reconstitute the input from the latent space representation.
• Classification: As its name indicates, the classification phase refers to predicting
which class a feature belongs to. In our system, we use SVM to classify extracted
features.
Monitoring module: We tested the proposed security system with UNSW NB-15
testing set. The performance of the enhanced IDS is directly related to anomaly detec-
tion algorithm. If the anomalies are detected correctly from the data set, it provides
a high detection rate and less false alarms. The anomaly detection has the ability to
detect novel attacks. In this phase, we tested the detection system with significant
features that were selected from the UNSW NB-15 data set. The behaviours are anal-
ysed, and then the IDS generates four types of alarms: true positive, true negative,
false positive and false negative.
Detection module: At this level, alarms are already generated and the detection
accuracy rate are used to measure the IDS performance. The detection phase has two
outputs which are: normal and attack.
4.2 Performance Metrics and Results
When referring to the performance of IDSs, the following terms are often used to
discuss their capabilities: True Positive (TP), False Positive (FP), True Negative
(TN), False Negative (FN). Figure 10 clearly explains the terms already mentioned.
Performance metrics The performance metrics calculated from performance
parameters are:
• Accuracy is a ratio of correctly predicted observation to the total observations.
Accuracy = TP+TN/TP+FP+FN+TN.
• Precision is the ratio of correctly predicted positive observations to the total pre-
dicted positive observations.
Precision = TP/TP+FP.
• Recall is the ratio of correctly predicted positive observations to the all observations
in actual class.
Recall = TP/TP+FN.
• F1-score is the weighted average of Precision and Recall.
F1-score = 2*(Recall * Precision) / (Recall + Precision).
Results
After applying our approach, the results obtained are detailed in Table 1).
Fig. 10 Performance patameters

Table 1 Experimental results

Obtained results
Parameters Values Parameters Values (%)
True Positive (TP) 99.3 % False Positive (FP) 5.78
True Negative (TN) 98.7 % False Negative (FN) 0.85
Accuracy 96.7%
Precision 94.4%
Recall 99.1%
F1-score 96.6%
Table 2 Comparison between approaches

Work DoS DDoS Black hole Wormhole Gray hole
Anzer et al. + − − − −
Zhao et al. + − − − −
Maglaras + − − − −
Zeng et al. + + + + −
Gao et al. + + − − −
Our approach + + + + +
We conclude that our enhanced IDS obtained good performances and reach its
design goals especially a high level of accuracy.
4.3 Security Analysis
Table 2 presents a comparison between previous solutions and our proposed one. We
observe that other solutions do not detect all attacks in VANET:
When using this approach, the five main attacks mentioned above are detected by
our enhanced IDS.
5 Conclusion
Vehicular Ad hoc NETwork (VANET) is more and more vulnerable to attackers

who try always to develop new techniques to hack it, by forging false information
which may result in accidents. This evolution in attacks must leads to an evolution in
protection methods such as intrusion detection system (IDS) which is defined as the
last line of defense in network during an attack. At the same time, neural networks
have seen a great evolution and have been able to penetrate several fields: health,
transport, . . . with different utilities such as: intrusion detection, object detection, . . ..
In this paper, we designed an enhanced IDS based on AutoEncoder (AE) network
and support vector machine (SVM) in order to take advantage of their benefits and
detect attacks in an effective manner. The experiment reflects the performance of
the enhanced IDS in reaching high level of accuracy. Furthermore, this approach
detected the various attacks that VANET may face such as DoS, DDoS, Wormhole,
Black and Gray hole attack.
As future work, we could apply this solution and test its feasibility in different
networks such as Flying Ad hoc NETworks (FANET) and Sea Ad hoc NETworks
(SANET).
References
1. Alamiedy, T., Anbar, M., Al-Ani, A., Al-Tamimi, B., Faleh, N.: A review on feature selection
algorithms for anomaly-based intrusion detection system. In: Proceedings of the 3rd Inter-
national Conference of Reliable Information and Communication Technology (IRICT 2018)
(2019). https://doi.org/10.1007/978-3-319-99007-1-57
2. Al-Sultan, S., Al-Doori, M.M., Al-Bayatti, A.H., Zedan, H.: A comprehensive survey on vehic-
ular Ad Hoc network. J. Netw. Comput. Appl. (2014). https://doi.org/10.1016/j.jnca.2013.02.
036
3. Kumar, G., Saha, R., Rai, M.K., Kim, T.: Multidimensional security provision for secure com-
munication in vehicular ad hoc networks using hierarchical structure and end-to-end authenti-
cation. IEEE Access (2018). https://doi.org/10.1109/ACCESS.2018.2866759
4. Zaidi, T., Syed, F.: An overview: various attacks in VANET. In: 4th International Conference on
Computing Communication and Automation (ICCCA) (2018). https://doi.org/10.1109/ccaa.
2018.8777538
5. Deng, L.: A tutorial survey of architectures, algorithms and applications for deep learning.
APSIPA Trans. Signal Inf. Process. (2014). https://doi.org/10.1017/atsip.2013.9
6. Mohammadi, M., Al-Fuqaha, A., Sorour, S., Guizani, M.: Deep learning for IoT big data and
streaming analytics: a survey. IEEE Commun. Surv. Tutor. (2018). https://doi.org/10.1109/
COMST.2018.2844341
7. Yan, B., Han, G.: Effective feature extraction via stacked sparse autoencoder to improve intru-
sion detection system. IEEE Access (2018). https://doi.org/10.1109/ACCESS.2018.2858277
8. Zhiquan, Q., Ying, J.T., Yong, S.: Robust twin support vector machine for pattern classification.
Pattern Recogn. (2013). https://doi.org/10.1016/j.patcog.2012.06.019
9. Phan, T.N., Martin, K.: Comparison of Random Forest, k-Nearest Neighbor, and support vector
machine classifiers for land cover classification using Sentinel-2 imagery. Sensors (Basel)
(2018). https://doi.org/10.3390/s18010018
10. Anzer, A., Elhadef, M.: A multilayer perceptron-based distributed intrusion detection system
for internet of vehicles. In: 2018 IEEE 4th International Conference on Collaboration and
Internet Computing (CIC (2018). https://doi.org/10.1109/CIC.2018.00066
11. Sunil, M.S., Reena, B., Vidhya, N.G.: Intrusion detection system for detecting rogue nodes in
vehicular ad-hoc network. In: International Conference on Data Management, Analytics and
Innovation (ICDMAI) (2017). https://doi.org/10.1109/ICDMAI.2017.8073497
12. Zhao, G., Zhang, C., Zheng, L.: Intrusion detection using deep belief network and probabilistic
neural network. In: International Conference on Computational Science and Engineering (CSE)
and IEEE International Conference on Embedded and Ubiquitous Computing (EUC) (2017).
https://doi.org/10.1109/CSE-EUC.2017.119
13. Maglaras, L.A.: Intrusion detection using deep belief network and probabilistic neural net-
work. Int. J. Adv. Comput. Sci. Appl. (IJACSA) (2015). https://doi.org/10.14569/IJACSA.
2015.060414
14. Zeng, Y., Qiu, M., Zhu, D., Xue, Z., Xiong, J., Liu, M.: DeepVCM: a deep learning based
intrusion detection method in VANET. In: 5th International Conference on Big Data Security
on Cloud (BigDataSecurity), IEEE International Conference on High Performance and Smart
Computing, (HPSC) and IEEE Intl Conference on Intelligent Data and Security (IDS) (2019).
https://doi.org/10.1109/BigDataSecurity-HPSC-IDS.2019.00060
15. Gao, Y., Wu, H., Song, B., Jin, Y., Luo, X., Zeng, X.: A distributed network intrusion detection
system for distributed denial of service attacks in vehicular ad hoc network. IEEE Access
(2019). https://doi.org/10.1109/ACCESS.2019.2948382
16. Vatilkar, R.S., Thorat, S.S.: A review on intrusion detection system in vehicular ad-hoc network
using deep learning method. Int. J. Res. Appl. Sci. Eng. Technol. (IJRASET) (2020). https://
doi.org/10.22214/ijraset.2020.5258
17. Kang, M., Kang, J.: Intrusion detection system using deep neural network for in-vehicle network
security. PLOS ONE (2016). https://doi.org/10.1371/journal.pone.0155781
18. Moustafa, N., Slay, J.: UNSW-NB15: a comprehensive data set for network intrusion detec-
tion systems (UNSW-NB15 network data set). In: Military Communications and Information
Systems Conference (MilCIS) (2015). https://doi.org/10.1109/MilCIS.2015.7348942
Comparative Study of Keccak
and Blake2 Hash Functions
Hind EL Makhtoum and Youssef Bentaleb
Abstract Hashing is so crucial for integrity. Indeed, in order to respond to new

technologies requirements, hash functions have always been subject to evolution and
optimization. In the same context, NIST organized a competition in 2012 for a Sha-3
hashing function. BLAKE and KECCAK were two interesting functions in terms of
speed and security. Keccak was the winner of the competition. It is different from
the antecedent of SHA family because it combines the sponge construction to the
Keccak-f-permutations. Therefore, it is optimized for security and hardware use. The
Blake function is based on the same principle of the Sha-2, and it also presents many
advantages of security and speed. The Blake was improvised for the BLAKE2, which
is more optimized, faster, and more efficient. In this paper, we present the Keccak and
the BLAKE2 algorithms; then, we will compare the performance parameters of the
functions: Blake-256, Blake-512, Keccak-256, Keccak-512, Blake2b, and Blake2s.
1 Introduction
With the drastic evolution of technologies and the growth of resource-constrained

smart objects„ all the related fields are led to ensure a higher level of performance
and security. Data integrity is based on cryptographic hash functions, which play
an essential role in digital signatures, message authentication codes, file checksums,
and many other protocols in the security schemes.
In the same context, The National Institute of Standards and Technology (NIST)
opened a public competition on November 2, 2007, to develop the new Sha-3 hash
function, and the winner was the KECCAK[1600]. Another function of the fifth
finalist that presented pretty good results is the BLAKE that was then improvised to
BLAKE2 to provide better performance. Blake2 has also shown considerably less
H. EL Makhtoum (B) · Y. Bentaleb

Engineering Sciences Laboratory, ENSA, Ibn Tofail University, Kenitra, Morocco
e-mail: hind.elmakhtoum@uit.ac.ma
Y. Bentaleb
e-mail: youssef.bentaleb@uit.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_24
344 H. EL Makhtoum and Y. Bentaleb
energy consumption and less computational cost in resource-constrained devices

of IoT domain. This paper concentrates on the finalist of SHA-3 Keccak and the
improvised BLAKE2. We compared and analyzed Blake, Blake2, and Keccak. We
examined differences in terms of basic parameters, speed, and storage.
2 Hash Functions Algorithms
2.1 Keccak
Keccak is the cryptographic hash function that won the NIST SHA-3 hash function
competition. Keccak is a family of hash functions based on the sponge construction
and used as a building block of a permutation from a set of seven permutations
[1]. The basic component is the Keccak-f permutation, consisting of several simple
rounds with logical operations and bit permutations.
The fundamental function of Keccak is chosen from a set of seven Keccak-f permu-
tations, denoted by Keccak-f[b], with b ∈ {25, 50, 100, 200, 400, 800, 1600} is the
width of the permutation. “b” is also the state’s width in the sponge construction.
The state is an array of 5 × 5 [2].
Algorithm The internal state size is 1600 = r + c bits, and it starts with the value 0
for SHA-3, and r = 576bits; c = 1024bits in case of Keccak-512 because c =
512(digest size) × 2
In the absorbing phase, the input message is first padded and divided into small
parts of r bits: the first r bits are XORed with the internal state, then the first r bits of
the resulted in the internal state (padding) XOR with the first r bits of the precedent
message, as shown in Fig. 1.
The squeezing phase is a reverse operation to get the digest. It consists of applying
only permutations on the state and finally get parts of the digest. It is repeated up to
generate the necessary length of the output (Fig. 2).
The f function here represents permutation is the Keccak-f(1600). It consists of
5rounds which repeat 24 times.
Θ: Calculate the parity of the 5 × w columns (of 5 bits) of the state, then calculate
the or exclusive between two neighboring columns (Fig. 3).
ρ: Circularly shift the 25 words of a triangular number (Fig. 4).
Π: Permutation of the 25 words with a fixed pattern (Fig. 5).
χ: Combine bit by bit lines (Fig. 6).
ι: Calculate the XOR of a constant with a state word, which depends on the
iteration number (n): This last step aims to break the last symmetries left by the
previous ones.
Comparative Study of Keccak and Blake2 Hash Functions 345
Fig. 1 Absorbing phase [1]
Fig. 2 Squeezing phase [1]
2.2 Blake2
BLAKE2 [3] is an ARX cryptographic hash function, a successor of the Blake family.
It shares many similarities with the original design. However, differences occur at
every level: internal permutation, compression function, and hash function construc-
tion. It is designed to have the best performances on software implementations, and
it is claimed to be faster than sha-2 and sha-3.
Blake2 consists of two main algorithms: BLAKE2b that is optimized for 64-bit
platforms and produces digests of a size up to 64 bytes, and BLAKE2s is optimized
for 8- to 32-bit platforms and produces digests of a size up to32 bytes. [4].
Fig. 3 Round Θ [1]
Fig. 4 Round ρ [1]
Fig. 5 Round Π [1]

Fig. 6 Round χ [1]
They are portable to any CPU, but they are twice fast when used on the CPU size
for which it is optimized.
Algorithm
The message is divided into blocks(d) of 512bits. The internal state v is initialized
by affecting the matrice IV (values below) XORed with the digest length to the h
values, which is an 8words block.
h0 = IV ⊕ P,
for i = 0 to N−1,
hi + 1 = compress(h i,l i,m i),
return hN,
This way, the state matrice V is initialized with the first eight items of h, and the
remaining items are initialized from the initialization vector iv. For 10 (Blake2s) or
12 (Blake2b) rounds, the G function is applied to columns and diagonals of the v
matrice in combination with the message bloc (d0), so we get the h values, which
will be compressed once again with the next block of the message d1 (Fig. 7).
3 Performance
The main objective of BLAKE2 is to provide several parameters to use in applications

without the need for additional constructions and modes and also to speed-up the
hash function to a level of compression rate close to MD5. Added changes are the
Fig. 7 Blake2 Algorithm and the compress function of BLAKE2 [3]
following: 16-word constant was removed, simplified padding was used, and rotation
was reduce from 16 to 12 from BLAKE2b and from 14 to 10 in BLAKE2s. Also,
BLAKE2 supported additional functionalities such as keyed salt, personalization
block, and tree hashing models. The blake2 is an optimization of the Blake; it is
faster, and it consumes less memory than the Blake (32%).
Hardware vs Software While Blake offers excellent software performance, Keccak
uses less hardware area than BLAKE and less energy per hashed bit. Keccak also has
a “permutation” structure that allows the same hardware to be efficiently reused for
applications beyond hashing. Therefore, Keccak consumes fewer physical resources
than Blake.
Storage BLAKE used eight-word constants as IV plus 16-word constants [5] for
use in the compression function; BLAKE2 uses a total of 8-word constants instead of
24 [5]. It saves 128 ROM bytes and 128 RAM bytes in BLAKE2b implementations
and 64 ROM bytes and 64 RAM bytes in BLAKE2s implementations. It means that
Blake2 consumes 32% less memory than Blake [6].
In 64 platforms, Blake2s needs 168bytes, Blake2b 336 bytes [6], less space than
the Keccak, that needs 1600Bits RAM for operations, which means that the Keccak
consumes less storage than the BLAKE2b and more storage than blake2s.
Both functions are considered as lightweight protocols. So, the choice between
them is basically related to the application requirements.
Speed (fast) Many works compared BLAKE2 variants to Keccak variants in terms
of speed, and it proves that the Blake2 is faster. It presents better results in terms of
speed. [5].
Table 1 Performance of 512-bit versions of hash algorithms, speed(Intel Core i3-2310 M-Sandy
bridge)
Algorithm Keccak 512 Blake 512 BLAKE2b 512
Round of compression 24 16 12
Major operations in a AND, OR, XOR, Modular addition, Modular addition,
round ROT, SHR XOR, and rotate XOR, and rotate
operations operations
Speed cycle/byte [7] 20.46 cycle/byte 14.69 cycle/byte 9cycle/byte
Hash applications Lightweight Website links Argon2, WinRAR,
of Perl, PHP, OpenSSL
Java script
Table 2 Performance of 256-bit versions of hash algorithms, Speed(Intel Core i3-2310 M-Sandy
bridge)
Algorithm Keccak 256 Blake 256 BLAKE2s
Round of compression 7 14 10
Major operations in a AND, OR, XOR, rotate Modular addition, Modular addition,
round operations, shr XOR, and rotate XOR, and rotate
operations operations
Speed cycle/byte [7] 10.87 16.38 5.50
hash applications Ethereum Blakecoin WireGuard,
checksum, 8th,
peerio
Tables 1 and 2 compare the 512 versions of the hash functions Blake and Keccak,
and it shows with various parameters.
4 Conclusion
In this paper, we selected two hashing functions of The NIST competition. Both
functions are suitable for constrained objects, but each presents different parameters.
The comparison is necessary to make the right choice depending on the require-
ment of the application. Regarding critical applications, such as authentication of
the IOT constrained objects, that must consider the limited capacities in storage and
processing of objects. These operations need a less consuming, fast, and light hashing
function to ensure a high-level security and avoid infiltration attacks.
References
1. Dat, T.N., Iwai, K., Matsubara, T., Kurokawa, T.: Implementation of high-speed hash function
Keccak on GPU. IJNC 9(2), 370–389 (2019)
2. Kavun, E.B., Yalcin, T.: A lightweight implementation of Keccak hash function for radio-
frequency identification applications. In: Ors Yalcin, S.B. (ed.) Radio frequency identification:
Security and privacy issues, vol. 6370, pp. 258–269. Springer, Berlin Heidelberg (2010)
3. Ramos-Calderer, S., Bellini, E., Latorre, J.I., Manzano, M., Mateu, V.: Quantum search for
scaled hash function preimages. , arXiv:2009.00621[quant-ph], September 2020, Consulté le:
Décember 28, 2020
4. Sugier,J.: Implementation efficiency of BLAKE2 cryptographic algorithm in contemporary
popular-grade FPGA devices. In: Kabashkin, I., Yatskiv, I., Prentkovskis, O. (eds.) Reliability
and statistics in transportation and communication, vol. 36, pp. 456–465. Springer International
Publishing, Cham (2018)
5. Rao, V., Prema, K.V.: Comparative study of lightweight hashing functions for resource
constrained devices of IoT. In 2019 4th International Conference on Computational Systems and
Information Technology for Sustainable Solution (CSITSS), Bengaluru, India, pp. 1–5 (2019)
6. Aumasson, J.-P., Neves, S., Wilcox-O’Hearn, Z., Winnerlein, C.: BLAKE2: simpler, smaller,
fast as MD5. In: Jacobson, M., Locasto, M., Mohassel, P., Safavi-Naini, R. (eds) Applied
cryptography and network security, pp. 119–135, vol. 7954. Springer, Berlin, Heidelberg (2013)
7. Aumasson, J.-P., Meier, W., Phan, R.C.-W., Henzen, L: The hash function BLAKE. Springer,
Berlin Heidelberg (2014)
Cryptography Over the Twisted Hessian
3
Curve Ha,d
Abdelâli Grini, Abdelhakim Chillali, and Hakima Mouanis
Abstract In this paper, we will give some properties of the twisted Hessian curve
over the ring Fq [] denoted by Ha,d3
, with Fq is a finite field of order q = pb , where
p is a prime number ≥ 5 and b ∈ N∗ , and we prove that when p doesn’t divide
#(Ha0 ,d0 ), then Ha,d
3
is a direct sum of Ha0 ,d0 and F2q , where Ha0 ,d0 is the twisted
Hessian curve over Fq . Other results are deduced from, we cite the equivalence of
3
the discrete logarithm problem on the twisted Hessian curves Ha,d and Ha0 ,d0 , which
is beneficial for cryptography and cryptanalysis as well, and we give an application
in cryptography.
1 Introduction
In [1], Bernstein et al. introduced the twisted Hessian curves over a field. In [6, 7],
we defined this curve over the ring Fq [], 2 = 0, and in [8] we studied the coding
over twisted Hessian curves over the same ring. In this article, our objective is to
study the twisted Hessian curve defined over the ring Fq [], 3 = 0. The goal of this
work is the search for new groups of points of a twisted Hessian curve over a finite
ring, where the complexity of the discrete logarithm calculation is good for use in
cryptography.
We started this article by studying the arithmetic of the ring Fq [], 3 = 0, where
we establish some useful results which are necessary for the rest of this work. In
the third section, we will define the twisted Hessian curves over Fq [] and we will
3
classify the elements of the twisted Hessian curve Ha,d . Afterwards, we will define
A. Grini (B) · H. Mouanis

Faculty of Science Dhar El Mahraz-Fez, Sidi Mohamed Ben Abdellah University,
P.O. Box 1796, Atlas-Fez, Morocco
e-mail: abdelali.grini@usmba.ac.ma
A. Chillali
Sidi Mohamed Ben Abdellah University, FP, LSI, Taza, Morocco
e-mail: abdelhakim.chillali@usmba.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_25
352 A. Grini et al.
the group law of Ha,d 3 3

and we will show that Ha,d is a direct sum of Ha0 ,d0 and F2q ,
when p doesn’t divide #(Ha0 ,d0 ).
3
Other purpose of this paper is the application of Ha,d in cryptography. Thereby,
3
from Corollary 3 we deduce that the discrete logarithm problem in Ha,d is equivalent
to that in Ha0 ,d0 and #(Ha,d ) = p #(Ha0 ,d0 ).
3 2b
Other cryptographic applications are given in Section 4 which we will establish

3
the coding over twisted Hessian curve Ha,d .
2 Arithmetic Over the Ring Fq [], 3 = 0
Let p be a prime number ≥5 such that −3 is not a square in Fp . We consider the

quotient ring R3 = Fq [X ]/(X 3 ), where Fq is the finite field of characteristic p and
q elements. Then the ring R3 can be identified with the ring Fq [], 3 = 0. In other
words,
R3 = {a + b + c 2 /a, b, c ∈ Fq }.
Now, we will give some results concerning the ring R3 , which are useful for the rest
of this paper.
Let two elements in R3 represented by X = x0 + x1 + x2 2 and Y = y0 + y1 +
y2 2 with coefficients xi and yi are in the field Fq for i = 0, 1, 2.
The arithmetic operations in R3 can be decomposed into operations in Fq , and
they are computed as follows:
X + Y = (x0 + y0 ) + (x1 + y1 ) + (x2 + y2 ) 2
X .Y = (x0 y0 ) + (x1 y0 + x0 y1 ) + (x2 y0 + x1 y1 + x0 y2 ) 2
Similar as in [2, 9, 15], we have the following results:

• (R3 , +, .) is a finite unitary commutative ring.
• R3 is a vector space over Fq and has (1, , 2 ) as basis.
• R3 is a local ring. Its maximal ideal is M = () = Fq .
• An element X = x0 + x1 + x2 2 is invertible in the ring R3 if and only if x0 =
0 mod p. In this case we have:
X −1 = x0−1 + (−x0−2 x1 ) + (x0−3 x12 − x0−2 x2 ) 2
We denote by π the canonical projection defined by
R3 → Fq
a + b + c 2 → a
3
Cryptography Over the Twisted Hessian Curve Ha,d 353
3 Twisted Hessian Curves Over the Ring R3
Definition 1 A twisted Hessian curve over the ring R3 is a curve in the projective
space P2 (R3 ), which is given by the equation: aX 3 + Y 3 + Z 3 = dXYZ, where a, d ∈
R3 and a(27a − d 3 ) is invertible in R3 , and denoted by Ha,d
3
. So we have:
3
Ha,d = {[X : Y : Z] ∈ P2 (R3 )\ aX 3 + Y 3 + Z 3 = dXYZ}.
3
3.1 Classification of Elements of Ha,d
To have a clear idea on the twisted Hessian curves over the ring R3 , we can classify its
elements according to their projective coordinate. This is the subject of the following
proposition.
Proposition 1 Every element in Ha,d 3

is of the form [1 : Y : Z] (where Y or Z ∈
R3 \ M ) or [X : Y : 1] (where X ∈ M ), and we write:
3
Ha,d = {[1 : Y : Z] ∈ P2 (R3 ) \a + Y 3 + Z 3 = dYZ, and Y or Z ∈ R3 \ M } ∪ {[X :
Y : 1] \aX 3 + Y 3 + 1 = dXY , and X ∈ M }.
Proof Let [X : Y : Z] ∈ Ha,d

3
, where X , Y and Z ∈ R3 .
• If X is invertible, [X : Y : Z] = [1 : X −1 Y : X −1 Z] ∼ [1 : Y : Z]. Suppose that
Y and Z ∈ M ; since a + Y 3 + Z 3 = dYZ then a ∈ M , which is absurd. So, Y or
Z ∈ R3 \ M .
• If X is non invertible, then X ∈ M , so X = x1 + x2 2 , where x1 and x2 ∈ Fq . So
we have two cases for Z:
– Z invertible :
[X : Y : Z] = [XZ −1 : YZ −1 : 1] ∼ [X : Y : 1].
– Z non invertible:
We have X and Z ∈ M , since aX 3 + Y 3 + Z 3 = dXYZ, then Y 3 ∈ M and so
Y ∈ M . We deduce that [X : Y : Z] isn’t a projective point since (X , Y , Z) isn’t
a primitive triple [[11], pp. 104–105]
Lemma 1 Let [X : Y : 1] ∈ Ha,d

3
, where X ∈ M .
If X = x1 + x2 , then
2
1 1
[X : Y : 1] = [x1 + x2 2 : −1 − d0 x1 − (d1 x1 + d0 x2 ) 2 : 1].
3 3
Proof Let X = x1 + x2 2 , Y = y0 + y1 + y2 2 , d = d0 + d1 + d2 2 and a =

a0 + a1 + a2 2 .
354 A. Grini et al.
Since [X : Y : 1] ∈ Ha,d
3
, then [X : Y : 1] verifies the equation:
aX 3 + Y 3 + Z 3 = dXYZ (1)
So:
aX 3 + Y 3 + 1 = dXY (2)
implies that:
(a0 + a1 + a2 2 )(x1 + x2 2 )3 + (y0 + y1 + y2 2 )3 + 1

= (d0 + d1 + d2 2 )(x1 + x2 2 )(y0 + y1 + y2 2 )
This means that: ⎧

⎪
⎨y0 = −1
y1 = − 13 d0 x1
⎪
⎩
y2 = − 13 (d1 x1 + d0 x2 ).
Thus [X : Y : Z] = [x1 + x2 2 : −1 − 13 d0 x1 − 13 (d1 x1 + d0 x2 ) 2 : 1].
3
3.2 The Group Law Over Ha,d
3
After classifying the elements of twisted Hessian curve Ha,d , we will define the group
law on it.
We first consider the mapping defined by:
π̃ : 3
Ha,d → Ha0 ,d0
[X : Y : Z] → [π(X ) : π(Y ) : π(Z)]
where Ha0 ,d0 is the twisted Hessian curve over Fq .
3
Then, we are ready to define the group law on Ha,d .
Theorem 1 Let P = [X1 : Y1 : Z1 ] and Q = [X2 : Y2 : Z2 ] two points in Ha,d

3
.
1. Define:
X3 = X12 Y2 Z2 − X22 Y1 Z1 ,
Y3 = Z12 X2 Y2 − Z22 X1 Y1 ,
Z3 = Y12 X2 Z2 − Y22 X1 Z1 .
If π̃ ([X3 : Y3 : Z3 ]) = [0 : 0 : 0] then P + Q = [X3 : Y3 : Z3 ].

3
2. Define:
X3 = Z22 X1 Z1 − Y12 X2 Y2 ,
Y3 = Y22 Y1 Z1 − aX12 X2 Z2 ,
Z3 = aX22 X1 Y1 − Z12 Y2 Z2 .
If π̃ ([X3 : Y3 : Z3 ]) = [0 : 0 : 0] then P + Q = [X3 : Y3 : Z3 ].
Proof By using [1, Theorems 3.2 and 4.2] we prove the theorem.
Corollary 1 (Ha,d
3
, +) is an abelian group with [0 : −1 : 1] as identity element.
3
The group law is now defined on Ha,d , we will give some of its properties and
homomorphisms defined on it.
3.3 The π̃ Homomorphism and Results
Theorem 2 Let a = ã + a2 2 , d = d̃ + d2 2 , X = X̃ + X2 2 , Y = Ỹ + Y2 2 and

Z = Z̃ + Z2 2 be elements of R3 , which verified the equation of twisted Hessian:
aX 3 + Y 3 + Z 3 = dXYZ.
Then
ãX̃ 3 + Ỹ 3 + Z̃ 3 = d̃ X̃ Ỹ Z̃ + (D + AX2 + BY2 + CZ2 ) 2 ,
where
D = d2 X0 Y0 Z0 − a2 X03 ,
A = d0 Y0 Z0 − 3a0 X02 ,
B = d0 X0 Z0 − 3Y02 ,
C = d0 Y0 X0 − 3Z02 .
Proof Let a = ã + a2 2 , d = d̃ + d2 2 , X = X̃ + X2 2 , Y = Ỹ + Y2 2 and Z =

Z̃ + Z2 2 be elements of R3 .
Then
Y 3 = Ỹ 3 + 3Ỹ 2 Y2 2
Z 3 = Z̃ 3 + 3Z̃ 2 Z2 2
aX 3 = ãX̃ 3 + 3ãX̃ 2 X2 2 + a2 X̃ 3 2
dXYZ = d̃ X̃ Ỹ Z̃ + (d2 X̃ Ỹ Z̃ + d̃ X̃ Ỹ Z2 + d̃ X̃ Y2 Z̃ + d̃ Ỹ Z̃X2 ) 2 .
356 A. Grini et al.
Since [X : Y : Z] ∈ Ha,d
3
, then:
aX 3 + Y 3 + Z 3 = dXYZ,
so
ãX̃ 3 + Ỹ 3 + Z̃ 3 = d̃ X̃ Ỹ Z̃ + (d2 X0 Y0 Z0 − a2 X03 + (d0 Y0 Z0 − 3a0 X02 )X2

(d0 X0 Z0 − 3Y02 )Y2 + (d0 Y0 X0 − 3Z02 )Z2 ) 2 ,
thus:
ãX̃ 3 + Ỹ 3 + Z̃ 3 = d̃ X̃ Ỹ Z̃ + (D + AX2 + BY2 + CZ2 ) 2 ,
where,
D = d2 X0 Y0 Z0 − a2 X03 ,
A = d0 Y0 Z0 − 3a0 X02 ,
B = d0 X0 Z0 − 3Y02 ,
C = d0 Y0 X0 − 3Z02 .
Lemma 2 The mapping
π̃ : 3
Ha,d → Ha0 ,d0
[X : Y : Z] → [π(X ) : π(Y ) : π(Z)]
is a surjective homomorphism of groups.
Proof From Theorem 2; π̃ is well defined, and from Theorem 1 we prove that π̃ is
a homomorphism.
Let [X0 : Y0 : Z0 ] ∈ Ha0 ,d0 , then there exists [X : Y : Z] ∈ Ha,d
3
such that π̃([X :
Y : Z]) = [X0 : Y0 : Z0 ].
Indeed, by Theorem 2, we have
D = −(AX2 + BY2 + CZ2 )
Coefficients −A, −B and −C are partial derivative of a function
F(X , Y , Z) = aX 3 + Y 3 + Z 3 − dXYZ
at the point [X0 : Y0 : Z0 ] cannot be all three null. We can then, at last, conclude that
[X2 : Y2 : Z2 ].
Finally, π̃ is a surjective.
3
We define on the set F2q the law ∗ by:
1
(x1 , x2 ) ∗ (x1 , x2 ) = (x1 + x1 , x2 + x2 + d0 x1 x1 )
3
Lemma 3 (F2q , ∗) c (0, 0) as unity.

Lemma 4 Let [X : Y : 1] and [X : Y : 1] ∈ Ha,d
3
, where X , Y , X , Y are as in
Lemma 1, we have:
[X : Y : 1] + [X : Y : 1] = [(x1 + x1 ) + (x2 + x2 ) 2 + 13 d0 x1 x1 2 , −1 − 13 d0 (x1 +

x1 ) − ( 13 d1 (x1 + x1 ) + 13 d0 (x2 + x2 ) + 19 d02 x1 x1 ) 2 , 1]
(3)
Proof By using the Theorem 1, we prove the lemma.

Lemma 5 The subset G = {[X : Y : 1], X ∈ M } is a subgroup of Ha,d
3
and every
element in G, not unity, is of order p.
Proof Let P = [x1 + x2 2 , −1 − 13 d0 x1 − 13 (d1 x1 + d0 x2 ) 2 , 1] ∈ G, we denote
2P = P + P and (n + 1)P = nP + P for all n ≥ 2. We have from Lemma 4:
pP = [px1 + (px2 + p3 d0 x12 ) 2 , −1 − p3 d0 x1 − 13 (pd1 x1 + pd0 x2 + p3 d02 x12 ) 2 , 1]

= [0 : −1 : 1] mod p
So, we claim that pP = [0 : −1 : 1].

Lemma 6 The mapping
φ : (F2q , ∗) → (Ha,d 3
, +)
(x1 , x2 ) → [x1 + x2 2 , −1 − 13 d0 x1 − 13 (d1 x1 + d0 x2 ) 2 , 1]
is an injective homomorphism of groups.
Proof From Lemma 1, we deduce that φ is well defined.

We have φ(0, 0) = [0 : −1 : 1] and for all (x1 , x2 ) and (x1 , x2 ) ∈ F2q we have by
Lemma 4:
φ((x1 , x2 ) ∗ (x1 , x2 )) = [(x1 + x1 ) + (x2 + x2 ) 2 + 13 d0 x1 x1 2 , −1 − 13 d0 (x1 + x1 )

−( 13 d1 (x1 + x1 ) + 13 d0 (x2 + x2 ) + 19 d02 x1 x1 ) 2 , 1]
= [x1 + x2 2 , −1 − 13 d0 x1 − 13 (d1 x1 + d0 x2 ) 2 , 1] + [x1 + x2 2 ,
−1 − 13 d0 x1 − 13 (d1 x1 + d0 x2 ) 2 , 1]
= φ(x1 , x2 ) + φ(x1 , x2 ),
then φ is an homomorphism of groups.

358 A. Grini et al.
It remains to prove that φ is injective. Let (x1 , x2 ) ∈ F2q such that
φ(x1 , x2 ) = [0 : −1 : 1].
Then,
1 1
[x1 + x2 2 , −1 − d0 x1 − (d1 x1 + d0 x2 ) 2 , 1] = [0 : −1 : 1];
3 3
therefore x1 = x2 = 0. This prove that φ is injective.
Lemma 7 Ker π̃ = Im φ.
Proof Let [x1 + x2 2 , −1 − 13 d0 x1 − 13 (d1 x1 + d0 x2 ) 2 , 1] ∈ Imφ, then
1 1
π̃ ([x1 + x2 2 , −1 − d0 x1 − (d1 x1 + d0 x2 ) 2 , 1]) = [0 : −1 : 1]
3 3
and so, Imφ ⊂ Ker π̃ .

Conversely, let [X : Y : Z] ∈ Ker π̃, then
[x0 ; y0 ; z0 ] = [0 : −1 : 1],
so Z is invertible, and from Proposition 1: X ∈ M so, [X : Y : Z] ∼ [X : Y : 1]; and

from Lemma 1
1 1
[X : Y : Z] ∼ [x1 + x2 2 , −1 − d0 x1 − (d1 x1 + d0 x2 ) 2 , 1] ∈ Im φ.
3 3
So Ker π̃ ⊂ Im φ.
Finally: Ker π̃ = Im φ.
From Lemmas 2, 6 and 7, we deduce the following corollary.
Corollary 2 The sequence
i π̃
O Ker π̃ 3
Ha,d Ha0 ,d0 0
3
is a short exact sequence which defines the group extension Ha,d of Ha0 ,d0 by Ker π̃ ,
where i is the canonical injection.
Theorem 3 Let n = #(Ha0 ,d0 ) the cardinality of Ha0 ,d0 . If p doesn’t divide n, then
the short exact sequence:
i π̃
O Ker π̃ 3
Ha,d Ha0 ,d0 0
is split.
3
Proof of Theorem 3. p doesn’t divide n, then exists an integer f such that nf = 1

mod p. So, there is an integer m such that 1 − nf = p m.
Let [1 − nf ] the homomorphism defined by:
[1 − nf ] : Ha,d
3
→ Ha,d3
P → (1 − nf )P
There exists an unique morphism ϕ, such that the following diagram commutes:
3 [1−nf ] 3
Ha,d Ha,d
π̃ ϕ
Ha0 ,d0
Indeed, let P ∈ ker(π̃ ) = Im φ, then: ∃(x1 , x2 ) ∈ F2q such that:
1 1
P = [x1 + x2 2 , −1 − d0 x1 − (d1 x1 + d0 x2 ) 2 , 1].
3 3
We have from Lemma 5:
(1 − nf )P = pmP = [0 : −1 : 1],
then P ∈ ker([1 − nf ]). It follows that ker(π̃ ) ⊆ ker([1 − nf ]), this proves the above
assertion.
Now we prove that π̃ ◦ ϕ = idHa0 ,d0 . Let P ∈ Ha0 ,d0 , since π̃ is surjective, then
there exists a P ∈ Ha,d
3
such that π̃ (P) = P . We have nP = [0 : −1 : 1], then
nπ̃ (P) = [0 : −1 : 1] and π̃ (nP) = [0 : −1 : 1]
implies that nP ∈ ker(π̃ ) and so, nfP ∈ ker(π̃ ); therefore, π̃ (nfP) = [0 : −1 : 1].
Moreover,
ϕ(P ) = (1 − nf )P = P − nfP,
then
π̃ ◦ ϕ(P ) = π̃ (P) − [0 : −1 : 1] = P
and so, π̃ ◦ ϕ = idHa0 ,d0 .

Finally, the sequence is split.
Corollary 3 If p doesn’t divide #(Ha0 ,d0 ) then, Ha,d

3
is isomorphic to Ha0 ,d0 × F2q .
360 A. Grini et al.
Proof From the Theorem 3, the sequence
i π̃
O Ker π̃ 3
Ha,d Ha0 ,d0 0
3 ∼
is split then, Ha,d = Ha0 ,d0 × ker(π̃ ), and since ker(π̃ ) ∼
= Im φ ∼
= F2q , then the corol-
lary is proved.
4 Cryptographic Applications
Let P ∈ Ha,d
3
of order k, we will use the subgroup < P > of Ha,d
3
to encrypt message,
and we denote E = <P>.
4.1 Coding of Elements of E
We will give a code to each element Q = rP ∈ Ha,d 3

, where r ∈ 1, 2, . . . , k.
Let Q = [X : y0 + y1 + y2 : z0 + z1 + z2 ], where yi , zi ∈ Fpb , for i = 0, 1, 2.
2 2
We set:
yi = c0,i + c1,i α + c2,i α 2 + ... + cb−1,i α b−1 = c0,i c1,i c2,i ...c(b−1),i
where α is a primitive root of an irreducible polynomial of degree b over Fp , then,

we code Q as it follows:
• If Q = [1 : y0 + y1 + y2 2 : z0 + z1 + z2 2 ], then Q = 10...0y0 y1 y2 z0 z1 z2

3×b×3
• If Q = [x1 + x2 2 : y0 + y1 + y2 2 : 1], then Q = 0...0x1 x2 y0 y1 y2 1...0

3×b×3
4.2 Exchange of Secret Key
Ali and Badr want exchange the secret key, for this they start publicly with integer
3
b, a twisted Hessian curve Ha,d , a point P ∈ Ha,d
3
of order k and the coding method
over E = <P>:
• Ali chooses a secret random integer sA ∈ [O, k − 1], computes KA = sA P and

sends KA to Badr.
3
• Badr chooses a secret random integer sB ∈ [O, k − 1], computes KB = sB P and

sends KB to Ali.
• Ali computes S = sA KB .
• Badr computes S = sB KA .
• Their secret common key is then S = sB sA P. Choose the p code of point S as a
private key, which transformed on the decimal code SD .
Ali and Badr can encrypt and decrypt the message (m), with the secret key SD such
as the decimal code of point S.
4.3 Twisted Hessian Curve Key Generation Block Diagram
3
The procedure to generate a public key in Ha,d is outlined as follows:
• The users Ali and Badr compute their secret common key S = sB sA P which
transformed on the decimal code SD .
• Encode the message (m) on point Pm ∈ Ha,d3
.
• Choose a random integer t ∈ [0, k − 1] and compute Q = tPm .
• Compute R = SD Q
Then, the public key is {a, d , P, Q, R}, and the private key is {sA , sB , t, SD }.
This operation is shown in Fig. 1.
Fig. 1 Twisted Hessian curve Key generation

362 A. Grini et al.
Fig. 2 Twisted Hessian curve encryption
Fig. 3 Twisted Hessian curve decryption
4.4 Twisted Hessian Curve Encryption Process Block

Diagram
As shown in Fig. 2, to encrypt Pm , a sender chooses an integer h at random and sends

the point (hQ, Pm + hR).
4.5 Twisted Hessian Curve Decryption Process Block

Diagram
To decrypt this message, a receiver multiplies the first component of the received
point by the secret key SD and subtract it from the second component:
(Pm + hR) − SD .hQ = Pm + hSD .Q − SD .hQ = Pm . (4)
This operation is shown in Fig. 3.
5 Conclusion
In this work, we have extended the results of twisted Hessian curves over R3 and
3
we have proved the bijection between Ha,d and Ha0 ,d0 × F2q . In cryptography appli-
3
cations, we have established the coding over twisted Hessian curves Ha,d ; further-
3
3
more, we deduce that the discrete logarithm problem in Ha,d is equivalent to that in
Ha0 ,d0 × Fq and #(Ha,d ) = p #(Ha0 ,d0 ). Our future work will focus on the generalist;
2 3 2b
these studies for all integers n > 3, n = 0, which are beneficial and interesting in
cryptography.
References
1. Bernstein, D.J., Chuengsatiansup C., Kohel D., Lange T.: Twisted Hessian curves. In: Lauter,
K., Rodrguez-Henrquez, F. (eds.) Progress in Cryptology—LATINCRYPT 2015. Lecture Notes
in Computer Science, vol. 9230, pp. 269–294. Springer, Cham (2015). https://doi.org/10.1007/
978-3-319-22174-8_15
2. Chillali, A.: Elliptic curves of the ring Fq [], n = 0. Int. Math. Forum (2011)
3. Chuengsatiansup, C., Martindale, C: Pairing-friendly twisted Hessian curves. In: Chakraborty,
D., Iwata, T. (eds.) Progress in Cryptology INDOCRYPT 2018. INDOCRYPT 2018. Lecture
Notes in Computer Science, vol. 11356. Springer, Cham. https://doi.org/10.1007/978-3-030-
05378-9_13
4. Diffie, W., Hellman, M.: New directions in cryptography. IEEE Trans. Inf. Theory 22(6), 644.
https://doi.org/10.1109/TIT.1976.1055638
5. ElGamal, T.: A public key cryptosystem and a signature scheme based on discrete logarithms.
In: Blakley, G.R., Chaum, D. (eds.) Advances in Cryptology. CRYPTO 1984. Lecture Notes
in Computer Science, vol .196. Springer, Berlin, Heidelberg. https://doi.org/10.1007/3-540-
39568-7_2
6. Grini, A., Chillali, A., Mouanis, H.: The binary operations calculus in Ha,d 2 . Boletim
da Sociedade Paranaense de Matematica. http://www.spm.uem.br/bspm/pdf/next/next.html

(2020)
7. Grini, A., Chillali, A., ElFadil, L., Mouanis, H.: Twisted Hessian curves over the ring Fq [e],
e2 = 0. Int. J. Comput. Aided Eng. Technol. https://www.inderscience.com/info/ingeneral/
forthcoming.php?jcode=ijcaet (2020)
8. Grini, A., Chillali, A., Mouanis, H.: Cryptography over twisted Hessian curves of the ring
Fq [], 2 = 0. Adv. Math.: Sci. J. 10(1), 235–243 (2021). https://doi.org/10.37418/amsj.10.1.
24
9. Hassib, M.H., Chillali, A., Elomary, M.: Abdou: elliptic curve over a chain ring of characteristic
3. In: International Workshop of Algebra and Applications, 2014, FST Fez, Morocco. Journal
of Taibah University for Science (2015). https://doi.org/10.1016/j.jtusci.2015.02.001
10. Joye, M., Quisquater, J.: Hessian elliptic curves and sidechannel attacks. In: Cryptographic
Hardware and Embedded Systems—CHES 2001, Lecture Notes in Computer Science, vol.
2162. Springer, pp. 402–410 (2001) https://doi.org/10.1007/3-540-44709-1_33
11. Lenstra, H.W: Eliptic curves and number-theoretic algorithms. In: Processing of the Interna-
tional Congress of Mathematicians, Berkely, California, USA, (1986)
12. Sahmoudi, M., Chillali, A.: Key Exchange over Particular Algebraic Closure Ring. Tatra Moun-
tains Math. Publ. 70, 151–162 (2017). https://doi.org/10.1515/tmmp-2017-0024
13. Silverman, J.H.: The Arithmetic of Elliptic Curves. GTM, vol. 106. Springer, New York (2009).
https://doi.org/10.1007/978-0-387-09494-6
14. Smart, N.: The Hessian form of an elliptic curve. Cryptographic hardware and embedded
systems-CHES: (Paris), 118–125. Lecture Notes in Computer Science, vol. 2162. Springer,
Berlin (2001)
15. Tadmori, A., Chillali, A., Ziane, M.: Coding over elliptic curves in the ring of characteristic
two. Int. J. Appl. Math. Inf. 8 (2014)
16. Zeriouh, M., Chillali, A., Boua, A.: Cryptography based on the matrices. Boletim da Sociedade
Paranaense de Matematica 37(3), 75–83 (2019). https://doi.org/10.5269/bspm.v37i3.34542
Method for Designing Countermeasures
for Crypto-Ransomware Based
on the NIST CSF
Hector Torres-Calderon , Marco Velasquez , and David Mauricio
Abstract Crypto-ransomware are malicious programs that encrypt the data of an

infected machine, making it a hostage until the owner of the device decides to pay the
fee to recover their information. This has become a complex cybersecurity problem
causing more and more economic damage. Crypto-ransomware has rendered cyber-
security models not adequate since they do not establish specific guidelines for the
design of countermeasures. This paper proposes a method for the design of counter-
measures related to crypto-ransomware attacks based on the NIST 800–53 revision
4 standard and the Information Security Maturity Model published by ISACA in the
COBIT Focus magazine. The model consists of five phases: identify vulnerabilities,
assess vulnerabilities, propose countermeasures, implement countermeasures, and
evaluate countermeasures. This allows an organization to measure its current cyber-
security state, know cybersecurity measures oriented to crypto-ransomware and its
prioritization through criticality indexes in a simple, adaptive and easy to implement
way. A case study in a Peruvian company shows the simplicity and ease of use of
the method, which allows the design of countermeasures with which the level of
cybersecurity can be improved by 55.6%.
1 Introduction
The inclusion of sensitive data in digital platforms within companies has prompted
crime around the world to migrate from a face-to-face to virtual context. Since the 90s,
malicious programs have been developed in order to perform some kind of damage to
organizations in order to benefit the criminals [1]. One of the most popular malicious
programs is crypto-ransomware or encryption ransomware, a malicious program that
encrypts all personal data on the infected machine, holding it hostage until the owner
of the device decides to pay the fee and obtain the means to recover their information
[2].
H. Torres-Calderon (B) · M. Velasquez · D. Mauricio

Universidad Peruana de Ciencias Aplicadas, Prolongación Primavera 2390, Lima, Perú
e-mail: u201510124@upc.edu.pe
https://doi.org/10.1007/978-981-16-3637-0_26
366 H. Torres-Calderon et al.
A well-known crypto-ransomware is SamSam, a highly complex program which

deactivates the protection of computers, allowing information to be encrypted quickly
and subsequently requesting a reward of up to fifty thousand dollars per attack,
achieving an estimated total of 6.5 million dollars between 2016 and 2018 [3]. For
this reason, over time, multiple companies have been forced to invest more and
more in resources for the defense of their information, operational continuity and
reputation [4]. However, 87% of organizations do not have the budget to offer levels
of cybersecurity and resilience that satisfy optimal protection against possible attacks
[4]. In Latin America, the periodic training rates in companies on social engineering
incidents, malicious code incidents, access control to sensitive information reaches
18%, 28%, 20% and 27%, respectively [5].
Various companies and organizations have shown their interest in the common
good in terms of cybersecurity by providing alternative frameworks and standards
in order to mitigate the risk of cyber-attacks. One of them is the National Institute
of Standards and Technology (NIST) which created the Cybersecurity Framework
(CSF), one of the most used frameworks worldwide, created to improve the crit-
ical infrastructure of companies based on a series of recommendations and refer-
ences according to the categorization of different information security functions and
cybersecurity outcomes that allow raising the protection index of companies [6].
However, the CSF does not show the specific guidelines to design countermea-
sures against a specific type of malware, specifically crypto-ransomware, making the
design arduous, requiring more time, money and specialists. This article proposes a
method for the design of countermeasures against crypto-ransomware attacks, which
consists of five steps: test, analysis, diagnosis, recommendations and continuous
improvement. The proposal is based on the CSF, an analysis of intrusion methods,
characteristics of a crypto ransomware, and the COBIT maturity model [7].
The article is organized into 5 sections. In Sect. 2, a literature review on crypto
ransomware and its detection is made. The proposed method is described in Sect. 3,
and its validation through a case study is presented in Sect. 4. Finally, the conclusions
follow in Sect. 5.
2 Literature Review
2.1 Methods for Designing Cybersecurity Countermeasures
The methods for designing cybersecurity countermeasures are composed of a series

of phases that allow users to identify vulnerabilities and security risks in order to
establish countermeasures. In general, these include the following phases: iden-
tify vulnerabilities, evaluate vulnerabilities, propose countermeasures, implement
countermeasures and evaluate countermeasures, as described in Table 1.
An inventory of some methods for designing cybersecurity countermeasures and
their phases are presented in Table 2, where it is observed that most of them contem-
Method for Designing Countermeasures for Crypto-Ransomware … 367
Table 1 Description of the identified phases

Phase Description References
Identify vulnerabilities In this phase, it is explored which elements, mainly [8–13]
vulnerabilities, related to cybersecurity the
organization has
Asses vulnerabilities In this phase, the elements identified later are [1, 8–13]
analyzed and evaluated to determine the impact
they have on the organization
Propose countermeasures After evaluating, countermeasures should be [1, 8–14]
considered to control and mitigate the impact
defined in the previous phase
Implement countermeasures The implementation of the proposed [1, 11, 14]
countermeasures should be planned to depend on
the needs and resources of the organization
Evaluate countermeasures After a period, the effectiveness, efficiency and [14]
efficacy of the implemented countermeasures must
be evaluated in order to identify opportunities for
improvement
Table 2 Phases of the methods studied

Author Identify Asses Propose Implement Evaluate
vulnerabilities vulnerabilities countermeasures countermeasures countermeasures
[1] Understand risks Develop Establish best
adequate practices for
policies users
[14] Implementation Training Evaluation of
of the awareness focused on effectiveness
program countermeasures
[8] Identify Monitor risks Deploy
vulnerabilities countermeasures
[9] Identify Identify the threat Define response
enablers of of tools
change crypto-ransomware
[10] Identify the Identify the impact Define a
risks of a of the cyber attack response
cyber attack strategy
[11] Identify new Explore identified Exploit trends Adapt trends to
trends trends explored business
[12] Identify Measure risks Implement
vulnerabilities countermeasures
[13] Identify Assess security Plan an
processes maturity improvement
plan
plate only three out of the five phases, and only the work of Torten et al. [14] considers
the countermeasures evaluation phase.
2.2 Mechanisms and Cases of Infection
Infection mechanisms are those vectors that are taken advantage of or exploited
by threats to infect a victim’s devices. These seek to violate the confidentiality,
integrity and/or availability of the different information and technology assets. A
group of these intrusion vectors is related to social engineering since they depend on
misinformation and ignorance about information security on the part of their victims.
Some of them are presented in Table 3.
Table 3 Intrusion vectors found

Intrusion vectors Description Articles
Phishing A set of techniques that seeks to deceive a victim [15–21]
through identity theft to manipulate him/her and
make them disclose confidential and sensitive
information
Drive-by download Mode related to the involuntary download of [15, 16, 22]
malicious components that endanger the
integrity, confidentiality or availability of a
device
Malvertising Modality that consists of the use of online [15, 16, 18–20, 22]
advertising for the distribution of malware
through failures or security breaches
Social networks Modality related to the access to unvalidated [15, 18, 20]
malicious links shared through publications on
social networks
Cloud applications Phishing-related modality that consists of the use [15]
of cloud storage applications to share malicious
files
Critical vulnerabilities Set of functionalities or characteristics of a [15, 20–23]
software that have errors or failures not foreseen
by the developer that cause security breaches
waiting to be discovered and exploited
Brute force attacks Trial-and-error-based mode that seeks to find [15, 22]
keys or access vectors by testing all possible
related combinations
Macros A mode that takes advantage of algorithms and [21–23]
autorun programs to infiltrate malicious
programs
3 Method
To solve the problem presented above, we developed a method for the design of
countermeasures against computer attacks of the crypto-ransomware type in order
to allow companies to have a tool to increase the effectiveness of their protection
schemes. The method is based on the NIST CSF to know the cybersecurity outcomes
associated with crypto-ransomware, the NIST standard 800-53 revision 4 [24] to
know how to obtain these outcomes and the published “Information Security Maturity
Model (ISMM)” maturity model by ISACA [7] which is used to measure the level
of implementation of controls in an organization.
Figure 1 shows the five phases that make up the flow of activities related to the
application of the method based on what has been analyzed in the literature review.
These consist of answering the questions that make up the questionnaire, which are
designed considering the controls. Then, the current state must be constructed based
on the responses to the questionnaire, which is stipulated in the controls and the
reference processes. Next, the current state must be contrasted with the desired state
to identify the gap that exists between the maturity levels and the opportunities for
improvement. Later, these are analyzed to define the recommendations to increase
the maturity level of the current state. Next, an implementation plan for the recom-
mendations should be defined and its viability should be analyzed. Depending on the
feasibility, adjustments will have to be made or the implementation should be started.
After a period stipulated by the user, they will have to evaluate the performance of
the countermeasures implemented to determine their effectiveness and efficiency.
3.1 Fundamentals
The method is based on the NIST CSF to know the cybersecurity outcomes associated
with crypto ransomware, the NIST standard 800–53 [6] to know how to obtain
these outcomes and the published “Information Security Maturity Model (ISMM)”
maturity model. by ISACA [7] which is used to measure the level of implementation
of controls in an organization.
We use the NIST CSF because, among cybersecurity frameworks, it best fits the
objectives we seek to achieve with the method:
• Easy to use
• Specialization in cybersecurity
• Flexibility
To verify this, we performed benchmarking (Table 4) where we compared four
widely used cybersecurity frameworks, considering three criteria that cover these
objectives and two predefined criteria associated with their age and maturity. The
criteria associated with the objectives had the same weighting among themselves,
equivalent to the other two criteria. Next, an assessment is used using the Likert scale
Fig. 1 Flow chart of method application
(1: Totally disagree; 2: Disagree, 3: Neither agree nor disagree, 4: Agree, 5: Totally
agree). For example, in usability, the NIST CSF, ISO 27,000 and COBIT 5 are used
and recognized internationally by different organizations and government agencies;
however, they present weights of 5, 5 and 3 respectively because the first two present
ease of use, and the last one because it is more complex despite its details; in IASME
governance, the valuation was 3 because it is little known, and the studies are limited
to the UK.
Table 4 Benchmarking results

Criteria Weight (%) NIST CSF COBIT 5 for SI IASME ISO 27,000
governance family
Usability 25 5 3 3 5
Cybersecurity 25 5 4 5 5
Specialty
Antiquity 12.5 2 5 3 5
Flexibility 25 5 4 3 1
Maturity 12.5 2 5 5 4
Total 4.3 4 3.8 3.9
For the standard used, we decided to use one of the tools provided by the CSF.
This tool is known as the core and provides different bibliographic references that
act as guides on how to achieve the outcomes. In this way, we selected the NIST
800–53 revision 4 standard because it is open access and is published by the same
body that published the framework.
For the maturity levels, we decided to use those proposed by the information
security group or ISG of the largest bank in India as of today [7]. These maturity
levels have been published and approved by ISACA, so we consider that they are the
indicated tools to allow us to evaluate the maturity of the controls published in the
selected standard.
3.2 Cybersecurity Controls
The NIST 800–53 standard initially has 18 families that group approximately 180
controls in total. A first analysis of different case studies was carried out to identify
the main intrusion methods and vulnerabilities related to crypto-ransomware attacks.
With this analysis, we were able to put together a list of 25 vulnerabilities present
in these types of attacks, and we were able to identify the frequency of each of
these. With this information, we were able to select the outcomes of the core of the
NIST framework related to crypto-ransomware attacks, which allowed us to select
the standard controls related to vulnerabilities, reducing the number to 144 controls.
Subsequently, a second analysis was carried out that consisted of identifying
the relationship between the controls and the vulnerabilities in order to know how
important the controls are depending on how many vulnerabilities they cover. In this
way, we were able to assign an expected maturity level to each control depending
on its importance, and it allowed us to reduce the number of controls to the 20 most
important in relation to crypto-ransomware, which are found in Table 5.
To calculate the criticality index shown in Table 5, a study was carried out where
the weight of each of the 25 vulnerabilities mentioned above was determined based
on their frequency. A matrix was used to define which vulnerabilities each control
Table 5. 20 most important controls.

Control Name Criticality Index (%)
IR-6 Incident reporting 100
PL-2 System security plan 100
PM-7 Enterprise architecture 100
RA-3 Risks evaluation 100
IR-5 Incident monitoring 93
PL-8 System security architecture 93
SA-11 Developer security testing and evaluation 91
RA-5 Vulnerability Scanning 90
IR-4 Incident handling 90
PM-8 Critical infrastructure plan 88
CA-7 Continuous monitoring 82
AU-6 Report, analysis and review of audits 79
AU-13 Information disclosure monitoring 79
PM-5 Inventory of information systems 79
SI-4 Information system monitoring 79
SI-5 Security alerts, warnings and policies 79
SI-7 Software, firmware and information integrity 79
CA-8 Penetration testing 70
RA-2 Categorization of security 64
SI-2 Failure remediation 64
covers, where a control can cover several vulnerabilities and a vulnerability can be
covered by several controls. Once we defined the relationship between controls and
vulnerabilities, we proceeded to calculate the criticality index as shown in Eq. (1).
In this, iC is the criticality index, x takes the value of 1 if the control covers the
vulnerability and 0 if not and pV is the weight of the vulnerability.

iC = x × pV (1)
Using the criticality index, the expected maturity levels were determined by
control. For this, we divide 100% by five equal parts and distribute them between the
criticality indexes. So, the expected maturity level is five for controls with a criticality
greater than 80%.
3.3 P1: Answering the Questionnaire
The questionnaire is a document that contains between three to five questions for
each control to determine the current state of the organization and is at https://bit.
ly/3fyJCDa. The questions for control are of the form: Is there control? Is there a
process or tool for reporting? Who is responsible? The questionnaire is applied to
the head of cybersecurity.
3.4 P2: Perform Diagnosis
The objective of the diagnosis is to show the contrast between the current level of
maturity of each control and the desired level of maturity. For this, the degree of
compliance of each control in the organization is analyzed according to the domains
and levels of the maturity model.
For the analysis of the fulfillment of each cybersecurity function, a table is consid-
ered that relates the controls with the functions of the CSF where the frequency with
which the former are cited as a reference is identified and a weight is calculated based
on this. An example is seen in Table 6 for the “Protect” function.
Also, by presenting the diagnosis grouped by the functions of the framework used,
we allow the organization to carry out its own analysis to determine which aspects
are most important to its business and decide which are more prudent to improve
first.
For the protection indexes, five risk scenarios were proposed based on the five
most frequent vulnerabilities identified in the initial analysis of the research. These
scenarios include a description, the information asset violated, the compromised
Table 6 Control weights according to the protect function of the NIST CSF
NIST function Control code Frequency Weight Current state of Target maturity state
maturity
Protect CA-7 2 0.18 5
PL-2 1 0.09 5
PL-8 1 0.09 5
RA-3 1 0.09 5
RA-5 1 0.09 5
SA-11 1 0.09 5
SI-2 1 0.09 4
SI-4 1 0.09 4
SI-7 2 0.18 4
aspect of security, the associated vulnerability, the threat that exploits the vulnera-
bility, the dimension where the scenario impacts and the controls related to managing
the scenario.
With this last variable, it is possible to calculate the protection index of the orga-
nization related to the said scenario according to the level of maturity that the organi-
zation possesses with respect to the related controls. This index is calculated through
the average of the current maturity of the controls related to each scenario contrasted
with the maximum maturity level.
3.5 P3: Define Recommendations
The recommendations provided by the method are based on the maturity levels, and
their objective is to improve the current levels of the organization using the tool.
It is convenient for the organization to review the content of the standard since
it provides the basic guidelines related to each control and a list of improvement
proposals for these. In addition, it is important to note that the maturity levels are
effective in identifying the lack of basic functionalities stipulated by the standard in
relation to the controls; however, it does not take into consideration the proposals for
improvement in the analysis, so these are optional.
3.6 P4: Implement Countermeasures
The implementation is a phase designed to help organizations adopt the recommen-

dations so that they can improve the level of maturity of their controls. It is made up
of three sub-phases.
3.6.1 P4.1 Define the Implementation Plan
The plan consists of 3 deliverables: a work-breakdown structure (WBS) to define the

scope of the implementation, a Gantt chart to define the time, budget and resources
allocated to WBS activities, and a matrix of risks associated with the implementation
that allows the organization to identify risk exposure and a response plan. Subse-
quently, this plan must be reviewed and evaluated to obtain the agreement of the
sponsors.
3.6.2 P4.2 Adjust the Plan
If there is an observation that results in the modification of the previously stipu-

lated plan, this must be managed using formal documentation and the respective
communication channels.
3.6.3 P4.3 Implement Countermeasures
After obtaining the final version of the implementation plan, it must be imple-
mented. Once completed, a period must be stipulated that will define the performance
evaluation of the controls and recommendations implemented.
3.7 P5: Evaluate Countermeasures
After the period defined in phase 4, a work team assigned to evaluate the performance
of the improved controls must be defined. This must determine if the efficiency
and effectiveness of the controls were improved through the implementation of the
recommendations and if this improvement is reflected in the organization’s security
indicators.
4 Validation
The validation of the proposed method was carried out through a case study in a
company in Peru.
4.1 Case Study
For the application of the method, a Latin American organization that specializes
in the trade of construction machinery and mining equipment was selected. This
organization has different companies located in various Latin American countries.
Its headquarters are located in Peru, which has a security area where a security officer,
specialists and experts in information security operate.
The security team that participated in the validation was made up of three special-
ists with an average of 10 years of experience each in auditing and information
security. One of them has been president of a chapter of ISACA and another is a SAP
security specialist.
4.2 Method Application
Regarding the application of the model in the organization presented, we followed

the flow shown in Fig. 1. For the first phase, it began with the request for approval for
the use of company information for the validation of the method. Once obtained, we
proceeded to coordinate meetings by video calls with the work team. During these,
questions from the questionnaire were asked to obtain the different points of view
and to be able to build a current state more faithful to reality.
In the second phase, we used the responses obtained in the meetings to analyze
each control and determine the maturity levels of the associated domains. We then
averaged the maturity levels of each domain to determine the maturity level of the
control. We locate this in Table 6, in the column of current maturity level. Likewise,
the calculation of Table 8, Fig. 2 and the protection indices was carried out. The third
phase consisted of consolidating a report requested by the organization made up of
three main sections:
• Executive summary: Summary of the results obtained and how to interpret the
recommendations.
• Results of the analysis: Report by control on the maturity of the related domains
that allows the analysis on where to begin the improvement of the controls.
• Recommendations: Consolidated of the recommendations provided by the matu-
rity model and the standard improvement proposals.
Due to the nature of the situation generated in the country and the organiza-
tion’s policies, in the fourth phase, it was decided to present recommendations for
implementation in the form of an EDT and a matrix of possible risks.
Fig. 2 Current maturity

level vs. target maturity level
Table 7 Expected maturity

Control Current maturity level Expected maturity level
level vs. current maturity
level per control AU-13 2 4
AU-6 5 4
CA-7 1 5
CA-8 4 4
IR-4 3 5
IR-5 3 5
IR-6 3 5
PL-2 4 5
PL-8 4 5
PM-5 2 4
PM-7 4 5
PM-8 4 5
RA-2 3 4
RA-3 4 5
RA-5 4 5
SA-11 1 5
SI-2 4 4
SI-4 3 4
SI-5 4 4
SI-7 3 4
4.3 Results
The results shown below are the product of processing the information obtained from
the questionnaire. In Table 7, we determined that the organization does not meet the
expected level of maturity in 16 of the 20 controls that were studied. Also, it was
observed that the organization did not meet the expected maturity levels by function
of the NIST CSF as shown in Table 8.
Table 8 Expected maturity

NIST CSF Role Current maturity level Expected maturity
level vs. current maturity
level
level by NIST CSF role
Identify 3.3 4.5
To protect 2.9 4.6
Detect 2.8 4.6
Reply 3.2 4.8
Recover 2.9 5.0

Function MA = control MA × Control weight per function (2)
The calculation shown in Table 8 was given through Eq. (2) using the values shown
in Tables 6 and 7 where “MA” refers to the current maturity level. Additionally, Fig. 2
was constructed to visually contrast the gap between the current maturity level and the
target maturity level and understand approximately how much improvement could be
made in the organization with respect to cybersecurity crypto-ransomware scenarios.
After obtaining and analyzing the diagnosis, a report was prepared and divided into
3 main segments, an executive summary, the results of the analysis and recommenda-
tions based on the results. This report was sent to the employees of the organization
who worked together with us, and an attempt was made to schedule a final meeting
to present the results and obtain feedback.
Likewise, a work-breakdown structure (WBS) was built with a series of activities
and recommended work packages for a subsequent implementation plan. Addition-
ally, a matrix of risks associated with the implementation of the recommendations
was developed.
5 Conclusions
In the present work, a method has been proposed for the design of countermeasures
related to crypto ransomware attacks based on the NIST 800–53 standard and the
Information Security Maturity Model published by COBIT and consists of 5 phases:
identify vulnerabilities, evaluate vulnerabilities, pose countermeasures, implement
countermeasures and evaluate countermeasures. The method allows an organization
to measure its current cybersecurity status, learn about cybersecurity measures in
the form of the 20 most important controls and prioritize these through criticality
indexes in a simple, adaptable and easy way.
The implementation of the proposed method in a Peruvian capital goods company
oriented to the trade of machinery for the agricultural and construction sectors shows
that the proposal is easy to apply. It is agile because it required 15 days of work.
In addition, it identified an intermediate level to more (3.02 out of 5) of maturity
in cybersecurity and that this can be improved up to a 55.6% by implementing
countermeasures.
Among the challenges to be developed, the method must be extended to consider
other malicious programs and cybersecurity standards; likewise, a tool must be
developed that supports the application of the method.
Acknowledgements The authors thank the organization of the case study and the UPC for the
partial funding of the following research.
References
1. Richardson, R., North, M.: Ransomware: Evolution, mitigation and prevention. Int. Manag.
Rev. 13 (2017)
2. Hassan, N.A.: Ransomware Revealed, 1st ed. Springer Science, New York (2019)
3. Toapanta Toapanta, S.M., Mafla Gallegos, L.E., Benavides Quimis, B.S., Huilcapi Subia, D.F.:
Approach to mitigate the cyber-environment risks of a technology platform. Proc 3rd Int.
Conf. Inf. Comput. Technol. ICICT 2020, 390–396 (2020). https://doi.org/10.1109/ICICT5
0521.2020.00069
4. EY: ¿La ciberseguridad es algo más que protección? Quito (2018)
5. Jurado Pruna, F.X., Yarad Jeada, P.V., Carrión Jumbo, J.L.: Análisis de las características del
sector microempresarial en latinoamérica y sus limitantes en la adopción de tecnologías para
la seguridad de la información. Rev. Científica. Ecociencia. 7,1–26 (2020). https://doi.org/10.
21855/ecociencia.71.303
6. NIST: Marco para la mejora de la seguridad cibernética en infraestructuras críticas (2018).
https://doi.org/10.6028/NIST.CSWP.04162018
7. Salvi, V., Kadam, A.W.: Information Security Management at HDFC Bank: Contribution of
Seven Enablers. COBIT Focus 1, 8 (2014)
8. Tolubko, V., Vyshnivskyi, V., Mukhin, V., et al.: Method for determination of cyber threats
based on machine learning for real-time information system. Int. J. Intell. Syst. Appl. 10,
11–18 (2018). https://doi.org/10.5815/ijisa.2018.08.02
9. Connolly, L., Wall, D.S.: The rise of crypto-ransomware in a changing cybercrime landscape:
Taxonomising countermeasures. Comput. Secur. 87, 101568 (2019). https://doi.org/10.1016/j.
cose.2019.101568
10. Mañas-Viniegra, L., Niño González, J.I., Martínez Martínez, L.: Transparency as a reputational
variable of the crisis communication in the media context of wannacry cyberattack. Rev. Comun
la SEECI., 149–171
11. Carayannis, E.G., Grigoroudis, E., Rehman, S.S., Samarakoon, N.: Ambidextrous cybersecu-
rity: The seven pillars (7Ps) of cyber resilience. IEEE Trans. Eng. Manag., 1–12 (2019). https://
doi.org/10.1109/TEM.2019.2909909
12. Rea-Guaman, A.M., Mejía, J., San Feliu, T., Calvo-Manzano, J.A.: AVARCIBER: a framework
for assessing cybersecurity risks. Cluster. Comput. 23, 1827–1843 (2020). https://doi.org/10.
1007/s10586-019-03034-9
13. Al-Matari, O.M.M., Helal, I.M.A., Mazen, S.A., Elhennawy, S.: Adopting security maturity
model to the organizations’ capability model. Egypt. Informatics. J. (2020). https://doi.org/10.
1016/j.eij.2020.08.001
14. Torten, R., Reaiche, C., Boyle, S.: The impact of security awarness on information technology
professionals’ behavior. Comput. Secur. 79, 68–79 (2018). https://doi.org/10.1016/j.cose.2018.
08.007
15. Sood, A.K., Bajpai, P., Enbody, R.: Evidential Study of Ransomware. 5, 1–10 (2018)
16. Ali A (2017) Ransomware: A Research and a Personal Case Study of Dealing with this Nasty
Malware. Issues Informing Sci Inf Technol 14:087–099. https://doi.org/10.28945/3707
17. Sipior, J.C., Bierstaker, J., Borchardt, P., Ward, B.T.: A ransomware case for use in the
classroom. Commun. Assoc. Inf. Syst. 43, 598–614 (2018). https://doi.org/10.17705/1CAIS.
04332
18. Thomas, J.E.: Individual Cyber Security: Empowering Employees to Resist Spear Phishing to
Prevent Identity Theft and Ransomware Attackss. Int. J. Bus. Manag. 13, 1 (2018). https://doi.
org/10.5539/ijbm.v13n6p1
19. Gupta, B.B., Tewari, A., Jain, A.K., Agrawal, D.P.: Fighting against phishing attacks: state of
the art and future challenges. Neural. Comput. Appl. 28, 3629–3654 (2017)
20. Hull, G., John, H., Arief, B.: Ransomware deployment methods and analysis: views from a
predictive model and human responses. Crime. Sci. 8, 2 (2019). https://doi.org/10.1186/s40
163-019-0097-9
21. Patyal, M., Sampalli, S., Ye, Q., Rahman, M.: Multi-layered defense architecture against
ransomware. Int. J. Bus. Cyber. Secur. 1, 52–64 (2017)
22. Wilner, A., Jeffery, A., Lalor, J., et al.: On the social science of ransomware: Technology,
security, and society. Comp. Strateg. 38, 347–370 (2019). https://doi.org/10.1080/01495933.
2019.1633187
23. Watson, F.C.: Petya/NotPetya Why It Is Nastier Than WannaCry and Why We Should Care.
ISACA 6, 1–6 (2017)
24. NIST: NIST Special Publication 800–53: Security and Privacy Controls for Federal Information
Systems and Organizations. NIST SP-800–53 Ar4 400+ (2013). https://doi.org/10.6028/NIST.
SP.800-53Ar4
Comparative Study Between Network
Layer Attacks in Mobile Ad Hoc
Networks
Oussama Sbai and Mohamed Elboukhari
Abstract Amid the most recent decade, a few research endeavors have explored
developing Internet of Things (IoT) and Mobile Ad hoc Networks (MANETs) appli-
cation situations in a new concept called IoT-MANETs. One of the constraints of these
applications is the security of the communication between the nodes. In this article,
we analyze and compare a simulation result of the impact of DATA Flooding, Link-
spoofing and Replay attacks with Optimized Link State Routing Protocol (OLSR)
routing protocol (RFC 3626) and DATA Flooding, RREQ Flooding and HELLO
Flooding attacks with Ad hoc On-Demand Distance Vector (AODV) routing protocol
(RFC 3561) on using ns-3 simulator. In this comparison, we took into consideration
the density of the network by the number of nodes included in the network, the
speed of the nodes, the mobility model, and even, we chose the IEEE 802.11ac wire-
less local-area network (WLAN) standard for the MAC layer, in order to have a
simulation, which deals with more general and more real scenarios.
1 Introduction
Mobile Ad Hoc Networks (MANETs) alludes to a class of wireless networks that

can be shaped dynamically and randomly without the need for infrastructural setups.
Such networks are capable of adapt and reconfigure themselves at the fly in keeping
with node mobility and changing network topologies [1]. Each wireless node can
be a simple device or even a router. Once the node is a simple device it could send
messages to any specified destination node through some route or perhaps receive
messages from other nodes. When the node functions as a router, it can relay the
packet to the destination or next router in the route. Whenever important, each node
can buffer packets awaiting transmission [2]. Present and future MANET applications
cover a variety of areas [1]:
O. Sbai (B) · M. Elboukhari

Department of Applied Engineering, ESTO (Higher School of Technology), Mohammed 1st
University, Oujda, Morocco
e-mail: o.sbai@ump.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_27
382 O. Sbai and M. Elboukhari
Fig. 1 The fields of application of MANETs
• Home applications: smart wireless sensor nodes and actuators embedded in

consumer electronics.
• Environmental applications including tracking movements of animals, chem-
ical/biological detection, precision agriculture, tracking of weather and earth
activities.
• Body area networks (BAN).
• Tactical networks (military communications and automated battlefields) (Fig. 1).
This article is a continuation of the work started in the following articles [3–6].
We will compare and analyze the impact of attacks targeting the OLSR protocol
[7] of one party: Link-spoofing and Replay and DATA Flooding attacks, and the
AODV protocol [8] of another party: DATA Flooding, RREQ Flooding, and HELLO
Flooding attacks. In the scenarios simulated for each protocol, we tried to set the
simulation parameters most appropriate to the use case of each protocol, and close
to one of the real cases.
The organization of the paper is as follows: Sect. 2 presents and defines the
different attacks studied and their implementation protocols; also, this section cites
some related work. The following section presents the experimental environment
set-up (Simulation time, IEEE 802.11ac, mobility model, etc.). Performance metrics
are presented in Sect. 4. Sections 5 and 6 present simulation comparison of Link-
spoofing, Replay, DATA Flooding, RREQ Flooding, and HELLO Flooding attacks
in ns-3 simulator. Finally, a conclusion is given in the last section.
Comparative Study Between Network Layer Attacks … 383
Fig. 2 Link-spoofing attack [3]
2 Background
2.1 Link-Spoofing Attack in MANETs
This attack is implemented in the network using the OLSR protocol. The link-
spoofing malicious nodes claim that they are its one-hop neighbor node, by the injec-
tion of wrong information about non-existent nodes or (and) insert false information
about neighbors using Hello messages [9, 10] (Fig. 2).
2.2 Replay Attack in MANETs
When the networks using OLSR protocol, infected by the replay attack, the attackers
replicate expired TC messages and forward them to the network’s nodes afterwards.
With the nature of MANETs: network topology on permanent changing, this
control packet inject by malicious nodes will be expired. Therefore, the routing
tables of nodes will be updated on using wrong information, so the routing operation
is disrupted (Fig. 3).
2.3 HELLO Flooding Attack in MANETs
In AODV protocol, the HELLO-Flooding malicious nodes modify the value of

HELLO_INTERVAL on reducing the interval time in HELLO-message sanded
to their nodes neighbors so that they can inject a grand number of hello packets.
Fig. 3 TC message format [7]
Therefore, the operation of other nodes in the network will be negatively impacted.
[11].
2.4 RREQ Flooding Attack in MANETs
The RREQ Flooding Attack is dedicated to the reactive routing protocols in

MANETs. In this attack, a malicious node injects a huge number of RREQ packets
to an unknown destination (non-existent IP address in the network) (Fig. 4).
Therefore, no node in the network will respond with a RREP packet, and the
DATA packets cannot be transmitted, because the network gets saturated by the
flooded RREQs packets.
Fig. 4 Route Request (RREQ) message format [8]

In addition, the route table of the intermediate nodes will overflow, so that no
new RREQ packet can be received, and in result, the network services are denied
(Denial-of-Service attack).
Besides, unnecessarily forwarding these fake RREQs packets causes a serious
loss of node resources like energy and bandwidth.
2.5 DATA Flooding Attack in MANETs
The malicious DATA Flooding node(s) send or inject in network a great volume of
futile DATA packets. The immoderate injected DATA packets engorge the network,
what makes the communication between the nodes of the network cannot be
completed, because the available network bandwidth will be exhausted.
The destination node gets submerged and cannot work normally, because of the
excessive packets generated by the malicious node(s).
2.6 Related Works
In our previous works [3, 5, 6], we studied and analyzed the performance of AODV
and OLSR protocols in MANET, the malicious nodes are introduced in the network to
analyses the impact of attacks on network’s performances. We used ns-3 simulation
tool, by taking into consideration the use cases of each protocol by a specific config-
uration of the simulation scenarios. The presence of malicious nodes in the network
decreases the performance of this last, and we compare the packet delivery ratio,
routing overhead, normalized routing load, and end-to-end delay with and without
the presence of the attacker node in the network.
In this research [12], the authors had analyzed three types of routing protocols in
MANET witch are AODV, ZRP [13] and LAR [13], by implementing a DDoS using
CBR traffic flooding. The simulator tool using id ns-2. Additionally, the efficiency
of routing protocols was analyzed using performance parameters of the throughput
and end-to-end delay.
In this paper [14], the authors had focused their study on Blackhole, Grayhole,
and Wormehole attacks in OLSR protocol. And in [15], the authors had analyzed the
impact of Selfishness attack on AODV and DSDV protocols, on using ns-2 simulator,
The network’s performance metric used are throughput, packet delivery ratio, and
average delay.
Table 1 Simulation parameters

Parameter Value
Dimension 1000 × 1000 m
Simulation time 60 s
Number of packet 60
Packet size 56 bytes
Mobility model Random Way point Mobility Model
MAC Layer IEEE 802.11ac
Mobility speed Uniform Random Variable [Min = 0|Max = 30]
Protocol AODV OLSR
Attack Link-spoofing DATA Flooding
Replay RREQ Flooding
DATA Flooding HELLO Flooding
Pause Constant Random Variable [Constant Constant Random Variable [Constant
= 0] = 30]
3 Experimental Environment Set Up
In this section, we discuss the experimental environment setup and the values taken
for different parameters in ns-3 simulator [16]. The network area consists of 1000 m
× 1000 m where nodes are randomly distributed and mobile on used the Random
Waypoint Mobility Model to have more general node mobility [17].
The simulation is started during 60 s, on sending one packet per second. We use
the protocol IEEE 802.11ac in MAC layer (recent version of 802.11n) thanks to the
advantage (higher data throughput, high capacity, low latency, efficient use of power,
etc.) defined in [18].
The main experimental parameters used are presented in Table 1 and all
experimental value refers to average values of experiments.
4 Performance Metrics
4.1 Packet Delivery Rate
The Packet delivery rate (PDR) is the division of the total number of packets received
by the destination (PR), by the total number of packets sent by the source (PS),
multiply 100%.

n
PRj
j=1
PDR = × 100 (1)
n
P Si
i=1
4.2 Normalized Routing Load
Normalized Routing Load (NRL) represents the total number of control packets (C),
divided by the total number of DATA packets received by the destination (PR).

n
Cj
j=1
NRL = (2)

n
PRj
j=1
5 Comparison Between Types of Flooding Attack
Figures 5 and 6 presents a comparison’s PDR between the implementation of single

and multiple of three types of Flooding attack, the results show that the impact
Fig. 5 Single malicious node’s PDR simulation result

Fig. 6 Multiple malicious nodes PDR simulation result
of DATA Flooding attack that is more negative than HELLO Flooding and RREQ
Flooding attacks for the two cases.
Figures 7 and 8 presents a comparison’s NRL between the implementation of three

types of Flooding attack in of single and multiple attacker, these results show that,
the NRL of DATA Flooding attack is superior than normal network and HELLO
Flooding and RREQ Flooding attacks for the two cases (single and multiple attack).
Therefore, the DATA flooding attack is more dangerous than other attacks.
Fig. 7 Single malicious node’s NRL simulation result

Fig. 8 Multiple malicious nodes NRL simulation result
6 Comparison Between Link-Spoofing, DATA Flooding

and Replay Attacks
Figures 9 and 10 presents a comparison’s PDR between the implementation of single

and multiple Link-spoofing, DATA Flooding and Replay attacks, the results show
that the three attacks have the same negative impact on network with single malicious
node, but for multiple malicious nodes, the simulation results show that the DATA
Flooding attack is more disruptive than others.
Fig. 9 Single malicious node’s PDR simulation result

Fig. 10 Multiple malicious nodes PDR simulation result
Figures 11 and 12 presents a comparison’s NRL between the implementation of

single and multiple Link-spoofing, DATA Flooding and Replay attacks, the results
show that, the NRL of both Replay and Link-spoofing attacks in single malicious
node has approximately the same impact on network, and the DATA flooding attack
is the least vulnerable. In case of multiple attack, the results indicate that the DATA
Flooding attack is the most defective of the network then others attack.
Fig. 11 Single malicious node’s NRL simulation result

Fig. 12 Multiple malicious nodes NRL simulation result
7 Conclusion
In this paper, we have compared the impact of RREQ Flooding, HELLO Flooding,
and DATA Flooding attacks in AODV routing protocols, and Link-spoofing, DATA
Flooding, and Replay attacks OLSR routing protocols against MANETs. These
attacks have been implemented in ns-3 simulator.
After comparison between the three derivatives of Flooding attack, DATA
Flooding attack is more vulnerable than others attacks: low PDR and high NRL;
and RREQ Flooding attack has less negative impact than others. For the difference
between the Link-spoofing, DATA Flooding, and Replay attacks, we conclude that
the three attacks have approximately the same negative impact on network’s perfor-
mances in single malicious node case, but multiple malicious node case, the DATA
Flooding attack is the most defective and disruptive of the network.
Consequently, the RREQ Flooding, HELLO Flooding, DATA Flooding, Link-
spoofing, and Replay attacks have a higher significant effect on the network
performance.
References
1. Datta, R., Marchang, N.: Security for mobile Ad Hoc Networks. In: Handbook on securing
cyber-physical critical infrastructure. Elsevier, pp. 147–190 (2012)
2. Fazeldehkordi, O.A., Elahe, A.„ Sadegh, I.,: Effect Of Black Hole Attack On AODV Routing
Protocol In MANET. A study of black hole attack solutions: On aodv routing protocol in manet.
Syngress, vol. 4333, March (2015)
3. Sbai, O., Elboukhari, M.: A simulation analysis of MANET’s link-spoofing and replay attacks
with ns-3. In Proceedings of the 4th International Conference on Smart City Applications—
SCA ’19, pp. 1–5 (2019)
4. Sbai, O., Elboukhari, M.: Simulation of MANET’s Single and Multiple Blackhole Attack with
NS-3. Colloq. Inf. Sci. Technol. Cist. 2018, 612–617 (2018)
5. Sbai, O., Elboukhari, M.: A Simulation Analyses of MANET’s Attacks Against OLSR Protocol
with ns-3. In: Ben Ahmed, M., Boudhir, A.A., Santos, D., El Aroussi, M. (eds.) Innovations in
smart cities applications edition 3, pp. 605–618. Springer International Publishing, Cham
6. Sbai, O., Elboukhari, M.: A simulation analyse of MANET’s RREQ flooding and HELLO
flooding attacks with ns-3, pp. 1–5 (2019)
7. Clausen, T., Jacquet, P.: RFC3626: Optimized link state routing protocol (OLSR). RFC Editor
(2003)
8. Perkins, C., Belding-Royer, E., Das, S.: RFC3561: Ad hoc on-demand distance vector (AODV)
routing. RFC Editor (2003)
9. Jeon, Y., Kim, T.H., Kim, Y., Kim, J.: LT-OLSR: Attack-tolerant OLSR against link spoofing,
pp. 216–219. Proc. Conf. Local Comput. Networks, LCN (2012)
10. Desai, V.: Performance evaluation of OLSR protocol in MANET under the influence of routing
attack, pp. 138–143 (2014)
11. Madhavi, S., Duraiswamy, K.: Flooding attack aware secure AODV. J. Comput. Sci. 9(1),
105–113 (2013)
12. Abdelhaq, M., et al.: The resistance of routing protocols against DDOS attack in MANET. Int.
J. Electr. Comput. Eng. 10(5), 4844–4852 (2020)
13. Vinet, L., Zhedanov, A.: A ‘missing’ family of classical orthogonal polynomials. Antimicrob.
Agents Chemother. 58(12), 7250–7257 (2010)
14. Bhuvaneswari, R., Ramachandran, R.: Denial of service attack solution in OLSR based manet
by varying number of fictitious nodes. Cluster Comput. 22(S5), 12689–12699 (2019)
15. Abdelhaq, M., et al.: The impact of selfishness attack on mobile ad hoc network. Int. J. Commun.
Networks Inf. Secur. 12(1), 42–46 (2020)
16. Kristiansen, S.: Ns-3 Tutorial (2010)
17. Nisar, M.A., Mehmood, A., Nadeem, A., Ahsan, K., Sarim, M.: A two dimensional performance
analysis of mobility models for MANETs and VANETs, 3(5), 94–103 (2014)
18. Paper, T.W., 802.11ac: The Fifth Generation of Wi-Fi, March, pp. 1–25 (2014)
Security of Deep Learning Models in 5G
Networks: Proposition of Security
Assessment Process
Asmaa Ftaimi and Tomader Mazri
Abstract 5G networks bring a new design paradigm that will revolutionize telecom-
munications and other sectors, such as industry 4.0, smart cities, and autonomous
vehicles. However, with the inherent advantages, many challenges can emerge.
Today, the community is increasingly motivated to address these challenges by lever-
aging deep learning models to improve the 5G end-user quality of experience (QoE).
However, this alternative approach would exhibit network assets to a series of security
threats that could compromise the availability, integrity, and privacy of 5G architec-
tures. This paper will extensively examine the vulnerabilities in the 5G models to draw
the community’s attention to the threats they may involve when integrated without
sufficient prevention. Its main contribution lies in the comprehensive and adaptive
approach it proposes to appraise deep learning models’ security by assessing and
managing the vulnerabilities they may present before their implementation in 5G
network architectures.
1 Introduction
With the arrival of smart cities, connected objects, and augmented reality, several
constraints have emerged. Indeed, the 4G network can no longer meet today’s needs
in throughput, latency, communication reliability, and connected objects’ density.
Hence, the arrival of 5G networks has taken over to offer a distributed and flexible
architecture providing high performances.
5G is a new generation of mobile networks offering performance far beyond the
one provided by the LTE network [1]. Nevertheless, several techniques are needed
to deal with the complicated challenges associated with 5G services, e.g., to cope
with the massive amount of data traffic carried in the 5G network, the classical
traffic management and resource allocation features must be deployed automatically
[2]. Thus, 5G network components require the involvement of high analysis and
A. Ftaimi (B) · T. Mazri

Laboratory of Advanced Systems Engineering, Ibn Tofail Science University, Kenitra, Morocco
https://doi.org/10.1007/978-981-16-3637-0_28
394 A. Ftaimi and T. Mazri
prediction capabilities and must be equipped with intelligence to ensure the necessary
performance, hence the new trend of deploying deep learning models in 5G networks.
Deep learning models have received increasing interest in the research commu-
nity. Their ability to predict, personalize, and detect behavior has prompted several
researchers to integrate them into various applications. Their deployment in 5G
networks is no exception, and the results obtained to date are promising, whether in
the optimization of resource allocation, orchestration between different NFVs and
SDNs, prediction of traffic and user mobility according to user behavior and histor-
ical data, and automation of complex and laborious tasks. Deep learning has enabled
the transformation of 5G into cognitive networks with high adaptability to end-user
demand to ensure a better quality of user experience [3].
Nevertheless, several studies have been conducted to examine the potential vulner-
abilities that may emerge from the expressiveness and the flexibility aspect of deep
learning models [4]. It has been shown that deep learning models enfold flaws that
can be harnessed by attackers to carry out illicit activities. Indeed, Papernot et al.
[5] have carried out an experiment to test machine learning models’ resilience to
adversarial attacks. They have successfully performed malicious attacks toward deep
learning models without needing specific access to the model’s features or dataset.
The obtained unexpected results have motivated the research community to focus on
the field of adversarial learning to develop secure and robust deep learning models
[6].
Although, the research community has been interested in deep learning models’
security without examining this aspect in a specific domain of application. However,
this approach can be misleading when assessing the impact of attacks and their
robustness [7, 8]. For instance, the analysis of the impact of attacks targeting deep
learning models cannot be dissociated from the system’s criticality built on this
model. Indeed, a less robust attack targeting a model exploited in e-health or 5G
network is extremely critical and has a huge impact compared to a highly robust
attack aiming models used in video games. Hence, the importance of this work
focuses on examining the security of deep learning models in their application in the
5G network.
In this paper, we will first present an overview of 5G networks and the innovative
features they import. Then we will discuss the challenges associated with these
functionalities in terms of management and computational and storage resources
and the contribution of deep learning models in this area as well as the efficient and
powerful solutions they provide. Afterward, we will study the challenges in terms
of security resulting from artificial intelligence integration in 5G networks. We will
scrutinize the vulnerabilities that are enfolded in deep learning models. Finally, we
will propose a method to evaluate the models’ security before incorporating them in
the 5G network components.
Security of Deep Learning Models in 5G Networks … 395
2 Overview of 5G Architecture
Today, 5G offers a variety of services designed to meet the needs of several sectors.
Indeed, besides the classic telephony and file transfer service, 5G provides the
connectivity of smart cities and autonomous vehicles, connected objects, augmented
reality, and the 4.0 industry. All these market segments require different qualities of
service. Hence, experts have defined key performance indicators to make a taxonomy
of the envisaged 5G services. As shown in Fig. 1, in addition to reliability, eight other
key performance indicators have been used to highlight the 5G market’s different
needs clearly. However, meeting all these criteria would be unfeasible, which explains
the approach followed by the 5G network that considers a polymorphic system to
define several declinations; each can fulfill a certain number of constraints while
using a well-determined configuration of the 5G network. The three declinations of
the 5G network are as follow.
Fig. 1. 5G KPI’s and declinations

2.1 Enhanced Mobile BroadBand (eMBB)
This configuration was specified in release 15, which was produced in 2018 during
the first phase of the 3GPP standardization process [9]. In this case, user throughput
is privileged over latency, the density of terminals, and reliability. This declination
allows peak data rate up to 20 Gbps, and user experienced data rate up to 100 Mbit/s.
It was designed to satisfy the need for high throughput required in Mobile Cloud
Computing, smart devices, and UHD streaming applications.
2.2 Massive Machine Type Communication (mMTC)
It was specified during 2020 in release 16. This configuration is not as concerned
with peak data rate and traffic capacity but rather with the density of terminals to be
managed and energy efficiency. It allows one million connections per kilometer and
a battery lifetime of ten years. It was proposed to offer an adequate solution to the
expanding number of connected devices in smart homes, smart cities, e-health, and
wearables [10].
2.3 Ultra-Reliable Low Latency Communication (URLLC)
This configuration is also specified in release 16. In this case, the focus is no longer
on the throughput and density of the terminals but mainly on the network’s reliability,
mobility, and latency. URLLC can reach a latency of one millisecond and reliability
of 10–9 error rate. This configuration is more adapted to smart vehicles and industrial
automation.
5G is empowered by several technologies. Nevertheless, above all, its revolu-
tionary approach comes from its reliance on an important trio: Cloud computing,
virtualization, and softwarization [2]. The conventional architecture of the access
network and the core network is no longer maintained in the 5G network [11, 12].
Instead, the 5G network’s cloudification has allowed the merging of the system’s
control and management functionalities and has brought the computing capacities
closer to the end-user to meet his needs in terms of latency and throughput, and
availability at the edge of the network [1]. In parallel to network cloudification,
softwarization is a new approach implemented in 5G networks to deliver great flex-
ibility and high networking components’ adaptability [13]. It involves employing
software programs to provide the functionalities offered by network equipment and
services. This approach requires a complete redesign of the networks [14]. However,
it provides many advantages for simplifying complex management operations, such
as coordination and load balancing between network components [2].
Virtualization has contributed significantly to the transformation of physical nodes

performing complex operations into software blocks that can be easily clustered and
combined to create sophisticated network operations [15]. In this way, Network
Function Virtualization (NFV) technology alleviates the dependency that has always
existed on hardware components in the network to ensure improved flexibility and
high network functionality and efficiency [16]. However, these technologies are
highly correlated in the 5G network. They require powerful computing resources
to orchestrate between NFVs, optimize resources and automate complex tasks effi-
ciently, hence the crucial necessity to harness deep learning networks’ potential to
fulfill such tasks.
3 Deep Learning Models Applications in 5G
Deep learning models significantly exceed many of the traditional methods of

automation and prediction. They are characterized by their great potential for
generalization and their high expressivity, enabling them to substitute conventional
approaches in several disciplines such as telecoms and networking. In the following
sub-sections, we highlight several potentials of deep learning models and the signif-
icant profits they can deliver to 5G networks, prominently in terms of predicting
behaviors, helping in optimizing resources, and detecting events in the network [17]
as illustrated in Fig. 2.
Fig. 2 Deep learning applications in 5G

3.1 Deep Learning model’s Capabilities of Forecasting in 5G
In the next paragraphs, we will explore the predictive potentials and capabilities
offered by deep learning models in 5G networks.
Traffic Prediction. With the growing volume of managed traffic in the 5G
network, traffic prediction becomes crucial to optimizing resource allocation. The
role of DL in such a context is primordial. Several research projects have focused
on traffic prediction through the deployment of deep learning models for network
slicing [18, 19] and congestion prevention [20]. In [21–23], the authors adopted an
interesting approach by examining the temporal and spatial properties of end-users
to predict the traffic they generate in the network.
Handover Prediction. In telecom networks, the handover is an essential process
to guarantee service quality while ensuring mobility for network users. The handover
process consists in swiping the user from one base station to another based on prede-
fined values of the Reference Signal Receiver Power (RSRP), the Reference Signal
Receiver Quality (RSRQ). However, this process is not as simple as that. Errors can
occur during handovers, and communications can sometimes be interrupted, hence
using deep learning to facilitate this operation [24, 25]. According to the work of
Khunteta et al. [26], a well-trained deep learning model can predict the success or
failure rate of the next handover based on previous experiences and several data
collected in the network. Ozturk et al. [27] used a deep learning model to predict the
likelihood of a handover occurrence to prepare the network in advance.
Device Location Prediction. Predicting the location of users is an important
feature in telecom networks. This functionality appears to be complex in 5G networks
due to users’ high mobility. However, using deep learning models can overcome these
challenges as, in most cases, user movements are generally characterized by a high
degree of repeatability. End-users frequently attend the same places very often [28].
Deep learning models can leverage this aspect of repeatability to establish a pattern of
end-user mobility behavior and predict future locations with high accuracy [29, 30].
In this way, the involvement of DL models in location prediction can significantly
enhance services depending on end-user location, such as computational and storage
resource management [31, 32] and handover prediction [33].
3.2 Deep Learning Model’s Capabilities of Optimization

and Adaptability in 5G
In the following paragraphs, we will outline the power of deep learning models in
helping to resolve optimization problems in 5G network components by adapting
their dimensioning according to the end-user’s requirements.
Optimization of Resource Management and Allocation. 5G is deeply based
on network slicing technology. The latter leverages the network infrastructure to
satisfy the various needs and requirements of multiple market segments. Gutterman
et al. [34] have examined the utilization of DL models to forecast resource allo-
cation measures to align each network slice with the intended usage category. Yan
et al. [35] have demonstrated the relevance of DL in historical data processing to
help in decision making and therefore optimizing resource allocation planning for
a network slice following Service Level Agreement (SLA) requirements. In [36], a
deep learning model was proposed to minimize energy consumption during resource
allocation for network slices. In this study, the authors approached the subject of
energy consumption by employing a DL model and considering several parame-
ters such as channel characteristics, end-user requests, duration of communication,
etc. Ahmed et al. [37] have studied the integration of a DL model in a multi-cell
network to achieve maximum throughput while helping in optimizing the allocation
of resources in the network regarding the CSI channel and the user’s location.
Optimization of Beamforming. Maksymyuk et al. [38] studied the use of a
reinforcement learning model in the beamforming technology that marks the 5G
network. The model evaluates each antenna’s required phases and amplitudes in the
MIMO technology to estimate the signal coverage according to the users’ location.
When users are scattered in separated areas, strong signal coverage is needed to reach
the maximum of the end devices. When they are concentrated in the same location,
the signal needs to be focused and aimed to reach them all.
3.3 Deep Learning model’s Capabilities of Event Detection

in 5G
This section highlights the key applications of deep learning in event detection,
especially regarding anomalies and failures that may occur in 5G networks.
Anomaly Detection. Anomaly detection is an important aspect that must be
optimized in a 5G network to ensure its service continuity. It consists in detecting
malicious activities occurring in the network and eventually identifying where the
problem is originating. However, the network’s heterogeneity and the large volume
of traffic it handles contribute considerably to the complexity of detecting anomalies
in the 5G network [39, 40]. Several works have proposed the introduction of deep
learning models for the detection of cyber-attacks. Many researchers have designed
models to extract normal traffic properties to detect malicious threats [41]. Others
have rather focused on learning normal network behavior to alert in case of antenna
failure or network congestion [41, 42].
Fault Detection. Fault detection is important to increase network performance
and reduce latency. Ensuring fault detection is critical for the URLLC component
of the 5G network. This mission is both vital and complex because when a failure
occurs in an antenna, for example, the fault must first be detected, and then the source
of the problem needs to be located. This task seems simple; however, it hides a great
complexity, especially in 5G networks because of the large number of antennas used
and the employed equipment’s heterogeneity [43, 44]. Thus, Chen et al. [45] proposed
a method based on a neural network model to detect and localize antenna faults in
mmWave systems in 5G networks.
After delving into the potentials delivered by deep learning models when
employed in 5G networks, we will present in the next section the challenges that may
be met when incorporating Artificial intelligence in 5G. We will mainly focus on the
security aspects and the vulnerabilities that can reside in deep learning models with
an extensive study of the impact of adversarial attacks targeting deep learning-built
systems in 5G networks.
4 Vulnerabilities in Deep Learning Models Employed in 5G

Networks
Today deep learning networks have attracted growing interest from the research
community. Several implementations of DL networks have been developed. While
some have already been implemented, others are being proposed. However, various
works have examined these models’ security aspects and have shown that they are
vulnerable [46]. Indeed, several studies have demonstrated that such models could be
targeted by adversarial attacks using carefully and precisely designed perturbation
injected into the dataset to mislead the model [47].
Several attack strategies have been developed to assess machine learning models’
robustness toward adversarial manipulations in this context. Papernot et al. [48] have
conducted adversarial attacks on machine learning models requiring zero accessi-
bility to the model’s dataset or features following a Jacobian saliency map-based
attack (JSMA). In 2013, Szegedy et al. [46] have examined adversarial attacks in
deep neural networks by introducing noise into the input dataset and have succeeded
in reducing the model’s classification meticulousness. Goodfellow et al. [49] have
proposed a new approach known as Fast Gradient Sign Method to generate adversarial
examples, Carlini and Wagner [50] and Liu et al. [51] have suggested optimization-
based methods to craft noisy perturbations that could be utilized in attacks targeting
machine learning models.
In [52], Usama et al. have carried out a white-box Carlini and Wagner attack [50]
against a CNN model used in an end-to-end Modulation classification system. They
used the L2 norm constraint as a metric of perturbations introduced into the input
dataset. They easily succeeded in diminishing the classifier accuracy dramatically.
Their findings have revealed the severity of the deep learning models’ vulnerability
in 5G networks. Usama et al. [52] have also examined the fragility of the Channel
Autoencoder unsupervised model deployed in 5G architectures. They introduced
an additive white Gaussian noise (AWGN) as an adversarial disturbance and have
witnessed a high drop in the model accuracy indicated by a large expansion of block
error rate (BLER). The conducted study has highlighted the flaws embedded in the
unsupervised autoencoders model, which have drastically compromised its integrity.
Besides, Goutay et al. [53] have developed an attack against autoencoders using
deep reinforcement learning models with noisy feedback. In an attempt to approxi-
mate the real scenario, the researchers have performed a black-box attack where no
knowledge of the targeted system is assumed, using a substitution model instead and
taking advantage of the adversarial examples’ transferability property. This approach
is founded on multiple studies that have highlighted the similarity in several models’
behavior against additive perturbations to the input data set, even if they appear to
be different. Usama et al. [52] have succeeded in dropping the model’s accuracy rate
from 95 to 80%, thereby reducing its confidence in its results.
Suomalainen et al. [54] have also scrutinized the criticality of harnessing inherent
vulnerabilities in deep learning models to cause large-scale damage to 5G networks.
Indeed, deep learning models can be leveraged in the load management of network
resources. The importance of such functionality is high since it accommodates the
end-users’ demand without compromising the optimal and efficient use of network
resources [55]. Load balancing combines criticality and complexity as it requires
resource orchestration and planning, traffic prioritization, classification, and predic-
tion [56]. However, Jani Suomalainen et al. have suggested that attacker risks carrying
out a DOS attack toward load balancing models and influences them to redirect traffic
to certain resources, causing an overload of some components while others remain
unused.
In addition to the aforementioned use cases, we have encountered several other
scenarios describing the harness of inherent vulnerabilities and flows in deep learning
models incorporated in 5G network. Therefore, we propose in the section below
to develop a taxonomy of groups and categories of different threats identified in
these models. Indeed, system security’s classical approach inspired us to design
classification of threats that may reside in deep learning models applied in 5G. This
classification is founded on three key categories that are confidentiality, integrity,
and availability as illustrated in Fig. 3.
1. Confidentiality threats: Regularly, this class of threat involves obtaining unau-
thorized access to information transmitted through the network by an adversary.
This menace can lead to harmful effects, including leakage of sensitive informa-
tion such as revealing information about end-user behavior or divulging critical
data. An opponent can potentially corrupt the model and even escalate privileges
to gain unauthorized access to network resources.
2. Availability threats: Attackers can jeopardize network availability by performing
denial of service attacks, either by causing network congestion or by overloading
network infrastructure components. The opponent can also conduct denial of
detection to prevent the network from detecting failures, allowing the attacker
to interrupt its normal operation.
3. Integrity threats: This threat is essentially related to traffic interception and the
modification of data transmitted in 5G networks. Indeed, the attacker could
inject carefully crafted infinitesimal perturbations in the traffic to mislead the
model. Other opponents possessing strong capabilities can completely alter the
model’s behavior by influencing its decisions.
Fig. 3 Threats in deep learning model in 5G architectures
Therefore, we can assume that several scenarios exist. An attacker can leverage
the vulnerabilities inherent in deep learning models to accomplish a malicious intent
threatening the availability, integrity, or confidentiality of 5G network services.
Consequently, it has become apparent that deep learning models’.
5 Proposition of Security Assessment Process for Deep

Learning Models
The integration of deep learning models in 5G networks provides optimal solutions

to the challenges confronted in such innovative generation of telecom networks.
However, it also imposes additional challenges related to their security and resilience
to adversarial attacks. Indeed, there is no doubt about the automation and forecasting
benefits that the application of deep learning in the 5G network can bring. Never-
theless, it is noteworthy that the 5G network would be exploited in a wide range of
application areas, including e-health, autonomous vehicle, smart homes, and many
other extremely critical structures. Hence, the necessity to examine their security
very carefully to avoid unexpected security incidents. So far, there are no reliable
method to evaluate the security of machine learning models. For this reason, we have
proposed in this section a new approach to assess deep learning models’ security and
their robustness toward adversarial attacks before they can be applied in 5G archi-
tectures. Our approach, built on our previous works in [57], proposes a process of
vulnerabilities assessment and management articulated around three essential steps
to obtain a comprehensive evaluation of the model from a security perspective as
shown in Fig. 4.
Fig. 4 The process of Assessment and management of deep learning model vulnerabilities
This process is initiated with the identification of existing vulnerabilities in the

model under consideration. In this framework, a security test of vulnerabilities is
highly recommended. According to Tian-yang et al. [58], security vulnerabilities
testing involves a real simulation of attack scenarios that the attacker may follow.
In this case, acting as an attacker will permit us to observe the system through the
adversary’s eyes and thus will help to examine the system effectively and identify the
different vulnerabilities that could potentially compromise the model and succeed
the attack. In this regard, we strongly recommend testing several types of attacks
among those existing in the literature, with a particular focus on the most robust
and adaptative ones. This step leads to the identification of corrupted assets and
successful attacks that can exploit them.
Following the identification of vulnerabilities and the attacks that could potentially
harness them, a work of expurgation needs to be performed. Indeed, an estimation of
the impact of successful attacks and their success rate is necessary and can contribute
considerably to the prioritization of vulnerabilities. In this step, the criticality of the
assets containing the vulnerabilities can significantly influence the decision to be
taken regarding the attack’s impact. The assessment must consider three important
criteria: the success rate of the attack, its impact, and the criticality of the assets
affected by the attack. The job is rounded off by drawing up a final prioritized list of
detected breaches.
The process concludes with the application of the necessary corrections to reme-
diate previously detected critical flaws. The choice of mitigation techniques must
be based primarily on their effectiveness against the attacks being tested. After
completing the application of fixes, a testing operation is required to ensure the
effectiveness of the chosen mitigation methods in the context of the tested model.
This could be accomplished by repeating the first step of the process. The evaluation
may be subjected to several iterations before converging to a secure model.
6 Conclusion
The aspect of adaptability and generality distinguishes the approach we have

proposed in this paper. Indeed, all the process steps can be implemented in any model
and deal with any flaws and attacks. Moreover, its adaptability derives from its effi-
ciency when deployed for any strategies followed by attackers. Among the positive
grants of this approach, it is essential to mention its contribution to integrating secu-
rity in the model implementation cycle in the 5G network architectures resources.
The results obtained during the implementation of the vulnerability assessment and
management pattern are realistic. They are extremely approaching the scenarios that
can be encountered during the production stage because all the data used during
the test and the vulnerability scan are consistent with the one transmitted over the
network. The attacks carried out during the identification of vulnerabilities simulate
the attacker’s behavior and approximate the real world’s security incidents. Finally,
in addition to its apparent flexibility, this process designed as an iterative wheel adds
a great advantage to this approach since it allows continuous improvement of the
models to enhance their security and reduce their flaws.
References
1. Chang, C.-Y., Nikaein, N.: Cloudification and slicing in 5G radio access network, http://www.
theses.fr/2018SORUS293/document (2018)
2. Barakabitze, A.A., Ahmad, A., Mijumbi, R., Hines, A.: 5G network slicing using SDN and
NFV: A survey of taxonomy, architectures and future challenges. Comput. Netw. 167, 106984
(2020). https://doi.org/10.1016/j.comnet.2019.106984
3. Santos, G.L., Endo, P.T., Sadok, D., Kelner, J.: When 5G meets deep learning: A systematic
review. Algorithms. 13, 208 (2020). https://doi.org/10.3390/a13090208
4. Barreno, M., Nelson, B., Sears, R., Joseph, A.D., Tygar, J.D.: Can machine learning be secure?
In: Proceedings of the 2006 ACM Symposium on Information, computer and communications
security - ASIACCS ’06. p. 16. ACM Press, Taipei, Taiwan (2006). https://doi.org/10.1145/
1128817.1128824.
5. Papernot, N., McDaniel, P., Goodfellow, I., Jha, S., Celik, Z.B., Swami, A.: Practical Black-Box
Attacks against Machine Learning. arXiv:1602.02697 [cs] (2017)
6. Huang, L., Joseph, A.D., Nelson, B., Rubinstein, B.I.P., Tygar, J.D.: Adversarial machine
learning. In: Proceedings of the ACM Conference on Computer and Communications Security.
pp. 43–57. ACM Press, New York, New York, USA (2011). https://doi.org/10.1145/2046684.
2046692
7. Chakraborty, A., Alam, M., Dey, V., Chattopadhyay, A., Mukhopadhyay, D.: Adversarial
Attacks and Defences: A Survey. arXiv:1810.00069 [cs, stat] (2018)
8. Mello, F.L. de: A survey on machine learning adversarial attacks. J. Inf. Secur. Cryptogr. 7,
1–7 (2020). https://doi.org/10.17648/jisc.v7i1.76
9. Tani, N.: IoT-driven evolution and business innovation. NTT DOCOMO Technical J. 19, 82
(2018)
10. Sabella, D., Vaillant, A., Kuure, P., Rauschenbach, U., Giust, F.: Mobile-Edge Computing
Architecture: The role of MEC in the Internet of Things. IEEE Consumer Electron. Mag. 5,
84–91 (2016). https://doi.org/10.1109/MCE.2016.2590118
11. Kekki, S., Featherstone, W., Fang, Y., Kuure, P., Li, A., Ranjan, A., Purkayastha, D., Jiangping,
F., Frydman, D., Verin, G., Wen, K.-W., Kim, K., Arora, R., Odgers, A., Contreras, L.M.,
Scarpina, S.: ETSI White Paper No. 28 MEC in 5G networks (2018)
12. Hassan, N., Yau, K.-L.A., Wu, C.: Edge computing in 5G: A review. IEEE Access. 7, 127276–
127289 (2019). https://doi.org/10.1109/ACCESS.2019.2938534
13. Sayadi, B., Gramaglia, M., Friderikos, V., von Hugo, D., Arnold, P., Alberi-Morel, M.-L.,
Puente, M.A., Sciancalepore, V., Digon, I., Crippa, M.R.: SDN for 5G Mobile networks:
NORMA perspective. In: Noguet, D., Moessner, K., and Palicot, J. (eds.) Cognitive radio
oriented wireless networks. pp. 741–753. Springer International Publishing, Cham (2016).
https://doi.org/10.1007/978-3-319-40352-6_61.
14. Trivisonno, R., Guerzoni, R., Vaishnavi, I., Soldani, D.: SDN-based 5G mobile networks: archi-
tecture, functions, procedures and backward compatibility: SDN-based 5G mobile networks:
architecture, functions, procedures and backward compatibility. Trans. Emerging Tel. Tech.
26, 82–92 (2015). https://doi.org/10.1002/ett.2915
15. Giannoulakis, I., Kafetzakis, E., Xylouris, G., Gardikis, G., Kourtis, A.: On the Applications of
Efficient NFV Management Towards 5G Networking. In: Proceedings of the 1st International
Conference on 5G for Ubiquitous Connectivity. ICST, Levi, Finland (2014). https://doi.org/10.
4108/icst.5gu.2014.258133.
16. Siddiqui, M.S., Escalona, E., Trouva, E., Kourtis, M.A., Kritharidis, D., Katsaros, K., Spirou,
S., Canales, C., Lorenzo, M.: Policy based virtualised security architecture for SDN/NFV
enabled 5G access networks. In: 2016 IEEE Conference on Network Function Virtualization
and Software Defined Networks (NFV-SDN). pp. 44–49. IEEE, Palo Alto, CA (2016). https://
doi.org/10.1109/NFV-SDN.2016.7919474
17. McClellan, M., Cervelló-Pastor, C., Sallent, S.: Deep Learning at the Mobile Edge: Opportu-
nities for 5G Networks. Appl. Sci. 10, 4735 (2020). https://doi.org/10.3390/app10144735
18. Bega, D., Gramaglia, M., Fiore, M., Banchs, A., Costa-Perez, X.: DeepCog: Cognitive Network
Management in Sliced 5G Networks with Deep Learning. In: IEEE INFOCOM 2019 - IEEE
Conference on Computer Communications. pp. 280–288. IEEE, Paris, France (2019). https://
doi.org/10.1109/INFOCOM.2019.8737488
19. Guo, Q., Gu, R., Wang, Z., Zhao, T., Ji, Y., Kong, J., Gour, R., Jue, J.P.: Proactive Dynamic
Network Slicing with Deep Learning Based Short-Term Traffic Prediction for 5G Transport
Network. In: Optical Fiber Communication Conference (OFC) 2019. p. W3J.3. OSA, San
Diego, CA (2019). https://doi.org/10.1364/OFC.2019.W3J.3
20. Zhou, Y., Fadlullah, ZMd., Mao, B., Kato, N.: A Deep-Learning-Based Radio Resource Assign-
ment Technique for 5G Ultra Dense Networks. IEEE Network 32, 28–34 (2018). https://doi.
org/10.1109/MNET.2018.1800085
21. Chen, L., Yang, D., Zhang, D., Wang, C., Li, J., Nguyen, T.-M.-T.: Deep mobile traffic forecast
and complementary base station clustering for C-RAN optimization. J. Netw. Comput. Appl.
121, 59–69 (2018). https://doi.org/10.1016/j.jnca.2018.07.015
22. Huang, C.-W., Chiang, C.-T., Li, Q.: A study of deep learning networks on mobile traffic
forecasting. In: 2017 IEEE 28th Annual International Symposium on Personal, Indoor, and
Mobile Radio Communications (PIMRC). pp. 1–6. IEEE, Montreal, QC (2017). https://doi.
org/10.1109/PIMRC.2017.8292737
23. Zhang, C., Zhang, H., Yuan, D., Zhang, M.: Citywide Cellular Traffic Prediction Based on
Densely Connected Convolutional Neural Networks. IEEE Commun. Lett. 22, 1656–1659
(2018). https://doi.org/10.1109/LCOMM.2018.2841832
24. Hosny, K.M., Khashaba, M.M., Khedr, W.I., Amer, F.A.: New vertical handover prediction
schemes for LTE-WLAN heterogeneous networks. PLoS ONE 14, e0215334 (2019). https://
doi.org/10.1371/journal.pone.0215334
25. Svahn, C., Sysoev, O., Cirkic, M., Gunnarsson, F., Berglund, J.: Inter-Frequency Radio Signal
Quality Prediction for Handover, Evaluated in 3GPP LTE. In: 2019 IEEE 89th Vehicular Tech-
nology Conference (VTC2019-Spring). pp. 1–5. IEEE, Kuala Lumpur, Malaysia (2019). https://
doi.org/10.1109/VTCSpring.2019.8746369
26. Khunteta, S., Chavva, A.K.R.: Deep Learning Based Link Failure Mitigation. In: 2017 16th
IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 806–811.
IEEE, Cancun, Mexico (2017). https://doi.org/10.1109/ICMLA.2017.00-58
27. Ozturk, M., Gogate, M., Onireti, O., Adeel, A., Hussain, A., Imran, M.A.: A novel deep
learning driven, low-cost mobility prediction approach for 5G cellular networks: The case
of the Control/Data Separation Architecture (CDSA). Neurocomputing 358, 479–489 (2019).
https://doi.org/10.1016/j.neucom.2019.01.031
28. Xiong, H., Zhang, D., Zhang, D., Gauthier, V., Yang, K., Becker, M.: MPaaS: Mobility predic-
tion as a service in telecom cloud. Inf Syst Front. 16, 59–75 (2014). https://doi.org/10.1007/
s10796-013-9476-z
29. Cheng, Y., Qiao, Y., Yang, J.: An improved Markov method for prediction of user mobility. In:
2016 12th International Conference on Network and Service Management (CNSM). pp. 394–
399. IEEE, Montreal, QC, Canada (2016). https://doi.org/10.1109/CNSM.2016.7818454
30. Qiao, Y., Yang, J., He, H., Cheng, Y., Ma, Z.: User location prediction with energy efficiency
model in the Long Term-Evolution network: User location prediction with energy efficiency
model. Int. J. Commun. Syst. 29, 2169–2187 (2016). https://doi.org/10.1002/dac.2909
31. Gante, J., Falcao, G., Sousa, L.: Beamformed Fingerprint Learning for Accurate Millimeter
Wave Positioning. In: 2018 IEEE 88th Vehicular Technology Conference (VTC-Fall). pp. 1–5.
IEEE, Chicago, IL, USA (2018). https://doi.org/10.1109/VTCFall.2018.8690987
32. Gante, J., Falcão, G., Sousa, L.: Deep Learning Architectures for Accurate Millimeter Wave
Positioning in 5G. Neural Process Lett. 51, 487–514 (2020). https://doi.org/10.1007/s11063-
019-10073-1
33. Wang, C., Zhao, Z., Sun, Q., Zhang, H.: Deep Learning-Based Intelligent Dual Connectivity for
Mobility Management in Dense Network. In: 2018 IEEE 88th Vehicular Technology Confer-
ence (VTC-Fall). pp. 1–5. IEEE, Chicago, IL, USA (2018). https://doi.org/10.1109/VTCFall.
2018.8690554
34. Gutterman, C., Grinshpun, E., Sharma, S., Zussman, G.: RAn resource usage prediction for a
5G slice broker. In: Proceedings of the International Symposium on Mobile Ad Hoc Networking
and Computing (MobiHoc). pp. 231–240. Association for Computing Machinery, New York,
NY, USA (2019). https://doi.org/10.1145/3323679.3326521
35. Yan, M., Feng, G., Zhou, J., Sun, Y., Liang, Y.-C.: Intelligent Resource Scheduling for 5G
Radio Access Network Slicing. IEEE Trans. Veh. Technol. 68, 7691–7703 (2019). https://doi.
org/10.1109/TVT.2019.2922668
36. Luo, J., Tang, J., So, D.K.C., Chen, G., Cumanan, K., Chambers, J.A.: A Deep Learning-
Based Approach to Power Minimization in Multi-Carrier NOMA With SWIPT. IEEE Access.
7, 17450–17460 (2019). https://doi.org/10.1109/ACCESS.2019.2895201
37. Ahmed, K.I., Tabassum, H., Hossain, E.: Deep Learning for Radio Resource Allocation in
Multi-Cell Networks. IEEE Network 33, 188–195 (2019). https://doi.org/10.1109/MNET.
2019.1900029
38. Maksymyuk, T., Gazda, J., Yaremko, O., Nevinskiy, D.: Deep Learning Based Massive MIMO
Beamforming for 5G Mobile Network. In: 2018 IEEE 4th International Symposium on Wireless
Systems within the International Conferences on Intelligent Data Acquisition and Advanced
Computing Systems (IDAACS-SWS). pp. 241–244. IEEE, Lviv (2018). https://doi.org/10.
1109/IDAACS-SWS.2018.8525802
39. Fernandez Maimo, L., Perales Gomez, A.L., Garcia Clemente, F.J., Gil Perez, M., Martinez
Perez, G.: A Self-Adaptive Deep Learning-Based System for Anomaly Detection in 5G
Networks. IEEE Access. 6, 7700–7712 (2018). https://doi.org/10.1109/ACCESS.2018.280
3446
40. Parwez, M.S., Rawat, D.B., Garuba, M.: Big Data Analytics for User-Activity Analysis and
User-Anomaly Detection in Mobile Wireless Network. IEEE Trans. Ind. Inf. 13, 2058–2065
(2017). https://doi.org/10.1109/TII.2017.2650206
41. Fernández Maimó, L., Huertas Celdrán, A., Gil Pérez, M., García Clemente, F.J., Martínez
Pérez, G.: Dynamic management of a deep learning-based anomaly detection system for 5G
networks. J. Ambient. Intell. Human Comput. 10, 3083–3097 (2019). https://doi.org/10.1007/
s12652-018-0813-4
42. Hussain, B., Du, Q., Zhang, S., Imran, A., Imran, M.A.: Mobile Edge Computing-Based Data-
Driven Deep Learning Framework for Anomaly Detection. IEEE Access. 7, 137656–137667
(2019). https://doi.org/10.1109/ACCESS.2019.2942485
43. Hu, P., Zhang, J.: 5G-Enabled Fault Detection and Diagnostics: How Do We Achieve Effi-
ciency? IEEE Internet Things J. 7, 3267–3281 (2020). https://doi.org/10.1109/JIOT.2020.296
5034
44. Yu, A., Yang, H., Yao, Q., Li, Y., Guo, H., Peng, T., Li, H., Zhang, J.: Accurate Fault Location
Using Deep Belief Network for Optical Fronthaul Networks in 5G and Beyond. IEEE Access.
7, 77932–77943 (2019). https://doi.org/10.1109/ACCESS.2019.2921329
45. Chen, K., Wang, W., Chen, X., Yin, H.: Deep Learning Based Antenna Array Fault Detection.
In: 2019 IEEE 89th Vehicular Technology Conference (VTC2019-Spring). pp. 1–5. IEEE,
Kuala Lumpur, Malaysia (2019). https://doi.org/10.1109/VTCSpring.2019.8746510
46. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.:
Intriguing properties of neural networks. arXiv:1312.6199 [cs]. (2014)
47. Ibitoye, O., Abou-Khamis, R., Matrawy, A., Shafiq, M.O.: The Threat of Adversarial Attacks
on Machine Learning in Network Security—A Survey. arXiv:1911.02621 [cs]. (2020)
48. Papernot, N., McDaniel, P., Sinha, A., Wellman, M.: Towards the Science of Security and
Privacy in Machine Learning. arXiv:1611.03814 [cs]. (2016)
49. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and Harnessing Adversarial Examples.
arXiv:1412.6572 [cs, stat]. (2015)
50. Carlini, N., Wagner, D.: Towards Evaluating the Robustness of Neural Networks. In: 2017
IEEE Symposium on Security and Privacy (SP). pp. 39–57. IEEE, San Jose, CA, USA (2017).
https://doi.org/10.1109/SP.2017.49
51. Liu, Y., Ma, S., Aafer, Y., Lee, W.-C., Zhai, J., Wang, W., Zhang, X.: Trojaning Attack on
Neural Networks. Department of Computer Science Technical Reports. (2017).
52. Usama, M., Mitra, R.N., Ilahi, I., Qadir, J., Marina, M.K.: Examining Machine Learning for
5G and Beyond through an Adversarial Lens. arXiv:2009.02473 [cs]. (2020)
53. Goutay, M., Aoudia, F.A., Hoydis, J.: Deep Reinforcement Learning Autoencoder with Noisy
Feedback. arXiv:1810.05419 [cs, math]. (2019).
54. Suomalainen, J., Juhola, A., Shahabuddin, S., Mammela, A., Ahmad, I.: Machine Learning
Threatens 5G Security. IEEE Access. 8, 190822–190842 (2020). https://doi.org/10.1109/ACC
ESS.2020.3031966
55. Le, L.-V., Sinh, D., Lin, B.-S.P., Tung, L.-P.: Applying Big Data, Machine Learning, and
SDN/NFV to 5G Traffic Clustering, Forecasting, and Management. In: 2018 4th IEEE Confer-
ence on Network Softwarization and Workshops (NetSoft). pp. 168–176. IEEE, Montreal, QC
(2018). https://doi.org/10.1109/NETSOFT.2018.8460129.
56. Zhang, S., Zhang, N., Zhou, S., Gong, J., Niu, Z., Xuemin, Shen: Energy-Sustainable Traffic
Steering for 5G Mobile Networks. arXiv:1705.06663 [cs, math]. (2017).
57. Ftaimi, A., Mazri, T.: Analysis of Security of Machine Learning and a proposition of assessment
pattern to deal with adversarial attacks. E3S Web Conf. 229, 1004 (2021). https://doi.org/10.
1051/e3sconf/202122901004.
58. Tian-yang, G., Yin-sheng, S., You-yuan, F.: Research on Software Security Testing. Interna-
tional Journal of Computer and Information Engineering. 4, 9 (2010). https://doi.org/10.5281/
zenodo.1081389.
Effects of Jamming Attack
on the Internet of Things
Imane Kerrakchou , Sara Chadli , Mohammed Saber ,

and Mohammed Ghaouth Belkasmi
Abstract The capacity of the Internet to be connected to everything, everywhere,

and at every time, thus offering great ease to users, has gained popularity over the
years. The automation of daily tasks is no longer a dream. The Internet of things
has reached a peak, where a wealth of gadgets and applications are successfully
connecting to sensors and the Internet, offering a new age of relief and communica-
tions. The heterogeneous combination of sensors, things, and the Internet together
has resulted in a variety of protection and privacy problems, making security difficult.
Since IoT is used for commercial purposes, it has paved the way for attackers and
hackers to enter the private lives of the public at large, as well as to organizing crimes
and attacks. In this paper, we focus on wireless sensor networks (WSN) because
malicious nodes easily attack these networks. We perform an implementation of the
Jamming attack and a detailed analysis of the results obtained. The protocol chosen
to evaluate the performance of the WSN is the S-MAC protocol. Different scenarios
are proposed in order to evaluate the performance of an attacked network and the
severity of the Jamming attack.
1 Introduction
Today, we live in a rapidly changing world. Rightly so, because technological

advances made what might once have been a fairy tale a reality. Internet of things
I. Kerrakchou (B) · S. Chadli · M. Saber · M. G. Belkasmi

Mohammed First University Oujda, ENSA Oujda, SmartICT Lab, Oujda, Morocco
e-mail: i.kerrakchou@ump.ac.ma
S. Chadli
e-mail: s.chadli@ump.ac.ma
M. Saber
e-mail: m.saber@ump.ac.ma
M. G. Belkasmi
e-mail: m.belkasmi@ump.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_29
410 I. Kerrakchou et al.
(IoT) is a new technology that will make human life smarter and much simpler.
Utilizing IoT, a user can monitor everything and everywhere to accommodate his or
her convenience. Smart objects can connect through many heterogeneous network
technologies like wireless local area networks (WLAN), radio frequency identifica-
tion (RFID), cellular services (3G, 4G, LTE, and 5G), and wireless sensor networks
(WSN) [1]. The WSN name is a three-word combination. These sensors imitate
human sensors. In other words, they are able to collect information on sound, sight,
smell, and temperature. Wireless sensor networks are composed of sensor nodes with
a low cost, limited capacity, and communicate with other nodes over short distances
utilizing a significantly low rate of power. WSNs are usually randomly scattered in
the destination area, and they execute a combined strategy to send predefined param-
eters from the environment to the target field. Due to their large variety of uses,
they could be utilized in many different applications varying from the military fields
to the health system. In the majority of cases, sensor nodes operate in fairly diffi-
cult environments and thus present a higher risk of physical harm than conventional
networks. The vulnerabilities of sensor networks in terms of security can be further
exploited to generate many different forms of threats [2]. For this reason, the secu-
rity of the WSN must achieve several objectives such as confidentiality, availability,
authenticity, integrity, etc., in order to protect the network against all types of attacks.
The attacks known in sensor networks can be divided into various categories based
on various taxonomic models. In this article, the wireless sensor network attacks are
classified according to the several layers of OSI whose function and operations are
affected, damaged, or destroyed. Attacks at the MAC layer have generated a great
deal of attention, and researches exist in this regard because the MAC protocol has a
significant impact on network performance, such as throughput, energy, etc. Attacks
on the MAC layer are primarily aimed to achieve denial of service (DoS). Attackers
typically aim to limit the access of authorized users to the wireless canal by affecting
the operation of the system thus affecting the availability and connectivity of the
network. They can also aim to unfairly consume the canal resources, causing severe
damages in the real world, where the attacker follows the media access control
(MAC) protocol and sends over the common canal, periodically or continuously, to
respectively target all the communications. The following is a brief description of
certain types of attacks for the three layers (application layer, network layer, and
perception layer). Then, we will implement a Jamming type attack that targets the
perception layer, more precisely the MAC layer using the S-MAC protocol, in order
to visualize the effects of this attack on the network performance [3].
The remainder of this paper is presented as follows. A detailed description of the
protocol used (S-MAC) is provided in Sect. 2. Attacks on IoT and their classifications
are presented in Sect. 3. In Sect. 4, we simulate and analyze the effects of the attack
on network performance. Finally, the conclusion will be presented in Sect. 5.
Effects of Jamming Attack on the Internet of Things 411
2 MAC Protocols for WSN
2.1 MAC Protocols
MAC protocols enter into play once a node is authorized to send its packet in order
to avoid collisions on the receiver end and to control any physical layer access
consistently. When a physical canal is used by all the nodes in a network, the MAC
protocols are decisive to ensure that access to the canal is controlled and coordinated
between the different nodes so that the information could be transmitted between
nodes. Different MAC protocols with various goals are suggested to support wireless
sensor networks. The primary objective of the MAC protocol on the focus of the
authors is energy efficiency. Other major attributes are extensibility and the ability
to adapt to changes such as node density, network size, and topology. The second
objective is latency and throughput. The MAC layer provides two primary criteria
for the design of an effective MAC protocol. The first is to identify the causes of
energy loss and the second is the communications model that the network uses to
decide the traffic behavior that the MAC protocol will process [4]. There are many
MAC protocols, including S-MAC, T-MAC, D-MAC, X-MAC, B-MAC, and others.
2.2 S-MAC Protocol
Sensor-MAC is the first protocol in the MAC layer for sensor networks that takes
into account the energy limits of battery-powered sensor nodes. The principal idea
to conserve energy is to switch off the radio transmitters when no pertinent commu-
nication is happening. In periodic listening/sleeping cycle Tc (see Fig. 1), the nodes
exchanging synchronization packets (SYNC) in the period of synchronization Tsync ,
Fig. 1 Frame for S-MAC protocol

sending data packets during the data period Td and turning off the radio in the sleeping
period Ts , i.e., Tc = Ta + Ts with the period of activity Ta = Tsync + Td .
If a node decides to transmit data, it must first transmit a request (RTS) while the
receivers are listening. Once the receiver responds with a CTS packet, the node sends
the data. In the end, an ACK will be sent by the receiver to indicate that reception is
established with success. In determining an energy-efficient listening/sleeping cycle
when entering a network zone, the node will first listen to its neighbors for a while.
If a SYNC packet is received, the node will follow the listen/sleep cycle defined by
the SYNC packet. This node is named follower. If the node does not receive a SYNC
packet, it selects its proper listen/sleep cycle and begins to broadcast SYNC packets.
This node is named synchronizer (see Fig. 2). Each node establishes a schedule table
that contains the schedules of all neighboring nodes [5].
3 Attacks on IoT
In this part, we will describe the basic architecture of an IoT system that typically
includes three layers: perception layer, network layer, and application layer. Then,
we are going to address some of the common issues and threats related to all of those
layers (see Fig. 3).
3.1 Attacks Perception Layer
This layer is often identified as the layer of a sensor. With the support of different
technologies such as wireless sensor network (WSN), RFID sensor network (RSN),
and radio frequency identification (RFID), this layer is responsible for identifying
things, capturing, and collecting data from the sensors [6]. These sensors are selected
according to application specifications. The data that these sensors collect can be
about the position, air changes, temperature, vibration, acceleration, etc. The percep-
tion layer is vulnerable to different types of attacks that affect nodes in the sensors.
Popular forms of these attacks are described below:
– Eavesdropping attack: IoT systems are mostly, made up of multiple nodes
distributed in open environments. As a consequence, certain IoT implementations
are subject to eavesdroppers. Attackers can listen and capture data throughout
various processes, such as authentication or data transmission.
– Jamming attack: It is a DoS type attack that can be very dangerous. By simply
transmitting an interference signal, the attacker will essentially interrupt commu-
nication on a wireless canal, stop normal service, create performance failures, and
even destroy the control system.
– Booting attack: All services of security are activated while the system is in
operating mode. But, at the moment of start-up, there’s a window for the attackers
Fig. 2 Diagram of S-MAC protocol
to target the nodes of the system. Devices that are low energy consumption have
consistent sleep–wake cycles making them more vulnerable to this attack [7].
Fig. 3 Layered classification of IoT attacks
3.2 Attacks Network Layer
Network layer is often referred to as the access gateway layer or transport layer. The
main objective of this layer includes handling information with message routing,
subscription and message publishing management, data transmission, etc. By way
of differing communication canals such as GSM, Wi-Fi, Ethernet, etc., the data is
obtained from the layer below it. Current access technologies and protocols, such as
IPV6/ IPV4/ LoWPAN, are used by the network layer [8]. Among the most popular
types of network attacks, we can cite:
– Sybil attack: On this, one malicious node takes different identities (defined as
Sybil node) and localizes itself at various positions within the network. This
conducts to an enormous redistribution of resources arbitrarily.
– Denial of Services (DoS) attack: This attack may be utilized in several ways
to make the system ineffective. It is not utilized to steal or alter the information,
but to target system availability and disable it. A system that can transmit a huge
amount of radiofrequency signals may disrupt and stop the operation of every
sensor node, thus, triggering a DoS attack.
– Sinkhole attack: In this attack, the attacker targets a node nearer to the SINK
(identified as the sinkhole node), and makes it appear attractive for the rest of the
nodes on the system, and therefore, draws network traffic to it [9].
3.3 Attacks Application Layer
– It is the upper layer representing the applicable elements of the system. Its key
role is to give the service needed by IoT-specific users. This layer uses several
protocols that usually include constrained protocol for application (CoAP) and
message queue telemetry transport (MQTT). These protocols help to easily supply
the user with the desired service [10]. As a result, it presents specific security
problems that are not presented in the rest of the layers, like privacy issues and
data theft. The main security problems faced by the application layer are examined
below.
– Data theft: The information or data obtained by sensors from the devices of
IoT is most sensitive while in transit. Attackers with the intentions of utilizing
credentials for private use or reselling it to the highest buyer will steal the data
quite easily if appropriate security procedures are not respected.
– Sniffing attack: Data packets could be captured and sensible data could be
removed by sniffing if the data packets have little or no encryption during
transmission [11].
– Unauthorized access: Access control involves providing access to legitimate
users and refusing access to non-authorized users. With non-authorized access,
attackers can steal data or get access to confidential data.
4 Jamming Attack Implementation on S-MAC Protocol
4.1 Proposed System
In this part, we will implement the Jamming attack in a WSN at the MAC layer, in
order to visualize the effect and severity of this type of attack on network performance.
The system we proposed in our work consists of the SINK, i.e., the base station and
a group of nodes. The role of the SINK is to collect and process the packets sent by
the nodes in a centralized model.
In this article, we will implement two types of Jamming attacks that target the
base station. In the first type, the attacker will only use request to send (RTS) control
packets to trigger the attack in the network. While in the second type, the attacker
will use data packets. Once the malicious node is implemented in the system, it
analyzes the traffic to detect the protocol used, which in our case is S-MAC, so that
it can synchronize with the network. Then, it continuously transmits a large number
of packets in the transmission channel to the SINK.
The objective of the simulation of the control packet Jamming attack (RTS) and
data packet Jamming attack (DATA) is to visualize and analyze the impact of each
of these two attacks and compare their gravity on the network in order to develop a
precise mechanism that protects against the most effective attack. Fig. 4 shows the
activity modeling of Jamming attack.
Fig. 4 Jamming attack

flowchart
4.2 Simulation Experiment
The simulation is realized in the OMNeT++ environment under Linux. We used a set
of twenty-five nodes, where node 0 is the SINK. The channel utilized in simulation is
a wireless channel. The MAC protocol used is S-MAC. Therefore, the nodes update
their sleep schedule using this protocol. The simulation period was set to 200 s. The
initial energy of each node is 18720 J. We utilized a star architecture in which all
the nodes send their packets to the SINK. The number of packets sent equals five
packets per second with a transmission power set at 36.3 mW. Table 1 illustrates the
simulation parameters.
In our simulation, two scenarios are proposed. The first scenario (see Fig. 5)
represents the normal case of the system, i.e., no malicious node is implemented in
the network. The nodes synchronize with each other and then transfer control and
data packets to the base station.
The second scenario (see Fig. 6) represents the Jamming attack by adding a
malicious node, which is node 25, in the network. In this scenario, we will simulate
the two types of Jamming attacks we explained above. Both have the same network
architecture and the same simulation parameter values. The only difference between
Table 1 Simulation
Parameters Values
parameters
Simulation time (s) 200
Simulation area (m) 60 × 60
Number of nodes 25
Mobility model No Mobility
Topology Star
Transmit power (mW) 36.3
Packet rate (pps) 5
Data packet size (bytes) 100
End time End of simulation
Protocol S-MAC
SYNC packets size (bytes) 11
RTS packets size (bytes) 13
Contention period (ms) 10
Frame time (ms) 610
Fig. 5 Scenario 1—Normal case
Fig. 6 Scenario 2—Network with jamming attack

Table 2 Simulation
Parameters Values
parameters for Jammer node
Number of jammers 1
Trajectory Fixed
Transmit power (mW) 57.42
Packet rate (pps) 50
Protocol S-MAC
SYNC packets size (bytes) 11
RTS packets size (bytes) 200
Data packets size (bytes) 200
Frame time (ms) 610
Contention period (ms) 10
End time (s) End of simulation
these two types of attack is the attacker node. In the first case, the malicious node
will use large control packets (RTS) to damage the network, while in the second case
the attacker will use data packets.
In order to study the difference in the impact of each of these two attacks under
the same conditions, the two attacking nodes use the same values of simulation
parameters, and to make the attack even more efficient, we have increased the packet
size and the rate of sending packets per second compared to the normal case. The
Jamming launched in our network is represented by a fixed node, which allows the
sending of a large number of packets saturating the transmission channel. Table 2
shows the parameters of the jammer node.
4.3 Analysis of the Results
After simulating the scenarios described above, we will first analyze the network
behavior under normal conditions, i.e., when no attacks are implemented in the
system. The objective of this simulation of Scenario 1 is to compare its performance
with an attacked network. Then, we will analyze the performance of the system and
the severity of the damage when the network is under attack (Scenario 2). The result
obtained is presented in the following figures.
Figure 7 represents the number of packets sent by all nodes and the number of
packets received by the base station for the three simulations. For the normal case,
we notice that the number of packets sent is almost the same as the number of packets
received. So the number of lost packets is very low, which explains that the network
is working correctly. When the attack is implemented, either for the RTS control
packet attack or the DATA packet attack, we notice that the number of packets sent
by the nodes is decreased compared to the normal case. This decrease in the number
of packets sent is due to the high-speed broadcast of false packets generated by the
Fig. 7 Number of packets sent and received
attacker in the network, which increased the traffic, made the transmission channel
busy, and ultimately prevent other nodes in the network from sending their packets
correctly.
Concerning the reception of packets, for the attacked network, we can see that a
very limited number of packets was received compared to the number sent. Figure 8
shows more precisely the packet loss rate in the network. As we noticed at the
beginning, the packet loss rate for the normal case is too low, but it is very high in
the case of an attack. Most of the packets sent to the SINK were not received. The
reason for the decrease in the number of packets received is the high traffic generated
by the malicious node and the number of false packets sent to the SINK, which made
the SINK busy by the reception of these packets and unable to receive all the packets
intended for it by the legitimate nodes.
Figure 9 shows the energy consumption by each node of the network for the
three simulations. For Scenario 1, when no attacks are implemented, the network
functions correctly, and the S-MAC protocol has a better energy efficiency thanks to
the implementation of SYNC packets. Once the attack is implemented (Scenario 2),
we notice that the power consumption has doubled compared to the normal case.
Fig. 8 Packet loss rate per node

Fig. 9 Energy consumption per node (J)
The reason for the high consumption at the base station level for both simu-
lations (attack with RTS packets and attack with DATA packets) is the reception
and processing of a large number of false packets sent by the attacker. This repeated
sending of the false packets resulted in high traffic in the transmission channel, which
caused the legitimate nodes to consume more power in unnecessary retransmissions
of their packets to the SINK.
Energy is a term that is often used synchronously with the lifetime of the network.
This means that maximum energy consumption leads to a minimum network life.
Figure 10 shows the estimated network lifetime for the three simulations. It can be
seen that the battery of the attacked networks is quickly depleted compared to the
normal case because of the Jamming attack which resulted in high energy waste and
thus a short lifetime.
As mentioned above, the goal of simulating the two types of attack (RTS and
DATA) is to analyze the effect of each of these two attacks and to compare their sever-
ities on the network. According to the results obtained, we notice that the Jamming
attack by data packets (DATA) is more efficient compared to the attack by control
packets (RTS).
Fig. 10 Estimated network

lifetime (days)
The attack by data packets (DATA) affects the entire network and degrades its
performance in a very significant way, which subsequently leads to the destruction
of the system.
5 Conclusion
As the emergence of IoT has occurred, several vulnerabilities, varying from attacks
on the devices to attacks on the data in transits, have attracted the attention of the
research community. Besides, the inexpensive design of sensor nodes and the facility
to reprogram them will make sensor networks highly vulnerable to intentional attacks.
In this paper, we have given an overview of the use of the internet of things, the objec-
tives of security, the MAC protocols as well as the classification of attacks according
to the different layers of an IoT application. Then, we analyzed the effects of the
Jamming attack on the network using the S-MAC protocol. Two types of Jamming
attacks were analyzed: The first is the attack by control packets (RTS) and the second
is the attack by data packets (DATA). The parameters used to determine system effi-
ciency are the number of packets delivered, packet loss rate, power consumption, and
network lifetime. In the end, the results showed that the Jamming attack, specifically
by DATA packets, is a very dangerous type of attack that can be effectively used to
deteriorate network performance and quality of service and then damage the device.
In the future work, we will try to design and implement a mechanism to protect
networks from Jamming attacks by data packets.
References
1. Hasan, Ali , Khattak: Munam Ali Shah, Sangeen Khan, Ihsan Ali, Muhammad Imran,
Perception layer security in Internet of Things. Futur. Gener. Comput. Syst. 100, 144–164
(2019)
2. Mohanta, B.K., Jena, D., Satapathy, U., Patnaik, S.: Survey on IoT security: Challenges and
solution using machine learning, artificial intelligence and blockchain technology. Internet of
Things 11, 100227 (2020).
3. Jagriti, D.K.: Energy consumption reduction in S-MAC protocol for wireless sensor network.
Procedia Comp. Sci., 143, 757–764 (2018).
4. Ouaissa, M., Ouaissa, M.: Rhattoy, Enhanced and Efficient Multilayer MAC Protocol for M2M
Communications. Adv. Intell. Syst. Comput. 1165, 539–547 (2021)
5. Sakya G., Singh P.K.: Medium access control protocols for mission critical wireless sensor
networks. In: Singh P., Bhargava B., Paprzycki M., Kaushal N., Hong WC. (eds.) Handbook of
wireless sensor networks: issues and challenges in current scenario’s. Advances in Intelligent
Systems and Computing, vol 1132. Springer, Cham (2020).
6. Aarika, K., Bouhlal, M., Ait Abdelouahid, R., Elfilali, S., Benlahmar, E.: Perception layer
security in the internet of things. Procedia Comp. Sci. 175, 591–596 (2020)
7. Hassija, V., Chamola, V., Saxena, V., Jain, D., Goyal, P., Sikdar, B.: A Survey on IoT Security:
Application Areas, Security Threats, and Solution Architectures. IEEE Access 7, 82721–82743
(2019)
8. Jha, R.K., Puja, H.K., Kumar, M., Jain, S.: Layer based security in Narrow Band Internet of
Things (NB-IoT), Comp. Netw. 107592 (2020).
9. Sengupta, J., Ruj, S., Das Bit, S.: A comprehensive survey on attacks, security issues and
blockchain solutions for IoT and IIoT, J. Net. Comp. Appl., 149, 102481 (2020).
10. da Cruz, M.A.A., Joel, J.P.C., Lorenz, P., Solic, P., Al-Muhtadi, J., Albuquerque, V.H.C.: A
proposal for bridging application layer protocols to HTTP on IoT solutions. Future Generation
Comp. Sys., 97, 145–152 (2019)
11. Anand, S., Sharma, A.: Assessment of security threats on IoT based applications. Materials
Today: Proceedings (2020)
H-RCBAC: Hadoop Access Control
Based on Roles and Content
Sarah Nait Bahloul, Karim Bessaoud, and Meriem Abid
Abstract Social networks, smartphones, mobile applications... produce an

avalanche of data on a large-scale and in an unstructured way. The phenomenon
Big Data was born in order to address the different challenges including data stor-
ing, data analysis, data querying and so on. Technological advances always carry new
security vulnerabilities that are not taken into consideration at the beginning. Security
aspects usually require time to be addressed. The information system security is the
set of measures to prevent any failure or threat including unauthorized accesses. Per-
fect protection must contain the four basic building blocks that are: authentication,
access control, auditing, and encryption. In our work, we are specifically interested
in access control. We first analyze the well-known access control models that were
applied to Big Data. We then investigate the most important security projects. Most of
these approaches and projects rely mainly on coarse-grained access control policies.
In this work, we propose a novel approach called H-RCBAC that relies on two known
models: the role-based access control (RBAC) and the content-based access control
(CBAC). H-RCBAC is a new architecture that refines the access control process by
considering a set of taboo words to guarantee fine-grained access control.
1 Introduction
Digital data explosion has forced researchers to find new ways to analyze and exploit
the world, to manage new scalability issues about the capture, the storage, the anal-
ysis, and the representation of data. Big Data has quickly become an unavoidable
S. Nait Bahloul (B)

Computer Science Department, LSSD Laboratory, University of Science and Technology of Oran
Mohamed Boudiaf, Oran, Algeria
e-mail: sarah.naitbahloul@univ-usto.dz
K. Bessaoud · M. Abid
Computer Science Department, University of Mostaganem, Mostaganem, Algeria
e-mail: karim.bessaoud@univ-mosta.dz
M. Abid
e-mail: meriem.abid@univ-mosta.dz
https://doi.org/10.1007/978-981-16-3637-0_30
424 S. Nait Bahloul et al.
Fig. 1 Big data 3Vs
trend for many industrial players because of the contribution it offers in terms of stor-
age, processing, data analysis, and decision support tools, by its computing power
which is distributed on a cluster that groups together a set of machines. It provides
a response to very complex queries and supports a variety of data sources (Fig. 1).
The research community agrees that the Big Data is characterized by the 3Vs
problem [1]. (1) Volume refers to the amount of data, (2) Variety refers to several
types of data, and (3) Velocity refers to the speed of data processing. According to
the 3Vs model, the challenges of big data management result from the expansion of
all three properties, and not just the volume.
Big Data projects require the choice of a storage method, operating technology,
and data analysis tools to optimize processing time on large data. In this context, one
solution has emerged among the different works, namely, Hadoop. Apache Hadoop
is an open-source software framework used for distributed storage and processing of
very large datasets. Various other frameworks have emerged since. The most known
and used is Apache Spark. It is significantly faster than Hadoop since it uses memory
to execute the treatments.
As a first step, we focus in this work on the Hadoop Framework. Despite its
performance compared to Spark, Hadoop is largely used as a solution for storage
and processing large distributed data.
Hadoop, as many new incoming technologies, was not fitted to address security
threats at the beginning. Hadoop was first utilized to manage a large amount of non-
sensitive data. The success of Hadoop and its adoption in various utilization contexts
raised new data security issues. One question then emerges: How to process and
exploit such a large volume of data while ensuring their security?
Like any information system, for enhanced protection, Hadoop must contain the
four basic bricks of security [2], namely:
• Authentication: To protect an information system, it is essential to integrate a tool

allowing a person (or a process) identification before allowing it to connect. There
are several authentication techniques that can be classified into two categories:
H-RCBAC: Hadoop Access Control Based on Roles and Content 425
– Basic authentication: Who am I (login), prove it (authenticator (password)).

– Strong authentication: simple authentication reinforced by the addition of
“locks”: biometric fingerprint, smart card, retinal recognition, . . .
• Authorization: Once connected, the person is not allowed to do everything or
to reach all the resources. It is, therefore, necessary to control its access. Several
techniques and approaches are proposed to meet the different needs of companies
and/or applications. In our work, we will focus more specifically on this security
brick. We will see later the different approaches used to ensure access control in
the Big Data.
• Audit: Although the person is identified, authenticated, and authorized to reach
certain resources, a well-secured system must track all the operations. This is
tractability, an essential process to recover problems.
• Encryption: In order to effectively combat ever-evolving threats, a data-centric
approach is also needed to protect sensitive information. To do this, we use encryp-
tion for certain data, stored safely, and made unintelligible via an encryption algo-
rithm preserving their confidentiality. The encryption key must be strictly protected
to make this process effective.
We focused on our work on access controls on Hadoop. Although many approaches
were proposed to enable Hadoop’s access control policy, most of them rely mainly
on coarse-grained access control policies.
In this paper, we propose a novel approach to define and manage fine-grained
access control rules on Hadoop.
The rest of the paper is organized as follows: Sect. 2 presents the framework
Hadoop. Section 3 presents the proposed access control models on Big Data. In
Sect. 4, we present the different security projects to enhance security on Hadoop and
the Big Data ecosystem. We present and detail our approach in Sect. 5. We conclude
in Sect. 6.
2 Hadoop
Hadoop is an open-source framework providing tools to use a cluster to store and

manipulate large volumes of data quickly and optimally. It exploits the computing
power of the cluster machines. It is managed by the Apache Foundation.
The two main components of Hadoop are the HDFS and the MapReduce pro-
gramming model [3].
2.1 HDFS: Hadoop Distributed File System
HDFS is a file system allowing to store a large volume of data on a cluster by

disregarding the capacities of the machines or their states. The system is fault-tolerant.
Fig. 2 HDFS architecture
Its architecture is based and thought according to this. The system uses data blocks of
a larger size than the known file systems. This value is 64 MB (but can be changed).
Each file is thus divided into blocks, and these blocks are distributed on the cluster.
There is a replication rate of the default blocks set to 3 (but there again modifiable).
Each block is present on different nodes at a time. The loss of a node is not a problem
since the lost blocks are replaced by other nodes.
HDFS consists of a main node, called the NameNode. This node is very important,
and it manages the data location. It matches a file with its associated blocks (the
metadata of a file) and knows also on which nodes each block is located. The second
type of node is DataNodes. A DataNode handles the data blocks. The DataNode very
often keeps the NameNode aware of the blocks it manages. So, the NameNode can
detect problems and ask block replication. DataNodes do not support files, but only
blocks. The NameNode manages the file. It will be able to open, close, drop files,
and send these changes to the relevant DataNodes. It will ask DataNodes to create
blocks, drop, read, or write them. Figure 2 summarizes the HDFS architecture.
2.2 MapReduce
MapReduce [3] is a programming model introduced by Google for the processing of

a large volume of data. MapReduce divides a query into tasks and assigns them to
many computers. Later, the results are collected in one place and integrated to form
the final result. The MapReduce algorithm has two important tasks, namely Map and
Reduce.
• The Map task takes a set of data and converts it into another set of data, where
each element is transformed into key–value pairs.
• The Reduce task takes the Map output as an input and combines those key–value
pairs into a smaller set of tuples. The Reduce task is always performed after the
Map task.
Fig. 3 MapReduce step
Figure 3 details a MapReduce processing that calculate the occurrences of the words
in a file.
• Split: The file is divided into blocks.

• Map: It is a user-defined function, which calculates the occurrences of words in
the bloc. Keys are created with an associated value. In this example, a key is a
word and the value is a number n which signifies that the word is present n time.
• Shuffle: All identical keys are grouped.
• Reduce: A processing is carried out on all the values of the same key. In this
example, the values are summed, which makes possible to get the number of
occurrences of the words in the file.
The final result is stored in a file and returned to the requester.
3 Access Control Models on Big Data
Access control is one of the most commonly used methods for managing autho-
rizations. Data security is likely to be compromised due to a poor access control
strategy.
Added to this, the data source multiplicity can generate problems in choosing
a good strategy of access control. We will show in this section a survey on the
adaptation over time of access control approaches in Big Data.
3.1 Access Control Lists (ACLs)
Before the ACL implementation, the HDFS authorization model was equal to tradi-
tional UNIX permissions (Permission Bits). In this template, permissions for each
file or directory are managed by a set of three distinct user classes: Owner, Group,
and Other. There are three permissions for each user class: Read, Write, and Execute.
Thus, for any file system object, the permissions can be encoded in 3 * 3 = 9 bits.
When a user tries to get access to a file, HDFS applies permissions based on the most
specific class of users applicable to that user. If the user is the owner, HDFS checks
the rights of the owner class. If the user is a member of the file group, HDFS checks
the permissions group class. Otherwise, HDFS checks for other class permissions.
This model can adequately handle a large number of security requirements.
In Hadoop, HDFS supports POSIX ACLs [4]. By default, ACL support is disabled,
and the NameNode prohibits ACLs creation. It must be activated by reconfiguring
the NameNode.
POSIX ACLs allows to assign different permissions to different users or groups,
even if they are not the original owner. For example, user John creates a file. It does
not allow anyone in the group to get access to the file it owns, except for another user,
Antony (even if other users belong to the John group). This means that the owners
or groups of the file can agree to other users and groups to manage access to that file
through POSIX ACLs.
3.2 Role-Based Access Control (RBAC)
The role-based access control model (RBAC) uses the role as an intermediary
between users and permissions [5]. It simplifies administration tasks by reducing
the number of assignments to be handled. This model is widely adopted by com-
panies and manufacturers. There is no need to change the permission each time a
person joins or leaves an organization. By this characteristic, RBAC is considered
as an “ideal” model for companies with high turnover rates. It minimizes the risk of
errors and unintended permissions by reducing the workload of administrators. The
RBAC standard was proposed by the National Institute of Standard and Technology
(NIST) in 2001 and formally adopted by the ANSI standard in 2004.
The RBAC model provides the necessary security functionality for the Big Data
environment. The most well-known security project using RBAC is Sentry (we will
detail this project below) which allows fine granularity access control for Impala,
Apache Hive, MapReduce, Apache Pig. . .
Some works have been based on the concept of role to propose more advanced
solutions. We can cite the work of Gupta et al. [6] who proposed a formal multi-layer
access control model (called HeAC) for Hadoop ecosystem. They extend HeAC base
model to provide a cohesive object-tagged role-based access control (OT-RBAC)
model, consistent with generally accepted academic concepts of RBAC. Besides
inheriting advantages of RBAC, OT-RBAC offers a novel method for combining
RBAC with attributes.
3.3 Content-Based Access Control (CBAC)
To satisfy the new access control needs in Big Data without explicitly identifying
each subject and object, a new approach has emerged, named the content-based
access control (CBAC) [7]. The approach is still in the proposal phase.
The CBAC is added as a second layer of access control and added to the base
layer of Hadoop. The CBAC model offers finer access granularity. For a user query,
the CBAC function evaluates f (u, d) to true or false. u represents the subject and
d the data object. The function is evaluated during query processing. If the function
returns True, access to the file will be granted. Otherwise, access will be denied.
In the model, a subject u is represented by a set of objects it possesses. The objects
are represented by a bag of words. The decision function will receive the maximum
similarity between the object of the candidate and all the files of the owner and will
compare it with a planned threshold:
f (u, di ) = max j (SIMd (du , j, di )) ≥ T
This threshold will determine the number of files that will be accessed by the
candidate. So, a poor choice of the threshold can be fatal. The approach Top-K
similarity is proposed for the choice of a perfect and dynamic threshold.
In [8], the authors propose an access control framework, which enforces access
control policies dynamically based on the sensitivity of the data. The framework
enforces access control policies by harnessing the data context, usage patterns and
information sensitivity.
3.4 Attribute-Based Access Control (ABAC)
In the ABAC approach [9], access to protected resources is based on the user attributes
(name, date of birth, . . .). This approach makes possible to associate the user attributes
with other attributes (IP address, time of day, . . .) to make an access decision. Rather
than using the role of a user to decide whether to grant access to a resource, ABAC
can combine several attributes to make a contextual decision.
ABAC uses attributes as building blocks to define access rules. This is done
through a structured language called eXtensible Access Control Markup Language
(XACML). For example, a rule could declare: Allow managers to get access to
finance-type data only if they come from the Department of Finance. This would
allow users with role attributes = “Manager” and Department = “Finance” to access
data with the category attribute = “Finance”.
Other works are based on the ABAC approach. We can cite among others, the
work of Gupta et al. [10]. They proposed a fine-grained attribute-based access control
model, referred to as HeABAC, catering to the security and privacy needs of the
multi-tenant Hadoop ecosystem.
4 Security Projects on Big Data
Several projects have contributed to the improvement and strengthening of Hadoop

security. We detail in the following the most important projects: Sentry, Apache
Ranger, and Rhino. We first introduce the basic security in Hadoop. We will focus
on the access control aspect of these projects.
4.1 Basic Hadoop Security
At the beginning of Hadoop, developers did not focus on data security because of
the specific use of Hadoop. Having become democratized, security has become one
of the main goals of the various actors. Various tools within the platform, such as
Hive, Impala, Pig . . . have drawn up their own security needs.
Security mechanisms have emerged, including the authentication to verify the
identities of users. The method chosen for Hadoop was Kerberos, a well-established
protocol and is common in business systems such as Microsoft Active Directory.
After authentication came the authorization system. Hadoop uses (as seen above)
ACLs, coarse access permission to HDFS files. In other words, a user has the right to
get access to either the complete document or nothing. Hadoop added the encryption
of the data transmitted between the nodes, as well as the data stored on the disk [11].
4.2 Apache Sentry
Apache Sentry [12] is one of the first security projects that offer fine-grained autho-
rization. Sentry is integrated with SQL query framework, Apache Hive, and Impala
de Cloudera.
Sentry provides the ability to control and enforce specific levels of privileges
to users or applications that authenticate to a Hadoop cluster. It is designed to be
an authorization engine for Hadoop. It allows fine-grained authorization rules to
be defined and allows a user or application to get access to Hadoop’s requested
resources. Sentry can authorize permissions for various varieties of data models.
Sentry aims to make the authorization for components of the Hadoop ecosystem
in a harmonized way, so that, security administrators can easily control what users
and groups have access without needing to know the ins and outs of every single
component in Hadoop.
A data processing tool (example, Hive) identifies the user who requests access to
a data item (such as reading a row from a table). The tool instructs Sentry to process
the user’s query to enable or deny access. Sentry uses rules to define permissions
and roles to combine or consolidate rules, making it easy and flexible to administer
group permissions for different objects.
Fig. 4 Sentry architecture
Fig. 5 Integration of sentry into the hadoop ecosystem
As shown in Fig. 4, the Sentry architecture consists of an authorization provider

and a link layer (binding layer) which is the bridge between the tools (Hive, Impala,
. . .) and Sentry.
The authorization provider consists of two components: The policy engine is
responsible for validating and evaluating the policy and then determines whether the
requested action is allowed by checking the provider (Policy Provider). The latter is
the storage mechanism for policies.
The Sentry architecture is extensible, and any developer can develop a link for
different components (e.g., Pig. . .). Each mandatory component in Sentry imple-
ments a privilege model for specific binding. For example, Hive binding implements
a privilege model to perform fine-grained permissions.
As shown in Fig. 5, Apache Sentry works with multiple components of the Hadoop
environment. At the core, there is Server Sentry that stores authorization metadata
and provides tools to retrieve and modify these metadata.
The actual authorization decision is made by a policy engine that runs in data
processing applications like Hive or Impala. Each component loads the Sentry plug-
in. This component contains the client service to process the permissions by using
Sentry and the policy provider.
4.3 Apache Ranger
Apache Ranger [13] provides a centralized security framework for managing fined-
granularity access control. Security administrators can easily manage policies for
accessing files, folders, databases, tables, or columns. These policies can be defined
for individual users or groups, and then applied inside Hadoop.
Formerly known as Apache Argus, Apache Ranger concurrence Apache Sentry
since it also handles permissions. It adds a layer of permissions in Hive, HBase, and
Knox. It has the advantage over Sentry in defining column-level permissions in Hive.
The architecture of Apache Ranger consists of:
• The Ranger portal: is the central interface of security for administration. Users can
create and update policies, which are then stored in a database. The portal also
consists of a verification server that sends audit data collected from plug-ins for
storage in HDFS or a relational database.
• Ranger Plug-ins: are lightweight Java programs that are integrated into the pro-
cesses of each cluster component. These plug-ins use policies to determine whether
an access request should be granted or not. When a request is delivered through the
component, these plug-ins intercept the request and evaluate it against the security
policy; they also collect data from the user’s request that is sent to the audit server.
• User group sync: Apache Ranger provides a synchronization utility to users and
groups from Unix or LDAP or Active Directory. The user or group information is
stored in Ranger portal and used to define the policy.
In the Apache Ranger version 0.5.0, the community made the first step toward a real
Attribute-Based Access Control (ABAC), providing an access control framework
based on dynamic rules.
4.4 Rhino Project
Rhino [14] is an open-source project developed and maintained by Intel, which aims
to improve the data protection mechanism in Hadoop at different levels. The aim is to
fill the gaps representing insecurity in the Hadoop environment and provide several
safety components. The various improvements consist of:
• Encryption: It provides the ability to encrypt or decrypt data stored in HDFS.

• Authentication: It supports multiple authentication mechanisms, such as public
key cryptography. In the current Hadoop security system, users and services are
authenticated with Kerberos. The Rhino project enables authentication of users in
a centralized service by developing Hadoop with single sign-on.
• Access Control: It provides better access granularity (at the cell level) in HBase. It
is more comfortable with column-oriented databases. Each cell is associated with
a visibility label. When the user queries the data, it provides the group to which
it belongs and the role it has. This information will be verified with the visibility
label by concluding whether or not the user can access this cell.
After presenting the different access control models applied in the Big Data and
the different implemented projects, we present in the following table a summary by
highlighting the relationship between the approaches and projects.
ACL RBAC ABAC CBAC

Access control Files, Tables Columns Cells Cells
granularity
Projects \ Sentry, Rhino Ranger \
Tools HDFS, HBase, Hive, Impala, HDFS, Hive, \
Pig, Zookeper, HBase, Sqoop, HBase. . .
Hue. . . ...
5 H-RCBAC: Hadoop Access Control Based on Roles

and Content
The evolution of the security needs and the technological environments to be con-
trolled means that the authorization systems must be adapted constantly. We defined
in the previous sections the different access control models implemented in the Big
Data (ACLs, RBAC, CABC, ABAC) and the projects that participate in the evolution
and reinforcement of security in Hadoop (Sentry, Rhino, Apache Ranger). Never-
theless, these projects don’t answer all the access control issues in Big Data. For
example, the access control model implemented in HDFS provides coarse-grained
access control. Indeed, we recall that the mechanism set up for HDFS (ACLs) allows
only to define whether the user accesses to the full document or not. This coarse-
grained access control particularly attracted our attention. Such strict access control
(full access or no access) can affect the rights of users. In fact, in some cases, a user
can access a part of the document without threatening the data confidentiality.
To address this problem, we propose an approach combining the RBAC and CBAC
models with an algorithm that filters the taboo words to offer a fine-grained access
control.
It is worth noticing that there are two ways to control access: (1) Assume that the
user has the default right to access everything and set (via rules) interdictions or (2)
assume that user has no right to access and set (via rules) access permissions. In our
approach, we opted for the first assumption.
The RBAC model in our approach simplifies administrative tasks by reducing
the assignments’ management. Each user is assigned to a role that allows access to
certain documents. However, this is not sufficient in the context of HDFS since the
permissions will always be based on the ACLs. The use of this model does not offer
enough flexibility. Let us take an example of a hospital database. We would like to
give physicians access to a certain document but which contains some sensitive data
that must be anonymous. By using ACLs, we have two choices: either take the risk,
make the document accessible, or prohibit all physicians to access this document.
To handle this rough decision, we propose to combine the RBAC with the CBAC
model by integrating an algorithm that filters the taboo words. This solution applied
to our example allows physicians to access the document parts that do not contain
sensitive data (taboo words).
We based our approach on the Discretionary Access Control (DAC) model. This
model is based on the “owner” concept. Each user is responsible for its documents.
In other words, it is up to him to define the access rules of his data. He can give
access to the complete document to a role X. Prohibit access to a role Y. It can define
finer accesses by filtering certain parts of the document through a set of taboo words
and the model CBAC.
We recall here that the CBAC rules are based on a decision function which takes
as input a document and a list of words and compares the words’ occurrences in the
document with a threshold.
The idea of our approach is as follows: Each user is allowed (through his role)
to access a certain number of objects. These accesses are defined via rules based on
the CBAC model. We have added a layer of restrictions to the CBAC that allows the
user to access only to a “subset” of the document. For a user u with a role r that
sends a query q for access to a document d. It will be necessary to calculate, via the
decision function, the taboo words’ occurrences of the role r in the document d. If
it is greater than the defined threshold, the user u will not access the document d. If
the threshold is not reached, the user u will access a filtered document d (without
the taboo words).
5.1 H-RCBAC Implementation
The implementation of H-RCBAC is divided into two parts (Fig. 6):

• The first step is to give the possibility to the document owner (when he puts the
document in HDFS) to define his prohibitions to a document d (define the taboo
words). At this step, it is necessary to pre-calculate (with the known word count
MapReduce) the frequency of each word in the document. This will prevent us
from calculating, at each query, the occurrence of taboo words throughout the
Fig. 6 Access rules definition
document. Indeed, by pre-calculating the frequency of each word at the document

input, we will execute only one time the word count algorithm.
• The second part deals with the execution or not of a query user.
Access Rules Definition
Our approach facilitates decision making. When a document has loaded, the inter-
dictions that are specific to it are loaded into the access policies. Each policy must
provide all necessary information to enable decision making. To achieve this, we
have to integrate a plug-in in HDFS that allows the user to manage their access
policies without going through an administrator.
The plug-in offers the possibility to the owner to:
• Choose roles: These roles are defined upstream by the administrator and assigned
to the different users of the system. When creating a document, the owner retrieves
the list of roles and assigns a set of prohibitions for each role.
• Define taboo words: Lists taboo words (forbidden access) for one or more roles.
• Execute the word count algorithm on the document.
Queries Answering
After defining the access rules, it is important to intercept the queries and filter the
results. For this, we have implemented our approach in the NameNode of HDFS.
When a user requests access to an HDFS file, the NameNode will use our plugin.
The latter will check if the user (who sent the request) has the right to access this
file. Figure 7 details how and when our algorithm interacts to answer to queries.
We suppose that a user u sends a query. For our example, the query is a simple
word count algorithm on a document d. The first step is to verify if the query can
be executed or not. For that, our plug-in will calculate the decision function of the
Fig. 7 MapReduce step with taboo words filter
model CBAC, based on the taboo words of the role r of the user u. The result will be
compared to the threshold. If it is higher, the query will not be executed. Otherwise,
it will be executed after filtering the document.
Let’s detail the second case via an example. Figure 7 summarizes the MapReduce
step. We suppose that the user has no right to access to the taboo word “Pie”.
• In the mapping phase: The query and the list of taboo words will be sent to
the various nodes containing the file partitions. Before executing the query, each
partition (a copy) is filtered (deletion of taboo words). In our example, the word
“Pie”. The result of the map phase will be stored on each node.
• In the Reduce phase: The query will continue and the final result will be stored in
the HDFS file system.
In this case, the query will be executed on a filtered document. This offers fine-
grained access control and better data confidentiality.
The Big Data offers computing power due to its speed of processing queries. It
supports a wide variety of data from heterogeneous environments while delivering
high performance for loading and analysis. However, all this accumulated data from
various sources requires protection. By default, HDFS works with ACLs, but this
is not enough to effectively protect data. Various security solutions are offered to
Hadoop and its ecosystem thanks to the various existing projects. That said, there
are limits to each approach.
We have proposed in our work a new approach that allows reinforcing the access
control in Hadoop. Our approach is based on the combination of two existing models:
RBAC and CBAC. This solution offers a fined-grained access control mechanism
by taking into consideration a set of taboo words before the querying process. H-
RCBAC offers a supplementary access control compared to CBAC. Indeed, when
CBAC offers access to the whole document, our solution excludes the taboo words
before authorizing the access. Our solution gives good results when the query is
independent of the taboo words list. If the taboo words are used in the calculation of
the query result, the result can be inconsistent in addition to being incomplete.
As perspectives, we aim to design a module that identifies the query feasibility
and the results consistency, based on the nature of the query and the taboo words list.
We can easily distinguish between three categories of results: correct and complete
results, correct but incomplete results, and the impossibility to execute the query.
This work can also be improved by integrating it into existing projects. Indeed,
the implemented projects (Sentry, Rhino . . .) offer an already effective framework
and present in several tools of the Hadoop ecosystem.
References
1. Chen, M., Mao, S., Liu, Y.: Big data: a survey. Mobile Netw. Appl. 19(2), 171–209 (2014)
2. Sharma, P.P., Navdeti, C.P.: Securing big data hadoop: a review of security issues, threats and
solution. Int. J. Comput. Sci. Inf. Technol. 5 (2014)
3. Guo, C., Wu, H., Tan, K., Shiy, L., Zhang, Y., Luz, S., Mapreduce: Simplified data processing
on large clusters. In: Proceedings of OSDI, San Francisco, CA, USA (2004)
4. Apache Hadoop. https://hadoop.apache.org/docs/current/hadoop-project-dist/hadoop-hdfs/
HdfsPermissionsGuide.html. Last accessed on 22 Feb 2021 (2017)
5. Ferraiolo, D.F., Sandhu, R., Gavrila, S., Kuhn, D.R., Chandramouli, R.: Proposed NIST stan-
dard for role-based access control. ACM Trans. Inf. Syst. Secur. (TISSEC) 4(3), 224–274
(2001)
6. Gupta, M., Patwa, F., Sandhu, R.: Object-tagged RBAC model for the hadoop ecosystem. In:
IFIP Annual Conference on Data and Applications Security and Privacy, pp. 63–81. Springer
(2017)
7. Zeng, W., Yang, Y., Luo, B.: Access control for big data using data content. In: 2013 IEEE
International Conference on Big Data, pp. 45–47. IEEE (2013)
8. Ashwin Kumar, T.K., Liu, H., Thomas, J.P., Hou, X.: Content sensitivity based access control
framework for hadoop. Digital Commun, Netw. 3(4), 213–225 (2017)
9. Cavoukian, A., Chibba, M., Williamson, G., Ferguson, A.: The importance of ABAC: Attribute-
based access control to big data: privacy and context (2015)
10. Gupta, M., Patwa, F., Sandhu, R.: An attribute-based access control model for secure big data
processing in hadoop ecosystem. In: Proceedings of the Third ACM Workshop on Attribute-
Based Access Control, pp. 13–24 (2018)
11. Das, D., O’Malley, O., Radia, S., Zhang, K.: Adding security to apache hadoop. Hortonworks,
IBM (2011)
12. Apache Sentry: https://sentry.apache.org/. Last accessed on 22 Feb 2021 (2016)
13. Apache Ranger: http://ranger.apache.org. Last accessed on 22 Feb 2021 (2014)
14. Rhino Project: https://github.com/intel-hadoop/project-rhino/ Last accessed on 22 Feb 2021
(2015)
Toward a Safe Pedestrian Walkability:
A Real-Time Reactive Microservice
Oriented Ecosystem
Ghyzlane Cherradi, Azedine Boulmakoul, Lamia Karim,

and Meriem Mandar
Abstract Mobility is one of the key factors to consider in order to make cities
more efficient, a necessity taking into account the millions of citizens travel on a
daily basis to places known or unknown. Road safety, in particular, is of tremendous
importance. Pedestrian accidents that cause too much injury or even death are serious
problems in cities. In this work, we present a real-time reactive system, the aim of
which is to provide the safest route among all possible routes for given source and
destination at a particular time period, based on a modeling of the network by a
fuzzy graph. Its main advantages over the proposed solutions lie in the robustness to
incomplete data modeled as fuzzy information. The system involves the pgRouting
opensource library that extends the PostGIS/PostgreSQL geospatial database, to
provide geospatial routing functionality. So, we offer a web-location based service
allowing pedestrians to enter their destination, then select a route uses an intelligent
algorithm, providing them with the safest route possible instead of the fastest route.
This service will certainly help save lives and, to a certain extent, reduce pedestrian
accidents.
1 Introduction
The advent of smart mobility has enabled the creation and development of applica-
tions for mobility support addressed to different categories of road users, including
pedestrians. Indeed, each of us has different preferences when it comes to transporta-
tion, but at one time or another everyone is a pedestrian. Unfortunately, pedestrian
accidents are on a rise all over the world. Statistics of accidents in the world indicate
G. Cherradi · A. Boulmakoul (B)

LIM/IOS FSTM, Hassan II University, Casablanca, Morocco
L. Karim
LISA Laboratory ENSA Berrechid, Hassan 1st University, Settat, Morocco
M. Mandar
ENS, Hassan II University, Casablanca, Morocco
https://doi.org/10.1007/978-981-16-3637-0_31
440 G. Cherradi et al.
that despite the efforts made, pedestrians remain the most vulnerable road users, with
road accidents taking place through the fault of both drivers and pedestrians. Since
a significant part of pedestrian crossings is non-signalized, the number of accidents
on them, including accidents where people died, is much higher. At the same time,
people die as a result of road traffic crashes outside the city almost 20 times less
than in the cities themselves. This is due to the ratio of the intensity of pedestrian
and traffic flows. As pedestrians, children are at even greater risk of injury or death
from traffic crashes due to their small size, inability to judge distances and speeds,
and lack of experience with traffic rules.
The capacity to respond to pedestrian safety is an important component of efforts
to prevent road traffic injuries. With advancements in technology and the avail-
ability of geolocation solutions in every field of mobility is fostering the creation
of novel approaches, such as the so-called Mobility-as-a-Service (MaaS) paradigm
[1]. The MaaS concepts aim to offer users an integrated, comprehensive and simple
service. The concept of walkability is gradually acquiring worldwide a key position
in mobility planning. Pedestrian dynamics are difficult to characterize as they have
many effects from different sources. Walking, unlike other models of travel, is not
tied to a vehicle on a lane, and the underlying infrastructure is very heterogeneous.
In addition, environmental factors (traffic lights, public furniture, advertising, etc.),
as well as the total waiting time of pedestrians, the average distance between the
vehicles following one after another, and the atmospheric conditions (wind, rain,
etc.) directly affect walking. In this context, modeling pedestrian safety with respect
to the uncertain nature of these factors is an important research objective.
This article presents an integrated platform to collect, identify, and provide a
variety of pedestrian-related communication and control functions to demonstrate
how it is possible to create a scalable and integrated mobility service for safe pedes-
trian walking, through the coordination of reusable components on a microservice
architecture. In addition, the article proposes a risk evaluation model using fuzzy
set theory [2, 3] to consider the incomplete and uncertain nature of the risk param-
eters. The system involves the open-source pgRouting library which extends the
PostGIS/PostgreSQL geospatial database to provide geospatial routing function-
ality. Thus, it offers a web location-based service allowing pedestrians to enter their
destination; then select a route using an intelligent algorithm; provide them with the
safe and short paths to take when they navigate the city. The remainder of this paper
is organized as follows. After discussing some related works in Sect. 2, we intro-
duce our microservice development approach in Sect. 3, focusing in particular on the
advantages that this paradigm introduces on the basis of a real, currently developing
infrastructure. Then, Sect. 4 details the proposed model of pedestrian safety for a
given urban area. Finally, Sect. 5 concludes the paper with some final remarks and
directions for future work.
Toward a Safe Pedestrian Walkability: A Real-Time Reactive … 441
2 Related Works
Within the field of urban computing, our work is related mainly to research on pedes-
trians’ risk modeling, on pedestrian safety systems, and on urban navigation. Here,
we review these lines of research and their connections to our study. Pedestrians’ risk
modeling: Traditionally, pedestrian risk is generally assessed using crash frequency
models, based on historical data [4–7]. Crash frequency models have been devel-
oped using spatial zones or intersection-level data. However, crash frequencies can
be affected by different factors including traffic, speeds, geometry, and built environ-
ment, etc. Authors in [8] proposed a composite indicator of pedestrian’s exposure,
taking into account pedestrian characteristics, road and traffic conditions, as well
as pedestrian compliance with traffic rules. The traffic conflicts technique has also
been used for measuring the exposure of pedestrians at specific crossing locations
[9]. Lassare et al. [10] developed an approach of accident risk based on the concept
of risk exposure used in environmental epidemiology, such as in the case of expo-
sure to pollutants. Saravanan et al. [11] present an accident pre-diction approach
based on fuzzy logic. Their study evaluates the accident risk for pedestrians as well
as vehicles and meets the accident zones on a given road network. Mandar et al.
[12] present a new measurement of virtual pedestrians and vehicles mutual accidents
risk indicator. Pedestrians’ dynamics are modeled using the basic fuzzy ant model
[13], to which they have integrated artificial potential fields. Another factor of risk
is the conflict between the vehicular and pedestrian flows at left-hand corners [14].
Different models have been proposed for risk assessing and working out of measures
to improve the pedestrian safety. Thus, the model proposed in the research [15] allows
to measure the impact of potential risk factors on pedestrians’ intended waiting times.
In research [16], the author proposes a multivariant method of risk analysis consisting
of two hierarchically generalized linear models, characterizing two different facets of
unsafe crossing behavior, and uses a Bayesian approach with the data augmentation
method to draw statistical inference for the parameters associated with risk exposure.
To date, several pedestrian safety systems have been proposed; David et al. [17]
discuss the response time for smartphone applications that warn users of collision
risks, analyzing approaches based on a centralized server and on ad-hoc communica-
tion connections between cars and pedestrians. The use of mobile phones by pedes-
trians (for talking, texting, reading) affects their awareness of the surrounding envi-
ronment, hence augmenting the risk of incidents. Authors in [18] proposed an android
smartphone application (WalkSafe) that aids people that walk and talk, improving the
safety of pedestrian mobile phone users. It uses the back camera of the mobile phone
to detect vehicles approaching the user, alerting the user of a potentially unsafe situ-
ation. In research [19], authors assessed two systems using “standard tests” (vehicle
driving toward a pedestrian dummy positioned on the course). The results attest that
the functionality of the systems depends on the vehicle speed to avoid collisions, but
it is limited at a certain speed (40 km/h). Although these systems are intended to
effectively reduce pedestrian injury outcomes, the collision avoidance performance
of these systems remains limited.
Urban navigation: To support pedestrian route choices to minimize potential

dangers, route qualities (including safe crossing facilities, motor traffic volume and
speeds) need to be considered in the design of pedestrian navigation systems [20].
Systems with new navigation objectives that go beyond the shortest path have recently
appeared in the literature, mainly tailored to the personal interests and spatiotemporal
constraints of users [16, 21–23], or routes that take into account traffic information
gathered from historical speed patterns and GPS devices [21–24]. Indeed, pedestrian
navigation and routing systems need to be developed in a user-friendly manner that
enhances road safety and provides an optimal user experience.
3 System Model
We consider the system model as an integrated solution offering various services:

geolocation, real-time traceability, and pedestrians routing. By leaning toward a
microservice architecture [25], the intention is to develop and execute a plurality of
microservices, which are developed, tested, deployed, and run isolated from the rest.
Moreover, the implementation reveals benefits of microservice architecture regarding
scalability, extensibility, flexibility, and fault tolerance. Figure 1 concisely shows the
main components and the architecture of the proposed system and their connections.
Data collection and pre-processing: To improve the quality of analysis and decision
making, a powerful, scalable, and interoperable data collection service is required.
Therefore, a smart data collection service has been developed in our system. Based
on concepts of IoT allied to wireless sensor networks, the proposed service uses
sensors to obtain information about environment, vehicle, and pedestrians in which
they are inserted and have great importance to analyze several factors that influence
pedestrian’s walkability.
Pedestrian risk calculator: This microservice devoted to calculate in real time the
risk related to a road. Risk indicators are valid and depend on how they are read.
There is no absolute risk indicator. Nevertheless, the indicator we propose estimates
the journey time. The residence time of a pedestrian on a section of road increases
the exposure of risk. We postulate that the value of risk must be a function of the total
time it takes to cross the segment. Therefore, our exposure model suggests that risk
poses appositive correlation with two factors: probability of an accident and duration
of risk exposure.
Pedestrians routing: The routing of the pedestrians is the service responsible for
finding point-to-point safest path in a constructed time-dependent network with
fuzzy risk. The implemented routing algorithm is designed to compute an opti-
mized shortest path that surpasses the performance of classical algorithms in terms
of computational time and space. In general, two kinds of data are needed as input
for the proposed service, namely: (a) spatial data, which represents the transporta-
tion network as a collection of arcs with nodes loaded from OSM (OpenStreetMap)
Fig. 1 The system architecture
into the PostgreSQL/PostGIS database. (b) Non-spatial attribute information for the
attribute data for road such as name, category, length, risk cost.
4 Modeling Pedestrian Risk
Throughout the paper, we will represent the urban network using an undirected graph
G (V, E). The nodes of the graph, V, represent intersections and the edges, E, represent
the road segments that connect intersections. Each road segment s ∈ E is associated
with two (unrelated) weights: its length, denoted by ls , and its risk, denoted by rs .
The length of segments s corresponds to the actual distance between the two intersec-
tions connected by the associated street segment, while the risk of segment s can be
defined as the multiplicative product of the probability of an accident and the gravity
(the pedestrian at risk factor) to pedestrian if it does occur. In fact, risk assessment
of pedestrian must not only derive the probability of a crash, but must estimate the
severity to which pedestrian is at risk from such events. One way to analyze a pedes-
trian safety issue is to identify the significant factors affecting pedestrian crash injury
severity. In the following subsections, we describe how the proposed model can be
obtained. Specifically, we explain how we combine information from different data
sets to build a pedestrian risk model.
4.1 Urban Road Networks
We extract the urban road network using OpenStreetMap (OSM) [26]. OSM is a
crowd-sourced project that provides free geographic data such as road maps. The
OSM community has developed a number of tools that facilitate the integration of
OSM maps with a variety of applications. For our purposes, we make use of the
osm4routing parser [27], that exports the map of interest from an OSM format to a
graph-based format (appropriate for graph-based algorithms). In the exported graph
G (V, E), each vertex from V corresponds to a designated pedestrian crossing, road
intersection, or dead-end. Whereas each edge from E is undirected and corresponds
to a road segment connecting the corresponding vertices. Osm4routing provides
supplemental information about the intersections and the street segments as node
and edge attributes. Especially, the geographical location (latitude/longitude pair)
of each node is provided. Each edge e is annotated with the physical length of the
corresponding road segment. It is the length value we associate with every edge in
the input graph. All road networks considered in this work correspond to urban areas
in the Mohammedia city.
4.2 Pedestrians Risk
Preliminaries. In this section, we recall some definitions and the main results about
the fuzzy set theory which will be used later. The first person who introduced fuzzy
Set theory is L. Zadeh. He introduced his article regarding “Fuzzy Sets” [11].
Definition 1 (Fuzzy set) [11] is defined in terms of a membership function, which

is a mapping from the universal set U to the interval [0, 1]. Then, a fuzzy set Ã in U
is a set of ordered pairs:

Ã(x) = x, μ Ã (x)|x ∈ U, μ Ã (x) ∈ [0, 1] (1)
A membership function is a function from a universal set U to the interval [0,

1]. A fuzzy set Ã is defined by its membership function μ Ã (x) over U. When fuzzy
set theory was presented, researches considered decision making as one of the most
attractive application fields of that theory [2, 3].
Definition 2 (Fuzzy number) [2, 3] a fuzzy subset Ã of the real line R with member-
ship function μ Ã : R → [0, 1] is a fuzzy number if its support is an interval [a, b] and
there exist real numbers s, t with a ≤ s ≤ t ≤ b fulfilling:
i. μ Ã (x) = 1 for s ≤ x ≤ t
ii. μ Ã (x) ≤ μ Ã (y) for s ≤ x ≤ y ≤ t
iii. μ Ã (x) ≥ μ Ã (y) for t ≤ x ≤ y ≤ b
iv. μ Ã (x) is upper semicontinuous
We will denote the set of fuzzy numbers by FN and a fuzzy number by Ñ . We
observe that every real number N is a fuzzy number whose membership function is
the characteristic function:

1 if x = N
μ N (x) = (2)
0 if x = N
Definition 3 (Triangular fuzzy number (TFN)), a triangular fuzzy number Ñ can be

defined by a triplet (a, b, and c). The triangular fuzzy number is used to represent
uncertainty. The membership function is [2].
⎧
⎪
⎨ 0 x < a, c < x
μ N (x) = b−a
x−a
a<x <b (3)
⎪
⎩ x−c b < x < c
b−c
where 0 ≤ a ≤ b ≤ c ≤ 1, a and c stand for the lower and upper values of the support
of N, respectively, and b stands for the modal values. The graphical representation
of a triangular fuzzy number is shown in Fig. 2.
Definition 4 (Gaussian fuzzy number (GFN)), compared to traditional triangular

fuzzy number (TFN), GFN is more realistic to represent uncertainty because the
Fig. 2 The graphical

representation of a triangular
fuzzy number (TFN)
Fig. 3 Representation of a
Gaussian fuzzy number
(GFN)
distribution is assumed to be Gaussian. The Gaussian fuzzy function transforms

the original values into a normal distribution. The fuzzy Gaussian function is given
below:
2
1 x −a
μ N (x) = exp − , x ∈ R, b > 0 (4)
2 b
A Gaussian Membership function is specified by two parameters: a Gaussian

membership function is determined complete by a and b; a is the expectation and b
is the standard deviation. Figure 3 shows a graphical representation of a Gaussian
fuzzy number (GFN).
Pedestrians risk Model. A risk model for the urban road network is essentially an
assignment of a risk score r to each segment s that is proportional to the probability of
an accident happening on the corresponding road segment and the pedestrians-at-risk
factors. In the proposed model, the duration of risk exposure is estimated by travel
time of the road segment traveled. During this time, the pedestrian is exposed to risk
depending on risk factors. This travel time is one of several sources of uncertainty in
pedestrian’s walkability, which cannot be known or expressed exactly. This uncer-
tainty is represented by Gaussian fuzzy number Ts∼GFN (ts , γts ) where ts , γts are
the mean and the standard deviation, respectively, and γ is a proportionality param-
eter. The pedestrian crash injury severity associated with each factor is multiplied
by the probability of accident occurrence, and the travel time to give an estimate of
risk exposure along each segment of the road network. The expression used in this
study is of the following form:
Rs = Ps j *F j * Ts (5)
j
where:
Rs : the risk on a segment road s;
Ps j : the probability of accident occurrence on road segment s with respect of a
risk factor j;
F j : the severity of pedestrian crash injury on a segment s with respect to factor j;

Ts : The travel time associated with segment s.
In expression (5), some inputs parameters are GFNs, and some are constant there-
fore, the resulting risk is also obtained in the form of a GFN whose membership
function normalized is expressed as:
2
1 x −a
μ R (x) = exp − , x ∈ R, b > 0 (6)
2 b
The total risk of a path, denoted by R(x), is the summation of the risk weight of
segment sk . Where sk is the kth segment for a path x.
n
R(x) = Psk j *F j *Tsk
k j
⎛ ⎞ (7)
n n
= N⎝ Psk j *F j * Tsk , γ Psk j *F j * Tsk ⎠
k j k j
The risk R is calculated for each arc in the network N using (7). R(R ∼ N (0, 1))
is the normalized value of the expected fuzzy risk of a segment s as shown in Fig. 4.
For each segment s, there is a variety of risk values according to an α (prede-
termined confidence level) to be specified according to the risk threshold not to be
exceeded μ R (x < λi ) = αi . (See Fig. 5).
We need to search values of λi the evaluation of edge according to an αi risk level
and a departure time t. Finally, the problem becomes to define the safest path in a
time-dependent network as shown in Fig. 6.
Fig. 4 Fuzzy risk graph

Fig. 5 Acceptable risk value

λi according to αi level
Fig. 6 Discrete time-dependent risk graph
4.3 Safest Route Determination
We now turn to an application exploiting the model presented above to provide safe
pedestrian navigation in urban environments. The safest path problem takes as an
input the road network G (V, E); together with a pair of source–destination nodes,
(s, d), and its goal is to provide to the user a short and safe path between s and d. The
algorithm for finding the safest path is based on a network composed of vertices and
edges (routes), defined by pairs of vertices. Each edge has a cost which, in our case,
represents the pedestrian risk. As this attribute is known, the problem of pedestrian
routing consists in finding the minimum cost from a vertex of source A to a vertex of
destination B specifies. We used the pgr_bdAstar () function (which implements the
bidirectional A * algorithm) to find the optimal path. Figure 7 illustrates the result
of the safest routing service.
Fig. 7 An illustrative example of safest paths. The routes depicted as Paths 1–2 offer various
alternative for traveling between a starting point and a destination
5 Conclusion
Smart walkability is a key element to support pedestrians in their daily activities and
to offer them a livable smart city. Information about urban transportation, pedes-
trian crossings, and safest paths would be of great benefit in this context. In order
to provide, an ecosystem to manage such services and features, we designed and
prototyped a distributed information system and proposed a fuzzy pedestrian risk
model under the consideration of travel time. As for the pedestrian routing problem
in practice, data are very important to evaluate risk as well as cost. The accuracy and
quality of the data could have significant impact on the result. Currently, data for
pedestrian are insufficient and incomplete. Future research directions include how
to establish sufficient databases and how to validate the proposed model.
Acknowledgments This work was partially funded by Ministry of Equipment, Transport, Logis-
tics, and Water−Kingdom of Morocco, The National Road Safety Agency (NARSA) and National
Center for Scientific and Technical Research (CNRST). Road Safety Research Program# an intel-
ligent reactive abductive system and intuitionist fuzzy logical reasoning for dangerousness of
driver-pedestrians interactions analysis.
References
1. MaaS Global. http://whimapp.com/. Last accessed 28 Nov 2020

2. Zadeh, L.A.: Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets Syst. 1(1), 3–28 (1978)
3. Zadeh, L.A.: Probability measures of fuzzy events. J. Math. Anal. Appl. 23(2), 421–427 (1968)
4. Brüde, U., Larsson, J.: Models for predicting accidents at junctions where pedestrians and
cyclists are involved. How well do they fit? Accid. Anal. Prev. 25, 499–509 (1993).
5. Cameron, M.: A method of measuring exposure to pedestrian accident risk. Accid. Anal. Prev.
14, 397–405 (1982)
6. Lee, J., Abdel-Aty, M., Xu, P., Gong, Y.: Is the safety-in-numbers effect still observed in areas
with low pedestrian activities? A case study of a suburban area in the United States. Accid.
Anal. Prev. 125, 116–123 (2019)
7. Ni, Y., Wang, M., Sun, J., Li, K.: Evaluation of pedestrian safety at intersections: a theoretical
framework based on pedestrian-vehicle interaction patterns. Accid. Anal. Prev. 118–129 (2016)
8. Van der Molen, H.H.: Child pedestrian’s exposure, accidents and behavior. Accid. Anal. Prev.
13(3), 193–224 (1981)
9. Gårder, P.: Pedestrian safety at traffic signals: a study carried out with the help of a 601 traffic
conflicts technique. Accid. Anal. Prev. 21(5), (1989)
10. Lassarre, S., Papadimitriou, E., Yannis, G., Golias, J.: Measuring accident risk exposure for
pedestrians in different micro-environments. Accid. Anal. Prev. 39(6), 1226–1238 (2007)
11. Saravanan, S., Sabari, A., Geetha M.: Fuzzy-based approach to predict accident risk on road
network. Int. J. Emerg. Technol. Adv. Eng. 4(5), (2014). ISSN 2250-2459, ISO 9001:2008
Certified Journal
12. Mandar, M., Boulmakoul, A., Lbath, A.: Pedestrian fuzzy risk exposure indicator. Transp. Res.
Procedia 22, 124–133 (2017)
13. Boulmakoul, A., Mandar, M.: Fuzzy ant colony paradigm for virtual pedestrian simulation.
Open Oper. Res. J. 2011, 19–29 (2011). ISSN: 18742432. https://doi.org/10.2174/187424320
1105010019
14. Di Stasi, L.L., Megías, A., Cándido, A., Maldonado, A., Catena, A.: The influence of traffic
signal solutions on self-reported road-crossing behavior. Span. J. Psychol. 17(103), 1–7 (2014)
15. Quistberg, D.A., Howard, E.J., Ebel, B.E., Moudon, A.V., Saelens, B.E., Hurvitz, P.M.,
Rivara, F.P.: Multilevel models for evaluating the risk of pedestrian–motor vehicle collisions
at intersections and mid-blocks. Accid. Anal. Prev. 84, 99–111 (2015)
16. Ayala, I., Mandow, L., Amor, M., Fuentes, L.: An evaluation of multi-objective urban tourist
route planning with mobile devices. LNCS Ubiquitous Comput. Ambient Intell. 7656(2012),
387–394 (2012)
17. David, K., Flach, A.: Car-2-x and pedestrian safety. Vehic. Technol. Mag. 5(1), 70–76 (2010)
18. Wang, T., Cardone, G., Corradi, A., Torresani, L., Campbell, A.T.: WalkSafe: a pedestrian
safety app for mobile phone users who walk and talk while crossing roads. In: Proceedings of
the Twelfth Workshop on Mobile Computing Systems & Applications, pp. 1–6 (2012)
19. Ando, K., Tanaka, N.: An evaluation protocol for collision avoidance and mitigation systems
and its application to safety estimation. In: Proceedings of the 23rd International Technical
Conference on the Enhanced Safety of Vehicles. Seoul, Republic of Korea (2013)
20. Czogalla, O., Herrmann, A.: Parameters determining route choice in pedestrian networks. In:
TRB 90th Annual Meeting Compendium of Papers DVD, pp. 23–27. Washington, DC (2011)
21. Gonzalez, H., Han, J., Li, X., Myslinska, M., Sondag, J.: Adaptive fastest path computation
on a road network: a traffic mining approach. In: 33rd International Conference on Very Large
Data Bases (2007)
22. Joo, J.Y., Kim, S.H.: A new route guidance method considering pedestrian level of service
using multi-criteria decision-making technique. J. Korea Spat. Inf. Soc. 1983–1991 (2011)
23. Kanoulas, E., Du, Y., Xia, T., Zhang, D.: Finding fastest paths on a road network with speed
patterns. In: IEEE International Conference on Data Engineering (2006)
24. Yuan, J., Zheng, J., Xie, X., Sun, G.: T-drive: enhancing driving directions with taxi drivers’
intelligence, In: IEEE Transactions on Knowledge and Data Engineering (2012)
25. Newman, S.: Building Microservices: Designing Fine-Grained Systems. O’Reilly Media, Inc.
(2015)
26. Openstreetmap Homepage. http://www.openstreetmap.org. Last accessed 28 Nov 2020
27. Osm4routing: https://github.com/tristramg/osm4routing. Last accessed 28 Nov 2020
28. Baibing, L.: A model of pedestrians’ intended waiting times for street crossings at signalized
intersections. Transp. Res. Part B. 51, 17–28 (2013)
29. Routledge, D., Repetto-Wright, R., Howarth, I.: Four techniques for measuring the 605 expo-
sure of young children to accident risk as pedestrians. In: Proceedings of the 606 International
Conference on Pedestrian Safety, Haifa, Israel (1976)
30. Routledge, D., Repetto-Wright, R., Howarth, I.: The exposure of young children to 603 accident
risk as pedestrians. Ergonomics 17(4), 457–480 (1974)
Image-Based Malware Classification
Using Multi-layer Perceptron
Ikram Ben Abdel Ouahab, Lotfi Elaachak, and Mohammed Bouhorma
Abstract Classification of malware variants is the most challenging task in the

cybersecurity landscape. Malware developers keep one step ahead of defenders for
the sake of using advanced artificial intelligence techniques. That is why we are in
extreme need of an efficient malware classifier. In this paper, we proposed and exper-
iment a malware classifier able to affect each inputted malware into its corresponding
family. To do so, we use the multi-layer perceptron algorithm with malware visu-
alization technique. It refers to converting a malware binary into grayscale images.
Besides, to reach a great accuracy, we experiment different architectures by varying
hidden layers, neurons and activation functions. Then, we obtained an accuracy of
97.6%. At the end, we compare the obtained results with literature, and we conclude
that the multi-layer perceptron algorithm is a good malware classifier with specified
hyperparameters that were used.
1 Introduction
Malware is malicious software like adware, ransomware, virus, worm, etc. There
are a variety of malware types, and each one of them have a special goal and target.
Common targets are industrial networks, personal computers, mobile phones and
IoT devices.
The damage caused by a malware attack varies and depends on the hacker’s
attention. Malware attacks could bypass access controls, gather sensitive information,
disturb operations, block access and loss information. Sometimes the damage is very
costly.
I. Ben Abdel Ouahab (B) · L. Elaachak · M. Bouhorma

Computer Science, Systems and Telecommunication Laboratory (LIST), Faculty of Sciences and
Techniques, University Abdelmalek Essaadi, Tangier, Morocco
e-mail: ibenabdelouahab@uae.ac.ma
M. Bouhorma
e-mail: mbouhorma@uae.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_32
454 I. Ben Abdel Ouahab et al.
Most recently, dealing with malwares has been in the top of interest of every
cybersecurity researcher [14]. In fact, we found various works related to malware
analysis, detection and classification.
Otherwise, with the widespread use of artificial intelligence (AI) and its sub-
fields as machine learning (ML) and deep learning (DL) in several domains [13],
malware classification has also been taken advantage of intelligent algorithms [12].
For malware classification task, many ML and DL algorithms are used, giving
impressive results.
To classify malware samples, features are required to characterize each family.
For this purpose, we have a use of different features as API calls sequences [5],
behavioral reports based on dynamic analysis [4] and others extracted from static
analysis. A new malware feature inspired from computer vision converts malware
binaries into images and extract images features. The malware visualization tech-
nique shows similarities in samples belonging to the same family which can be used
in the classification task.
In this paper, we propose a malware classifier based on the multilayer perceptron
algorithm and using the visualization technique. We present an experimental study
by varying different hyperparameters of the artificial neural network algorithm.
The remainder of this paper is organized as follows. Section 2 presents related
works of malware classification task. Then, in Sect. 3, we present MLP algorithm that
we will use after. Section 4 focused on the dataset and data processing. Afterward, we
give experimentation use cases in Sect. 5. Then, the results are discussed in Sect. 6.
Finally, we conclude and give some of our future works.
2 Related Works
In literature, we found a use of many algorithms for malware classification task. For
instance, K-nearest neighbor (KNN), random forest (RF), support vector machine
(SVM), multi-layer perceptron (MLP), convolutional neural networks (CNN),
recurrent neural networks (RNN), etc.
In [7], authors proposed a deep learning-based malware classification model by
using an 18-layers deep residual network. Experiment result reaches an average accu-
racy of 86%. In [6], researchers performed CNN classifier with Malimg database
using visualization technique. Their proposed model achieved an accuracy of
98.52%. In addition, using the same CNN model architecture with another dataset
named Microsoft malware dataset gives highest accuracy of 99.97%.
Otherwise, researchers in [9] combined two feature named Gabor wavelet trans-
form (GWT) and GIST, then build a malware classifier using feed forward artificial
neural network (ANN). Experiment results give an accuracy of 96.35%.
In a previous work [1], we performed a KNN malware classifier based on the
visualization technique. By varying many parameters, we reached an accuracy of
97.92%. In addition, a similar classifier was made [2] using KNN and the visualization
technique, but this time, we have the use of only 50-GIST features instead of 320. The
Image-Based Malware Classification Using Multi-layer Perceptron 455
given result is very interesting. The accuracy reached using 50 features is 97.67% in
less duration.
Another deep architecture was adopted in [8], where authors proposed CNN
malware images classifier. They classify malware images into 32 different fami-
lies, by extracting local binary pattern (LBP) feature. Finally, they got an accuracy
of 93.92%.
3 Multi-layer Perceptron
Multi-layer perceptron (MLP) algorithm is a part of artificial neural networks (ANN)

[3]. The basic unit of an ANN is a network with only one node called a perceptron. The
perceptron is made of inputs, weights, activation function and output. The activation
function takes each input and its weight, multiplies them and gives the result as
output.
While the MLP is a class of ANN, it should have at least three nodes. Nodes
of MLP are arranged in layers: the input layer, the hidden layers and the output
layer. In general, there must be one or more hidden layers. For the training, MLP
uses backpropagation. The problem with MLP is to find a suitable number of hidden
layers and choose an appropriate number of neurons in each layer. So using evaluation
metrics, the MLP architecture is evaluated each time we change parameters.
Returning to the activation functions, which are very important in ANN, because
they act directly in the learning phase by mapping inputs to outputs using nonlinear
complex functions. In practice, these are the common activation functions used with
MLP hidden layers: identity, logistic sigmoid function, hyperbolic tan function (tanh)
and the rectified linear unit function (relu). Corresponding equation used with each
activation function is given in Table 1.
Hence, in this paper, we use two evaluation metrics: accuracy and confusion
matrix. Accuracy is obtained by dividing correct classifications by the total number
of samples. Also, the confusion matrix gives a clearer view of predicted values and
real ones. At the end, we added more evaluation metrics as precision, recall, f1-score
and hamming loss for comparison.
Table 1 Activation functions

Activation function Mathematical formula
definitions
(1) identity f (x) = x
(2) logistic f (x) = 1/(1 + exp(−x))
(3) tanh f (x) = tanh(x)
(4) relu f (x) = max(0, x)
4 Data Processing
Malimg dataset is one of the most known malware datasets released by the vision
research lab of UCSB. The Malimg dataset contains 9339 malware images belonging
to 25 different families (Fig. 1). The resulting malware images used visualization
technique that was firstly proposed in [10] in 2011. Visualization technique refers
to the process of converting a binary into grayscale image; an extract of the used
database is presented in Fig. 2.
In order to extract features from images, there are several descriptors that have
been used in many researches. For instance, the local binary pattern (LBP), discrete
wavelet transformation (DWT), Gabor wavelet transform (GWT) and GIST are the
most used in malware image processing.
Our proposed classifier uses GIST descriptor as malware image features. Global
texture GIST descriptor feature is a 320-dimensional vector that was firstly proposed
by AUDE OLIVA and their colleagues in [11]. Then it is been used in several
applications.
Fig. 1 Malware families Allaple.L

present in the Malimg Allaple.A
database
Yuner.A
Lolyda.AA 1
Lolyda.AA 2
Lolyda.AA 3
C2Lop.P
C2Lop.gen!g
Instantaccess
Swizzot.gen!I
Swizzor.gen!E
VB.AT
Malware
Fakerean
families
Alueron.gen!J
Malex.gen!J
Lolyda.AT
Adialer.C
Wintrim.BX
Dialplatform.B
Dontovo.A
Obfuscator.AD
Agent.FYI
Autorun.K
Rbot!gen
Skintrim.N
Fig. 2 Extract from the malware images database
5 Experimentation
The main goal in this experimentation is to have an efficient malware classifier using
advanced techniques. To do so, we chose to use the multi-layer perceptron algorithm
due to its significant performances in many applications. In addition, the visualization
technique allows us to visualize a malware binary, as a grayscale image. We adopted
this method because it is an easy, efficient and rapid manner to deal with malwares.
Furthermore, we propose two MLP architectures presented in two use cases:
• CASE 1: MLP architecture with one hidden layer is used for classifier malware.
Then, we variate number of units in this unique hidden layer.
• CASE 2: Another MLP architecture using 2 hidden layer is performed to clas-
sify malwares into their corresponding families. Then, we evaluate the model by
varying the number of units in each hidden layer.
The proposed MLP architecture of the case 1 is illustrated in the scheme of Fig. 3,
where input layer presents malwares images features. In other words, the input layer
is the GIST descriptor vector of malware images. Then, we add one hidden layer.
And we have the result as output layer.
In the same manner, we perform the case 2 which concerns the MLP classifier
architecture using 2 hidden layers (Fig. 4). As mentioned before, the input layer
regards malware features, and the output layer presents the classification result.
For each case, we change the number of units in every hidden layer, and we vary
the activation function with all instances. Activation functions used are: identity,
logistic, relu and tanh.
Fig. 3 MLP model Input 1 Hidden Output

architecture with one hidden Layer Layer
Layer
layer (Case 1)
features X n units Family(X)
Fig. 4 MLP model Input Output

architecture with two hidden 2 Hidden Layers Layer
Layer
layers (Case 2)
features X n1 units n2 units Family(X)
In this section, we present and discuss the obtained results. As the experimentation,
the results are given for each case: case 1 using MLP with 1 hidden layer and case 2
using MLP classifier with 2 hidden layers.
The neurons for our architecture were set in range of 1 to 101 units in a one-
hidden layered architecture. Table 2 presents the results that we obtain. It gives
accuracy variation each time we change the units in the hidden layer, with different
activation functions. Likewise, Fig. 5 shows more clearly the ups and downs of
the accuracy while increasing units in the hidden layer using the four mentioned
activation functions.
Results of case 1 show that the MLP malware classifier using 1 hidden layer
reached a highest accuracy of 0.9745. This highest value is obtained when we use
Table 2 Accuracy by units’ variation using different activation function for one hidden layer
architecture
Number of neurons Accuracy
Identity Logistic Relu tanh
1 0.4980 0.3983 0.3222 0.4133
11 0.9623 0.9010 0.9595 0.9584
21 0.9684 0.9555 0.9638 0.9656
31 0.9699 0.9591 0.9656 0.9738
41 0.9702 0.9591 0.9673 0.9702
51 0.9720 0.9612 0.9656 0.9720
61 0.9709 0.9666 0.9677 0.9724
71 0.9727 0.9648 0.9691 0.9738
81 0.9720 0.9663 0.9706 0.9738
91 0.9699 0.9677 0.9720 0.9717
101 0.9731 0.9670 0.9702 0.9745
0.98
0.97
0.96
0.95
Accuracy
0.94 identity
logistic
0.93
relu
0.92
tanh
0.91
0.9
11 21 31 41 51 61 71 81 91 101
Number of neurons in the Hidden layer
Fig. 5 Representation of accuracy during units’ variation in case 1
101 neurons in the single hidden layer architecture with the hyperbolic activation
function.
Moreover, we can see that the classifier had a smaller accuracy when using few
units in the hidden layer. For instance, when using only 11 neurons, the accuracy
Table 3 Important results in

Accuracy Activation function Neurons
case 1
Best 0.974524578 tanh 101
Worse 0.900968784 logistic 11
Table 4 Important results in

Accuracy Activation function Neurons
case 2: MLP with two hidden
layers Best 0.976059813 tanh [30, 100]
Worse 0.821672049 logistic [10, 10]
obtained is 0.9009 with logistic sigmoid function. A summary of highest and lowest
accuracy obtained is given in Table 3.
Overall, the variation of activation function shows significant gap. In our case, the
malware classification using visualization technique, and using these combinations
of parameters, we can say that the most suitable activation function is tanh.
Then again, for the second experimental case, we use MLP malware classifier
using two hidden layers. Results are presented in Table 4, Figs. 6 and 7. In this case,
we variate both neurons of hidden layer 1 named n1 and neurons of the hidden layer 2
named n2, which form the couple [n1, n2]. The neurons were set in range of 10 to 100
for both hidden layers. In addition, in every case, we use four activation functions:
identity, logistic, relu and tanh.
As shown in the graphic of Fig. 6, the logistic activation function curve is below all
the others. So to compare the remaining function, we extract in Fig. 7 only identity,
relu and tanh activation functions’ curves.
0.98
0.96
0.94
0.92
Accuracy
0.9
0.88 identity
0.86 logistic
0.84 relu
0.82 tanh
0.8
[10, 10]
[10, 50]
[10, 90]
[20, 30]
[20, 70]
[30, 10]
[30, 50]
[30, 90]
[40, 30]
[40, 70]
[50, 10]
[50, 50]
[50, 90]
[60, 30]
[60, 70]
[70, 10]
[70, 50]
[70, 90]
[80, 30]
[80, 70]
[90, 10]
[90, 50]
[90, 90]
[100, 30]
[100, 70]
Neurons couple for the hidden layers
Fig. 6 Representation of accuracy during units’ variation in case 2 with all activation functions
0.98
0.975
0.97
0.965
Accuracy
0.96
0.955
identity
0.95 relu
0.945 tanh
0.94
[100, 30]
[100, 70]
[10, 10]
[10, 50]
[10, 90]
[20, 30]
[20, 70]
[30, 10]
[30, 50]
[30, 90]
[40, 30]
[40, 70]
[50, 10]
[50, 50]
[50, 90]
[60, 30]
[60, 70]
[70, 10]
[70, 50]
[70, 90]
[80, 30]
[80, 70]
[90, 10]
[90, 50]
[90, 90]
Neurons couple for the hidden layers
Fig. 7 Representation of accuracy during units’ variation in case 2, with the three activation
functions
It is obvious that activation functions presented in Fig. 6 give close accuracy

values. On top, we find the tanh curve in most cases. The highest accuracy reached
is 0.9760 when we use: 30 neurons in hidden layer 1, 100 neurons in hidden layer 2
and the hyperbolic tan as activation function. The lowest accuracy is obtained when
we use relu activation function. A summary of highest and lowest accuracy obtained
in case 2 is given in Table 4.
To sum up, the proposed malware classifier based on MLP network and visualiza-
tion technique gives great accuracy when we use architecture of two hidden layers
and the hyperbolic tan activation function. The classifier reached 0.9760 accuracy
value with a network composed of 30 neurons in the first hidden layer, and 100
neurons in the second hidden layer.
Confusion matrix in Fig. 8 presents more clearly the result of this specific case.
The classifier failed totally in the family ‘Autorun.K,’ based on test data. Generally,
the MLP malware classifier performs correctly.
To summarize briefly the above-obtained results, we present in Table 5 the detailed
MLP architecture and hyperparameters as well as other’s evaluation metrics.
Let’s compare results we got with literature review. We present in Table 6 a
summary of related works showing the used classifier and the resulting accuracy
sorted.
Obtained results depend on several factors such as the nature and structure of
the database, features, classifier algorithm and others hyper parameters. That is why
you can see very different results in [6] and [8] while using CNN. In addition, KNN
classifier in [1] outperforms the current MLP classifier with small difference.
Fig. 8 Confusion matrix of the best case
Table 5 MLP architecture,

Model architecture and hyper parameters
hyper parameters and
evaluation metrics Number of hidden layers 2
Neurons per hidden layer [30, 100]
Activation function tanh
Solver adam
Learning rate 0.001 (constant)
Other evaluation metrics
Accuracy 0.976089343
F1-score macro 0.912807811
F1-score weighted 0.967928184
Hamming loss 0.026910657
Recall macro 0.91549644
Precision macro 0.910930803
Precision weighted 0.961172179
Table 6 Summary of related

Reference Method Accuracy (%)
works
[6] CNN 98.52
[1] KNN + visualization 97.92
ours MLP + visualization 97.60
[9] Feed forward ANN 96.35
[8] CNN 93.92
[7] Deep residual network 86
7 Conclusion and Perspective
In conclusion, this paper proposed a malware classifier, and its main goal is to classify
malware samples into their families effectively in short timing. For this purpose, we
performed a multi-layer perceptron classifier using malware visualization technique
based on grayscale images. In order to have the highest accuracy, we made several
experimentations include variations of: hidden layers, neurons in each hidden layer
and activation functions. At the end, the highest accuracy was reached when using
two hidden layers with, respectively, 30 and 100 neurons each, and by using the
hyperbolic tan as activation function. The performance of our classifier was evaluated
using accuracy which gives 0.9760 and also confusion matrix.
As future work, we intend to use the proposed classifier to defend against malware
attacks in real environments following a special process. Also, to improve the clas-
sifier accuracy, we could use a hybrid solution. In other words, we suppose that
a combination of machine learning and deep learning algorithms could give better
results.
Acknowledgements We acknowledge financial support for this research from the “Centre National
pour la Recherche Scientifique et Technique”, CNRST, Morocco.
References
1. Ben Abdel Ouahab, I., et al.: Classification of grayscale malware images using the K-nearest
neighbor algorithm. In: Ben Ahmed, M., et al. (eds.) Innovations in Smart Cities Applications,
3rd edn., pp. 1038–1050. Springer International Publishing, Cham (2020). https://doi.org/10.
1007/978-3-030-37629-1_75
2. Ben Abdel Ouahab, I. et al.: Speedy and efficient malwares images classifier using reduced
GIST features for a new defense guide. In: Proceedings of the 3rd International Confer-
ence on Networking, Information Systems & Security. Association for Computing Machinery,
Marrakech, Morocco (2020). https://doi.org/10.1145/3386723.3387839
3. Bishop, C.M., Bishop, P., of N.C.C.M.: Neural Networks for Pattern Recognition. Clarendon
Press (1995)
4. Galal, H.S., et al.: Behavior-based features model for malware detection. J. Comput. Virol.
Hack. Tech. 12(2), 59–67 (2016). https://doi.org/10.1007/s11416-015-0244-0
5. Jerlin, M.A., Marimuthu, K.: A new malware detection system using machine learning tech-
niques for API call sequences. J. Appl. Secur. Res. 13(1), 45–62 (2018). https://doi.org/10.
1080/19361610.2018.1387734
6. Kalash, M., et al.: Malware classification with deep convolutional neural networks. In: 2018 9th
IFIP International Conference on New Technologies, Mobility and Security (NTMS), pp. 1–5
(2018). https://doi.org/10.1109/NTMS.2018.8328749
7. Lu, Y., et al.: Deep Learning Based Malware Classification using Deep Residual Network, p. 7
(2019)
8. Luo, J., Lo, D.C.: Malware image classification using machine learning with local binary
pattern. In: 2017 IEEE International Conference on Big Data (Big Data), pp. 4664–4667 (2017).
https://doi.org/10.1109/BigData.2017.8258512
9. Makandar, A., Patrot, A.: Malware analysis and classification using artificial neural network.
In: 2015 International Conference on Trends in Automation, Communications and Computing
Technology (I-TACT-15), pp. 1–6 (2015). https://doi.org/10.1109/ITACT.2015.7492653
10. Nataraj, L., et al.: Malware images: visualization and automatic classification. In: Proceedings
of the 8th International Symposium on Visualization for Cyber Security—VizSec ’11, pp. 1–7.
ACM Press, Pittsburgh, Pennsylvania (2011). https://doi.org/10.1145/2016904.2016908
11. Oliva, A., Torralba, A.: Modeling the shape of the scene: a holistic representation of the spatial
envelope. Int. J. Comput. Vision 42(3), 145–175 (2001). https://doi.org/10.1023/A:101113963
1724
12. Sikos, L.F.: AI in Cybersecurity. Springer (2018)
13. Soufyane, A., et al.: An intelligent chatbot using NLP and TF-IDF algorithm for text under-
standing applied to the medical field. In: Ben Ahmed, M., et al. (eds.) Emerging Trends in
ICT for Sustainable Development, pp. 3–10. Springer International Publishing, Cham (2021).
https://doi.org/10.1007/978-3-030-53440-0_1
14. Souri, A., Hosseini, R.: A state-of-the-art survey of malware detection approaches using data
mining techniques. Hum. Cent. Comput. Inf. Sci. 8(1), 3 (2018). https://doi.org/10.1186/s13
673-018-0125-x
Preserving Privacy in a Smart
Healthcare System Based on IoT
Rabie Barhoun and Maryam Ed-daibouni
Abstract Internet of things (IoT) refers to a new technology made up of many

heterogeneous resources that are interconnected, controlled, monitored, and anal-
ysed through the Internet. These connected devices and resources are part of the
smart environment. Recently, many IT organizations and companies adopted this
technology. Today, modern societies are highly dependent on technology informa-
tion and communications, especially, based on IoT which becomes an emerging IT
technology being used increasingly in industry, government and healthcare. In a
smart healthcare system based on IoT, there are usually many distributed care units
with shared resources. One of the most challenging tasks in a smart healthcare system
based on IoT is the preservation of privacy. In fact, protecting personal information
from unauthorized persons is considered an essential requirement for many smart
systems. The traditional access control models do not support privacy requirements
and inflexible. In this paper, we propose a new access control model, called activity
and attribute-based access control (ABAC), which can efficiently preserving privacy
in a smart healthcare system based on IoT.
1 Introduction
Today’s smart environments are highly dependent on information and communica-

tion technologies based on the Internet of Things, which is becoming an emerging
technology and increasingly used in smart industry, smart government, smart educa-
tion and smart healthcare. In a recent report [1], considering that “things” in connected
smart environments are growing enormously, it is predicted that there will be up to
billions of devices by 2025.
In a smart healthcare environment, IoT “things” are computationally constrained
devices, such as sensors, that can detect, measure and extend connectivity between
systems and medical actors via the Internet in an omnipresent manner. Figure 1
R. Barhoun (B) · M. Ed-daibouni

Faculty of Science Ben M’sik, University Hassan II, Casablanca, Morocco
https://doi.org/10.1007/978-981-16-3637-0_33
466 R. Barhoun and M. Ed-daibouni
Fig. 1 Smart healthcare system based on IoT
presents a typical smart healthcare environment, where several heterogeneous

objects/nodes are installed.
In a smart healthcare system based on IoT, there are usually many distributed
care units with shared resources. These generate many problems such as degradation
or loss of control over access to information and privacy issue. One of the most
challenging tasks in a smart healthcare system based on IoT is privacy concerned with
protection of the personal information (PI) against unauthorized, which is considered
a critical requirement for many smart environments. The traditional access control
models do not support privacy requirements and inflexible. In this paper, we first
present an overview of access control models in a distributed smart environment to
retrieve a suitable model for such an environment based on IoT. Secondly, we propose
a new access control model, called activity-attribute-based access control (AABAC),
which can effectively enhance privacy in the smart healthcare system and produce
a more perfect and flexible mechanism of access control. In this model, the concept
of activity is crucial for the preservation of privacy. The principle of our approach
is that the tasks performed, and actions are determined by the activity purpose, and
therefore, access outside the purpose will be considered a violation of privacy. For
example, access to a patient’s medical record during a consultation activity, each
access to this record outside the purpose activity will be considered a violation of
the patient’s privacy.
The rest of the paper is organized as follows: Sect. 2 introduced an overview of the
access control models. In Sect. 3, we have a comparative analysis of access control
models for distributed smart environments. In Sect. 4, we propose an improvement
of ABAC model for a smart healthcare system based on IoT by integrating activity
Preserving Privacy in a Smart Healthcare System Based on IoT 467
concepts into the basic ABAC model. Section 5 presents the design and implemen-
tation of our AABAC model in smart healthcare systems based on IoT, and Sect. 6
concludes this paper.
2 Access Control
In computer security, access control is the enforcement of a predefined access policy

on objects to define the actions of an entity to only those requests and object to
which it is entitled. A fine-grained access control system gives control over methods,
objects and collections of resources. The dissertation thesis [2] presents an access
control policy for the IoT environment that applies a fine-grained access control
model at both the function and object-level to ensure that users with permission to
call a function also have appropriate access to target objects that are required to
successfully perform a particular task.
The access control is the mechanism by which services decide whether to accept
or deny requests according to the permissions affected to users. There are four parts
to the problem:
A. Identification: Specifying a responsible party for actions. A responsible party
may be a person or a non-person entity (NPE), such as a computer or a router.
We will use the term user to cover both cases.
B. Authentication: Used to prove the right to use an identity, take on a role or
prove possession of one or more attributes.
C. Authorization: Paper [3] express the access policy by explicitly granting a
right.
D. Access Decision: Applying some combination of the other three to decide
whether a request should be honored.
Many access control models are used, including mandatory access control (MAC),
discretionary access control (DAC), identity-based access control (IBAC), role-based
access control (RBAC) and attribute-based access control (ABAC).
2.1 Dictionary Access Control (DAC)
DAC is a method of restricting access to an object based on the entity identity (e.g.,
user, process and group). This type of access control is called discretionary because
the entity with a certain privilege can pass its privileges or security attributes to
another entity. This results in, however, the problem of loss of confidentiality of
information. DAC is widely used in many networks and operating systems.
2.2 Mandatory Access Control (DAC)
MAC is a method of constraining an entity from accessing or performing operations

on a resource, based on predefined security attributes or labels assigned to the entity
and the target object.
2.3 Identity-Based Access Control (IBAC)
Access control based on the identity (IBAC—identity-based access control) is among

the first access control models. This model introduces the basic concepts of subject,
action and object. The objective of this model IBAC is to control direct access of
subjects to objects using actions. This control is based on the identity of the subject
and the object identifier.
2.4 Role-Based Access Control (RBAC)
The papers [4–7] proved that the RBAC model is versatile and conforms closely to
the organizational model used in firms. RBAC meets this requirement by separating
users from roles. Work [8] discusses that access rights are given to roles, and roles are
assigned to users. Here, the role combines users and privileges. Roles are created for
various job functions, and users are assigned roles based on their qualifications and
responsibilities. RBAC is thus more scalable than user-based security specifications
and greatly reduces the cost and administrative overhead.
This model is very simple and easy to use. And, it is considered to be the best model
of access control for the local domain. The roles are assigned to the user statically
by the security administrator that is not preferable in a dynamic environment. Also,
it is difficult to change the privilege of the user without changing the role of the user.
Furthermore, RBAC becomes problematic in a distributed and dynamic environment
and has no delegation models which are required in such an environment. Also, work
[4] discusses that problem of “Role Explosion” that can arise when used to support
dynamic attributes in large organizations, in thousands of separate roles for different
collections of permissions.
2.5 Attribute-Based Access Control (ABAC)
In traditional policies, the users can obtain their privileges through roles or directly,
but it is possible that users can give certain privileges that they do not really need. This
is contradictory to the least privilege principle which requires a subject that must be
able to access only the information and resources that are necessary for its purpose.
Currently, how to assign privileges to subjects so as to achieve this principle is still not
solved. Papers [9, 10] present that ABAC is a recent policy that has drawn particular
attention; its principle of decision is based on taking into account the attributes of
different actors (subject, object and environmental conditions) before giving access
to resources.
The overview of AABAC model is shown in Fig. 2, the permissions containing
the combination of an object and operations, where the subject access to the object
according to certain conditions. The operation describes the instructions to execute
on the objects. Access rights can be defined in a subject attribute and permission.
Paper [9] explains that ABAC model can dynamically assign permissions to subjects
and objects. ABAC uses subject, object and their environmental attributes. Papers
[11, 12] show that in model ABAC before using the attributes to make an access
control decision, the integrity and validation of these attribute resource is verified.
The ABAC model is very flexible and supportive in an environment which is
large, open, distributed, sharable and collaborative and where the numbers of users
are very high and most of the users are unknown before and the roles of the users are
statically or not defined in advance. Furthermore, it supports the global agreement
features such as the user attributes that are provided in one domain can be forwarded
to the other domain at the point of domain-to-domain interaction.
This model has the high complexity due to the specification and maintenance of
the policies. In addition, there is a problem of mismatching and confusing attributes
especially when those attributes provided by the user do not necessarily match to
those used by the service provider of a web-based system or service. Furthermore,
it increases the privacy, flexibility, sharing and global agreement and provides the
interoperability among several service providers which can use these attributes data
dynamically and can decide upon user rights.
Subject Environment&
Object
attributes conditions
attributes
attributes
Subject Object Operations Environment

Permissions
Fig. 2 Attribute-based access control

3 Comparative Analysis of Access Control on Distributed

Smart Environment
In this section, we present a comparative analysis of MAC, DAC, IBAC, RBAC,

ABAC and AABAC with some point of issues. This comparison is based on different
factors, which will help us to choose the most efficient access control model suitable
for a smart environment based on IoT. Those factors are dynamicity, distributed
systems, global agreement, flexibility, simplicity, authorization decision, granularity,
manageability, trust, changing privileges, policies specification and maintenance, role
explosion, contextual information and scalability.
The following table summarizes this comparison:
MAC DAC RBAC IBAC ABAC

Dynamicity No No No No Yes
Distributed systems No exactly No exactly No exactly No exactly Yes
Global agreement No No No No Yes
Flexibility No No No No Yes
Simplicity Yes Yes Yes Yes No
Authorization Locally Locally Locally Locally Globally
decision
Granularity Low Low Low Low High
Manageability Simple Simple Simple Complex Complex
Trust Locally Locally Locally Locally Globally
Changing privileges Simple individual user Simple Complex Simple Simple
cannot change access
rules
Policies specification Low Low Simple Low Complex
and maintenance
Role explosion No No Yes No No
problem
Contextual Low Low Low Low High
information
Scalability Low Low Low Low Yes
From the above table, it is clear that attribute-based access control (ABAC) is the
most suitable for dynamic distributed smart environment. This model granted access
to users based on the attributes of the requesting user. It uses multiple attributes
for authorization decision, which enables the system to be highly flexible, scalable,
interoperable, and multifunctional access control that may deal with diverse security
requirements in distributed environment based on IoT.
4 Our Proposed AABAC Model Based on IoT
The ABAC model has drawbacks with reference to privacy concerns. Indeed, because
of the descriptive nature of subject attributes, implementation of attributes by sharing
capabilities causes a problem of increasing the risk of privacy violation of personally
identifiable information by dint of involuntary exposure of attribute data to untrusted
third parties or aggregation of sensitive information in environments less protected.
A second consideration is that releasing attributes to the policy evaluating engine is
a sensitive activity as the third party may not be trusted.
According to security constraints, the principle of least privilege as an access
control policy allows the assignment of least rights to different actors. This prin-
ciple is important in preserving privacy and sensitive resources in the design of an
access control policy in a smart environment. In our previous work [13], we proposed
a model called medical activity-attribute-based access control model (MA-ABAC)
which extends the functionality of ABAC by introducing the concept of medical
activity, defined by the purposes of treatment, to meet privacy concerns in a collab-
orative healthcare environment. In MA-ABAC model, if a privilege is used outside
the purpose of the activity, then it is considered a violation of the principle of least
privilege and then a violation of privacy. To take into account the smart environment
based on IoT, we propose the activity-attribute-based access control (AABAC) model
shown in Fig. 3.
Fig. 3 Activity-attribute-based access control (AABAC) model for smart healthcare system based
on IoT
(b) Hierarchical structure of activities (a) Linear structure of activities (c) Hybrid structure of activities
Fig. 4 Different structures of activities based on IoT
In this design, activity (A) is an abstraction of the collaborative works, which

provides a scope for resources, privileges and roles defined by a medical activity
purpose (P) in a medical unit, in which the participants (collaborators) realize their
tasks in order to achieve the medical activity purposes. Indeed, according to our
approach, the smart healthcare system allows a smart group of medical actors to
work together in order to treat a patient admitted to a hospital.
In a smart collaborative healthcare system, many medical activities exist, and
new activities are created. An activity can also span across multiple environments
based on IoT. In most cases, the distributed collaborative environments need to be
structured hierarchically. A medical collaboration specification needs to support the
decomposition of tasks into activities/sub-activities and planning how such activities
can perform. Activity concept, planning, and its decomposition are widely used
in collaboration systems, specifically in workflow environments. In our study, the
activities/sub-activities can be classified or nested in the same way that human work
activities are created. These activities can be mainly structured in different manners:
hierarchical, linear or hybrid as illustrated in Fig. 4a, b and c.
5 Design and Implementation of the AABAC Model

in a Distributed Healthcare Environment Based on IoT
In a smart healthcare system, we consider, for example, that a general physician

trying to find a diagnosis (treatment) for a patient. This process is typically lengthy
and goes through several medical activities, see Fig. 5, which updates the patient’s
medical record in each medical activity.
The physician enters notes in the smart medical record (MR). During the radiology
activity (or biology analysis activity), the general physician studies x-ray images (or
biology analysis) with a radiologist (or with a biologist). At the expertise activity,
the general physician discusses proper medication with colleagues while browsing
medicine catalogs. Otherwise, it is necessary to start another medical activity, as
shown in Fig. 6.
Consultation Activity Biology Activity
Radiology Activity Expertise Activity
Fig. 5 Hybrid medical activities of general diagnosis in smart healthcare based on IoT
Consultation Activity: General Physician St:X1 Ft:Y1

MR1
MP: General consultation
General Physician Patient Nurse Administrator
Radiology Activity: Radiologist St:Xb Ft:Yb Biology Activity: Biologist St:Xa Ft:Ya
MR1b MR1a
MP: Radiology session MP: Biology analysis session
Radiologist Patient Nurse Administrator Biologist Patient Nurse Administrator
Expertise Activity: General Physician St:X2 Ft:Y2 MR2=1a+1b

Unchanged entity
MP: Expertise session
Changed entity
General Physician Patient Nurse Administrator Radiologist Biologist
Creator
Fig. 6 The activities of general diagnosis process of a patient in smart healthcare
The proposed architecture of our mechanism for preserving privacy and protected
resources in smart healthcare system is illustrated in Fig. 7.
When initiating an activity, the activity decision module (ADM) retrieves the
attributes of the triggered activity from policy information point (PIP) entity. The
policy activity server will generate an XML file containing all the permissions for the
activity created. The ADM manager will decide whether an action (executed in the
medical activity) is allowed or not by consulting the XML file relating to the active
medical activity on the one hand, and on the other hand based, on other attributes
relating to subject, environments and object retrieved from PIP entity, respectively,
by the managers: subject manager module (SMM), environment manager module
(EMM) and object manager module (OMM). The administration activity policy
(AAP) is used to define medical activities and their policies, see Fig. 8.
Fig. 7 Architecture of AABAC mechanism in smart healthcare system
The prototype developed was based on the Java language and API framework on
Linux Ubuntu 16.04 (32 bits). The choice of the Java language is based on the fact
that it is platform-independent; therefore, it runs on a heterogeneous platform.
6 Conclusion
In this paper, we have presented a critical review of the main access control models.
We have examined the advantages and disadvantages of each model. This investiga-
tion led us to choose the ABAC model as the most suitable model for all distributed
smart environments, although this model presents some drawbacks such as privacy
violation. We then improved this weakness by proposing a new model called AABAC
based on activity concept. In this model, if a privilege is used outside the activity
purpose, then this access is considered as a violation of privacy. Our access control
model not only guarantees the properties of access control but also preserving privacy
in IoT environment.
We presented the design of our implementation in the smart healthcare system
based on IoT. In the current implementation, we focused on the medical activity
concept; this concept encompasses healthcare requirements as well as the imple-
mentation of the principle of least privileges, and then the preservation of privacy.
We believe that our approach can be adapted to support any distributed smart
environment.
PDP
Subject PEP ADM SMM EMM OMM PIP PAA Resource
Access request
to the ressource
Access request to Load Medical-
resource in the Medical-Activity Activity policy
medical activity policy request
Reponse
Medical-activity request Subject attributes

for subject attributes Subject attributes
request retrieves
Reponse
Rponse
Medical-Activity request for Environment

evironment attributes Environment attributes attributes retrieves
request
Reponse
Reponse
Preserving Privacy in a Smart Healthcare System Based on IoT
Medical-Activity request
for object attributes Object attributes Object attributes
request retrieves
Reponse
Reponse
evaluates
access policy
granted access
to resource
Fig. 8 The AABAC authorization process—diagram sequence

475
References
1. Lee, Y.T., Hsiao, W.H., Lin, Y.S., Chou, S.C.T.: Privacy-preserving data analytics in cloud-
based smart home with community hierarchy. IEEE Trans. Consum. Electron. 63(2), 200–207
(2017)
2. Faraji, M.: Identity and access management in multi-tier cloud infrastructure. Doctoral
Dissertation, University of Toronto (2013)
3. Karp, A.H., Haury, H., Davis, M.H.: From ABAC to ZBAC: the evolution of access control
models. J. Inf. Warfare 9(2), 38–46 (2010)
4. Ahn, G.J., Sandhu, R.: Role-based authorization constraints specification. ACM Trans. Inf.
Syst. Secur. (TISSEC) 3(4), 207–226 (2000)
5. Bertino, E., Bonatti, P.A., Ferrari, E.: TRBAC: a temporal role-based access control model.
ACM Trans. Inf. Syst. Secur. (TISSEC) 4(3), 191–233 (2001)
6. Joshi, J.B., Bertino, E., Latif, U., Ghafoor, A.: A generalized temporal role-based access control
model. IEEE Trans. Knowl. Data Eng. 17(1), 4–23 (2005)
7. Li, N., Tripunitara, M.V.: Security analysis in role-based access control. ACM Trans. Inf. Syst.
Secur. (TISSEC) 9(4), 391–420 (2006)
8. Kalajainen, T.: An Access Control Model in a Semantic Data Structure: Case Process Modelling
of a Bleaching Line. Department of Computer Science and Engineering (2007)
9. Hu, V.C., Kuhn, D.R., Ferraiolo, D.F., Voas, J.: Attribute-based access control. Computer 48(2),
85–88 (2015)
10. Brossard, D., Gebel, G., Berg, M.: A systematic approach to implementing ABAC. In:
Proceedings of the 2nd ACM Workshop on Attribute-Based Access Control, pp. 53–59 (2017)
11. Biswas, P., Sandhu, R., Krishnan, R.: Label-based access control, an ABAC model with enumer-
ated authorization policy. In: ABAC ’16 Proceedings of the 2016 ACM International Workshop
on Attribute Based Access Control, pp. 1–12 (2016)
12. Mukherjee, S., Ray, I., Ray, I., Shirazi, H., Ong, T., Kahn, M.G.: Attribute Based Access (2017)
13. Barhoun, R., Ed-daibouni, M., Namir, A.: An extended attribute-based access control (ABAC)
model for distributed collaborative healthcare system. Int. J. Serv. Sci. Manage. Eng. Technol.
(IJSSMET) 10(4), 81–94 (2019)
Smart Digital Learning
Extracting Learner’s Model Variables for
Dynamic Grouping System
Noureddine Gouasmi, Mahnane Lamia, and Yassine Lafifi
Abstract Collaborative learning is of great help in improving the training of

learners in e-learning systems. But the efficiency of collaborative learning depends
on the quality of the learning groups formed. The question is therefore: Which vari-
ables should be used to group learners in order to reach the most effective col-
laboration between them? In this paper, we present a literature review of variables
describing learner’s model, which can be used for grouping learners in collabora-
tive e-learning systems. Variables are ranged in 11 classes, and possible associations
between variables are considered, after extracting frequent patterns from variables
using FP-growth algorithm, and generating association rules. The results show that
four variables from learner’s are mostly used with association with other variables.
These variables are: “Number of accesses to the platform”, “Number of complete
activities”, “Platform time” and “Final Grades”.
1 Introduction
Collaboration is the act of working with one or more other people. It involves a
cyclical process of renegotiation. Collaboration evolves as parties interact over time
[12].
In the learning domain, “collaborative learning” is any learning activity carried out
by members of a group of learners having a common objective, in order to succeed
in their learning. Each learner being a source of information, motivation, interaction,
N. Gouasmi (B) · Y. Lafifi

Labstic, 08 Mai 45 University, P.O. Box 401, 24000 Guelma, Algeria
e-mail: gouasmi.noureddine@univ-guelma.dz
Y. Lafifi
e-mail: lafifi.yacine@univ-guelma.dz
M. Lamia
LRS Laboratory, University of Badji Mokhtar, P.O. Box 12, 23000 Annaba, Algeria
e-mail: mahnane_lamia@yahoo.fr
https://doi.org/10.1007/978-981-16-3637-0_34
480 N. Gouasmi et al.
mutual assistance and each benefiting from the contributions of others and the help
of a trainer to facilitate individual and collective learning [14].
Collaborative learning systems help learners to work in groups. It allows document
sharing between members and provides tools to upload, download and archive several
types of documents [11]. It also provides means for exchanging information between
learners of the same group [7], enables learners to develop cognitive capacity and
knowledge necessary for the development of collaborative skills [3], gives learner
the possibility of managing his time and providing a wider choice of activities more
suited to their needs and interests, increases reflection time during collaborative
online learning and allows to combine learner’s contributions [14].
For a collaboration to be effective, it is important to choose the “best” members
of a collaborative group. According to Bekele [2], the process of grouping learners
is focused on three key points: how to assign a learner to a group, size of groups and
whether the group is heterogeneous or homogeneous [15].
The paper is organized in two main parts. The first one shows variables used to
describe learner’s model, which can be used for grouping learners. The second part
presents the analyses of learner’s model variable, by describing the methodology
used in this work, and after, by showing results of data analysis.
2 Variables Used to Describe Learner’s Model
This research analysed 63 primary studies published between 2001 and 2019. After
analysis of the 63 articles approved, it was possible to find 76 variables. The variables
were categorized into seven layers for a better understanding, and each layer repre-
sents only the subject that variable approaches. A variable can be categorized in more
than one layer. For example, the variable “area of employment” could fit in layers
“User’s Information” and “Psychological”; however, in this work each variable will
be presented in only one category. The layers are detailed in Table 1.
Table 1 Variables’ layers

Layer Number of variables
Communication 14
Access 8
Activities 8
Actions 16
Time 10
User’s information 8
Psychological 12
Extracting Learner’s Model Variables for Dynamic Grouping System 481
Table 2 Variables found that are referring to communication

Sublayer Variables
General Which communication resource was used
Communication level
Forum Participation
Total amount of posts
Posts per forum
Total amount of replies
Number of access
Access frequency
Added files
Content of the message
Chat Number of messages
People the student interacted with
Comments Number of comments made
Number of comments read
• Communication
The first layer, which addresses communication, represents all the data found that
refers to some conversation, whether by posts in forums or by direct messages, this
communication can also be considered between students or between the student and
the teacher.
Communication is an essential element that must be analysed. According to Takaf-
foli, 2012 [9], in order to fully appreciate the participation of students, we need to
understand their patterns of interactions and answer questions like who is involved
in each discussion, who is the active/peripheral participant in a discussion thread.
The number of interactions between students or teachers can represent how
engaged the student is. According to Zhang [16], “for using the discussion board, the
total number of postings (all were qualified ones) by each student on the discussion
board was used as the measure for participation or communication in class. This
measure reflects how involved a student was in his or her learning.”
During the analysis, 14 variables emerged, referring to communication. These
variables divided into four sublayers relating to the origin of this communication,
and these variables are presented in Table 2.
• Access
The second layer is referring to access. Social learning systems provide data about
students’ access, and it means that it is possible to know when students were online
and which pages they were accessing. It is important to detect what are the primary
students’ access pages to understand what type of content they are accessing.
Table 3 Variables found that are referring to access

Layer Variables
Access Number of accesses per day to the platform
Number of accesses to the platform
Access platforms in specific sessions
First and last file accessed
Single views to specific pages
Number of pages visited
Number of views per page
Time viewing specific pages
Table 4 Variables found related to the activities

Layer Variables
Activities Number of complete activities
Answers given in the activity
Tests done
Reading of activities
Grades in activities
Time spent on activities
Activities done correctly
Number of attempts on the same activity
The number of visualizations is one order of magnitude above creations and updates, which
in turn are on order above deletions. […]. This situation clearly derives from two facts: the
first is that for accessing every creation, update or delete page, we need first to visualize it;
the second, relates to the natural curiosity of people, associated with the fear to ‘act’ in spite
of just ‘observe’. (Figueira 2017) [4]
Eight variables were found, referring to access. These variables are presented in
the following Table 3.
• Activities
As in the face-to-face learning mode, the activities done by the student are a way
to measure his interest, performance and learning. Social learning systems provide
information to the teacher about the proposed activities. It is possible to analyse not
only activities done by students but also the period in which the student has delivered
it and whether it is delayed. This layer presents eight variables that can be found in
Table 4.
All the record of actions made by the user within the platform is detected through
logs, clicks or events. This information allows identifying how the student behaves
inside the platform.
Figueira [4] explains “We have counted the number of distinct students accesses,
of different resources being used; of events logged by the system. Then, we were
Table 5 Variables found that are referring to the actions of users

Sub-layer Variables
Events Affected user
Which event
Context of the event
Complete events
Time the event occurred
Number of events
Logs General logs
Keywords
Number of logs
Clicks General clicks
Clicks by activity
Hyperlinks
Number of clicks
Number of clicks per session
Downloads Downloaded files
Size of downloaded files
able to differentiate between four types of actions performed by students: creations,

visualizations, updates, and deletions.”
• Actions
This layer can be divided into four sublayers, information about logs, clicks, down-
loads and events made by the user (Table 5).
• Time
Information about the student’s study time is essential to detect how engaged that
student is in his or her studies. While, in the face-to-face modality, it is not possible to
identify how much time the user is studying at home, social learning system provides
data reporting about how long the student is online.
A student who devotes his time to studies may present better results, and this rela-
tionship is confirmed by Romero and Barbera [10] “We observed a slight correlation
between the time-on-task devoted by students on a weekly basis and their academic
performance.”
Table 6 presents the variables of this layer divided into two categories, the one in
which it considers the duration and the one that considers the schedule of the action.
• User’s Information
Personal information about the student is often used to better identify each student’s
profile. Primary studies currently point some data like gender, where the students
live. The eight variables found in this layer are presented in Table 7.
Table 6 Variables found related to time

Sub-layer Variables
Duration Platform time
Time between activities
Time between the original post and the activity
done
Time between posting activity and deadline
Time per session
Average idle time
Average time per week
Schedule Time of day
Weeks without posting
Date of access
Table 7 Variables found related to the user’s information

Layer Variables
User’s information ID
Full name
Age
Gender
IP address used
Appearance
Culture
Student profile
• Psychological Layer
The psychological layer uses data related to the psychological profile of the learner
(Table 8).
However, it is clear that none of these variables can be considered singly and sep-
arately. All studies approved for this research used the variables together to achieve
the desired result and a better understanding of the data.
3 Data Analysis
After establishing a list of variables used to describe the learner’s model, we analyse
these variables to determine if it is possible to extract any relationships or associations
between them. For this, we will start by extracting frequent patterns combining
the variables, and then from these patterns, we will generate association rules. The
extraction of frequent patterns is carried out with the FP-growth algorithm, and then
from the result obtained, we generate association rules.
Table 8 Variables found related to psychological layer

Layer Variables
Psychological Group work attitude
Sociability
Openness
Extroversion
Agreeableness
Neuroticism
Engagement
Self-confidence
Conscientiousness
Area of employment
Intent
Learner’s preferences
3.1 Frequent Patterns Mining
Research on frequent pattern algorithms started in the 1990s. The goal was to detect
sets of data that appears recurrently on a database, not to classify instances, but to
determine which items emerge frequently and which items are associate [5].
Frequent pattern mining is mostly used in customer transaction analysis. It
attempts to identify associations, or patterns, between various items that have
been bought by a particular costumer, which tends to costumer behaviour analy-
sis, where we seek to extract associations such as: “a customer who buys item x, also
buys item y”.
Extraction of frequent patterns consists in searching in a dataset (itemset) for
groups of items which appear s times, at least. s is the minimum support, correspond-
ing to the smallest value from which a set of items is considered to be frequent [1].
FP-Growth Algorithm
FP-growth instead of Apriori algorithm is not based on candidate generation. It stands
on two paradigms [3]:
• A compact representation of the database, with a structure called FP-tree.
• A divide-and-conquer strategy for exploring the data.
The FP-growth algorithm can be described as follows [1]:
1. Constructing FP-tree by compressing the DB representing frequent items,
2. Extracting conditional FP-tree for a selected item,
3. From the conditional FP-tree, generate frequent itemsets.
The main advantage of FP-growth algorithm is that it reduces the scan time com-
pared to Apriori algorithm [6], while its disadvantage is that it requires large memory
space for conditional FP-tree in the worst case [8].
Fig. 1 The analysis process
Association Rules
Frequent patterns are often used to generate association rules. They describe corre-
lations among items in a pattern.
An association rule is an implication X => Y, where X and Y are sets of items.
The confidence value is the ratio of the support of X U Y to that of the support of X.
It means that if the antecedent X is satisfied, then it is probable that the consequent
Y is satisfied too [13].
Association rules are generated in three steps [1]:
1. Generate all the frequent patterns in the itemset at a minimum support level,
2. Extract all the rules from the frequent patterns,
3. Keep only rules whose confidence is greater than a minimum level of confidence.
3.2 Analysis and Results
For our experimentation, we want to find possible association rules between 76

learner’s model variables. A PHP program, implementing FP-growth algorithm and
association rules generation, was executed on the variable. Figure 3 shows interfaces
of the application. FP-growth algorithm and rules association were applied with a
min support equal to 2, which will allow us to take into account only the variables
appearing in at least two papers. The confidence value is set to 0.75.
For implementing FP-growth algorithm, we first transform the values to constitute
a binary vector corresponding to a transaction of the itemset.
The analysis process steps are (see Fig. 1):
1. Construct a table which columns are student model variables.

2. Each row of the table matches a referenced paper, so if a variable is cited in
the paper, its corresponding column receives the value 1, 0 otherwise. Figure 2
shows a representation the variables vectors, where the first paper cites the first
variable, but not the second one, and so on.
Fig. 2 Variables’ vectors
Fig. 3 Interface of FP-growth and AR generation program
Fig. 4 Extracted frequent patterns histogram
3. After having created all vectors of the student model variables, they are merged
to constitute a transaction. The set of transactions represents the itemset to be
analysed with the FP-growth algorithm.
4. Frequent patterns in the itemset are generated using the FP-growth algorithm.
5. Finally, the association rules are extracted from frequent patterns.
Variables, with their relative supports for 63 papers (for min support = 2), are
shown in Fig. 4.
After applying FP-growth algorithm, we obtained only 52 frequent patterns (with
a support value great or equal to 2). Table 9 shows the frequent patterns extracted.
From these frequent patterns, 36 rules were generated (Table 10).
Table 9 Extracted frequent patterns

No. Frequent pattern Support
1 Number of accesses to the platform & Total amount of posts 8
2 Number of accesses to the platform & Platform time 7
3 Platform time & Final grade 6
4 Total amount of posts & Total amount of replies 5
5 Total amount of posts & Number of views per page 5
6 Number of accesses to the platform & Final grade 5
7 Platform time & Number of complete activities 5
8 Number of accesses to the platform & Number of views per page 4
9 Number of accesses to the platform & Total amount of posts & Final grade 4
10 Total amount of posts & Final grade 4
11 Number of accesses to the platform & Number of complete activities 4
12 Sociability (Big5 Factor) & Extraversion (Big5 Factor) 3
13 Sociability (Big5 Factor) & Group work attitude 3
14 Number of accesses to the platform & Number of pages visited 3
15 Platform time & Number of complete activities & Final grade 3
16 Number of complete activities & Final grade 3
17 Number of accesses to the platform & Total amount of posts & Platform 3
time
18 Total amount of posts & Platform time 3
19 Number of complete activities & Tests done 2
20 Number of accesses to the platform & Number of comments read 2
21 Sociability (Big5 Factor) & Group work attitude & Extraversion (Big5 2
Factor)
22 Group work attitude & Extraversion (Big5 Factor) 2
23 Platform time & Full Name 2
24 Clicks by activity & Number of clicks 2
(continued)
Table 9 (continued)
No. Frequent pattern Support
25 Total amount of posts & Total amount of replies & Content of the message 2
26 Total amount of posts & Content of the message 2
27 Total amount of replies & Content of the message 2
28 Platform time & Number of complete activities & Final grade & Time 2
spent on activities
29 Platform time & Number of complete activities & Time spent on activities 2
30 Platform time & Final grade & Time spent on activities 2
31 Platform time & Time spent on activities 2
32 Number of complete activities & Final grade & Time spent on activities 2
33 Number of complete activities & Time spent on activities 2
34 Final grade & Time spent on activities 2
35 Total amount of posts & Participation 2
time & Total amount of replies
37 Number of accesses to the platform & Total amount of posts & Total 2
amount of replies
38 Number of accesses to the platform & Platform time & Total amount of 2
replies
39 Number of accesses to the platform & Total amount of replies 2
40 Total amount of posts & Platform time & Total amount of replies 2
41 Platform time & Total amount of replies 2
42 Number of accesses to the platform & Total amount of posts & Number of 2
views per page
43 Number of accesses to the platform & Number of pages visited & Number 2
of views per page
44 Number of pages visited & Number of views per page 2
45 Platform time & Number of complete activities & Number of pages visited 2
46 Platform time & Number of pages visited 2
47 Number of complete activities & Number of pages visited 2
time & Final grade
49 Number of accesses to the platform & Platform time & Final grade 2
50 Total amount of posts & Platform time & Final grade 2
51 Number of accesses to the platform & Total amount of posts & Number of 2
complete activities
52 Total amount of posts & Number of complete activities 2
Table 10 Extracted frequent patterns

No. Association rule Confidence
1 Group work attitude & Extraversion (Big5 Factor) ⇒ Sociability 1
(Big5 Factor)
2 Extraversion (Big5 Factor) ⇒ Sociability (Big5 Factor) 0.75
3 Group work attitude ⇒ Sociability (Big5 Factor) 0.75
4 Total amount of replies & Content of the message ⇒ Total amount of 1
posts
5 Total amount of posts & Content of the message ⇒ Total amount of 1
replies
6 Final grade & Time spent on activities ⇒ Platform time & Number of 1
complete activities
7 Number of complete activities & Time spent on activities ⇒ Platform 1
time & Final grade
8 Number of complete activities & Final grade & Time spent on 1
activities ⇒ Platform time
9 Platform time & Time spent on activities ⇒ Number of complete 1
activities & Final grade
10 Platform time & Final grade & Time spent on activities ⇒ Number of 1
complete activities
11 Platform time & Number of complete activities & Time spent on 1
activities ⇒ Final grade
12 Number of complete activities & Time spent on activities ⇒ Platform 1
time
13 Platform time & Time spent on activities ⇒ Number of complete 1
activities
14 Final grade & Time spent on activities ⇒ Platform time 1
15 Platform time & Time spent on activities ⇒ Final grade 1
16 Final grade & Time spent on activities ⇒ Number of complete 1
activities
17 Number of complete activities & Time spent on activities ⇒ Final 1
grade
18 Platform time & Total amount of replies ⇒ Number of accesses to the 1
platform & Total amount of posts
19 Total amount of posts & Platform time & Total amount of replies ⇒ 1
Number of accesses to the platform
20 Number of accesses to the platform & Total amount of replies ⇒ Total 1
amount of posts & Platform time
21 Number of accesses to the platform & Platform time & Total amount 1
of replies ⇒ Total amount of posts
22 Number of accesses to the platform & Total amount of posts & Total 1
amount of replies ⇒ Platform time
(continued)
Table 10 (continued)
No. Association rule Confidence
23 Number of accesses to the platform & Total amount of replies ⇒ Total 1
amount of
24 Platform time & Total amount of replies ⇒ Number of accesses to the 1
platform
25 Number of accesses to the platform & Total amount of replies ⇒ 1
Platform time
26 Platform time & Total amount of replies ⇒ Total amount of posts 1
27 Number of pages visited & Number of views per page ⇒ Number of 1
accesses to the platform
28 Number of complete activities & Number of pages visited ⇒ Platform 1
time
29 Platform time & Number of pages visited ⇒ Number of complete 1
activities
30 Number of complete activities & Final grade ⇒ Platform time 1
31 Total amount of posts & Platform time & Final grade ⇒ Number of 1
32 Number of accesses to the platform & Platform time & Final grade ⇒ 1
Total amount of posts
33 Total amount of posts & Final grade ⇒ Number of accesses to the 1
platform
34 Number of accesses to the platform & Final grade ⇒ Total amount of 0.8
posts
35 Total amount of posts & Number of complete activities ⇒ Number of 1
36 Total amount of posts & Platform time ⇒ Number of accesses to the 1
platform
4 Discussion and Proposed Model
The first observation to be made about variables is that the most frequent variables
only belong to six layers: grades layer, time layer, activities layer, access layer,
communication layer (forum sublayer). This leads to the conclusion that in the papers
studied, the profile of the learner is built mainly from his grade and his acts on the e-
learning systems (activities, access and communication). Naturally, studying learner
activities and communication leads to study his accesses to the learning platform,
which is confirmed by rules (19, 24, 27, 31, 33, 35 and 36).
Other observation, papers using “Total amount of posts” as learner profile vari-
able (rules 21, 23, 26, 32 and 34) uses also access variable (“Number of accesses
to the platform”) coupled with other layer variable (time layer and communication
layer). For papers using “Platform time”, they mainly use “Number of complete
activities” or “Number of accesses to the platform” coupled with other variable
(rules 8, 12, 14, 22, 25, 28 and 30). For the variable “Number of complete activi-
ties”, the link can be explained by the fact that completing activities for a learner is
closely linked to deadlines, and therefore to platform time. We can also observe that
the counterpart is also true, while “Number of complete activities” is associated
with “Platform time” coupled with other variables (rules 10, 13, and 29).
“Sociability” variable is associated with two other variables (“Group work atti-
tude” and “Extroversion”—rules 1, 2 and 3). It is clear that sociabilty and extrover-
sion are very related while both are defined by social engagement, and thus attitudes
in group (the “Group work attitude” variable).
Finally, “Final Grades” variable is associated with combinations of three vari-
ables (Platform time, Number of complete activities and Time spent on activities)
in rule 11 (and also rules 15 and 17).
From this analysis, we can draw as a conclusion that to have learners profile
which can be used for grouping learners, we can build this model from the vari-
ables: “Number of accesses to the platform”, “Number of complete activities”,
“Platform time” and “Final Grades”. These variables appear mainly in the most
association rules with other variables. Moreover, only a few papers adopt multiple
intelligences theories in a dynamic grouping learner [17].
5 Conclusion
This paper presents a review of learner’s model variables which could be used for
dynamic group formation for collaborative learning. These variables were analysed
with frequent pattern extraction method (FP-growth) and association rule generation
concerning the variables. The results show that, from the analysed papers, only
four variables from learner’s are mostly used with association with other variables.
These variables are: “Number of accesses to the platform”, “Number of complete
activities”, “Platform time” and “Final Grades”. They can be good candidates as
criteria for grouping learners. The next step is to use them with multiple intelligences
theories in a dynamic grouping system.
References
1. Aggarwal, C.C., Bhuiyan, M.A., Al Hasan, M.: Frequent pattern mining algorithms: a survey.
In: Frequent Pattern Mining, pp. 19–64. Springer (2014)
2. Bekele, R.: Computer-assisted learner group formation based on personality traits. Ph.D. thesis.
Staats-und Universitätsbibliothek Hamburg Carl von Ossietzky (2005)
3. Benraouane, S.A.: Guide pratique du e-learning: stratégie, pédagogie et conception avec le
logiciel Moodle. Dunod (2011)
4. Figueira, Á.: Mining moodle logs for grade prediction: a methodology walk-through. In: Pro-
ceedings of the 5th International Conference on Technological Ecosystems for Enhancing
Multiculturality, pp. 1–8 (2017)
5. Fournier-Viger, P., Lin, J.C.W., Nkambou, R., Vo, B., Tseng, V.S.: High-Utility Pattern Mining.
Springer (2019)
6. Gu, X.F., Hou, X.J., Ma, C.X., Wang, A.G., Zhang, H.B., Wu, X.H., Wang, X.M.: Comparison
and improvement of association rule mining algorithm. In: 2015 12th International Computer
Conference on Wavelet Active Media Technology and Information Processing (ICCWAMTIP),
pp. 383–386. IEEE (2015)
7. Gweon, G., Jun, S., Lee, J., Finger, S., Rosé, C.P.: A framework for assessment of student
project groups on-line and off-line. In: Analyzing Interactions in CSCL, pp. 293–317. Springer
(2011)
8. Kavitha, M., Selvi, S.: Comparative study on apriori algorithm and FP growth algorithm with
pros and cons. In. J. Comput. Sci. Trends Technol. (IJCST) 4 (2016)
9. Rabbany K.R., Takaffoli, M., Zaiane, O.R.: Social network analysis and mining to support the
assessment of on-line student participation. ACM SIGKDD Explor. Newslett. 13(2), 20–29
(2012)
10. Romero, M., Barbera, E.: Quality of learners’ time and learning performance beyond quanti-
tative time-on-task. Int. Rev. Res. Open Distrib. Learn. 12(5), 125–137 (2011)
11. Stahi, G., Koschmann, T., Suthers, D.D.: Computer-supported collaborative learning. The Cam-
bridge Handbook of the Learning Science, pp. 409–425 (2006)
12. Thomson, A.M., Perry, J.L.: Collaboration processes: inside the black box. Public Admin. Rev.
66, 20–32 (2006)
13. Ventura, S., Luna, J.M.: Supervised Descriptive Pattern Mining. Springer (2018)
14. Walckiers, M., De Praetere, T.: L’apprentissage collaboratif en ligne, huit avantages qui en font
un must. Distances et savoirs 2(1), 53–75 (2004)
15. Zamani, M.: Cooperative learning: homogeneous and heterogeneous grouping of iranian efl
learners in a writing context. Cogent Educ. 3(1), 1149,959 (2016)
16. Zhang, X.: An analysis of online students’ behaviors on course sites and the effect on learning
performance: a case study of four LIS online classes. J. Educ. Library Inf. Sci. 57(4), 255–270
(2016)
17. Zheng, Y., Subramaniyan, A.: Personality-aware collaborative learning: Models and explana-
tions. In: International Conference on Advanced Information Networking and Applications,
pp. 631–642. Springer (2019)
E-learning and the New Pedagogical
Practices of Moroccan Teachers
Nadia El Ouesdadi and Sara Rochdi
Abstract The strategic vision 2015–2030 is the most recent school reform in
Morocco. It promotes the use of information and communication technologies in
the professionalization of teaching in all schools and in all disciplines. Teachers are
therefore called upon to develop professionally throughout their careers by innovating
their teaching practices. This article aims to answer the question: How could an online
device meet the needs of teachers in terms of ICT use in the classroom? As well as
to define distance learning education, its characteristics, namely the added value of
using ICT in the classroom to enhance learning and to stay in touch with the teacher-
learner through the Collab platform for teachers of all disciplines in Morocco. The
experimentation we have conducted, presents to teachers-learners an online device,
involving the consultation of content and optimization of interactive activities to be
developed through software. This is one of the innovative models developed with
this in mind to reveal the benefits on the implementation of a learning device and
apply it in the classroom.
1 Introduction
During the last few years, information and communication technologies (ICT) have
experienced a remarkable boom in all areas including that of teaching and learning.
The integration of ICT in education is one of the main concerns of the Ministry of
National Education in Morocco. So that they can improve the quality of teaching
and learning. The strategic vision 2015–2030 is the most recent reform in Morocco’s
education system. It promotes the use of technology. Cited in Lever 12 [1]. Develop-
ment of an open, diversified, efficient, and innovative pedagogical model: “Strengthen
the integration of educational technologies to improve the quality of learning, through
the implementation of a new national strategy, able to accompany and support inno-
vations likely to promote the development of institutions.” This will be achieved
N. El Ouesdadi · S. Rochdi (B)

CLHEN, Linguistics, Didactics and Communication, EST, University Mohammed First, Oujda,
Morocco
https://doi.org/10.1007/978-981-16-3637-0_35
496 N. El Ouesdadi and S. Rochdi
“by integrating digital media and interactive tools in teaching and learning activities,
research, and innovation.”
The development, new specific pedagogy in teaching/learning has appeared thanks
to technology and the explosion of the Internet, a new mode of distance learning has
emerged whether it is school, university or professional. Online education allows a
multitude of new possibilities such as learning, exchange, collaboration, etc.
In other words, it opens new avenues for learning in general and it corresponds
to the use of computer technologies that help learners to improve their performance
and knowledge through the exchange of necessary information which allows for total
autonomy of users. Some authors have proposed a typology of the pedagogical uses
of ICT in five categories, namely ICT to exchange, communicate, collaborate, and
cooperate, ICT to produce, create and publish, ICT to research and document, ICT
to train and self-train, ICT to animate and organize [2].
Teachers are invited to innovate and evolve their pedagogical concepts and prac-
tices. Indeed, “ICTs are powerful cognitive tools. But, if they offer multiple solutions
to many of today’s educational problems, they will only be truly useful if the trainer
agrees to transform, or even change his or her conceptions and practices” [3]. It is
therefore useful for teachers to be receptive and trained in the proper use of techno-
logical tools in classroom practices. They should also be trained in scriptwriting and
the creation of pedagogical content so that they can design content that meets their
pedagogical objectives and the needs of their learners. This will allow interactivity
and accessibility for the learner and the effective use of educational technologies.
Mangenot [4] asserts that we can only speak of an integration of ICTE when
“the computer tool is used effectively to support learning”. For all these reasons,
the Ministry of National Education has programmed a series of face-to-face training
sessions for teachers in information and communication technologies. And other
online training courses are set up for teachers to familiarize them with the use of
distance learning platforms, the use of the digital environment, the pedagogical
scripting, and to introduce them to the mediatization of online courses.
It is in this context that this article is written, with the aim of answering our
fundamental question: How could an online device meet the needs of teachers in
terms of using ICT in the classroom?
To answer our question, we have made available a set of questions:
1. What is distance education?
2. What is collaborative learning?
3. How can we learn online?
4. How can pedagogical scenarios be developed and used in the classroom?
5. What is the role of a tutor?
Based on a pre-investigation on the platform in which we have been beneficiaries
and subsequently tutors, we have drawn up the following hypotheses:
• An online device could help teachers to design their courses online and then apply
them in the classroom through educational software.
• An online device promotes collaboration and exchange between learners.
E-learning and the New Pedagogical Practices of Moroccan Teachers 497
In our research, we will initially define the key concepts of our research, namely
distance learning, educational technologies according to the strategic vision 2015–
2030, collaboration in e-learning, etc. Then we will discover the universe of the
Collab platform, its aims, its content, etc. Then we will discuss the results of our
experience with the Collab platform, and we will conclude with the contributions of
this eminent experience.
2 Methodology
The objective of this research is to identify the use of ICT use in the professional
process of primary and secondary school teachers in Morocco in order to answer our
main question: How could an online device meet the needs of teachers in terms of
ICT use in a classroom?
In order to do so, we opted for a quantitative approach to collect the data that we
will present and which are the result of a survey and research in the Collab platform
in the form of description, tutoring, and evaluation of the homework given by the
beneficiaries in the second part of the training which contains the development of
interactive exercises with the Rubis software. This method was chosen because it
will allow us to answer our research question.
3 Theoretical Framework
As mentioned above, the world of education today is witnessing transformations that

affect teaching and learning methods and modes with the emergence of the:
Educational technologies. These are defined by the CSEFRS [5] as follows:
Educational technologies are information and communication technologies
adapted to education. They intervene in learning, training and supervision, in
planning, organization and management, and in evaluation.
Educational technologies include interactive programs and software, digital
resources, various electronic tools and equipment, as well as networks and informa-
tion systems and the services they offer such as distance learning, video conferencing,
digital libraries, etc.
The integration of educational technologies aims:
• To raise the quality of education and training by facilitating the acquisition of
knowledge, increasing learner motivation, and enhancing the attractiveness of the
school;
• To qualify the learner to access the knowledge society, to master the strategies of
distance learning, and to build personal projects in research and innovation;
• To rationalize educational governance by relying on integrated information
systems;
• To create networks for exchange, sharing, and development of collective intelli-

gence and support for teamwork in education;
• To involve and help educational stakeholders in isolated areas.
Distance Learning
The notion of distance in training is opposed to presence, in geographical, psychic,
and relational terms in which the need for contact and trust comes into play [6];
thus, “distance training is defined as opposed to face-to-face training, by breaking
the spatial co-presence between the teacher and his learners” [7]. It “brings together
students, teachers, one or more knowledge objects, and technical supports such as
Moodle-type platforms, the Internet, electronic files, not forgetting traditional printed
courses” ([6]: 397).
Among the practices in use, Papi et al. [8] list four intentions of the devices:
– Reduce distance: counter geographical distance (virtual classroom);
– Enrich the experience: diversify learning experiences (wiki, webinar);
– Support interaction: connect to counter isolation (videoconferencing);
– Develop skills: achieve learning objectives through the uses of the social web
(forum).
Virtual learning spaces and third places

A distance course contains several virtual learning spaces: information, collaboration,
production, assistance, self-management, and evaluation spaces [9]. Some spaces
may be formal (the institution’s digital learning environment) or informal (social
networks, personal learning spaces). In organizing virtual spaces, one should not
ignore what happens outside the classroom; one can even seek to enhance these
informal spaces [10].
Distance learning is supervised and tutored. It is in fact a framework that seeks to
guide the learners in a device and offers them flexibility in teaching. Indeed, tutoring
is based on a follow-up of the learners to accompany and guide them with the help
of a tutor.
Tutoring
Distance tutoring is a pedagogical accompaniment in a device. According to Bourdet
[11], tutoring is seen as a “formative function which is characterized by the exercise
of varied and sometimes contradictory roles (joint empathy/validation of stages,
general guidance/adaptation to specificities).”
Today we talk about the active approach in the teaching–learning process. This
insists on the participation of the learner in the construction of his learning. The
teacher is no longer the person who holds the knowledge; he has become a mediator,
a coach, and a facilitator of learning.
The coaching scenario is developed in relation to the learning scenario and takes
into account several aspects [12]:
– Tutorial objective;
– Actors involved;
– Types of tutoring: individual, collective, peer tutoring, etc.;
– Learning support plans: cognitive, socio-affective, motivational, and metacogni-
tive;
– Tutoring methods: synchronous, asynchronous, proactive, and reactive;
– Frequency and positioning in the learning scenario;
– Tutor support and resources to be produced;—Reusability of resources and
approaches.
The role of the tutor

The role of the tutor is the person who “accompanies a learner or a group of learners
at a distance by the means of communication and training that computer, multi-
media, and the Internet allow today” Lisowski [13]. It has several functions within
a computer-mediated device. It has a social role because it ensures the organization
of the learners’ work and the constitution of groups to interact and exchange.
The tutor is the one who constitutes the groups, distributes the work tasks for
the learners. He is, according to Legendre, “a guide, an instructor who teaches one
person or a small group of students at a time: he is a student advisor” [14]. In addition
to this, the tutor animates the work of the group and encourages the learners to make
a collective production and work toward the same goal while working in a digital
environment and in professional learning conditions.
Collaborative distance learning

Thanks to the digital evolution in recent years, the avenues of virtual collaboration
have developed considerably. This new modality has undergone important theo-
rizations, such as the definitions of cyberculture [15] and the new learning theory,
connectivity [16] (Downes, 2012, May 21).
Henri and Lundgren-Cayrol [17] proposed a definition of collaborative learning:
“Collaborative learning is an active process in which the learner works on building
knowledge…The learner commits to working with group members toward a common
goal while reconciling personal interests and objectives…Exchanges with the group
and the regulation of a collective task allow the learner to share his or her discoveries,
negotiate the meaning to be given to his or her work, and validate newly constructed
knowledge".
Among the tools for collaborative work in an online device, we have.
The Wiki
The Wiki is a collaborative distance working tool; it allows learners to work together
on the same document, and they can edit or comment on it.
Its advantages:
– The teacher will be able to stimulate learners, correct, and orient their work.
– Allows learners to work together on course topics.
– Learners and teachers are notified of any new modifications or notifications.
The forum
The forum is also a tool for remote collaborative work. It is composed of a set of
online discussions on different topics, and it allows to generate debates in the form
of questions and answers between learners, or between learners and teachers.
Its advantages:
– Allows the teacher to structure and direct the exchange.
– Promotes communication between learners.
– Discusses course topics.
– Motivates learners.
Thanks to these tools that are easy to use, the shared knowledge allows the
construction; the collective co-construction of knowledge facilitates the interac-
tion between the teacher and the learners. We move from individual learning to
collaborative learning.
We are going to discover the environment of the Collab system.
4 Presentation of the Collab [18]
COLLAB is an e-learning device designed and produced by the National Center for
Pedagogical Innovations and Experimentation—Distance Learning Division—of the
MENFP. This platform offers e-learning on the use of free and open-source software
(Chaînes éditoriales scenari) to the entire Moroccan educational community (see
Fig. 1).
Its objectives:
– To help beneficiaries navigate the platform to take advantage of its training content.
– Navigate the platform.
Fig. 1 Collab platform web page

– Access the different courses.

– Locate course components: Blocks, Resource Activities…
– Use the activity modules presented in the course.
– Interact with tutors and other beneficiaries.
The trainings.
The Collab platform offers three types of training courses:
– Development of educational multimedia content
– Distance learning in team sports
– Tutor training.
The first training consists of two modules (see Fig. 2). In our research, we will
focus on the second module “Elaboration of interactive exercises.”
First module: development of digital resources.
SCENARI In this section, the learner provides an overview of the course environ-
ment. By clicking on the development of educational multimedia content (DCME)
course, learners can access the first module of the course, which is the SCENARI
software.
Beneficiaries are called to install the Opale Software: Opale advanced is a
complete version of Opale intended for the personnel of the ICTE units, Web
professionals, and people already autonomous on Opale starter (see Fig. 3).
When learners have completed the first module, they will find a link to the second
module.
Module 2: Developing Interactive Exercises.
RUBIS which is an editorial chain for the creation of exercisers and which allows
to create exercise activities in two modes: self-learning and evaluative. The RUBIS
course is subdivided into three weeks; each week contains courses to view or files to
download and an activity to complete (see Fig. 4).
Fig. 2 Development of educational multimedia content

Fig. 3 Opale advanced software
Fig. 4 Ruby software

Fig. 5 Homework assignment
The Forum
Each course has its own forum. The course forum is usually located at the beginning of
each course. The forums in this course are considered complementary collaborative
learning tools.
The DCME course is supervised under the heading “Homework,” with learners
receiving support and follow-up from online trainers called “tutors.” The trainers
schedule activities in the form of assignments at the end of each week of training
(see Fig. 5). These assignments are limited in time. This is indicated in the header of
each week, and the tutors indicate the final deadline for completing the assignment.
Each week of training contains one activity to be submitted by the end of that
week. By clicking on “Activity1” for example; a window will open showing the place
to click to hand in the assignment and by clicking on add (in the same way as posting
a topic on the forum).
5 Results of the Experiment
We monitored and tutored the learners and evaluated their performance through
the homework assignments handed in each weekend. The number of participants
registered in the training is 399.
– Homework assignment 1 Number of participants who handed in the assignment:

142 participants
Homework assignment 2 Number of participants who handed in the assignment:

178 participants
– Homework assignment #3 Number of participants who handed in the assignment:

177 participants
The submission of the project

– Number of participants who submitted the final project is: 178 participants
The discussion forum

We observed learners’ interactions in the forum and found that learners interact with
other beneficiaries and with the tutors on each week’s courses.
Example of a submitted job

6 Commentary
Conducting this experiment has shown us that the place of information and commu-
nication technologies has a great place in teacher training. The Ministry of National
Education in Morocco relies heavily on pedagogical innovation; but according to
our experimentation, this remains insufficient because we believe that the actors of
the educational system still show resistance to the use of technological tools. The
results above show a reticence and abandonment of the beneficiaries (399 registered
and 178 learners who were able to follow and submit their final project). Regarding
the interactions in the forum, we observed that the discussion space is a beneficial
and effective way to exchange, share, and collaborate between learners and learners
with their tutors.
The projects delivered at the end of the training show the involvement of the
teachers who benefit from distance learning and their desire to use ICT tools in their
classroom practices in order to innovate. The software proposed in the training (Opal
and Ruby) is tools that allow the designing teachers to realize a digital pedagogical
scenario as well as to evaluate the learners’ achievements with various types of
exercises. This has been approved through the feedback from the beneficiary teachers.
7 Conclusion
Today, digital technology has become an essential tool for all teacher training. This
is why the COLLAB e-learning system offers tools and learning methods free of
charge, so as to innovate, break with the traditional method and arouse the curiosity
of teachers. It is a method that enables a better understanding of how technology
can have a real impact on teaching and learning. In this way, it helps the teacher
understand that integrating ICT has a positive impact on student learning. This allows
us to confirm our hypotheses put forward above as the Collab platform has allowed
the beneficiary teachers to design their courses online, and then they can easily apply
them in the classroom through the educational software offered as well as the online
device has fostered collaboration and exchange between learners.
We believe then that the integration of technology in teaching and learning requires
a considerable effort on the part of all components of the educational system to
promote a better integration of the technological tool within the Moroccan school.
References
1. Levier 12, la vision stratégique 2015–2030

2. Basque, J., Karin, L.-C.: A typology of ICT applications in education. Sciences et techniquesé-
ducatives 9(3–4), 263–298 (2002)
3. Karsenti, Peraya, Viens: Review and prospects for research on teacher training for the pedagog-
ical integration of ICTs. Revue des sciences de l’éducation XVIII(2), 459.470 (2002). www.
erudit.org/revue/rse/2002/v28/n2/007363ar.pdf
4. Mangenot, F.: Classification des apports d’Internet à l’apprentissage des langues. p. 145sur:
www.alsic.univ-fcomte.fr (1998)
5. Conseil supérieur de l’éducation, de la formation et de la recherche scientifique, Pour une école
de l’équité de la qualité et de la promotion; la vision stratégique 2015–2030 p. 95
6. De la Berchoud, M.J.: distance - prendre en compte des publics loinignés et décentrer la
réflexion méthodologique: Sociodidactique ou sémiodidactique ? Ela. Etudes de linguistique
appliquée 4(168), 395–405 (2012)
7. Peraya, D.: Distances, absence, proximities and presences: concepts in motion. Distances
Knowl. Mediation [Online], 8 (2014)
8. Papi, C., Brassard, C., Bédard, J.L., Medoza, G.A., Sarpentier, C.: L’interaction en formation à
distance: entre théories et pratiques. TransFormations – Recherches en éducation et formation
des adultes (17) (2017)
9. Paquette, G.: L’ingénierie pédagogique: Pour construire l’apprentissage en réseaux. Sainte-Foy,
QC, PUQ (2002)
10. Stockless, A, Villeneuve, S.: Les compétences numériques chez les enseignants: doiton devenir
un expert? Dans Romero, M. et al (dir.), Usages créatifs du numérique pour l’apprentissage au
XXIe siècle, pp. 141–148. Québec, QC, PUQ (2017)
11. Bourdet, J.-F.: Tutorat en ligne et création d’un espace formatif. Alsic [Online] 1002(30), 23–32
(2007)
12. Rodet, J.: Propositions pour l’ingénierie tutorale. Tutorales, Revue de la communauté de
pratiques des tuteurs à distance 7, 6–28 (2010)
13. Lisowski, M.: L’e-tutorat. http://www.centreinffo.fr/IMG/pdf/AFP2204357.pdf (2010)
14. Legendre, R Dictionnaire actuel de l’éducation. p. 1378 (2005)
15. Lévy, P.: L’intelligence collective. Pour une anthropologie du cyberespace. La Découverte,
Paris (1994).
16. Siemens, G.: Connectivism. A Learning Theory for the Digital Age. Retrieved from http://
www.elearnspace.org/Articles/connectivism.htm (2004)
17. Henri, F., Lundgren-Cayrol, K. Collaborative distance learning: understanding and designing
virtual learning environments. Sainte-Foy: Presses universitaires du Québec., p. 42 (2003)
18. http://collab.men.gov.ma/pluginfile.php/12817/mod_resource/content/8/guide_papier.pdf
A Sentiment Analysis Based Approach
to Fight MOOCs’ Drop Out
Soukaina Sraidi , El Miloud Smaili , Salma Azzouzi ,

Abstract Over the past few decades, a new model of education has emerged in the
world of education under the acronym MOOC (Massive Open Online Courses). This
model has made it possible to create transitional educational solutions by attracting
large populations and ensuring a worldwide online presence. However, there is still
a dark and fuzzy interface regarding the high dropout rate of learners at the end of
the courses. This issue leads us to ask: How can we make online mega-courses more
reliable and attractive? In this article, we explore new tools and methods (quality and
machine learning) to analyze the different limitations and difficulties learners face
in MOOCs and then build a deeper and better understanding of the main causes of
the high dropout rate.
1 Introduction
The rapid rhythm of educational development as well as the emergence of new

learning styles have led human innovation to develop new tools and methods to
exchange knowledge. The aim is to provide an interactive and productive space for
learners in order to make the learning process more flexible and accessible to all. In
this context, a new style of distance learning cited as global, free, large-scale and
open—known as MOOC (Massive Open Online Courses)—has emerged in recent
years, taking advantage of new technologies.
Although the concept is attracting more and more learners thanks to the new flex-
ible features it offers to learners. However, there is still a dark and fuzzy interface
related to the dropout rate at the end of such courses. On the other hand, the prolifera-
tion of different means of communication and the advent of social networks (FACE-
BOOK, Twitter; ….) as well as the forums included in the MOOCs has increasingly
S. Sraidi (B) · E. M. Smaili · S. Azzouzi · M. E. H. Charaf

Faculty of Sciences, Ibn Tofail University, Kenitra, Morocco
e-mail: sraidi.soukaina@uit.ac.ma
E. M. Smaili
e-mail: smaili.miloud@uit.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_36
510 S. Sraidi et al.
facilitated the social interaction of different users who can express themselves, ask
questions or even share their experiences. Therefore, the huge data obtained through
the exchanges of learners on MOOCs’ forums and social networks (Facebook, or
twitter for example) present important information that can help to identify the main
causes of MOOCs dropping out by analyzing the difficulties the learners’ encounter
as well as their motivation.
To this end, we suggest in this paper to gather data from MOOC forums and asso-
ciated social media groups. Then, we use sentiment analysis to extract and analyze
useful data such as participants’ feedback, learning methodology, programs, and so
on. More precisely, a combination of a good number of quality tools (Pareto, cause
effects, etc.) and machine learning approaches (Naives bayes, k-medoid) can be used
on the basis of the comments of the forums, to determine the learners’ motivation
rate but also to identify and group the causes of failures. This information will help to
understand learners’ needs and the relationship between a problem and all possible
root causes. The idea is to anticipate the learners’ performance and then make the
appropriate changes to ensure continuous improvement.
The document is organized as follows: The second section describes some basic
concepts. Then, we describe materials and methods to implement our model in Sect. 3,
Afterwards, we introduce the problematic statement and some related works respec-
tively in Sects. 4 and 5. Then, we explain in Sects. 6 and 7 our MOOC analysis model
as well as our quality approach. In the last section, we give some conclusions and
identify future works.
2 Preliminaries
2.1 Massive Open Online Course (MOOC)
Recently, there has been a new style of distance learning, known as MOOC (Massive
Open Online Courses), emerging in the world of education [1]. It provides a space
for collaborative sharing of techno-pedagogical practices to promote both the indi-
vidual’s involvement as well as the common construction of knowledge [1]. The “
massive “ character of the MOOC refers to the large number of learners who partic-
ipate and enroll in these programs. These massive courses are made up of different
resources and activities that allow both communicating knowledge and learning as
well as monitoring and supervising learners during their training paths.
A Sentiment Analysis Based Approach to Fight MOOCs’ Drop Out 511
2.2 Quality Approach
In a complex world, competitive and especially exigent, the quality remains a way
to distinguish competitive organizations. Indeed, the quality is defined as the set of
activities that guide and control an organization in terms of quality. The principle is:
• Organizations understand present and future needs.
• The leaders establish the purpose and the orientations of the organization.
• People at all levels are the essence of an organization and full involvement on
their part allows their abilities to be used for the benefit of the organization.
• A desired result is achieved more efficiently when resources and related activities
are managed as a process
• Identify, understand and manage interrelated processes as a system contributes to
the effectiveness and efficiency of the organization to achieve its objectives.
• Continuous improvement of the overall performance of an organization should
be a permanent objective of the organization and effective decisions are based on
the analysis of data and information.
• An organization and its suppliers are interdependent and mutually beneficial
relationships increase the capacity of both organizations to create value.
3.1 Naïve Bayesienne Algorithm
Naive Bayesian is a very simple and powerful algorithm for supervised machine
learning, and it is one of the most popular classification algorithms, it allows to
predict the probability of an event based on the conditions we know for the events in
question.
The algorithm is easy to understand and guarantees good performance. It’s also
fast and easy to train even with a small data set.
3.2 K Medoids Algorithm
Description. k-medoids is a machine learning clustering algorithm that identifies

groups of observations with similar characteristics, as well as individuals in different
groups that stand out as much as possible.
Principe. The unsupervised k-medoids algorithm starts with a set of medoids, then
iteratively replace one by another if it helps to reduce the overall distance. It is more
efficient for small data. K-medoid is a robust alternative to k-means clustering (Fig. 1).
Fig. 1 k-medoid algorithm
3.3 Pareto
Description. The Pareto chart is a tool for showing graphically the problems affecting
a given situation, listed in descending order. It is used to prioritize issues based on
their frequency of occurrence (number of impressions). Pareto is useful to identify
the reasons on which to significantly improve the situation. This will avoid wasting
energy on things that have little impact.
Principe. In general, the method aims to sort any aggregate into two parts: the
vital problems and the more secondary problems. This tool highlights the 80/20 rule.
In other words, acting on 20% of the causes makes it possible to treat 80% of the
effects. Below is a representation that illustrates this principle (Fig. 2).
4 Problematic Statement
Today, the MOOCs offer a space for exchange that allows a large number of learners
around the world to easily access courses. They represent one of the best-known
means of transmitting and disseminating knowledge. However, such motivation
should be linked to the quality and performance of the training.
In fact, even of the MOOCs’ popularity, the platforms they proposed don’t face
many challenges such as the high dropout rate. Our objective is to suggest a quality
approach to assess the quality and performance of a MOOC. Therefore, the goal is
to:
Fig. 2 Pareto diagram
• Collect and analyze the information provided by the learners in order to classify
them according to three types (motivated, demotivated and neutral)
• Extract and group problems linked to a MOOC at all levels that could affect
the high rate of dropout. (Platform defaults, E-Course Quality default, (student)
Instruction Support defaults, Evaluation quality defaults, Infrastructure quality
defaults)
• Identify all the causes that have a more or less direct influence on an observed
problem in order to prioritize the efforts to be made for the problem resolution
• Highlight the most important causes on the total number of effects and thus take
targeted measures to improve the course.
To this end, we propose tousesomemethods to extract, analyze, and control
different constraints encountered by each learner in a MOOC in order to enhance the
quality of courses.
5 Literature Review
Currently, there is little recent research devoted to the correction and prevention of
the problems described in the previous section. Indeed, the authors in [2] present a
visual analysis system for the purpose of exploring anomalous learning patterns and
aggregating them into the data. The system integrates an anomaly detection algorithm
that allows the interactive detection of anomalies between and within groups on the
basis of semantic and interpretable data summaries by group.
On the other hand, the authors in [3] propose a MOOC prediction method based
on the combination of the two tensors (global, local), to tackle the task of predic-
tion of abandonment. In [4], the authors contribute to a deeper understanding of
learner engagement in a MOOC by identifying three influential parameters, namely
visits, attempts and feedback, which are sufficiently independent to allow grouping
of students in a MOOC.
The study in [5] describes the methodology for analyzing behavior in MOOCs
using the k-means algorithm and the “elbow method”. In [6], the authors suggest
to classify the dropout factors into seven major themes: learning experience, inter-
activity, course design, technology, language, time and situation. The authors in [7]
examine the general characteristics of large-scale MOOC courses and quantify the
influences of these characteristics on student performance. Furthermore, the authors
propose in [8] an analysis of the characteristics of MOOCs users using the unsu-
pervised k-means machine learning algorithm according to three stages: the first,
the weight calculation method is designed to select the important characteristics
according to the weight, the second is to optimize the algorithm for the initial cluster
center, and the third is to determine the optimal number of clusters.
The article [9] provides an analysis of the expected dropout rate for MOOC
students. This analysis automatically extracts functionality from click data and filters
functionality using clustering tools and weighted MaxDiff to improve the accuracy
of prediction. Moreover, the authors in [10] present an analysis with focus on three
dimensions of learner behavior: Course Activity Profiles; Test activity profiles and
the most relevant forum peers or best friends. The article [11] makes three major
contributions to the literature related to the design and evaluation of open online
courses: (1) an expanded assessment tool for MOOC teaching methods to be used
by learned designers and researchers in their own contexts, (2) an illustration of how
to use group analysis Close neighbors to identify educationally similar MOOCs,
and (3) a preliminary analysis of clusters to take into account characteristics and
factors contributing to educational similarity between massive, intra-clustered online
courses.
The work [12] presents a deeper and better understanding of the behavior of
MOOC actors by bringing together and analyzing the different objectives of these
actors. The main finding was a set of eight clusters, namely blended learning, flexi-
bility, high quality content, instructional design and learning methodologies, lifelong
learning, learning by network, openness and student-centered learning.
Another study [13] apply models of unsupervised students initially developed for
synchronous didactic dialogue to MOOC forums. They use a clustering approach
to group similar articles, compare clusters with manual annotations by MOOC
researchers. Besides, the authors in [14] review the factors that lead to a high number
of dropouts in order to predict, explain and solve the problem related to both students
and MOOCs. To this end, they suggest to use machine learning tools as well as arti-
ficial intelligence. Furthermore, the paper [15] suggests a CNN-LSTMATT method
based on MOOC dropout prediction time series. The aim is to improve student
course completion rate and to obtain a better prediction results using LSTM to extract
temporal characteristics and CNN to extract local abstract characteristics.
Another model for MOOC dropout prediction based on extracting learning behav-
iors features is introduced in [16]. Their work is based on click path data and opti-
mizing SVRmodel parameters using IQPSO. The approach presented in [17] gives
an analysis and interpretation of the dropout phenomenon. Their study is performed
on a dataset with four MOOC courses and 49,551 enrolled learners. They use also
feature selection methods and machine learning algorithms in order to ensure the
prediction and classification of learners in a MOOC.
Finally, we suggest in [18, 22] a new system to analyze the learner traces in the
forum posts in order to understand and predict the causes of failure in MOOC learning
environments. Moreover, we propose in [19] a system to analyze the performance
of a MOOC by determining the relation between the learners’ satisfaction rate and
the main drop out causes using the combination of both ISHIKAWA quality method
and a machine learning algorithms.
This paper can be considered as a continuity of previous works [18–21] where we
propose to increase the courses completion rate and fight against the dropping out.
The purpose is to make the MOOC adapted to each learner based on the knowledge
and preferences of each one.
6 MOOC Analysis Model
In this section, we describe our approach to analyze the learners’ engagement on

MOOCs. We aim to provide an appropriate system to identify the root causes of
dropouts. More specifically, we aim to analyze the comments left on social networks
concerning MOOCs in order to deal with the various dropout causes.
Therefore, we highlight below the basic concepts of our analysis model. As shown
in Fig. 3, the model is implemented in five steps:
6.1 Data Collection Approach
Data collection. We are moving from information societies to recommendation and

sharing societies and social networks offer a perfect space for sharing information.
In this context, we approach our subject by defining, collecting and storing MOOC
forum data (forum posts) from Facebook and Twitter groups in Coursera sessions.
Therefore, the dataset is obtained via application programming interfaces (APIs)
and also automatically via web scraping.
Preprocessing. As social media data is not standardized, we will need to prepro-
cess such collected information. In fact, we pre-process the textual data for further
analysis by replacing and removing all symbols with their emoticons (hashtags and
URLs, stop words, non-English words, punctuation marks, symbols and numbers).
Generally, the data obtained remains incomplete, noisy and inconsistent due to
the lack of attribute or errors and outliers for example.
Fig. 3 MOOC analysis model
6.2 Classification System
We aim through this step to detect the sentimental polarity (positive, negative, neutral)
of a given text using the supervised Naive Bayes algorithm. The name comes from
Bayes’ theorem, which can be written mathematically as follows:
P(B|A).P(A)
P(A|B) = (1)
P(B)
With:
• P(A) and P(B) are respectively the probability of events: A, B and
• P(B) is greater than 0.
The algorithm remains predictive despite the fact that the hypothesis of indepen-
dence of the explanatory variables conditioned by the classes is difficult to justify.
The extraction of opinion targets and the sentiment expressed towards these targets
will help us to extract the causes of success (positive ranking) and failure (negative
ranking) of the MOOC.
6.3 Clustering
The classification of the feelings expressed in the previous step will help us to
distribute the data concerning the demotivation of learners in clusters using the
unsupervised algorithm k-medoids with k = 5: (Platform defaults, E-Course Quality
default, (student) Instruction Support defaults, Evaluation quality defaults and Infras-
tructure quality defaults). In order to extract the causes of failure in the MOOC, the
algorithm works as follows:
• Choose k data points in the cloud as the starting points for the cluster centers.
• Calculate their distance from all points of the point cloud.
• Classify each point of the cluster where it is closest to the center.
• Select a new point in each cluster that minimizes the sum of the distances of all
the points in this cluster from itself.
• And at the end Repeat step 2 until the centers stop changing in a data set.
At this stage, we specify also the indicators needed to assess the obtained results.
After having the problemsgrouped, we move to the analysis of each problem using an
adequate quality tool. The aim is to identify the causes of a problem and to understand
the relationship between a problem and all the possible causes.
6.4 Causes Analysis
After collecting, classifying and grouping the causes into categories, we proceed
to the presentation of the most important causes of failure of a MOOC using the
PARETO quality control method.
The objective is to highlight the important causes regarding the total number
of effects which allow to take corrective and preventive actions and by the way to
improve the quality of the MOOC.
7 Proposed Quality Approach
7.1 Prototype
Figure 4 gives an overview of our prototype. It describes the proposed approach

starting with the data collection from the social networks until the detection of the
various causes of failure of the MOOCs.
The approach can serve as a system for detecting the root causes of failure but also
to establish corrective and preventive actions for help decision-makers to improve
the quality of MOOCs.
Fig. 4 MOOC quality analysis approach
7.2 Model Description
The aim of our approach is to extract the root causes of failures and provide an
overview of the most important causes in order to overcome the low completion rate
and improve the quality of MOOCs.
As shown in Fig. 4: All learners’ interactions (forum post) are defined, collected
and stored through application programming interfaces (APIs) or obtained automat-
ically through web scraping. Afterwards, the data is preprocessed and classified in
order to extract various characteristics and to determine polarities (motivate, demo-
tivate, or neutral) using the machine’s naive bayes algorithm. After classifying the
learners, we group the causes of success and failure into five groups ((Platform
defaults, E-Course Quality default, (student) Instruction Support defaults, Evaluation
quality defaults, Infrastructure quality defaults).
Finally, we calculate the degree of impact of each given phenomenon in relation
to other phenomena using the law of 20/80. The objective is to detect the 20% of the
causes that generates 80% of the consequences in order to enhance the performance
and the quality of the MOOCs.
8 Conclusion
As conclusion, the work presented in this article deals with the dropping out of
learners during their MOOC training. Therefore, the main objective is to analyze
the obstacles and problems encountered by learners during their online courses.
To this end, we follow in the footsteps of the learners during a course, then we
analyze their interactions on MOOCs forums as well as social media platforms using
machine learning algorithms. The basic idea is to classify learners’ data according to
their polarities (negative, positive or neutral) then grouping the difficulties encoun-
tered by the learners (with negative polarity) into five clusters (Platform defaults,
E-Course Quality default, Instruction Support defaults, Evaluation quality defaults
and Infrastructure quality defaults) which allow us to extract the causes of failure in
MOOCs.
Therefore, our approach aims to provide the necessary support to the deci-
sion makers to define the corrective and preventive actions to eliminate the causes
which generate a higher dropout rate. Finally, our approach is limited to the use
of data according to Latin language as our experience regarding the use of the
Arabic language presents many difficulties. As prospects, we plan to propose new
quality methods with other different strategies to evaluate the learning process within
MOOCs to further improve our approach.
References
1. Tahiri, J.S., et al.: MOOC…Un espace de travail collaboratif mature: Enjeux du taux de réussite.
La 2éme conférence francophone sur les systemes collaboratifs (SysCo’14). SEPTEMBER,
pp. 131–144 (2014)
2. Mu, X., et al.: MOOCad: visual analysis of anomalous learning activities in massive open
online courses. EuroVis (Short Papers). 2019, 91–95 (2019). https://doi.org/10.2312/evs.201
91176
3. Liao, J., et al.: Course drop-out prediction on MOOC platform via clustering and tensor comple-
tion. Tsinghua Sci. Technol. 24(4), 412–422 (2019). https://doi.org/10.26599/TST.2018.901
0110..
4. Shi, L., et al.: Revealing the hidden patterns: a comparative study on profiling subpopulations of
MOOC students. In: Proceedings of the 28th International Conference on Information Systems
Development: Information Systems Beyond 2020, ISD 2019 (2019)
5. Shi, L., et al.: Social interactions clustering MOOC students: an exploratory study. arXiv.
(2020).
6. Goopio, J., Cheung, C.: The MOOC dropout phenomenon and retention strategies. J. Teach.
Travel Tourism 00, 00, 1–21 (2020). https://doi.org/10.1080/15313220.2020.1809050.
7. Xing, W.: Exploring the influences of MOOC design features on student performance and
persistence. Distance Educ. 40(1), 98–113 (2019). https://doi.org/10.1080/01587919.2018.155
3560
8. Xiao, L.: Clustering research based on feature selection in the behavior analysis of MOOC
users. J. Inf. Hiding Multim. Signal Process. 10(1), 147–155 (2019)
9. Ai, D., et al.: A dropout prediction framework combined with ensemble feature selection. In:
ACM International Conference Proceeding Series (New York, NY, USA, Mar. 2020), 179–185
(2020)
10. Liu, Z., et al.: MOOC learner behaviors by country and culture; an exploratory analysis. In:
Proceedings of the 9th International Conference on Educational Data Mining, EDM 2016,
pp. 127–134 (2016)
11. Quintana, R.M., Tan, Y.: Characterizing MOOC pedagogies: exploring tools and methods for
learning designers and researchers. Online Learn. J. 23(4), 62–84 (2019). DOI:https://doi.org/
10.24059/olj.v23i4.2084.
12. Yousef, A.M.F., et al.: A cluster analysis of MOOC stakeholder perspectives. RUSC. Univ.
Knowl. Soc. J. 12(1), 74 (2015). https://doi.org/10.7238/rusc.v12i1.2253
13. Ezen-Can, A., et al.: Unsupervised modeling for understanding MOOC discussion forums: a
learning analytics approach. In: ACM International Conference Proceeding Series. 16–20-Mar
(2015), pp. 146–150. https://doi.org/10.1145/2723576.2723589.
14. Fisnik, D., et al.: MOOC Dropout Prediction Using Machine Learning Techniques: Review
and Research Challenges (2018). https://doi.org/10.1109/EDUCON.2018.8363340
15. Min. C., et al.: A dropout prediction method based on time series model in MOOCs. J.
Phys.: Conf. Ser. 1774(2021) 012065 IOP Publishing (2021). https://doi.org/10.1088/1742-
6596/1774/1/012065.
16. Cong, J.: (2020). MOOC student dropout prediction model based on learning behavior features
and parameter optimization. Interact. Learn. Environ. https://doi.org/10.1080/10494820.2020.
1802300
17. Youssef, M., et al.: (2019) A machine learning based approach to enhance MOOC users’
classification. Turk. Online J. Distance Educ.-TOJDE April 2020 ISSN 1302-6488
18. Soukaina, S., et al.: Quality approach to analyze the causes of failures in MOOC. Proceedings of
the 5th International Conference on Cloud Computing and Artificial Intelligence: Technologies
and Applications November 24–26, pp. 1–5 (2020). https://doi.org/10.1109/CloudTech49835.
2020.9365904
19. Soukaina, S., et al.: MOOCs performance analysis based on quality and machine learning
approaches. In: Proceedings of the 2nd IEEE International Conference on Electronics, Control
and Computer Science, 2–3 Dec 2020, Kenitra Morocco (2020). https://doi.org/10.1109/ICE
COCS50124.2020.9314606
20. Miloud, S., et al.: An adaptive learning approach for better retention of learners in MOOCs.
(2020). In: Proceedings of the 3rd International Conference on Networking, Information
Systems & Security (NISS2020), Article 26, pp. 1–5 (2020). https://doi.org/10.1145/3386723.
3387845
21. Smaili, E.M., et al.: An optimized method for adaptive learning based on PSO algorithm. In:
Proceedings of the 2nd IEEE International Conference on Electronics, Control and Computer
Science, 2–3 Dec 2020, Kenitra Morocco (2020). https://doi.org/10.1109/ICECOCS50124.
2020.9314617
22. Smaili E., et al.: (2021) Towards Sustainable e-Learning Systems Using an Adaptive Learning
Approach. In: Emerging Trends in ICT for Sustainable Devel Science, Technology & Innovation
(IEREK Interdisciplinary Series for Sustainable Development). Springer, Cham. https://doi.
org/10.1007/978-3-030-53440-0_38
The Personalization of Learners’
Educational Paths E-learning
Ilham Dhaiouir, Mostafa Ezziyyani, and Mohamed Khaldi
Abstract The personalization of learners’ educational paths in MOOC (Massive

Open Online Courses) is one of the major questions that help teachers in their func-
tions of teaching and monitoring of learners. Being intended for a huge and multiple
number of learners, MOOCs are problematic for teachers in monitoring and under-
standing each learner’s profile. It is for these reasons that we propose a semantic
web (SW) approach to build a recommendation system in a graphical interface that
we created using the NetBeans IDE. The goal is to enable learners who follow the
MOOC to directly head for modules that meet the characteristics and requirements
of their profiles. Moreover, the proposed approach is based on some essential tech-
nologies such as resource description framework (RDF), ontology, web ontology
language (OWL) and RDF query language (SPARQL). Thus, the main objective of
this paper is to present another beneficial technique aimed at improving distance
courses by allowing teachers and learners to work in better conditions. The primary
results showed that the performance of the (SW) based recommendation system is
effective.
1 Introduction
Online classes or E-learning represent a set of educational resources (courses, videos,

activities, projects, quizzes, etc.) that are accessible for everyone and free via Internet.
Indeed, it is the current most useful way of learning in universities and institutions
all over the world because of the current circumstances of the COVID-19 pandemic
we are experiencing so far.
In an online course, the learner is at the center of the learning process because he
must be autonomous in the organization of the learning process, as for the teacher who
I. Dhaiouir (B) · M. Khaldi

Laboratory of Mathematics and Application FST, Tangier, Morocco
M. Ezziyyani
Laboratory of S2IPU ENS, UAE, Tetouan, Morocco
https://doi.org/10.1007/978-981-16-3637-0_37
522 I. Dhaiouir et al.
provides supervision in a distance training, it plays a very important and primordial

role in the preparation of educational resources (courses, activities, exercises, videos)
and also in the simplification of interactions between them and the learners. But since
MOOCs are accessible to everyone and free, then this presents a real problem in terms
of the massive number of registered learners with different backgrounds, levels and
learning interests.
Indeed, according to the collected statistics during the ABC MOOC of Project
Management, we have noticed this variety [1] because each learner has different
expectations, basic comprehension, priorities or even ways of learning that varies
from one learner to another. But presently, we find that in most MOOCs, there is
only one path offered to learners, this latter does not necessarily meet the expectations
and needs of all the learner profiles. So, the apprentices who participate in the training
and pursue it to the end have a low rate of 10%. This issue is at the center of the current
research on MOOCs through the analysis of learner behavior [2–4] to improve this
retention rate [4–6]. So, as a solution to this problem, we need the personalization
of educational content to adapt to the characteristics of the learner.
The personalization of learning in a MOOC is a learning process that seeks to
improve the quality of online education and training. The latter has become very
important in the educational field because a good pedagogy requires a good design
of courses that adapt to the profiles of the learners.
The personalization operation obliges having an idea about the knowledge of
the learners [7] based on their experience, their level of knowledge, their format
and resource preferences [8] as well as the monitoring of learners in their learning
process to have any information that may be useful in defining the progress situation
of learners.
To guarantee the personalization of a MOOC, the teacher must present content
adjusted to the profile of each learner taking into account their characteristics and to
meet their needs as regards the learner, he must play an important role in the lead
and control of learning [9].
This is why our objective in this article is to create a graphical interface intended for
the use of learners who follow distance training by recommending them personalized
educational content appropriate to their profiles. First, we will propose the work
related to the question of the customization of courses in MOOCs. Secondly, in
our approach we will collect all the beneficial traces left by the learners during the
registration in the formation also during the passage of the diagnostic test. Thirdly,
on the basis of these traces we will determine the information on the basis of which
we will create ontologies thanks to the “Protégé” software. Fourth, on the basis of
these ontologies we create a recommendation system in a graphical interface that
we create using NetBeans IDE to be able to recommend personalized educational
content that is appropriate to the learning levels and preferences of each learner.
The Personalization of Learners’ Educational Paths E-learning 523
2 Related Work
Many approaches have been appointed by numerous author that allows personaliza-
tion either by focusing on the observation of the learner himself, or by recommending
or adapting the content to the profiles of the learners.
Nonetheless, most of the approaches have limits and, in most cases, they offer
learners internal resources for the followed MOOC that we find reasonably limited.
Because in case the MOOC does not manage to respond to a difficulty confronted by
the learner, then it is useful to offer them information on the web that can be useful
for their noted difficulty or to widen their knowledge.
The approach of [10], proposed a model named PERSUA2MOOC which allows
teachers and MOOC designers to personalize learners’ trails by offering them content
that fulfills their educational objectives, taking into consideration analyzing the traces
left by learners during their participation in the platform and carrying out activities.
As in the work of [11], which is formulated to guide and advise participants with
low knowledge and low know-how in the MOOC. My learning mentor intends to
increase the level of this section of learners by nudging them to be independent in
their learning and by empowering them with planning and cues to help them get
the most out of MOOCs by supporting self-learning. Regardless, this work is still
at an initial stage and needs to be implemented and evaluated with real MOOCs.
Moving on to another work, which is that of [12], who considers that the strategy of
personalization comprises in establishing a process of transformation of learning
scenarios, to resources for the semantic personalization of learning experiences
MOOC, via the plan of a method for estimating the relationships between abilities
based on a structured competency classification, whose purpose is to match the skills
of learners with those who are involved in other elements of the learning plan. This
comparison method has been executed within a professional platform called TELOS.
The example illustrated here is quite general, and thus, it is not that manageable to
implement.
The work of [13], has made it feasible to find the MOOCs which suit a learner, by
asking them what are their learning objectives, expressed throughout a taxonomy of
the studied field. This approach [13] is established on an internal study, by directly
asking questions to the learners to specify their necessities. But the difficulty that
occurs here is the massive number of learners, which makes it impossible to analyze
all the responses of the learners to find them MOOCs suitable, for their learning
objectives. Another work that was able to formulate the first adaptive MOOC platform
(MOOC) is that of [14]. The platform, offers a solid academic framework and a
personalized learning experience in a MOOC learning environment. [14] Expanded
the first MOOC in the field of computational molecular dynamics (CMD). In this
work, [14] we describe the design, improvement and deployment of this MOOC
which has managed to handle the heavy loads of the massive open online course and
the stress of the users. The convincingness of this approach [14] has not yet been
approved, because the data on learner behaviors obtained by the authors are under
study and analysis and have not been presented definitely. Another cited project is that
of [5], which recommended the examination of the requested activities to learners on

the MOOC platform, to detect the problems overlooked by learners which demotivate
them and which push them to with-draw the MOOC. This approach [5], consists of
estimating the basic deviation between the objectives set by the learners and the
current accomplished level, to offer corrective classes for learners, who have not
achieved their goals. This project had thresholds concerning the taking into account
of all the learners are given their massive number.
Our work is in the field of computer learning environments (EIAH) rather than
in the field of information and communication technologies applied to education
(NTICE) and pedagogical engineering which is defined in [15], as: “A method that is
supporting the analysis, design, control and planning of the dissemination of educa-
tion systems, combining the concepts, processes and precepts of educational design,
software engineering and cognitive engineering. The pedagogical design according
to [15] which is a form of engineering attempted to improve educational practices.”
The submitted protocol is an approach aimed at the creation of a graphical interface
facilitating the orientation of learners by recommending personalized content. On the
basis of a set of criteria: mark, level, sector, language, age … which must first meet
their needs and secondly facilitate teachers’ monitoring and supervision of learners.
Our interface also allows learners who need support the possibility of reuse of free
educational resources (OER) delivered on the web by other teachers and public insti-
tutions which, according to the report from the 2002 UNESCO forum, stand out as
follows: ‘The recommended definition of open educational resources is: The open
provision of educational resources, made possible by information technologies and
of communication, for consultation, use and adaptation of the user community for
non-commercial purposes’. It has a heterogeneous identity, which strives to help
learners in their learning, as well as to help teachers in the operation schooling. They
are privileged, which means no one can edit them, but they can be reused and shared.
A quality OER is very expensive in terms of preparation time and effort.
Thus, to regulate the diversity of OER, we will use the ontologies of the semantic
web and the LOM metadata standard to use the external educational resources of the
web when needed by the learner. Therefore, in our work, we will use the information
entered by the learners when registering for the training, then these learners will be
directed to take a diagnostic test to determine the prerequisites of each of them. On
the basis of these traces, we will create an ontology on “Protégé” which will allow
us subsequently by using the RDF / XML and OWL language in NetBeans IDE
to create a graphical interface which introduces a system of recommendation that
recommends to each learner, the modules followed in this training that fits with his
specialty, his level and his skills.
3.1 Research Framework and Context
The study involves a sample of 120 students with a bachelor’s degree and who have
graduated from the following three majors: Mathematical and Computer Sciences,
Physics and Biology.
This is a supplement concerning various modules appropriate to the three courses
mentioned above which was started in the 1st of December 2020. The content of
the modules was structured in six main sessions, corresponding to the six weeks of
lessons.
Each of these weeks was itself structured into short sub-sequences each
comprising a short video, a manuscript and a self-assessment quiz of the concepts
covered in the video and the manuscript.
In addition, each set of sequences was accompanied by an introduction that deter-
mines the objectives of the week, the necessary prerequisites that the learner must
have as well as the necessary time that must be devoted to each part.
The learners are led to read the manuscripts shared by the teacher, to see videos
which facilitate the comprehension of the lessons as well as to carry out individual
assignments in the form of exercises that are requested at the end of each week.
3.2 The Traces
The reuse of traces, within the framework of computerized human learning environ-
ments (EIAH), has emerged in a movement of increasing complexity of technologies
supporting ILE and their uses. The question of collecting traces, analyzing them and
using them is far from new. The problem is not only how to analyze the traces but
also how to really complete and exploit the traces resulting from the observations
and left by the learners in the platform to improve the learning of the learners.
The digital traces left by the students in the registration form were of a heteroge-
neous and massive character but useful for us, because thanks to them we can have
an impression on the profiles of the learners registered in the MOOC.
In our case, we will use all the traces found during the registration of learners in
the training by filling out the registration form. Once learners have completed their
registration they will automatically be taken for a diagnostic test, and on this basis
the teacher can get a general idea of each learner’s level, prerequisites, knowledge
and gaps in order to address them. Then, guide them to courses that meet their
expectations, preferences and in general their profile. These traces can be used to
create ontologies in the “Protégé” software (see Fig. 1).
According to the figure above, the processing of all the traces will be carried out
based on two types of data: The first type concerns the registration data retained
when filling out the registration form, the second type concerns the prerequisite data
retained during the diagnostic test. From this group of data, we will create ontologies
on the “Protégé” software in order to attribute to each student the modules and the
Registration data Complete the registration form Example

Processing
of all data
Prerequisite data Taking diagnostic test
Fig. 1 All the traces to be processed
course that are appropriate to their characteristics. The figure above shows all the
traces that seem useful to us in our study and which will help us subsequently to
create ontologies.
3.3 Semantic Web
The semantic web represents an evolution of the World Wide Web. This term of
(SW) involves specific techniques recommended by the World Wide Web Consortium
(W3C) that improves the current web and its use. These techniques provide a meaning
of the data without worrying about their representation [16] and both humans and
machines can understand it. Additionally, the objective of (SW) is to create a global
data base of connected data on the web [17]. Thus, the semantic web technologies are
ontology, resource description framework (RDF), web ontology language (OWL),
and SPARQL query language.
3.4 Web Ontology Language (OWL)
Web ontology language (OWL) allows users to represent and write their ontologies
in a specific domain, it is built on the top of RDF [18]. For creating these ontologies,
we can use a free open-source “Protégé” ontology editor, this platform is popular in
the semantic web field and it was developed in JAVA. With this editor, the user could
create and manipulate the ontologies in various representation formats.
The use of these traces allowed us to get an idea about each learner profile. We
will use the secure software for the creation of OWL2 (Web Language Ontologies)
which are defined by an IRI. Ontology is a central element of the semantic web
which seeks, on one hand, to rely on the modeling of web resources from concep-
tual representations of the concerned fields and on the other hand, enable to make
assignments [19]. An ontology is not only the identification and the classification of
theories but they are also elements which are attached to them and which one calls
here properties that can be evaluated. “Protégé” is a graphical ontology development
environment developed by Stanford SMI.
In “PROTÉGÉ” knowledge model, ontologies are made up of a skeleton of classes
that have attributes (slots), which can themselves have certain properties (facets) [20].
Fig. 2 Example of our ontology created by “Protégé”
The development of the list is done via the graphical interface without needing to
produce a formal language. In our case, we chose on three types of route (see Fig. 2).
In the figure above, we have created an ontology on “Protégé” for the three
specialties: Biology, Mathematical Computer Science and Physics.
For each specialty, we have proposed three modules. Learners are automatically
classified according to their specialty based on the information they entered, while
registering for the training.
3.5 Resource Description Framework (RDF)
Resource description framework (RDF) is a model to describe the resource on the

web and their metadata or semantics in order to interchange the data on the web.
The RDF consists of three concepts [21] such as: Resource: this term refers to a web
page, image, person, video, fragment page, which means anything that has a URI
can be considered as a resource. Description: for describing the resource in order
to be more understandable, also the description has attributes, features, and specific
relation between resources. Framework: for providing the language syntaxes of these
descriptions. The information on the web is structured in the form of the triple by
RDF [22] (see Fig. 3).
The RDF language which represents a formal framework for describing resources
according to a graph model, by expressing the relations between these resources in
the form of triples: Subject-Predicate-Object.
Predicate
Subject Object
Fig. 3 The RDF triples
The subject that represents the resource to be described is represented by the

ellipse.
The predicate means the type of property applicable to this resource is represented
by a rectangle.
The object that is the value of the property (the literal) is represented by an arrow.
3.6 RDFS Language (RDF Scheme)
RDFS adds to RDF the ability to define hierarchies of classes and properties whose
applicability and range of values can be constrained using the rdfs: domain and rdfs:
range properties. Each application domain can thus be associated with a scheme
identified by a particular prefix and corresponding to a URI.
The rdfs:Class: allows a resource to be declared as a class for other resources.
For example, we can define in RDFS the Biology class which describes one of
the streams of our training.
The subclassOf property is used to define class hierarchies for our example the
Biology sector is an existing course in our E-learning course.
<rdfs:Class rdf:ID="Biology">
<rdfs:subClassOf rdf:resource="#Elearning"/>
<rdfs:Class>
– RDFS clarifies the notion of property defined by RDF by allowing to give a type
or a class to the subject and to the object of triples. For this, RDFS defines:
– The rdfs property domain, which allows you to define the class of subjects
linked to a property.
– The rdfs property, which allows you to define the class or the data type of the
values of a property.
3.7 SPARQL Protocol and RDF Query Language (SPARQL)
SPARQL protocol and RDF query language are a query language used for querying
and updating the RDF documents [23]. The purpose of using the SPARQL query
language is that the RDF is based on the XML language, and for making the commu-
nication and exchanging the information, we need the SPARQL query. The query
PREFIX ns: <http://www.owl-ontologies.com/learning.owl#>

PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
SELECT ?elearning
WHERE{
?elearning ns:Module ?Module.
Fig. 4 Example of SPARQL query language
language must understand the syntax of RDF and also must understand the data
model and semantic vocabulary of RDF [24] (see Fig. 4).
Where, PREFIX ns: http://www.owl-ontologies.com/learning.owl# is the names-
pace of ontology, PREFIX rdf: http://www.w3.org/1999/02/22-rdf-syntax-ns# is the
namespace of RDF, rdf: type is the property, and ns: E-learning is the object. So,
when execute this query, will return all E-learning. Therefore, all the steps taken to
develop our recommendation system proposed, in our graphical interface based on
the semantic web approach, in order to recommend the appropriate modules for each
learner, during the training are illustrated in the flowchart below (see Fig. 5).
4 Result
In our work, we have been able to obtain satisfying results by creating a graphical
interface, incorporating a recommendation system that will facilitate the task of
monitoring and supervising the teachers who provide this training. Our web interface,
is based on the SW approach, developing HTML5, JavaScript, CSS and a technique
for dynamically create HTML code (JSP) “Java Server Pages,” were all used in order
to implement the web page, in addition the Bootstrap framework was used for the
design of our web interface.
And for the development of our recommendation system based on the semantic
web, eclipse IDE for Enterprise Java developers and the Apache Jena framework (for
creation of the semantic web application) were used for the purpose of implementing
all classes and methods by working with the ontology file created in the “Protégé
“software. This ontology is in RDF format and having semantic data.
To create our ontology, we based ourselves on traces left by learners when filling
out the registration form to participate in the training. These data consists of specific
parameters to recommend the modules that each learner must follow during the
training and which meet the characteristics of his learning profile. In Fig. 6, we find
the source code that allowed us to create our graphical interface in NetBeans IDE
(see Fig. 6).
Once our ontology had been created in the “Protégé” software, we will then
create a graphical interface by NetBeans IDE in order to recommend to the learners
the modules that they will follow during the training and which are suitable for their
profiles.
Fig. 5 Course
recommendation system Select a specific number of parcourse
based on the semantic web
approach design flowchart
This paper proposed
Modules parcourse
Knowledge base representation by
Ontology
Save it the Form
RDF/XML file adding semantic
Developing web interface of Recommen-

dation System based on SW Technologies.
Used by
Users (The Learners)
Submit
SPARQL Query Question (for
recommending a SPARQL query

question
SPARQL Query Answering on
Web Interface
To demonstrate our ontology based recommendation system approach we will

develop a computer application on Java. The software will be able to consult
and analyze our ontological model created in order to reveal the necessary data
to formulate the recommendation data to learners. For this, we use the following
elements.
The Sparkle Query language for the selection and the JenaOnthology package
to create the ontological model where we will load the data as well as a file for the
Fig. 6 The source code of our graphical interface created on NetBeans
rules. The data revealed by the recommendation will be interpreted in the form of a
text on a graphical interface see Fig. 7.
For example, according to the figure above, the 20-year-old student Karim
Zerhouni with a bachelor’s degree in MCS, was recommended to his three modules
in English to follow during the training:
– Module 1: Mathematical programming.
– Module 2: Commutative algebra.
– Module 3: Java Programming.
When we click on Submit SPARQL Query, the query uses all the parameters
inserted by the teacher (see Fig. 8).
When we click on submit response to the SPARQL request of the recommended
module according to these settings, that is to say: First Name = Karim, Last Name
= Zerhouni, Age = 20, Level = Bac + 3, Science Stream = MSC, Language =
English.
Then, according to this person’s settings, modules that meet the characteristics of
his profile will be recommended to him to follow during the online training.
5 Conclusion
In this article, we proposed a recommendation system based on the semantic web

approach in the field of E-learning to better help teachers to monitor and supervise
learners as well as to help them follow the MOOC until the end. In our graphical inter-
face, our recommendation system has the function of recommending the appropriate
Fig. 7 Web interface of recommendation system
PREFIX ns: <http://www.owl-ontologies.com/learning.owl#>

PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
SELECT ?elearning
WHERE{
?elearning ns:Age ?Age.
FILTER(?Age = 20)
?elearning ns:First_Name "Karim".
?elearning ns:Last_Name "Zerhouni".
?elearning ns:Level "bac+3".
?elearning ns:Language "English".
?elearning ns:Science_stream "MCS".
?elearning rdf:type ns:Elearning.
Fig. 8 SPARQL Query based on parameters provided by the teacher

modules for each learner, according to the specific criteria that have been dedicated
when registering learners in the training. Our objective was to help teachers as well
as learners by improving web research, improving efficiency and give the answer
quickly thanks to the technologies of the semantic web. Additionally, this system
facilitates the exchange of both human and mechanical information is why we can
say that this is the smart web. The results showed that the recommendation system
based on the SW approach is effective. Our perspective is to improve this system in
future work.
References
1. Cisel, M.: Qui étaient les participants du MOOC Gestion de Projet ? Blog La révolu-
tion MOOC. http://blog.educpros.fr/matthieu-cisel/2013/08/16/qui-etaient-lesparticipants-du-
mooc-gestion-de-projet
2. Baker, R., Evans, B., Dee, T.: Understanding persistence in MOOCs: descriptive & experimental
evidence. EMOOCs 2014, 5–10 (2014)
3. Willems, C., Renz, J., Staubiz, T., Meinel, C.: Reflections on enrollment numbers and success
rates at the openhpi MOOC platform. EMOOCs 2014, 101–106 (2014)
4. Halawa, S., Mitchell, J.: Dropout prediction in MOOCs using learner activity features.
EMOOCs 2014, 58–65 (2014)
5. Miranda, S., Mangioni, G., Orciuoli, F., Loia, V., Salerno, S.: The SIRET training platform:
facing the dropout phenomenon of MOOC environments. EMOOCs 2014, 107–113 (2014)
6. Liyanagunawardena, T.R., Parslow, P., Williams, S.A.: Dropout: MOOC participants’ perspec-
tive. EMOOCs 2014, 95–100 (2014)
7. Brusilovsky, P.: Adaptive and intelligent technologies for web-based education. http://www.
kuenstliche-Intelligenz.de/archive/ (2001)
8. Höök, K.: Steps to Take Before IUI Becomes Real. The Reality of Intelligent Interface
Technology, Edinburgh (1997)
9. Lebow, D.: Constructivist values for instructional systems design: five principles toward a new
midset. Educ. Tech. Res. Dev. 41(3), 3–16 (1993)
10. Lefevre, M., Guin, N., Jean-Daubias, S.: Personnaliser des activités pédagogiques de ma-nière
unifiée: une solution à la diversité des dispositifs. STICEF 19, (2012)
11. Gutiérrez-Rojas, I., Alario-Hoyos, C., Pérez-Sanagustin, M., Leony, D., Delgado-Kloos, C.:
Scaffolding self-learning in MOOC. EMOOCs 2014, 43–49 (2014)
12. Gilbert Paquette, O.M.: Competency-based personalization for massive online learning. Smart
Learn. Environ., pp. 1–19 (2015)
13. Gutiérrez-Rojas, I., Leony, D., Alario-Hoyos, C., Pérez-Sanagustin, M., Delgado-Kloos, C.:
Towards an outcome-based discovery and filtering of moocs using moocrank. EMOOCs 2014,
50–57 (2014)
14. Sonwalkar, N.: The first adaptive MOOC: a case study on pedagogy framework and scalable
cloud architecture—part I. MOOCs Forum, pp. 22–29 (2013)
15. Paquette, G.: L’ingénierie pédagogique: pour construire l’apprentissage en réseau. Presses de
l’Université du Québec, Québec (2002)
16. Robu, I., Robu, V., Thirion, B.: An introduction to the semantic web for health sciences
librarians. J. Med. Libr. Assoc. 94(2), 198–205 (2006)
17. Laufer, C.: Guia_Web_Semantica, p. 133 (2015)
18. McGuinness D.L., Van Harmelen, F.: OWL web ontology language overview. W3C Recomm.
2004, 10 (2004).
19. Charlet, J., Bachimont, B., Troncy, R.: Ontologies pour le Web sémantique, s. d., 31.2004.
20. https://protege.stanford.edu/index.shtml. Author, F.: Article title. J. 2(5), 99–110 (2016)

21. Gandon, F., Krummenacher, R., Han, S.-K., Toma, I.: The Resource Description. Framework
and its Schema. Handbook of Semantic Web Technologies (Issue January). http://www.spr
inger.com/us/book/9783540929123 (2011)
22. Cyganiak, R., Wood, D., Lanthaler, M.: RDF 1.1 concepts and abstract syntax. W3c Recomm.
2014; 25
23. Group WSW: SPARQL 1.1 Overview. W3C Recomm. W3C (2013)
24. Che,H.: A semantic web primer. J. Am. Soc. Inf. Sci. Technol. 57(8) (2006). https://doi.org/
10.1002/asi.20368
Formulating Quizzes Questions Using
Artificial Intelligent Techniques
Abdelali El Gourari, Mustapha Raoufi, and Mohammed Skouri
Abstract Formulating the question is one of the most important things for evidence-
based practice, regardless of the discipline in question. Formulating questions, and
even answering questions, is a powerful key to our profession, and our practice can
be seen through research in information science and by significant developments in
evidence-based practice. We should therefore develop a comprehensive classification
of the types of frequently asked questions and prioritize initial research. In this
paper, we will discuss how to create and make tests questions based on artificial
intelligence techniques and algorithms and make a simple comparison between the
most recent methods used to make questions, and finally, we will do an analytical
study at Cadi Ayyad University Marrakech Morocco by one of the ways based on
artificial intelligence (AI) algorithms.
1 Introduction
The process of making questions is to focus on using computer technology to create

questions from one or more sources of data. In other words, we are creating an
artificial intelligence-based system consisting of inputs and outputs that take existing
sources from data sources as inputs, as these sources can be organized by (datasets,
excel, tables, and texts…) or unorganized sources (such as texts: books, science
articles, email messages we send and receive…). In addition, other sources may
have expressed questions that are made by the teacher in order to obtain questions
similar to those he has entered, and when the system handling this data, it will provide
us with an output, which can be one type or divided into several types. So what do
A. El Gourari (B) · M. Raoufi · M. Skouri

Laboratoire de Didactique Et de Pédagogie Universitaire, Centre d’Etudes, d’Evaluation Et de
Recherches Pédagogiques, Université Cadi Ayyad Marrakech, Marrakech, Morocco
e-mail: abdelali.elgorari@uca.ac.ma
M. Raoufi
Laboratory of Materials Energy and Environment, Faculty of Sciences Semlalia, Cadi Ayyad
University, Marrakech, Morocco
https://doi.org/10.1007/978-981-16-3637-0_38
536 A. El Gourari et al.
Knowledge source Question bank The right

Question Template1 Q1 Q2 Q3 Question text
Q1 answer
generator Other Data
Template2 Q4 Q5 Q6 Metadata
Additional input Feedback
Fig. 1 Process based on the use of computer technologies (AI) to create questions
we mean by making the questions? The complex process is based primarily on the
use of computer technologies (AI) to create questions from one or more sources of
available data.
Through Fig. 1, we note that the makers of these questions are taking data sources
such as the reference book for the material, and other forms of data sources, which
can also be observed that the outputs are expressed from the back of these questions
[1]. If we take, for example, question one where it contains the text of the question,
the correct answer, and other data, such as the revised feed and the description data,
which determine how difficult the question or topic is.
2 Quizzes Questions Vocabulary Types and Standards

for Their Construction
Questions have an important role to play in assessing learning outcomes and assessing
the level of achievement of learning goals; the skill of question formulation is, there-
fore, one of the most important criteria to be included in the quality of the calendar,
and the test vocabulary is classified into two types [2].
2.1 Response Production Questions
This type requires the learner to write his answer to the problem posed to him
and was divided into (perfect questions—knowledge of terminology—pictures and
drawings—article questions (short—long)).
2.2 Response Selection Questions [3]
In this type of question, the learner is given several answers to the question or solutions
to the problem, and he has to choose (identify) the correct answer or solution among
them or the best ones. These questions are called objective questions because they
Formulating Quizzes Questions Using Artificial Intelligent … 537
are objective, that is, they do not differ from one individual to another. It is divided
into (choice of multiple—right and wrong—the mood—rearranging).
Here, we show each of these species, in terms of the educational output is measured
by each of these kinds of questions and the rules for preparing each type of question.
This is a frame of reference against which to measure the quality of the question in
achieving the educational output that it should measure.
Article Questions [3]

The vocabulary of the article is of the kind that allows the learner to answer it in
his own words. The learner asks to summon from memory the information relevant
to the question, and in this kind of question, he is required to organize and present
facts, terms, concepts, or ideas, that is, to engage in creative activity, and these
questions usually start with words like (discuss—explain—compare—write what
you know about—remember…). So, what characterizes a vocabulary is the freedom
of response it provides to the examiner, as this type of test presents a particular
question that requires an answer. The examiner is free to decide how to interpret
the problem and the information he uses, how to organize the answer, and how to
structure it. The questions of the article thus help to measure specific objectives, such
as the ability to innovate and organize and to integrate and express ideas, using the
specific scanned language, which other types of test singles cannot achieve to the
same degree.
Educational output for which the vocabulary of the article measures [4]
• The ability to express in writing, where the primary importance lies in this type
of capacity to produce, integrates, and expresses ideas.
• The ability to select, organize, and link information as the learner summons and
reorganizes the answer linking different elements of the decision and composes
it in the way he sees it.
• Makes the learner active in his choice of information on the problem asked by the
question and then organizes, links, and brings it out in an integrated subject.
• It is useful to verify higher mental processes because it requires conclusions,
comparisons, analyses, and judgments on knowledge of different types.
• If the article’s questions are well crafted, they lead learners to become accustomed
to good school habits that enable them to learn the important facts, to understand
the relationships between them, and to understand and absorb the subject.
Defects of article questions [3]

The formulation of the vocabulary may often lead students to differ in their under-
standing of the meaning of the article. This leads to a lack of compromise in arriving
at the required answer not because of their low level in the article, but because of
the lack of clarity in the question of the article. Some students may have a linguistic
skill in written expression, and in the way information is presented, linked, and orga-
nized. This affects the correction, giving a very high degree to a subject that may not
contain substantive ideas of value, in the sense that the framework within which the
answer is placed may affect the quality of the degree regardless of the integrity and
accuracy of the subjects. It takes a long time to correct the questions of the article
over and above the teacher’s stress because each student tries to write as many pages
as possible because he believes that quantity has a significant impact on the degree
to which it is obtained, even if it is unrelated to the substance of the subject.
Rules for preparing article questions

The question must be clear and specific so that the problem it poses is the same in
the minds of learners. This can be achieved by testing the exact terminology and
reviewing the question several times to make sure that it is clear. Also to measure the
output of education, the question must avoid words such as “who and what, when,
I remember and identify” and other words associated with remembering facts and
information. Other phrases that would measure the higher levels of goals, such as
“why, explain, compare, link, explain, analyze and criticize,” can be used.
Judging tests on article questions

• The question provides a great deal of freedom for the learner to express his views,
thinking, and conscience.
• The question presents the problem in a way that enables the learner to organize
and link information related to the problem to produce an integrated theme.
• The question achieves a focus on different mental processes, requiring conclu-
sions, comparisons, analyses, and judgments.
• The question encourages good school habits that enable students to understand
the important facts and the relationships between them.
• The question measures a learner’s creativity.
• The question is formulated in a clear and unambiguous manner.
Substantive questions [5]

Substantive questions mean questions that can be objectively valued, meaning that
there is an agreement in the provisions, if the substantive question is corrected, by a set
of corrections, where all arrive at the same scores without room for the intervention
of subjective judgments, and of the types of objective questions (complete—multiple
choice, right and wrong, identical).
Completion questions
Completion questions are words written by a teacher, from which one or more words
have been deleted, each deleted word has been placed one or several points, and the
learner is asked to place deleted words that make the meaning of the phrase complete
and clear. Merits of completion questions (ease of placement and correction—rela-
tively comprehensive for the scientific material to be tested—the field of speculation
is relatively weak).
Education output measured by completion questions

• Test of vocabulary and terminology.
• Test output on simple facts such as names, dates, events, places, and descriptions.
• Test output on principles.
• Test output on method and method knowledge.
• Test output for measuring simple interpretation of information.
Rules for preparing completion questions

• The missing phrase must be formulated in as brief a form as possible so that the
question is clear and unambiguous so that the answer is fully determined, and the
vacuum is filled only by the required answer.
• The phrase should not contain a large number of spaces because the presence of
many spaces leads to the ambiguity of the question and thus to the wide variety
of answers. Only keywords should be deleted from the phrase.
• As a general rule, the phrase should contain only one or two incomplete words so
that learners can understand and answer the question.
Questions of right and wrong

Questions of right and wrong are the type of language to which the learner responds
by one of the following means: (Right/wrong) (Yes/no) (True/wrong).
Merits of right and wrong questions

• Objective in its rectification and it takes no effort to correct it.
• Easy to develop, compared to other objective tests.
• It consumes a large area of the paper.
Educational output measured by right and wrong questions

• Measuring knowledge of simple facts and concepts and meaning of terminology.
• Measure the ability of the learner to detect common concepts that are incompatible
with scientific realities.
• Measure the ability of the learner to distinguish truth from opinion.
• Measuring the ability of learner’s to think critically. In this case, data on a particular
subject may be presented to the learner in the form of an experiment, map, graphs,
or tables.
• Measuring the ability of know the malaise relations. To measure this goal, we put
two issues into one phrase and the learner rules if the relationship between them
is true or wrong.
Rules for Building Right and Wrong Questions

• The phrase or question must contain only one idea, be key and important, and be
in a prominent place of the phrase.
• The phrase must be formulated tightly so that it is completely correct or false and
has no argument about its correctness or error.
• The phrases in the test must not be consistently longer than the wrong phrases,
but the correct phrases must be approximately equal in length.
Multiple selection questions [4]

In its simplest form, multiple-choice questions consist of a problem and several
alternative solutions. The problem arises either in an inquisitive form or in the form
of an imperfect phrase. The problem is called the origin of the question. Alternatives
include one correct answer, several false answers called distractions or camouflage,
and the function of distractions distracting learners who do not know the correct
answer.
The educational output measures

• Measure of remembrance: Such as the recollection of information, facts, termi-
nology, principles, procedures, methods, generalizations, theories, and other
first-level outputs of knowledge education objectives.
• Measuring understanding and assimilation: In this case, the position contained in
the question must be new to the learner, but if the question contains a position
identical to that experienced by the learners, the question will only measure the
level of recollection of information and facts.
• Measuring analytical capacity: Examples include the ability to infer function, the
implicit structure of a written text, the ideas used in a document, the ability to
analyze the elements of a situation new to the learner, or the use of certain criteria
to analyze a situation.
• Measurement of calendar capacity: Examples include the ability to judge a partic-
ular act or idea, to identify the values and views used in an act, to judge an action
by comparing it to another act, or to use a set of criteria for making a judgment.
Rules for Building Multiple-Choice Question

• The origin of the question must contain a well-defined problem so that the learner
understands the meaning of the problem without having to use alternatives to
clarify it. The main criterion for containing the origin of the question on an entire
problem is the ability to respond to the origin of the question without considering
alternatives.
• The problem must be carefully formulated: So that the learner does not have
to speculate on what he means by asking, so the question must contain all the
information needed for the solution, even if it seems self-evident.
• Negation formula must be avoided whenever possible, and a line must be placed
under the negation device so that the learner can pay attention to it and take it into
account when answering the question.
• Each alternative must be linguistically appropriate to the origin of the question.
If the first word of the correct alternative is an actor, all other alternatives must
start with an actor and so on. Some studies have indicated that respondents choose
long alternatives more than short ones.
• The connections between the origin of the question and the correct answer must
be avoided. A word in the correct answer is often marked because it is similar to
a word in the origin of the question. But it is okay to put such words in the wrong
answers.
• The use of this alternative has proved to be very easy if it is the correct alternative.
On the other hand, if it is wrong, it is sufficient to exclude it simply by suspecting
that one of the other alternatives is wrong.
3 The Required Background to Complete and Develop This

Field
We are dealing with sources of the content data organized and unorganized, so we
must have a background in a treatment of natural languages, texts, images, and for the
manufacturer of questions; we are in the process of using educational and evaluation
theories, so we must have a background in education and learning. We do not limit
ourselves to the direct application of these technologies while developing new ways
to benefit the educational and the evaluation communities, also, when we get products
we try to make as close as possible to the questions that people make. That means,
it must be free from the spelling mistakes, so we should have this background in the
treatment of natural language generation. Figure 2 illustrates in detail what we have
explained.
All of these areas that have been mentioned depend on machine learning, and
deep learning is employed very much. For this reason, the background required from
them must be present strongly. Also in terms of the answer-making must be objective,
credible, and must contain a goal of transparency and stability. But the question is
how to achieve credible and objective results? That is what we will talk about in
financial terms.
Image processing Education assessment Natural language generation
Knowledge source Question bank The right

Question Template1 Q1 Q2 Q3 Question text
Q1 answer
generator Other Data
Template2 Q4 Q5 Q6 Metadata
Additional input Feedback
Natural language processing

Knowledge engineering
Fig. 2 Sources of the content data organized and unorganized

Remember Understand Apply Analyze Evaluate Create
Fig. 3 Tools used to determine the cognitive level
4 Some of the Ingredients that Will Help Us Verify

the Credibility and Stability of the Question
There are many ingredients, but we have just focused on three ones because over the
past years, there has been researching interest in this area, we mean, most researchers
focus on them to make questions, which are as follows:
4.1 Cognitive Level
These are all the cognitive processes [6] that the answers to the question include,
for example, when a student answers the question, what are the processes that are
activated in his mind, is he trying to remind his or her courses? Or is trying to use
in-depth analysis, or what?
Figure 3 divides cognitive levels into six phases, beginning with a capacity to
remember the teachers that the student has studied and ended with the capacity to
invent and create new things.
4.2 Difficulties
There are many definitions of how to measure difficulty that has been addressed in
the literatures, but one of the most common definitions is that of the correct answers
to the question; the relationship (1) shows that [7]:
number of correct answers

Difficult question = × 100 (1)
total number of answers
There are several proposals for the acceptable range of question difficulties, which
are limited to 20 and 80%. This is, of course, according to the scenario in which the
question is used because these questions have many uses, they can be used, for
example, at the beginning of the classroom to know what information the student
knows and what information they do not know.
Example: We have a question for 60 students, among them 40 ones answered
correctly. Difficult question = 40/60 * 100 = 66.67%.
4.3 Discrimination
The question is how to distinguish between different-level tests; of course, there are
many equations to calculate the coefficient of differentiation [8], but there is a well-
known equation that depends on the order of students to step up by grade, where
artificial intelligence algorithms can be used, especially unsupervised learning algo-
rithms that classify students by each student’s level. We can calculate this discrimina-
tion rate (DR) with the following question (2), which (NCAT) is Number of Correct
Answers in the Top group and the (NCAL) is Number of Correct Answers in the
Lower group:
NCAT − NCAL NCAT − NCAL

DR = DR = (2)
individuals for one group individuals for one group
The acceptable range is as follows: The discriminating factor is positive, more

than 19.
Example: We have three questions for 100 students in the following distribution
of answers to the upper and lower groups (Table 1).
If we take the example of question three, 20% of students in the upper group
versus 20% of students in the lower group have answered the question, which makes
us unable to know and distinguish between high-level and low-level students.
5 Uses of Artificial Intelligence in the Industry of Questions
Artificial intelligence is used in many areas to solve some complex problems. Educa-
tion is one of the areas that it care about using artificial intelligence to solve its
problem. For example, it is used to guide and assist students [9]. Also, linking it
to the concepts of e-learning and distance learning to measure, evaluate and track
students during the educational process [10]. It can also be used to create questions.
These are techniques and algorithms of artificial intelligence to make high-quality
questions.
In Fig. 4, it can be seen that the use of artificial intelligence, especially deep
learning, in the creation of questions can give high accuracy and credibility to the
question, classifying good questions and false questions.
Table 1 Coefficient of
Question The upper The lower The coefficient
differentiation
number groups (%) groups (%) of differentiation
Question 1 14 4 +0.6
Question 2 4 14 −0.6
Question 3 20 20 0
Tribal Processing
Create a correct
(sentence Choose the question Choose the right answers to
question and answer
simplification, sentence topic remove and reprocess them
text
Classification)
Control the guess of the

quality of the
Feedback industry
processing (make sure questions(Calculate the
Choose questions to (providing additional
the question is error, coefficient of difficulty,
create a test with references if the
update the priority calculate the Coefficient of
specific specifications student does not have
questions) Cognitive level, other
the correct answer)
characteristics)
Fig. 4 Uses of artificial intelligence in the industry of questions
5.1 Methods for Formulating and Making Questions
Template-based methods [11]

The blocks identify the surface building for the questions also; templates contain
static text and variable parts. These templates describe the specification of terms
that can replace changing parts where they can be based on linguistic or moral
approaches.
Rules-based methods [12]

Rules often accompany text-based methods as inputs rely on append descriptive data
to the input; this metadata is used to match the entries to the pattern in the rules.
The rule is implemented to turn matching entries into questions.
Example: identify the appropriate questioning tool (where from, why?).
Methods based on statistical approaches [8]

Learning how to make questions from training data (pairs of questions and answers).
Example: Guess the location of the word or words to be replaced by blank (e.g.,
in questions to fill in the blank).
6 Results
All of these results were obtained using the following AI algorithms (artificial neural
network (ANN), convolutional neural network (CNN), and recurrent neural network
(RNN)) (Table 2).
In Fig. 5, we observe the quality of the accuracy for statistical method with a
scale ranging from 0.05 to 0.1% in each episode that we did compare to the other
two methods.
Table 2 AI algorithms used

Model ANN [13] CNN [14, 15] RNN [14]
to solve our problem
Use it for Tabular datasets Image data Text data
Classification Classification Classification
prediction prediction prediction
Regression Regression Regression
prediction prediction prediction
1,2
1
Rules method
0,8
Accuracy
0,6 Statistical method

0,4
0,2 Templates method

0
0 2 4 6 8 10
epochs
Fig. 5 Accuracy-based comparison of the three methods
Illustrations in Figs. 6 and 7 show how important it is to use advanced computer

technology instead of old traditional techniques to develop tests questions through
an analytical study we conducted at Cadi Ayyad University Marrakech, Morocco in
2020 when most questions were expressed about the University’s multiple-choice
questions. The results were analyzed after several criteria were taken into account,
including the average level of difficulty, and the quality of the questions and the
discrimination factor must be positive. Every time the rate of bad questions is
measured, the method based on artificial intelligence algorithms is much better than
Fig. 6 Ratio of bad 9

Number of studies
questions using the 8

traditional method
7
6
5
4
3
2
1
0
0-30% 31%-60% 61%-80% 81%-100%
Percentage of flawed questions
Fig. 7 Ratio of bad 9
Number of training
questions using AI 8
algorithms
7
6
5
4
3
2
1
0
5% 2,10% 0,75% 0,25%
Percentage of flawed questions
the traditional method. For example, when we studied nine studies, we found that
they represented 61–80% of the total number of questions analyzed there. This is a
very large percentage that was bad because it violated one of the criteria that was
set. But when we used modern methods in the same study (9 studies), the proportion
of bad questions was very low (0.25%) and so the good of the system in question
making compared to traditional methods.
Finally, we can conclude that there are many bad questions that are used to assess
students. What is the reason? Maybe one of the residents did not have the background
that we mentioned earlier, so to address these forms, we have decided to use and
develop a smart system based on AI (deep learning) techniques. This type of issue is
designed to improve a quality and avoid errors in the use of traditional techniques.
7 Conclusion
There are many areas in which we can make significant progress, especially in the
formulating quizzes questions, for example, enriching shapes and organizing ques-
tions mean designing a good and attractive structure to stimulate student skills and
improve the formulation of questions that can be demonstrated in a similar way to
human making, as well as creating and enriching data sources used. In this paper,
we study how to create good tests questions based on artificial intelligence tech-
niques and algorithms. Doing that is based on one of the most important criteria to
be included in the quality of the calendar and the test vocabulary. The comparison
between the most recent methods used to make traditional questions, and one of
the ways based on AI algorithms at Cadi Ayyad University Marrakech Morocco we
confirm that using the AI improve a quality and avoid errors in the use of tradi-
tional techniques. In particular, also organized data sources; explore other aspects of
controlling the difficulty of questions as one question has many characteristics such
as language or ethical qualities. Finally, this area should be developed so that only
a one-touch teacher can make a complete test, and this is the purpose we seek to
achieve with these modern technologies.
References
1. Bala, K., Kumar, M., Hulawale, S., Pandita, S.: Chat-Bot for college management system using
A.I. Int. Res. J. Eng. Technol. 2–6 (2017)
2. http://www.khayma.com/mohgan73/101msdcf/21.htm
3. Caldwell, D.J., Pate, A.N.: Effects of question formats on student and item performance. Am.
J. Pharm. Educ. 77, 1–5 (2013). https://doi.org/10.5688/ajpe77471
4. Bonham, S.W., Deardorff, D.L., Beichner, R.J.: Comparison of student performance using web
and paper-based homework in college-level physics. J. Res. Sci. Teach. 40, 1050–1071 (2003).
https://doi.org/10.1002/tea.10120
5. Jia, J., Chen, Y., Ding, Z., Ruan, M.: Effects of a vocabulary acquisition and assessment system
on students’ performance in a blended learning class for English subject. Comput. Educ. 58,
63–76 (2012). https://doi.org/10.1016/j.compedu.2011.08.002
6. Tobin, K.: The role of wait time in higher cognitive level learning. Rev. Educ. Res. 57, 69–95
(1987). https://doi.org/10.3102/00346543057001069
7. Mieloo, C., Raat, H., van Oort, F., et al.: Validity and reliability of the strengths and difficulties
Questionnaire in 5–6 year olds: differences by gender or by parental education? PLoS One 7
(2012). https://doi.org/10.1371/journal.pone.0036805
8. Tuwor, T., Sossou, M.A.: Gender discrimination and education in West Africa: strategies for
maintaining girls in school. Int. J. Incl. Educ. 12, 363–379 (2008). https://doi.org/10.1080/136
03110601183115
9. El Gourari, A., Raoufi, M., Skouri, M., & Ouatik, F. (2021). The Implementation of deep
reinforcement learning in e-Learning and distance learning: Remote practical work. Mobile
Information Systems. https://doi.org/10.1155/2021/9959954.
10. El Gourari, A. Skouri, M., Raoufi, M., & Ouatik, F. (2020) The future of the transition
to e-learning and distance learning using artificial intelligence. In: 2020 Sixth International
Conference on e-Learning (econf) pp. 279–284. https://doi.org/10.1109/econf51404.2020.938
5464.
11. Wu, S., Zhang, Y.: A comprehensive assessment of sequence-based and template-based
methods for protein contact prediction. Bioinformatics 24, 924–931 (2008). https://doi.org/
10.1093/bioinformatics/btn069
12. Douglas, K.M., Mislevy, R.J.: Estimating classification accuracy for complex decision rules
based on multiple scores. J. Educ. Behav. Stat. 35, 280–306 (2010). https://doi.org/10.3102/
1076998609346969
13. Agatonovic-Kustrin, S., Beresford, R.: Basic concepts of artificial neural network (ANN)
modeling and its application in pharmaceutical research. J. Pharm. Biomed. Anal. 22, 717–727
(2000). https://doi.org/10.1016/S0731-7085(99)00272-1
14. Retrieval, D.: Natural language processing for prolog programmers. Data Knowl. Eng. 12,
246–247 (1994). https://doi.org/10.1016/0169-023x(94)90017-5
15. Banerjee, I., Ling, Y., Chen, M.C., et al.: Comparative effectiveness of convolutional neural
network (CNN) and recurrent neural network (RNN) architectures for radiology text report
classification. Artif. Intell. Med. 97, 79–88 (2019). https://doi.org/10.1016/j.artmed.2018.
11.004
Smart Campus Ibn Tofail Approaches
and Implementation
Srhir Ahmed and Tomader Mazri
Abstract This paper describes the concept and studies of “smart campus” using
different methodologies and introduces a global strategy and use case for each envi-
ronment in the University Park. Our main goal is to define, present the general smart
campus principles and objectives which revolve around the different IOT and cellular
infrastructures in the university, introducing a general ICT architecture for their coor-
dination and also detailing their direct use in the management of the university and
applications for learning and research, reduce the cost and take one step forward a
university smart campus.
1 Introduction
The smart campus concept has been the main focus of many researchers recently
due to the valuable insights gained toward developing smart campus. The university
campus apparently is a small city where it delivers to all variant user’s a variant
service. There are several factors that attract the researchers to study the smart campus
including protecting the environment delivering high-quality services, protecting and
saving the cost. Smart campus is a related ecosystem, applications, services and use
cases, among the main standards that are directly related with focus on:
• Urban planning
• Transport
• Energy
• Education
• Health care
• Mobility
• Logistic
S. Ahmed (B)
Ibn Tofail University, Kénitra, Morocco
T. Mazri
National School of Applied Sciences of Kenitra, P.B: 242, 14000 Kénitra, Morocco
https://doi.org/10.1007/978-981-16-3637-0_39
550 S. Ahmed and T. Mazri
• E-government.
The Internet of Things is a fundamental part of the smart campus, and it is
inescapable getting it invoked. The Internet of Things is a communication model that
gained its tenacity from its capability of connecting variety of everyday life objects to
the Internet [1]. These objects include alarms, security locks, sensors, drones, appli-
ances, robots, office equipment and so on. Even though IOT is in its early stages,
there are many applications and standardization that has been developed in many
domains including home automation, smart grids, water and waste management,
traffic control, smart vehicles, healthcare assistance and industrial automation.
The IOT is involved everywhere in smart campus especially in:
• Facilities
• Electrical systems
• Safety
• Classroom technologies
• Tutoring spaces
• Residential services
• Physical and mental health.
This paper provides an overview of all the use case that we can implemented in
our smart campus.
2 Smart Campus Architecture
Smart campus is one of the innovations that will be developed at Ibn Tofail Univer-
sity, one of the main aspects that being considered in developing smart campus is
the infrastructure [2] that is why we introduce a technical architecture for smart
campus and proposes a local operational model. It also presents a case study of the
digitalization of the university.
The architecture of smart campus which described in (Fig. 1) defines three basic
aspects for smart campus:
1. Smart education that consists on: e-learning, RFID tags, virtual classroom
2. Smart parking, a parking system that provides a vehicle location tracking [21],
reporting information and also provides information when the parking is full
and specifying the number of availability places
3. Smart room: System that provides information related to the building tempera-
ture control systems, electrical systems and lighting systems.
As described in [26, 27], there are several concepts to be implemented to achieve
smart campus among them we find:
• Internet of Things (IoT)—The Internet of Things integrates sensors, controllers,
machines, people and things in a new way to realize intelligent identification,
location, tracking and monitoring
Smart Campus Ibn Tofail Approaches and Implementation 551
Fig. 1 System design of smart campus
• Big data technologies include mass data acquisition, mining, storage and
processing. Big data technology in the wisdom of the campus [28, 29] in all
aspects of the application will make its management services to a higher level
• Cloud computing: Cloud computing requires a combination of grid computing,
parallel computing, powerful integrated computing, distributed computing only
the use of open, integrated, high scalability, on-demand service cloud computing
model can provide good infrastructure support, collaborative information archi-
tecture and dynamically configurable resources
• Business intelligence: Business intelligence utilizes data warehousing and data
mining techniques to systematically store and manage user data, provide analysis
reports, provide decision-making information for a variety of university activities
[25], basis and analyze user data through various statistical analysis tools
• Social networking: The social network covers all forms of network services [14]
with the core of human society. It is a social or social characteristics of the network
services.
3 Related Work
In terms of applying the smart campus concept to learning activities, many research
projects have been proposed to develop many implementation of high-level use case
for smart capabilities that could improve education; the work in [3] presents a brief
description of the design of an IoT-based smart campus scheme focused on smart

parking, smart room and smart education. Also studies of an integrated platform for
all the service presented, where Wi-Fi is used to connect the different sensors and
end devices associated with the platform. Also different works aim for the smart
analysis.
And artificial intelligence support of teaching activities, such as following game-
based approaches [4], supporting multimedia conferences [5], smartphone applica-
tions [6], and students’ health [7]. In terms of smart mobility, Alvarez-Campana [8]
presented a university campus IoT platform for environmental and people flows moni-
toring, and Toutouh [9] describe a mobility prediction mechanism, implementing it
on the University of Málaga campus. The work in [10] focuses instead on IoT-
supported disaster management for smart campuses. The university of malaga has
identified the characteristics elements, solutions and key features of SmartUMA [11]
and present it assesses the impact and level of involvement of the campus community
in the different research and learning activities, and also there are numerous studies
conducted about smart campus realizations or concepts [9, 12–15], whereof many
are focusing on technological frameworks and architectures in particular use cases
the authors of [12] consider optimizing the management and the increasing number
of connected devices and propose a very precise model of dividing campus networks
according to roles; using authentication controllers based on Open Flow and virtu-
alization technologies throughout most of these work, particularly focused on IoT
and monitoring systems, the reader is confronted with a situation in which many do
not focus on security-related issues of these remote technologies, which will be a
significant concern for industries looking to utilize and exploit these technologies.
4 Proposed Method
Our work consists to present paper focuses on the description as a relevant smart
campus example case, of the characteristics, elements, solutions and key features
of outsmart campus. Secondly, it defines the impact and level of involvement of
the campus community in the different research and learning activities, providing
recommendations for all environment that we define in our university park.
Our approach consists in defining four functional definition parts (Fig. 2) that will
describe our fields of view, and then, we will define all the description relevant smart
campus use case with each part according to well-defined and detailed axes for each
field.
After the functional decomposition, we will associate each functional definition
among the four functions defined in (Fig. 2) by all the involved use cases.
Fig. 2 Functional definition

for smart campus
Academic Infrastruture
Business and
Residential
Leisure
4.1 Academic Function
As defined in the previous section, which defines our smart campus architecture
separate into three parts.
• Smart home
• Smart education
• Smart parking.
For the part of functions academic, we break it down into different use case as
mentioned below to cover our university environment:
• Conference room
• Laboratories
• Study place for individual use
• Library
• Academic hospital
• Special conference seating.
4.2 Infrastructure Function
A key benefit of smart campus infrastructure is that it helps universities which offer
incoming students the resources they want while keeping costs down, and it involves
all the logistic and transport parts as seen below:
• Parking spaces
• Means of transportation
• Accessibility (car, bus)
• Reserved space for bicycles.
4.3 Residential Function
This part will define everything that is related to the accommodation and housing of
students, both external and internal, as well as guests at university events
• Student accommodation (national)
• Student accommodation (short-term)
• Faculty housing
• Short stay apartment.
4.4 Business and Leisure
The goal in this part is to involve activities that will help the students to inte-
grate among themselves either through sports competitions or through paid activities
offered by companies in collaboration with the university.
• Eternal conference spaces
• Cultural centers associating
• Sports and technical competitions
• Activities that combine work and learning.
5 Use Case for Smart Campus
The subject and its exploitation possibilities being vast, we have chosen to focus on
four proposals as seen in (Fig. 3) where the installation of sensors seems interesting
as well as the corresponding scenarios. Also the use of the Internet of Things (IoT)
on the intelligent campus provides the ability to monitor every aspect of the work
environment such as the temperature, overcrowding, the availability of equipment
Fig. 3 Smart campus Ibn

Tofail use case Smart Home
System Use Smart

Monitoring Case Parking
Smart
Éducation
and catering facilities. This data can be combined with the onboarding process that
collects feedback data from staff to match with and improve the work experience,
and we have defined these scenarios for the use of our smart campus Ibn Tofail.
Below, we will define each use case separately and define the proposals and studies
for each case.
5.1 Smart Home
The trend in recent years has been toward “Smart Building” and “Smart Home,” but
the notion of a smart building is still rather vague, and there are many definitions
[12]. An overall definition could be as follows: automation implemented to make
the management and operation of the building more efficient. This automation can
take several forms as well as the final goal. Smart buildings are often associated with
buildings that are energy autonomous and regulate their consumption themselves
[13]. Thus, smart campus projects already existing in France such as the one in
Versailles are mainly concerned with this energy aspect. In the context of our project,
the meaning we are interested in is to make the Campus “Smart” by connecting it and
offering its users information on its current use, the conditions that prevail there [19].
But to offer this type of service, like smart buildings and smart home, it is necessary
to set up a network of sensors that allow the desired information to be transmitted
and processed [20, 22]; for this purpose, we have defined our work area for smart
home on the following points:
• Temperature control systems
• Electrical systems
• Lighting systems
• Water sensors
• Fire alarm systems
• Security alarms
• Fire detection
• Temperature monitoring
• Visual management.
5.2 Smart Parking
A new need was identified, with suggesting the idea of being able to find out the
number of spaces available in the parking lot.
In this section, we will install presence sensors in the parking lot and then register
them from a sensor API. In this way, the newly collected data will be stored and
can be evaluated. Following this, a user will be able to create a service that allows
users to view the number of spaces remaining in the parking lot [15]. For example,
a student driving to class in the morning would like to know whether there are any
free parking spaces left in the parking lot before he or she goes to class. If there
are none left, it is more interesting for him or her to go directly to another parking
lot. He therefore accesses a web application of his choice [17] (website, smartphone
application) connected to our platform and can visualize the number of remaining
spaces in the parking.
A little later, a study group wants to know the average occupancy of the parking
lot during the day in order to know if it is interesting to activate the barriers [18]. To
do this, this group logs on to the platform and retrieves information on the parking
lot occupancy over the last few months and can perform calculations using this data.
The work area for smart parking was defined on the following points:
• Number of cars
• Control departure/arrival
• Traffic control
• Monitoring
• Parking control.
Considering the increase of urban and traffic congestion in our smart campus,
smart parking is the best strategic issue to work on, not only in the research field but
also in economic interests.
5.3 Smart Education
Access to education is vital for the economic, social and cultural development of all
societies. In this section, we have set ourselves many goals to help to connect all
the students in the digital age. First of all, we want to provide higher education with
an online learning platform or e-learning. It should equip academic institutions and
open digital spaces with an autonomous virtual desktop infrastructure (VDI) with
its own processing, hosting and storage capacity. The smart education aims on the
one hand to modernize the digital communication infrastructure data processing and
storage, and on the other hand, to deploy technological platforms to improve teaching
and learning in universities, elementary schools, colleges and high schools, it will
enable as follows:
• The online learning of students through the “e-learning platform”
• The production and recording of courses and didactic and pedagogical content
through the “Virtual Classroom”
• The modernization of networks and strengthen the security of the connectivity of
the universities
• The deployment of a multimedia room in universities. To provide students with
connectivity and access to personal data processing and storage space through
VDI rooms.
• Laboratory access through campus cards only
• Smart printers in campus, access using campus cards
• Biometric/card-based attendance system for students

• Online booking facility for hall and classrooms
• Online leave application and tracking
• Online book recommendation from library for the faculty. e-Notice facilities over
Intranet.
5.4 System Monitoring
This section interacts in all the defined domains and functionalities that have been
defined in the previous sections (academic, infrastructure and residential) [16]. All
the equipment on our intelligent campus must be supervised in real time, hence the
need to set up certain points:
• IP video surveillance
• Fire alarm systems
• Access to electronic doors
• IP-based police and security teams
• Police vehicles over IP
• Smart monitoring of gas leakage in residences, flats.
Among the main advantages that we will propose for our smart campus defined
in the following:
• RFID tags
• Notification, centralization, alerts
• Physical safety
• Digitalization.
Smart campuses are safe and secure; it allows to ensure optimal student attendance
through the use of RFID tags also for each activity; a smart school gains a competitive
advantage through the use of SMS and email communication integrated into the
software [21, 22], and the management can send instant notifications and alerts, and
the installation of CCTV cameras and other surveillance systems on the premises
ensures total security for students, teachers, staff and school equipment. Also a variety
of sensors are used at the street layer for a variety of smart campus use cases. Here
is a short representative list [23]:
• Magnetic sensor
• Lighting controller
• Video cameras combined with video analytics
• Air quality sensor
• Device counters.
The magnetic sensor detects a parking event by analyzing changes, such as a car or
a truck, [24] comes close to it for the lighting controller which can dim and brighten
a light according to a combination of temporal and ambient light conditions to save
more energy, and for cameras combined with video analytics, video can detect faces,
vehicles and traffic conditions for a variety of traffic and security use cases for our
campus.
For the air quality sensor, it allows to detect and measure the concentrations of
gases and particles to give a perspective of the pollution in a given area and avoid
the risk of fire either on our parking lot or on our smart home, and for the device
counters, he gives an estimate of the number of devices in the area.
6 Conclusions
Certain requirements for the establishment of a smart campus pose several techno-
logical challenges, including the following:
• How to collect data?
• What are the different data sources, including hardware terminals and software?
• What type of network connectivity is best suited to each type of data to be
collected?
• What type of power availability and other infrastructure, such as storage, is
required?
• How can data from different sources be combined to create a unified view?
• How can the final analysis be made available to specialized intelligent campus
staff, such as traffic operators, parking control officers, lighting operators?
Each intelligent campus needs a suitable and structured computer model that
enables distributed data processing with the level of resilience, scale, speed and
mobility required to efficiently and effectively deliver the value that the generated data
can create when properly processed on the network. For this purpose, the principles
and driving characteristics of the University of Ibn Tofail Smart Campus approaches
learning and research activities are detailed. A general system design that describes
the main technological infrastructures of a smart campus is presented associated
with functional definition and will define a scheme for the implementation of smart
campus based on four functional definition parts that will describe our fields view and
define all the description relevant smart campus use case with each part according to
well-defined and detailed axes for each field:
1. Smart education
2. Smart parking
3. Smart home.
And describe their implementation all this field can be developed on IOT
Technology which is the main key for the deployment of smart campus.
References
1. Alghamdi, A., Thanoon, M., Alsulami A.: Toward a Smart Campus Using IoT: Framework for
Safety and Security System on a University Campus
2. Jurva, R., Matinmikko-Blue, M., Niemelä, V., Nenonen, S.: Architecture and Operational
Model for Smart Campus Digital Infrastructure
3. Sari, M.W., Ciptadi, P.W., Hardyanto, R.H.: Study of Smart Campus Development Using
Internet of Things
4. Zhai, X., Dong, Y., Yuan, J.: Investigating learners’ technology engagement—a perspective
from ubiquitous game-based learning in smart campus. IEEE Access 6, 10279–10287 (2018)
5. Zhang, W., Zhang, X., Shi, H.: MMCSACC: a multi-source multimedia conference system
assisted by cloud computing for smart campus. IEEE Access 6, 35879–35889 (2018)
6. Kim, T., Ramos, C., Mohammed, S.: Smart city and IoT. Futur. Gener. Comput. Syst. 78,
160–162 (2017)
7. Gao, X., Sun, Y., Hao, L., Yang, H., Chen, Y., Xiang, C.: A new soft pneumatic elbow pad for
joint assistance with application to smart campus. IEEE Access 6, 38967–38976 (2018)
8. Lin, Y.B., Chen, L.K., Shieh, M.Z., Lin, Y.W., Yen, T.H.: CampusTalk: IoT devices and their
interesting features on campus applications. IEEE Access 6, 26036–26046 (2018)
9. Alvarez-Campana, M., López, G., Vázquez, E.V., Villagrá, V., Berrocal, J.: Smart CEI Moncloa:
an IoT-based platform for people flow and environmental monitoring on a smart university
campus. Sensors 17, 2856 (2017)
10. Van Merode, D., Tabunshchyk, G., Patrakhalko, K., Yuriy, G.: Flexible technologies for smart
campus. In: 13th International Conference on Remote Engineering and Virtual Instrumentation
(REV), 2016.
11. Toutouh, J., Arellano, J., Alba, E.: BiPred: a bilevel evolutionary algorithm for prediction in
smart mobility. Sensors 18, 4123 (2018)
12. Hannan, A., Arshad, S., Azam, M., Loo, J., Ahmed, S., Majeed, M., Shah, S.: Disaster manage-
ment system aided by named data network of things: architecture, design, and analysis. Sensors
18, 2431 (2018)
13. Universidad de Málaga: Smart-campus, Vicerrectorado de Smart-campus. Available online:
https://www.uma.es/smart-campus
14. Chen, C., Chen, C., Lu, S.-H., Tseng, C.-C.: Role-based campus network slicing. In IEEE 24th
International Conference on Network Protocols (ICNP) Workshop on Control, Operation and
Application in SDN Protocols, 2016
15. Qian, Lv.: Establishment of smart campus based on cloud computing and Internet of Things.
Comput. Sci. 38(10), 18–21 (2011)
16. Nie, X.: Constructing smart campus based on the cloud computing platform and the internet
of things. In: Proceedings of the 2nd International Conference on Computer Science and
Electronics Engineering (ICCSEE), 2013
17. Bahl, P., Padmanabhan, V.N.: Radar: an in-building rf-based user location and tracking system.
In: IEEE INFOCOM 2000
18. Kaur, V., Tyagi, A., Kritika, M., Kumari, P., Salvi, S.: Crowdsourcing based android application
for structural health monitoring and data analytics of roads using cloud computing. In: 2017
International Conference on Innovative Mechanisms for Industry Applications (ICIMIA, 2017)
pp. 350–360
19. Sharma, K., Suryakanthi, T.: Smart system: IoT for university. In: International Conference on
Green Computing and Internet of Things (ICGCloT), pp. 1586–1593 (2015)
20. Wang, C., Vo, H.T., Ni, P.: An IoT application for fault diagnosis and prediction. In: IEEE
International Conference on Data Science and Data Intensive Systems, pp. 726–731 (2015)
21. Lee, N.K., Lee, H.K., Lee, H.W., Ryu, W.: Smart home web of object architecture. In:
International Conference on Information and Communication Technology Convergence,
pp. 1200–1216 (2015).
22. Hager, M., Schellenberg, S., Seitz, J., Mann, S., Schorcht, G.: Secure and QoS-aware commu-
nications for smart home services, in 2012 35th International Conference Telecommunications
and Signal Processing (TSP), 2012, pp. 10–19.
23. Luo, L.: Data acquisition and analysis of smart campus based on wireless sensor. Wirel. Pers.
Commun. 102, 2897–2911 (2018)
24. Prandi, C., Monti, L., Ceccarini, C., Salomoni, P.: Smart campus: fostering the community
awareness through an intelligent environment. Mob. Netw. Appl. 2019.
25. CiscoIOT https://idoc.pub/documents/ciscopressiotfundamentals-6nq80ejgd9nw,last accessed
2020/08/10
26. SmartBuilt: http://www.greenbang.com/from-inspired-to-awful-8-definitions-of-smart-buildi
ngs_18078.html. Last accessed 2020/07/20
27. Tien, J.M.: Big Data: unleashing information. J. Syst. Sci. Syst. Eng. 2013(02) (2013)
28. Guo, H., Wang, L., Chen, F., Liang, D.: Scientific big data and digital Earth. Chin. Sci. Bull.
2014(35) (2014)
29. Chen, J., Xiang, L.G., Gong, J.Y.: Virtual globe-based integration and sharing service method
of GeoSpatial Information. Sci. China (Earth Sci.) 2013(10) (2013)
Boosting Students Motivation Through
Gamified Hybrid Learning Environments
Bleurabbit Case Study
Mohammed Berehil
Abstract Gamification refers to the use of game elements in non-game context, it

has been highly used in different fields like business, where stakeholders try to cre-
ate enjoyable experiences and motivate the users. Moreover, gamification also have
been frequently applied in education, where many teachers applied game mechanics
to design learning activities that can increase the students’ engagement and motiva-
tion in the learning environment. In our research, we tried to assess the impact of
hybrid gamified leaning environment on student’s motivation. We have used the MDA
(Mechanics, Dynamics, Aesthetics) and the Hexad (Player’s type model) models to
create different enjoyable activities inside a gamified learning environment called
Bleurabbit. Furthermore, to respond to our research questions, we have designed a
case study based on applied research methodology and used interviews and obser-
vation to collect data. The collected results were very satisfactory and have demon-
strated the positive effect of gamification on boosting students’ motivation.
1 Introduction
The pandemic spread has been a big accelerator for online learning, many universities
have moved online, and in fact, it was considered as magical solution in difficult times.
Nevertheless, one of the main problems facing online or hybrid learning environments
is the lack of motivation, many students feel more frustrated and alone, as the physical
presence of the tutor and the classmates is partly or totally suppressed. So, the learning
environment plays a major in learning process, and gamifying it can solve the problem
of motivation, especially that many kinds of the literature [16] suggested that gamified
learning environment can enhance students’ motivation [2, 20].
In our research, we assumed that that changing the learning environment could
ameliorate the learning process. Consequently, we tried to combine two gamifica-
tion frameworks to create more enjoyable environments and bring knowledge to the
M. Berehil (B)
Faculty of Letters, Mohammed First University, Oujda, Morocco
e-mail: m.berehil@ump.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_40
562 M. Berehil
student’s favourite environment (which is related to games) [25]. In addition, we tried

to assess the effectiveness of gamified virtual learning environment on motivating
studentsusing a case study approach, so that we can observe closely the interaction
between students and different aspects of the gamified learning environment.
We divided our paper into nine parts organising it as follows: Sect. 2 was devoted
to related works, in Sect. 3, we provided working definition of gamification, in Sect. 4,
we discussed the used gamification frameworks, in Sect. 5, we gave an overview of
Bleurabbit, in Sect. 6, we highlighted the adopted methodological approach, and in
Sect. 7, we described our case study. While Sect. 8 was devoted to the results and
discussion, in the ninth section, we provided conclusion and perspectives.
2 Related Works
Many kind of the literature related to the gamified learning environments used quan-
titative research approach [20]; in a published article Gamification and student moti-
vation, researchers proposed a study where they assess the effect of the gamification
learning environment on students motivation. The researchers [2] used a pre- and
post-experimentation survey to test the knowledge acquisition in gamified learning
environment. The results have shown positive attitude from students toward the gam-
ified learning environment. Nonetheless, we consider that qualitative methods can
be more effective in dealing with gamification, especially that we consider it as a
human phenomenon that is related to a uniquely personal experience which differs
from one person to another.
The same vision is advanced in a study [3] where a qualitative study was led,
based on usage as cogent (College Enterprise), a gamified environment to simulate
a real enterprise environment using virtual money was developed. Researchers [3]
used focus group and interviews to collect data, and the results show that well-
implemented gamified learning environment has very positive effects on student’s
motivation and increases the engagement among them. Nonetheless, the proposed
study was conducted in game-based environment which differ from gamified learning
environment as it needs more financial and human resources, and additionally, it
cannot fit different pedagogical models.
In another paper [12], the researcher used the MDA framework to develop a design
model, for eLearning gamified environment. The researcher emphasises the impor-
tance of using gamification framework as it can bridge the gap between the game
elements and the design rules. As they used this model to develop the BlackSlash.com,
which is a website for children to learn HTML, researchers have proposed promising
results. This also can be observed in the work developed by researchers in Singa-
pore University, where they use octalysis model to develop a mobile application for
increasing students’ activity, outside the classroom. The results were so promising,
and student have shown a big motivation to such application [21]. Eventless, none
of the above-mentioned works have address the issue of player type, which can be
very important to create a better experience.
Boosting Students Motivation Through Gamified Hybrid Learning … 563
3 Gamification Working Definition
Gamification is a very hard term to be defined. While games are not an odd element
in education, they were highly adopted in the past and adopted in by many teachers
[15]; nevertheless, gamification represents much more complex concept than a sim-
ple educational game [19], as it was used in other fields rather than education like
healthcare and business [7] …
There is a big misunderstanding between different concepts, including games as
“a system in which players engage in an artificial conflict, defined by rules, that
results in a quantifiable outcome” [18], serious games “serious games describe the
use of complete games for non-Entertainment purposes” [14]Game-based learning
(GBL)called. GBL is a pedagogical approach to learning which consists of devel-
oping games to attend pedagogical goals; the developed games are part of a learning
process and are meant for developing new learning skills. [15].
The term gamification [4] was firstly used in 2008 in the digital media industry,
but was not popularised until 2010 in the field of interfaces design [5]. Dren defines it
as “the use of game mechanics in non-game context” [5, 24]. To apply gamification,
the designer works carefully on deciding how to bring game elements in a serious
context in a swift way and create a game-like experience[4, 22]. In education, we can
identify two kinds of gamification design [10], content gamification “the application
of game elements, game mechanics and game thinking to alter content to make it
more game-like” [9].
This kind of gameful design acts on the content and makes it more game-like,
so, for example, you can add a narrative element or make the course as story…The
second type is structural gamification which is “the application of game elements
to propel a learner through content with no alteration or changes to the content”
[9]. This design is about making learning environment gamified without interfering
with the content, for instance, you can add a leaderboard or badges or experience
points…in your classroom.
In our research, as shown in Fig. 1, we consider gamification as the intersec-
tion between playful design, game elements and motivational design, each of these
elements can be part of the gamification design process
4 Gamification Design Frameworks (Two Models

Approach)
The gamification frameworks are the result of transdisciplinary research in differ-

ent fields including psychology, neurosciences and computer science …[11, 17], in
fact, many researchers and game designers have tried to regroup the different game
elements into one design framework (framework for success, gamification design
process...) [11] that as a toolkit for games like experience designers. We can identify
two types of gamification frameworks:
564 M. Berehil
Fig. 1 Gamification
• Game-based frameworks: refer to gamification frameworks that are mainly con-

cerned with the game structure and elements, including mechanics, dynamics
…[10] the main concern of the designer is providing good game structure, which
facilitate player journey inside the game.
• Player-based gamification frameworks: a framework that is player oriented, in
other words, the development of the game is related to the player actions and
reactions inside the game.
4.1 M.D.A
Developed by robin Hunicke et al. [8], the MDA framework, (Mechanics, Dynam-
ics and Aesthetics), represents one of the earliest gamification frameworks used to
develop, understand and depict game elements in a context that is not a game. It
brokes games into three major elements
• Mechanics represent the different elements provided by the game designer that
includes the settings, the goals, the rules... Games mechanics are constant and do
not change [8].
• Dynamics Unlike mechanics, dynamics are produced by the player, more precisely
dynamics are the results of the player behaviour that occurs on the mechanics, and
in other words, dynamic are not fix matter but rather a variant that depends on the
player’s strategies of playing and the way he interacts with the game mechanics. [8].
• Aesthetics The last important component of games according to MDA creatorsis
what it called aesthetics which refers to the emotional response of the user to the
system; this could be experienced as passion, joy, eustress…Aesthetics are the
results of the interaction between the player and the game or between mechanics
and dynamics. Most of the time the wished result is FUN [8].
Fig. 2 MDA perspectives
4.1.1 Game Designer Versus Player Perspective
MDA represents a group of “connected lenses” [8] that work complementary; it helps
to analyse the game from both player and designer perspectives (Fig. 2).
First, we have the designer whose main interest is the mechanics, which in his
perspective gives rise to dynamics and aesthetics [8]. The second actor is the player,
and the actions of the player are driven by aesthetics which are primordial to the
success of gamification.
4.2 Player-Based Gamification Hexad Model
The Hexad model, which was developed by Andrzej Marczewsk [6], and based on
BRTEL model [1], was empirically proven and tested [13, 23], through a question-
naire. Marczewski offers a more detailed model and identified eight player types,
depending on the player’s behaviour and intentions inside the game [13]:
• Free Spirits are motivated by autonomy and self-expression. They want to create
and explore.
• Achievers are motivated by mastery. They are looking to gain knowledge, learn
new skills and improve themselves. They want challenges to overcome.
• Philanthropists are motivated by purpose and meaning. This group are altruistic,
wanting to give to other people and enrich the lives of others in some way, with
no expectation of reward.
• Disruptors are motivated by change. In general, they want to disrupt the estab-
lished system, either directly or through other users to force positive or negative
change.
• Players are motivated by extrinsic rewards. They will do what is needed to collect
rewards from a system and not much more. They are in it for themselves.
These types of players also fall in three big categories as shown in Fig. 3. Willing
to play, less willing to play and not willing to play.
In our research, we tried to combine both framework (MDA and Hexad design
frameworks) to create an adequate learning environment, which allows a fluid
players–mechanics interaction and takes into consideration subtleties between dif-
ferent types of players, aiming to generate the targeted dynamics and aesthetics.
566 M. Berehil
Fig. 3 HEXAD player type
5 Bluerabbit
BlueRabbit is a free Web-based gamified virtual learning environment upon which we

have based our case study, developed by a Mexican computer scientist, BlueRabbit
is built on the principles of MDA gamification design framework. The platform’s
first aim is to move beyond marks and grades, by establishing a system offers an
enjoyable learning environment. Additionally, it allows the use of Hexad player
type questionnaire for each student, which make it the propitious platform for our
experience.
The platform is equipped with a very colourful interface as shown in Fig. 4 and
does not need great technical skills, ergonomically speaking, it is very easy and
enjoyable, and it is also responsive which mean it offers the possibility for both
mobiles and computers usage. Furthermore, it is both responsive and user friendly
for course designer/tutor and the student, which means it allows the use of it in both
computers and mobile devices (tablet, smartphones …). We can divide the platform
componenets into three main elements:
Fig. 4 Overview of the platform interface
5.1 Narrative Elements
Narrative elements refer to the proposed way by the platform to exhibit the content
as a story, and in other words, these elements allow the course designer to transform
the content into a plot where students play the heroes, this includes:
• Quests are the most basic element of BlueRabbit, and they are the mandatory
activities of the learning process.
• Sidequest: The sidequest represents secondary activities, nonetheless the
sidequests represent another opportunity for the players to get money, badges
and XP.
• Characters are the learners inside the platform.
• Missions are special activities that are limited in time and have more constraints.
• Project is a space for cooperative activities, which allows the interaction between
different team members.
• Challenges are a rubric that allows you to propose activities which are challenging
to the students.
5.2 Feedback and Rewards
We can divide the reward system provided by the Bluerabbit platform into two types.
Extrinsic motivation includes the XP rewards and badges gained by students through
realising quests or sidequests or any other activities intrinsic motivation component
mainly related to the student’s progress, including levels and leaderboar.
568 M. Berehil
Fig. 5 Research design cycle
6 Methdological Approach
We have opted for applied research paradigm, as the nature of our research problem,
and our main goal was testing the use of gamification in particular learning environ-
ment and assessing its efficiency on motivating students, as shown in Fig. 5 which
we have designed our research iteratively.
We have adopted a qualitative research. So, we have chosen to design a case study
and collect data through interviews and observation
7 Case Study Design
We have opted for a case study as a research method, which allows “an in-depth
exploration of a single case” [26]. In other words, it helps studying and analysing
the behaviours and reactions of a student or a group of students inside a learning
environment. We constructed our case study as follows; first choosing the sample,
setting up the environment, designing the activities, implementing the environment
and finally observing the students.
7.1 Sampling
For our research sample, we have had sixteen students (Master1 students, Ingénierie
de formation et Technologies éducatives Mohammed first university) to whom we
introduced the Bluerabbit platform in one of the modules they studied in the second
semester, entitled Motivation, Objectifs et compétences. These students were familiar

with platform usage, and nonetheless, it was the first time they use a gamified learning
environment.
7.2 The Onboarding Process
The onboarding process is the process through which we have introduced students to
the learning environment and its functions. First, we started with face to face session
then a series of synchronous and asynchronous communication sessions were kept
to help students with any kind of issues.
To introduce the platform to the students, we have devoted a half-hour face to face
session. In the first 15 min, we have introduced the main functions and mechanisms
of the platform through a small practical presentation and simulation.
7.2.1 Synchronous Session
In addition to the face-to-face sessions, and as the platform did not offer the possibility
of synchronous communication, a Facebook special group also was created to ensure
fluid communication with the students and continuous feedback delivery, especially
during the first days.
7.2.2 Welcome Badges
The first offered badges were what we called welcome badges, and these badges were
offered to each student who successfully subscribes to the platform (Each student
can display the badge in his profile and can be seen by other students.). To push the
onboarding process, familiarise the students with the reward system.
7.2.3 Quest Design
The designed activities were part of a studied module where students were obliged
to have exams, grades and final mark to pass the module. Additionally, we have been
obliged to redesign the last two mandatory activities of the module in collaboration
with the tutor, so we can respect the module’s objectives, hierarchy and pedagogi-
cal design. The first activity was programmed a week before the second, and both
activities were accounted for the final mark of the module.
We have first changed the nature of the activities by moving from the traditional
notion of pedagogical activities into the realm of narrative as we identified them as
quests. Additionally, we have programmed the system, so that it offered 1500 Bloo
and 400 XP for submitting the task on time, and it was opened directly without the
570 M. Berehil
need of an extra sidequest to be done. Students were asked to write a synthesis about
motivation based on the multiple intelligence theory of Gardner. A leader board was
always kept available to show students who finished first the quest.
Unlike the first quest, the second quest did not open until you finish the first one
and reach the second level, to reach the second level inside the game you need to
acquire a given amount of XP points which can be collected through completing the
sidequests.
7.3 Sidequest
Simultaneously, we have also designed four sidequests which appeared successively

in the learning environment. The sidequests did not have any influence on the marks,
and nonetheless, they help to gather virtual money (which helps to buy deadlines),
open mandatory quests and reaching higher levels.
We have proposed four sidequests, each sidequest was rewarded differently
depending on the difficulty of it, the asked tasks were simple and do not need a
lot of time to be done, and no deadline was imposed. The students can directly
respond to the sidequests on the platform and get an immediate reward.
7.4 Deadlines
Even though we have put strict deadlines, we have offered the students the possibility
to buy new deadlines, in case they missed the first established deadline, which gave
them more possibilities for failing and redoing the activities. Besides, the deadline
buying would also push students into doing more activities (sidequest) for gathering
the money to buy the deadlines. The flexibility of deadlines was set up to offer student
several lives inside the learning environment.
7.5 Reward and Rating Mechanism
A special rating system was set up, making both the tutor and the students more
comfortable and raise the competitiveness among the students. The proposed system
offers different kinds of tokens including badges, marks, experience points and rating
leaderboard. Furthermore, some rewards were automatically given after submitting
the work, while other rewards were given by the tutor. This system was specially
and carefully designed to keep students motivated using the different gamification
features of the learning environment while designing it. We have tried to abide by
certain pedagogical constraint imposed by the tutor like the final mark and the final
exam.
7.5.1 Badges
in addition to the subscription badges, we have also provided other kinds of badges
calling them badges d’excellence, those badges were given to students who have pre-
sented the best works, respecting the deadlines. Badges represent immediate reward
as they were distributed shortly after the submission of the work. The badges were
also accompanied by Bloo money and XP.
7.5.2 Experience Points
Experience points were given to students after realising the asked tasks, and the
XP amount depends on the length and difficulty of each task. These points allow the
student to transcend in the levels and assess his level of mastery. Each quest, sidequest
or badge is rewarded with money, as shown in fig, the money allows students to buy
deadline or get unblocked. We have established this type of reward to push students
to do their best and maximum of the activities and collect the maximum of the money
as some students can be motivated with the collection of items.
7.5.3 Leader Boards
The second kind of proposed reward was leaderboards, so we had proposed two
kinds of leaderboards. The first leaderboard called performance leaderboard was a
personal board kept on the left of the screen of each student, the role of this board is
to remind him of his achievements, tickets, done quest, sidequests, and also it plays
the role of a to-do list by alerting students about what is left to be done (Fig. 6).
The second leaderboard was a general ranking board that exhibits student’s
achievements in the learning environment. This board includes the finished quests
and the sidequests. Additionally, the board shows the experience points gained by
each student and the position of the student among the classroom, which kept student
updated about their position among their classmates.
Fig. 6 Performance
leaderboard
572 M. Berehil
Fig. 7 Ranking leaderboard
As shown in Fig. 7 no GPA was calculated, there were no accumulated points to

be counted, and we opted for XP and levels to quantify the progress of the student
in the learning environment which was more practical.
8.1 Results
8.1.1 Quest Versus Sidequest
The comparison between the mandatory quests and optional sidequest completion
would allow us to assess the motivated/non-motivated students, 4 students have com-
pleted both the proposed quests and optional sidequests, 12 students only mandatory
quests.
8.1.2 Player Type Identification
The second criterion is the player type identification. We consider students who
identify their play as more involved with the game and believe in the game mechanics.
Four students identify their player item, two were explorers and two were free spirit.
Item 12 did not respond to the questionnaire.
8.1.3 Deadlines
The third criterion was the deadline respect, and this item will influence the other
mechanic as first we consider the student who respects deadlines more motivated,
nevertheless who grilled the deadline would use other game mechanics like the quests
or the missions which make him enjoys more the game. No deadline was grilled, all
students submitted their work on time, and they were very punctual.
8.1.4 Badges
Reward distribution represents one of the major criteria that we have considered
for assessing the students’ motivation and interaction with the game mechanics as
badges reflect the achievements of students and parallel the involvement with the
game mechanics which 16 first subscribers were distributed (for all students), 4 first
realisation, 4 hard workers and 0 philanthropist.
8.1.5 Communication
Communication emerges as a criterion in the middle of the experience, as we have

noticed that some students communicate constantly and were very interested in the
experience. So we considered it as influential in the motivation of students. Four
students out of 16 have contacted us. The five students who have communicated did
it through Facebook group rather than the official platform in both the synchronous
and asynchronous ways. Most of the discussion was around technical issues.
8.1.6 Interviews
Even though we inform the students beforehand about the interview’s time, we had
interviews with only six students, the only available students, as they were full-time
574 M. Berehil
students and do not work. Our questions aimed at collection the student’s impressions,
who were relaxed as we keep the informal statue.
8.2 Discussion
While conducting our interviews, we have noticed that students have shown more
interest in the platform visual elements, which reflects the importance of the visual
elements in the learning process and especially the motivational aspect.
The results show that the mandatory quests which were directly involved in the
final mark were all submitted while the optional side quests were done only by four
students, and this reflects the nature of the items that motivated the student. Besides,
it opens the debate about the perception of the reward and the culture of the final
test, having a good final mark can be influential and hinder doing other activities. In
other words, changing the nature of the reward can affect the student’s activity and
the learning process.
We have also noticed that only four players identified their player types by respond-
ing to the questionnaire, this was optional and not mandatory, and the same four
players are those who completed all the sidequests and win the biggest number of
badges, XP and Bloo. We can conclude that those four players were the most enthu-
siast about the platform and enjoy the different proposed game mechanics, as it was
confirmed during the short interviews. Those four players all had the same player
types: explorer and free spirits, which pushes us into thinking that the activities were
more oriented toward this kind of players.
The same four players were also the ones who frequently communicate and ask
more questions. Consequently, they have succeeded in the onboarding process and
have grasped the different components of the learning environment. We can con-
sider that communication is very important in any learning process, and ensuring a
fluid dialogue with students can guaranty their success. Most of the communication
process took place in Facebook, which is frequently used by student.
Deadlines and rewards distribution plays a major role in students’ motivation. As
we have seen, no deadline was grilled, which reflect the motivation of the student
to the activities. Nevertheless, the amount of gained rewards, which does not affect
the final mark, was poor; an only minority has done the sidequests, which means the
values of reward in the real world are very important to push the student’s motivation.
To sum up, the results extracted from this experience do not give obvious conclusions;
nevertheless, we can say that the gameful design have a positive impact on the
students who expressed their interest in the visual design of the platform and keep
constant communication with the environment. Additionally, we can say that the
visual elements play a major role in students’ motivation for learning, especially
when students start exploring the learning environment. Moreover, opting only for
structural gamification and not changing the nature of the content can still hinder
the gamification process. We can develop a more enjoyable environment applying
both structural and content gamification design effectively. We also can have a larger
sample and longer study, which will help us to have more exact results about gamified
learning environment.
References
1. Bartle, R.: HEARTS , Clubs, Diamonds, Spades: (1996) (2014)

2. Buckley, P., Doyle, E.: Gamification and student motivation. Interactive Learning Environments
24(6), 1162–1175 (2016). https://doi.org/10.1080/10494820.2014.964263
3. Chen, Y., Burton, T., Mihaela, V., Whittinghill, D.M.: Cogent: A case study of meaningful
gamification in education with virtual currency. International Journal of Emerging Technologies
in Learning 10(1), 39–45 (2015). https://doi.org/10.3991/ijet.v10i1.4247
4. Deterding, S.: Gamification: Designing for Motivation. Interactions 19(4), 14 (2012). https://
doi.org/10.1145/2212877.2212883
5. Deterding, S., Khaled, R., Nacke, L.E., Dixon, D., et al.: Gamification: toward a definition. In:
CHI 2011 Gamification Workshop Proceedings, vol. 12. Vancouver BC, Canada (2011)
6. Dixon, D.: Player Types and Gamification. CHI 2011 Workshop Gamification Using Game
Design Elements in NonGame Contexts, pp. 12–15 (2011). DOIACM 978-4503-0268-5/11/05.
http://gamification-research.org/wp-content/uploads/2011/04/11-Dixon.pdf
7. Hochleitner, W., Lankes, M., Nacke, L.E., Tscheligi, M., Busch, M., Mattheiss, E., Orji, R.,
Marczewski, A.: Personalization in serious and persuasive games and gamified interactions.
In: CHI PLAY 2015—Proceedings of the 2015 Annual Symposium on Computer-Human
Interaction in Play (May 2016), 811–816 (2015). https://doi.org/10.1145/2793107.2810260
8. Hunicke, R., Leblanc, M., Zubek, R.: MDA: A formal approach to game design and game
research. In: AAAI Workshop—Technical Report WS-04-04, pp .1–5 (2004)
9. Kapp, K.M.: Choose Your Level: Using Games and Gamification to Create Personalized
Instruction. Handbook on Personalized Learning for States, Districts, and Schools, pp. 131–143
(2016). http://www.centeril.org/2016handbook/resources/Cover_Kapp_web.pdf
10. Lamprinou, D., Paraskeva, F.: Gamification design framework based on SDT for student moti-
vation. In: Proceedings of 2015 International Conference on Interactive Mobile Communication
Technologies and Learning, IMCL 2015 (November), pp. 406–410 (2015). https://doi.org/10.
1109/IMCTL.2015.7359631
11. Maican, C., Lixandroiu, R., Constantin, C.: Interactivia.ro—a study of a gamification frame-
work using zero-cost tools. Comput. Hum. Behav. 61, 186–197 (2016). https://doi.org/10.1016/
j.chb.2016.03.023. http://dx.doi.org/10.1016/j.chb.2016.03.023
12. Malas, R.I., Hamtini, T.M.: A gamified e-learning design model to promote and improve
learning. International Review on Computers and Software 11(1), 8–19 (2016)
13. Marczewski, A., Holdings, M.: User Types HEXAD Gamification. Game Thinking and Moti-
vational Design, User Type HEXAD (2016)
14. Marsh, T.: Serious games continuum: Between games for purpose and experiential environ-
ments for purpose. Entertainment Computing 2(2), 61–68 (2011)
15. Muntean, C.I.: Raising engagement in e-learning through gamification. In: Proceedings of 6th
International Conference on Virtual Learning ICVL, vol. 1, pp. 323–329 (2011)
576 M. Berehil
16. Nah, F.: Gamification of Education: A Review of Literature—HCI in Business

vol. 8527, pp. 401–409. Springer (2014). https://doi.org/10.1007/978-3-319-07293-
7. http://link.springer.com/10.1007/978-3-319-07293-7%5Cnlink.springer.com/content/pdf/
10.1007/978-3-319-07293-7.pdf
17. Novak, D., Nagle, A., Riener, R.: Linking recognition accuracy and user experience in an
affective feedback loop. IEEE Transactions on Affective Computing 5(2), 168–172 (2014).
https://doi.org/10.1109/TAFFC.2014.2326870
18. Salen, K., Tekinbaş, K.S., Zimmerman, E.: Rules of Play: Game Design Fundamentals. MIT
Press (2004)
19. Stott, A., Neustaedter, C.: Analysis of Gamification in Education, vol. 8, p. 36. Surrey, BC,
Canada (2013)
20. Subhash, S., Cudney, E.A.: Gamified learning in higher education: A systematic review of the
literature. Comput. Hum. Behav. 87, 192–206 (2018). https://doi.org/10.1016/j.chb.2018.05.
028
21. Tan, J., Sockalingam, N.: Gamification to Engage Students in Higher Education, pp. 1–14.
Research Collection Lee Kong Chian School of Business (2015)
22. Tondello, G.F., Nacke, L.E.: Applying gameful design heuristics. In: Proceedings of the 2017
CHI Conference Extended Abstracts on Human Factors in Computing Systems, CHI EA ’17,
p. 1209–1212. Association for Computing Machinery, New York, NY, USA (2017). https://
doi.org/10.1145/3027063.3027116. https://doi.org/10.1145/3027063.3027116
23. Tondello, G.F., Wehbe, R.R., Diamond, L., Busch, M., Marczewski, A., Nacke, L.E.: The
gamification user types Hexad scale. In: CHI PLAY 2016—Proceedings of the 2016 Annual
Symposium on Computer-Human Interaction in Play, pp. 229–243 (2016). https://doi.org/10.
1145/2967934.2968082
24. Topics, T.: Using Game Elements in Non-game Contexts: Towards Determining the Effects
and Applicability of Gamification, pp. 3–4 (2011)
25. Yildirim, I.: Students’ perceptions about gamification of education: a Q-method analysis.
Egitim ve Bilim 42(191), 235–246 (2017). https://doi.org/10.15390/EB.2017.6970
26. Yin, R.K.: Case study Research and Applications: Design and Methods. Sage Publications
(2017)
An Analysis of ResNet50 Model
and RMSprop Optimizer for Education
Platform Using an Intelligent Chatbot
System
Youness Saadna, Anouar Abdelhakim Boudhir, and Mohamed Ben Ahmed
Abstract A chatbot is a software (or machine) that has the ability to talks with
a user: it is a virtual assistant that can answer a number of user questions, and
providing the correct responses. In the last few years, the use of chatbots is very
popular in various fields, such as health care, marketing, educational, supporting
systems, cultural heritage, entertainment, and many others. This paper proposes an
intelligent chatbot system that can give a response in the form of natural language or
audio to a natural language question or image input in different domains of educa-
tion and will support multiple languages (English, French, and Arabic). To realize
this System, we used different deep learning architectures (CNN, LSTM, Trans-
formers), computer vision, transfer learning to extract image features vector, and
natural language processing techniques. In the end, after the implementation of the
proposed model, a comparative study was conducted in order to prove the perfor-
mance of this system using image-response model and question-response model
using accuracy and BLEU score metrics.
1 Introduction
The advance in artificial intelligence technology has many benefits in various

domains, industry, economy, agriculture, education, and more.
Many researchers concluded the importance of artificial intelligence to improve
human life and particularly in the education domain. Thus, chatbots in industry and
education have increased considerably in the last few years. Many of them are used
as a tutor for students or as customer support. In each case, the chatbot is trained to
Y. Saadna (B) · A. A. Boudhir · M. Ben Ahmed

List Laboratory, Faculty of Sciences and Techniques, UAE University, Tangier, Morocco
A. A. Boudhir
e-mail: aboudhir@uae.ac.ma
M. Ben Ahmed
e-mail: mbenahmed@uae.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_41
578 Y. Saadna et al.
carry out a question-response task. Otherwise, chatbots are known to respond to user
questions on a given topic or level and do not support all types of questions (natural
language question, image). The purpose of this paper is to design and implement
a chatbot that has the ability to support all types of questions (natural language
question, image), support multi-language questions (English, French, Arabic), and
cover multiple levels of education. The system acts like an intelligent robot that
explains, in different languages, the given image or text in the input by giving the
response as text or audio in the output.
2 Related Works
In the literature, there are many approaches associated with chatbots, especially on
e-learning systems. From the start of the last decade, the usage of AI as e-learning
support has captured the interest of many researchers for its many implementa-
tions. One among these research works is [1], during which Nenkov N. et al. have
looked over the realization of intelligent agents on platform IBM Bluemix with IBM
Watson technology. These agents in the form of chatbots need to automate the inter-
action between the student and therefore the teacher within the frames of Moodle
learning management system. Watson is a cognitive system that merges capabilities in
analytics, NLP, and machine learning techniques. In this case, Facebook Messenger
Bot GUI Builder realizes a chatbot through Facebook Messenger to simplify commu-
nication between teachers and students: it might be arranged by acquiring Moodle
test basis. In Nordhaug et al. [2] proposed a game based e-learning tool named TFC
(The Forensic Challenger), used to teach digital forensic investigation. A chatbot
inside the learning platform helps students. A multiple-choice question based quiz is
realized for kinesthetic learners, and there is a pedagogical chatbot agent that helps
users. It provides easy navigation and interaction within the content. The chatbot
is implemented to be a pedagogical agent for the users, which is meant for discus-
sions and help with the topics. It also acts as a navigation tool and can play video
or use the advanced wiki if there are something to ask. In Satu et al. [3] many
chatbot applications based on AIML are analyzed: especially, an integrated platform
that consists of basic AIML knowledge is presented. During this project, chatbot is
named Tutor-bot because its functionality backing of didactics done in e-learning
environments. It contains some features as natural language management, presenta-
tion of contents, and interaction with a search engine. Besides, e-learning platforms
work is linked to indispensable services to web service. A continuous monitoring
service has been created on e-learning platform servers, which is another controlling
machine: Daemon. In Niranjan et al. [4] discussed an interesting approach using
Bayesian theory to match the request of students and provide the right response. In
particular, chatbot agent accepts to student’s answers and extracts the keywords from
the question with the use of a lexical parser, and then, the keywords are compared
with the category list database. The Bayesian probabilities are obtained for all cate-
gories within the list. Once the category is chosen, keywords are compared with
An Analysis of ResNet50 Model and RMSprop Optimizer … 579
the questions under the category using Bayesian probability theory. The answer to
the question, which has the biggest posterior probability, is then fed into the text to
speech conversion module and thus the student receives the answer to his question
as a voice response. In Farhan et al. [5] using a WebBot in an e-learning platform,
to deal with the lack of real-time responses for the students. In fact, when a student
asks a question on the e-learning platform the teacher could answer later. If there are
more students and more questions, this delay is going to be increased. WebBot is a
web-based chatbot that predicts future events based on keywords entered on the web.
During this work, Pandora is employed, a bot that saves the questions, and answers
them on XML style language,i.e., artificial intelligence markup language (AIML).
This bot is trained with a sequence of questions and answers: When it cannot furnish
a response to a question, a human user is in charge of responding.
3 System Architecture
The architecture of our system is composed of:

• Front-End
• Back-End
• Model
• Database (Fig. 1).
The first module is the presentation layer (front-end) where we provide a friendly
interface for the user; it consists of different kinds of devices like phones, tablets,
PCs, and so on.
The back-end used to handle operations that are not seen by end-user like:
• Users registration and authentication.
• Registration and authentication of users.
• Prediction of responses for question or image input.
• Conversion from text to speech to make an audio response.
• Handling multiple languages (English, French, Arabic) using translation API.
This module works in the background to better assure the demands of the user:
It manages business logic and data storage, working in collaboration with the model
part.
The model is where we have our models of deep learning.
The database used to store dataset after cleaning and user data.
4 Proposed Architecture
In this section, we discuss each tool that we use in the model and how they work
together for solving this heterogeneous problem.
Fig. 1 System architecture
4.1 Image-Response Model
This model combines between two families of artificial intelligence, natural language
processing, and computer vision. In this model, we have merged the results of two
different models:
– The responses are going to be pre-processed before being indexed and encoded
using our vocabulary built from the pre-processed response tokens of the entire
corpus.
– The images are going to be passed to a pre-trained convolutional neural network
(CNN), so as to extract the features vector of image using the transfer learning
fixed feature extracted method.
– Then, we will pass the vector of response and the image features vector to our
model encoder.
– The image feature vector will pass through a dropout layer to avoid model over-
fitting and a dense layer (fully connected layer) to obtain a 256-dimension vector
output.
– The response vector will go through an embedding layer to make the correlation
between the words, then a dropout layer to avoid model overfitting and an LSTM
layer to obtain an output vector of dimension 256.
– As the two outputs of the last two layers have the same 256 dimension, we will
merge them with an add layer, and the output is the output of our encoder that we
will pass to the decoder.
– The output of the encoder will be passed to the decoder of which we have two
dense layers (fully connected layer). The last dense layer contains a Softmax
activation function to generate the probability distribution for the 2124 words in
the vocabulary we have.
The main goal of this approach is to repeat the vector of image n times, where n
is the length of the response that is fixed for the entire responses corpus; then these
resulting vectors are going to be passed to an encoder and a decoder are going to
generate the response at the end. An encoder is generally used to encode a sequence, in
our case the sequence is the two vectors, first sequence of the response vector and the
image vector, to merge them and pass them to a decoder, in order to generate a prob-
ability distribution. To obtain the next word, we choose the word with a maximum
probability at each time step using greedy search algorithm (Fig. 2).
4.2 Question-Response Model
Here, we tend to use the transformers architecture [6] without making any change in
the global architecture. We make changes only in the hyperparameters until we get
high results and adapt the model to our problematic and dataset (Fig. 3).
5 Implementation
5.1 Dataset
• SciTail [7]: The SciTail dataset is an entailment dataset created from multiple-
choice science exams and web sentences. Each question and the correct answer
choice are converted into an assertive statement to form the hypothesis.
• ARC [8]: A new dataset of 7787 genuine grade-school level, multiple-choice
science questions, assembled to encourage research in advanced question-
answering. The dataset is partitioned into a challenge set and an easy set.
• SciQ [9]: The SciQ dataset contains 13,679 crowd-sourced science exam ques-
tions about Physics, Chemistry, and Biology, among others. The questions are in
multiple-choice format with four answer options each.
Fig. 2 Image-response model
• Question Answer Dataset [10]: There are three question files, one for each year
of students: S08, S09, and S10, as well as 690,000 words worth of cleaned text
from Wikipedia that was used to generate the questions.
• Physical IQA [11]: Physical Interaction QA, a new commonsense QA benchmark
for naive physics reasoning focusing on how we interact with everyday objects in
everyday situations.
• AI2 science [12]: The AI2 Science Questions dataset consists of questions used
in student assessments in the United States across elementary and middle school
grade levels. Each question is 4-way multiple-choice format and may or may not
include a diagram element.
Fig. 3 Question-response
model (transformers
architecture)
• Image Answer Dataset: this is a dataset that we collect using google forms and
it contains about 1200 image answer in different domain (Physics, Biology,
Computer science, …) and level (primary, school, high school and university)
of education.
5.2 Hardware
To train the two models, question-response and image-response models, we used a

machine with the following specification (Table 1).
Table 1 Hardware
Item Value
specification
Processor i7-8550U
RAM 24Go
Storage 1To HDD + 256Go SSD
GPU NVidia GeForce MX130
VRAM 2Go
Operating system Windows 10 Pro 64bit
5.3 Languages and Libraries Used
In this section, we are going to list the languages and libraries we used during the
development of the system. We will decompose it into four parts:
• Front-End
For the front-end, we used ReactJS framework with other front-end tools like
HTML5, Bootstrap, CSS3, and JavaScript.
• Back-End
For the back-end, we used Django REST Framework and Python language. We
used other libraries as gTTs to convert text to speech and translate-api to handle
the translation from language to other.
• Model
In this part, during preprocessing, creating models and training them, we used
Keras, TensorFlow, Pandas, NumPy, Scikit-Learn and we used NLTK library to
evaluate models with BLEU score.
• Database
To store cleaned dataset and data of users, we used MongoDB database.
5.4 Pre-processing
• Image-Response model
Images considered as entries (X) to the model. As we know, any entry to a model
must be in the form of a vector. Thus, we need to convert each image into a vector with
a fixed size, which can then be fed as an input to the neural network. To achieve this,
we choose transfer learning by using pre-trained models like VGG16 (Convolutional
Neural Network) to extract a vector of characteristics for each input image. For feature
extraction, we used the pre-trained model up to part 7 × 7 × 512. If we want to do
a classification task, we will need to use the entire model (Fig. 4).
The model accepts as input an image of size 224 × 224 × 3 and returns as output
a feature vector with a dimension of 7 × 7 × 512. Noting that the responses are
what we want to predict. Thus, during the learning period, the responses will be the
output variables (Y) that the model learns to predict. Nevertheless, the prediction of
Fig. 4 Architecture of the VGG16
the entire response is not done at once. The prediction of the response is done word-
by-word. Therefore, we need to code each word in a fixed size vector. Quite simply,
we will represent each unique word in the vocabulary by an index (integer). We have
a vocabulary with 2124 unique words in the entire corpus and thus each words are
going to be represented by an integer between 1 and 2124. We will take the following
example of the response: “angular is typescript based open source web application
framework." We will build our vocabulary by adding the two words “startseq” and
“endseq” to determine the start and end of the sequence: (Suppose we have already
done the cleaning steps).
vocab = {angular, is, endseq, typescript, based, opensource, web, application,
framework, startseq}
Let us give an index to each word in the vocabulary we get:
angular-1, is-4, endseq-3, typescript-9, based-7, opensource-8, web-10,
application-2, framework-6, startseq-5
Let us take an example where the first characteristics vector of the image Image_1
has a logo of the angular framework and the corresponding response is “startseq
angular is typescript based open source web application framework endseq.”
Keep in mind that the characteristics vector of the image is the input, and the
response is what we need to predict.
Although, we predict the response in the following way:
First, we provide the characteristics vector of the image and the first word as input
and we attempt to predict the second word, i.e.
Input = Image_1 + ‘startseq’; Output = ‘angular’.
We then provide the characteristics vector of the image and the first two words as
input and let us attempt to predict the third word, that is:
Table 2 Predicting a response

i Image feature First part of the response Target word
1 Image_1 startseq angular
2 Image_1 startseq, angular is
3 Image_1 startseq, angular, is typescript
… … … …
N-1 Image_1 startseq angular is typescript based opensource web framework
application
N Image_1 startseq angular is typescript based opensource web endseq
application framework
Input = Image_1 + “startseq angular”; Output = “is”.

And so on…
To sum up, the following table represents the data matrix for an image and its
corresponding response (Table 2).
The model will not accept the English text of the response, but rather the sequence
of indexes (integers) where each index corresponds to a unique word.
First, we will need to make sure that all the sequences are equal in terms of length.
That is why, we need to add zeros at the end of each sequence.
We will add zeros that will make each sequence will have the same length, which
is the maximum length of the largest response in our case 114 words (Table 3).
• Question-Response model
Since each model has an encoder-decoder architecture, they have the same basis
for predicting responses. The only difference is that they have a very different archi-
tecture of the encoder and decoder, and rather than giving an image feature vector in
this case, we have a vector that represents the asked question.
Table 3 Predicting a response using word indexes and sequence padding

i Image feature First part of the response Target word
1 Image_1 [5, 0, 0, … , 0] 1
2 Image_1 [5, 1, 0, 0, … , 0] 4
3 Image_1 [5, 1, 4, 0, 0, … , 0] 9
… … … …
N-1 Image_1 [5, 1, 4, 9, 7, 8, 10, 2, 0, 0, … , 0] 6
N Image_1 [5, 1, 4, 9, 7, 8, 10, 2, 6, 0, 0, … , 0] 3
5.5 Evaluation Metrics
We used two types of evaluation metrics: accuracy and BLEU score.

Accuracy is calculated by the following equation:
Item predicted correctly

Accuracy = (1)
All items predicted
BLEU (bilingual evaluation understudy) is an algorithm for evaluating the quality

of text that has been machine-translated from one natural language to another. Quality
is considered the correspondence between a machine’s output and that of a human.
BLEU’s output is always a number between 0 and 1. This value indicates how similar
the candidate text is to the reference texts, with values closer to 1 representing more
texts that are similar.
BLEU score is calculated by the following equation:
N

BLEU = BP.exp wn log( pn ) (2)
n=1
6 Results and Comparative Study
In this section, the results of our models are tested by using the test data of the datasets
we used during training. It represents a comparative study with different pre-trained
model and hyperparameters and with respect to accuracy and BLEU score.
• Image-Response model
For this model, we use an equivalent hyperparameters, but we modify the pre-trained
model used in each training (Tables 4 and 5).
Table 4 Used
Hyperparameter Value
hyperparameters
Dropout 0.5
Optimizer Adam
Split data 0.4
Batch size 128
Epochs 30
Loss function Categorical_crossentropy
Table 5 Results with

Pre-trained model Accuracy BLEU score
different pre-trained model
VGG16 99.03 86.67
VGG19 99.82 85.24
Xception 99.70 86.99
ResNet50 99.79 91.88
InceptionV3 99.58 85.22
Table 6 Fixed
Hyperparameter Value
hyperparameters
Dropout 0.5
Split data 0.2
Batch size 64
Epochs 150
Loss function SparseCategoricalCrossentropy
As shown in Table 2, the pre-trained model ResNet50 provides the best BLUE
score and the second best accuracy.
While the BLUE score is more valid as an evaluation metric than the accuracy
within the case of text generation, we decided to use the model with ResNet50 for
the deployment.
Question-Response model
For this model, we fixed some hyperparameters and changed the others (Tables 6, 7,
and 8).
Table 7 Results with

Number of heads Number of layer Accuracy BLEU score
RMSprop optimizer and
changing the other model 16 1 37.89 42.88
hyperparameters 2 37.01 27.44
8 1 36.67 41.63
2 36.36 32.13
Table 8 Results with Adam

Number of heads Number of layer Accuracy BLEU score
optimizer and changing the
other model hyperparameters 16 1 35.18 40.55
2 31.47 26.72
8 1 34.32 35.50
2 33.53 26.53
From Tables 4 and 5, we can see that the model with good results is with the
Optimizer RMSprop, 16 heads and 1 layer. Thus, this is the best one for deployment.
7 Conclusion
In this paper, we proposed a chatbot system for education applications, which can
support multiple languages (English, French, and Arabic) and different levels of
education. The concluded results show the best performance of the pre-trained model
ResNet50 compared to other models in the case of Image-Response. Otherwise, and
according to the question-response model, the results conducted to the relevance of
the RMSprop Optimizer compared to Adam Optimizer for well deployment interest.
The chatbot uses an API for translation to handle multiple languages. This is not
favorable, because there is a loss of technical words of a specific domain, to improve
this it is better to use a dataset, for example in Arabic, to train the model with it.
There is many things that we can add to improve this work. In this context, we have
plan in future work to improve the performance of the model in terms of accuracy
and response time and add the functionality of recording a question in audio format.
References
1. Nenkov, N., Dimitrov, G., Dyachenko, Y., Koeva, K.: Artificial intelligence technologies for
personnel learning management systems. In: Eighth International Conference on Intelligent
Systems, 2015
2. Nordhaug, Ø., Imran, A.S., Alawawdeh, Al., Kowalski, S.J.: The forensic challenger. In:
International Conference on Web and Open Access to Learning (ICWOAL), 2015
3. Satu, S., Parvez, H., AI-Mamun, S.: Review of integrated applications with AIML based
chatbot. In: First International Conference on Computer and Information Engineering (ICCIE),
2015
4. Niranjan, M., Saipreethy, M.S., Kumar G.T.: An intelligent question answering conversational
agent using naïve Bayesian classifier. In: International Conference on Technology Enhanced
Education (ICTEE), 2012
5. Farhan, M., Munwar, I.M., Aslam, M., Martinez Enriquez, A.M., Farooq, A., Tanveer, S.,
Mejia P.A.: Automated reply to students’ queries in e-learning environment using Web-BOT.
In: Eleventh Mexican International Conference on Artificial Intelligence: Advances in Artifical
Intelligence and Applications, Special Session—Revised Paper, 2012.
6. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polo-
sukhin, I.: Attention is all you need. In Advances in Neural Information Processing Systems,
pp. 5998–6008 (2017)
7. Khot, T., Sabharwal, A., Clark, P.: SciTaiL: A textual entailment dataset from science question
answering. In: AAAI (2018)
8. Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., Tafjord, O.: Think
you have solved question answering? Try ARC, the AI2 reasoning challenge. arxiv:1803.05457
(2018)
9. Welbl, J., Liu, N.F., Gardner, M.: Crowdsourcing multiple choice science questions. arxiv:
1707.06209 (2017)
10. Smith, N.A., Heilman, M., Hwa, R.: Question generation as a competitive undergraduate course
project. In: Proceedings of the NSF Workshop on the Question Generation Shared Task and
Evaluation Challenge (2008)
11. Bisk, Y., Zellers, R., Le Bras, R., Gao, J., Choi, Y.; PIQA: reasoning about physical
commonsense in natural language. arXiv:1911.11641 (2020)
12. Clark, P.: Elementary school science and math tests as a driver for AI: take the aristo challenge!
In: AAAI (2015)
Smart Information Systems
BPMN to UML Class Diagram Using
QVT
Mohamed Achraf Habri, Redouane Esbai,

and Yasser Lamlili El Mazoui Nadori
Abstract The business process model and notation (BPMN) standard provides nota-
tions in the form of diagrams, which are clearly legible for the needs of internal orga-
nizations and facilitate collaboration between enterprise components. The problem
is how to find a transformation solution between the BPMN and the UML (Unified
modeling language) to benefit from the simplicity of the BPMN on the one hand,
the stability and the widespread UML on the other hand. We work to have a high-
performance solution within the framework of model-driven architecture (MDA) that
saves time, cost and quality of the software. This article represents a transformation
method that allows us to go through the BPMN business process diagram to arrive at
the UML class diagram, finally, to generate the code automatically by using the query
views transformations (QVT) transformation language, and this transformation is a
fruitful combination between the trades side and the computer side.
1 Introduction
The computer modeling language at the base is not understandable by the majority
of the staff of all the specialties in the company that prevents a perfect collaboration
with the information system.
In this context, object management group (OMG) has adopted the business process
model and notation (BPMN), a standard for writing business processes by models,
and it offers a simple and clear graphical notation for the entire body of the company.
From business analysts through IT developers to simple users, this process modeling
standard incorporates new symbols for business process diagrams [1].
The OMG has previously created the model-driven architecture (MDA) standard
[2], and the goal is to migrate legacy information systems that exist in enterprises
to new platforms, adapt them with new IT components, to protect the investment,
maximize flexibility and save time and have a strong, efficient and up-to-date system.
The MDA standard is based on three models, namely the computing-independent
M. A. Habri (B) · R. Esbai · Y. Lamlili El Mazoui Nadori

Mohammed First University, Oujda, Morocco
https://doi.org/10.1007/978-981-16-3637-0_42
594 M. A. Habri et al.
model (CIM), the platform-independent model (PIM) and the platform-specific

model (PSM).
This convergence between these two OMG standards prompted us to think how
to move from the CIM model of the BPMN to a PIM model of another modeling
language, namely UML, for example, from a process diagram to a class diagram
through the use case diagram and also to consolidate the modernization already
carried out by MDA, our ultimate goal is to automatically generate the code from the
class diagram, so our operation is to start with a BPMN business process diagram
and end up with the code. This transformation requires the use of a transformation
language such as query view transformation (QVT), a language that belongs to the
OMG family.
Our paper rethinks the method of transformation presented in Addamssiri et al.
[3] “Generating the PIM Behavioral Model from the CIM using QVT,” Rhazali et al.
[4] “A New Methodology CIM to PIM Transformation Resulting from an Analytical
ĳ
Survey,” Bao et al. [5] “A proposal for a method to translate BPMN model into UML
activity diagram,” [6]. Ondrej Macek et al. “The BPM to UML activity diagram
transformation using XSLT” and Oualid et al. [7] “Transforming CIM to PIM Using
UML.”
This article is organized as following the introduction in the first place, related
work is presented in the second section, and the third contains the background of the
basics used in the work by offering explanations on BPMN, MDA and QVT.
The fourth section is the part that reveals the performed task, exactly starting with
the tracing of the BPMN diagram according to the case study passing through the
transformation rules and the use of the QVT arriving at the desired result.
Finally, we made a conclusion, and at the same time, a vision of the tasks that we
intend to accomplish in the future, which will enhance our work and even take it as
a reference.
2 Related Works
In this section, we present the related works of transformation between BPMN and
UML according MDA.
In [3], the authors present a method founded on a transformation and a modeliza-
tion of the model CIM to the model PIM; practically, in this case, they propose the
transformation of the BPMN (business process model and notation) to the UML, of
the diagram of business processes to the diagram of use case and later arriving at
the sequence diagram based on sequence business vocabulary rules (SBVR), and the
transformation language adopted is the query views transformation (QVT), in which
case the transformation rules respect the transformation rules of QVT.
The authors propose in [4] a methodology that makes it possible to transform the
CIM models into the PIM models into a model-driven architecture (MDA); the idea is
to first create transformable models at the CIM to facilitate the transformation using
the language ATLAS TRANSFORMATION LANGUAGE (ATL); in this case, the
BPMN to UML Class Diagram Using QVT 595
task is to go from the BPMN at the level of the CIM arriving at the UML to model
the PIM, and the example chosen is the booking service.
The authors present in [5] an idea that focuses on the conversion of BPMN into
UML, precisely to go from the business process diagram to the activity diagram by
concocting the context as it is, by specifying rules of direct, complex transformations
are specified according to the need and the case studied.
In [6], the authors propose a new type of diagram transformation in business
process modeling (BPMN) notation into a unified modeling language (UML) activity
diagram. Using the XSLT, a transformational language for transforming XML docu-
ments into other documents, includes a vocabulary, an XML document style and
specifying transformation rules that match the description of the BPMN in XMI and
the description of the UML DA in XMI.
The authors present propose in [8] an idea that is based on the creation of the
components of the platform trades Java ee 6 from the business processes modeled
by the business process model notation (BPMN) 2.0; this creation consists to carry
out three types of transformation, first transform the BPMN diagram into an UML
class diagram, then transform it into a UML model with Java Platform profiles. To
finally get a meta-object facility script (MOFScript) into Java EE components, the
transformation operations are done using query view transformations (QVT) and
MOF script.
3 Background Knowledge
3.1 The Business Processes Models: BPMN
A business process model and notation (BPMN) will provide companies with the
ability to understand their internal procedures via graphical notation and give orga-
nizations the ability to communicate these procedures in a standard way; BPMN was
originally developed by the business process management initiative (BPMI) and has
been maintained by the object management group (OMG) since the merger of these
two Consortia in June 2005 [1, 9].
The current version of BPMN is 2.0.2 and is from 2013. Since July 2013, it has
been an international standard ISO/IEC 195106. In this context, BPMN has been
used because of its clarity and simplicity, especially in the business world.
3.2 The Model-Driven Architecture MDA
The model-driven architecture is a software development approach, proposed and

supported by the OMG.
MDA provides a strong framework that enables system infrastructures to evolve

in response to an endless parade of platforms, while preserving and leveraging
existing technology investments. It enables more efficient, faster and less costly
system integration strategies [10].
MDA models:
• CIM (independent model computation) independent calculation model: describes
the flows and actions on the system.
• PIM (platform-independent model) platform-independent model: describes
business-oriented processes.
• PSM (platform-specific model) platform-dependent model: describes the tech-
nical details related to the implementation for a platform.
The MDA standard ensures the passage between the three types of models so it
helps us a lot for the realization of our work.
3.3 The Query View Transformation: QVT
As part of saving effort and reducing errors, OMG adopted the model transformation
language called QVT in 2005 [1]. Model transformation into MDA is an automated
way of modifying and creating models. A model transformation usually specifies the
acceptable input models and the models that it can output by specifying the meta-
model that a model must conform to. The QVT standard introduces the means to
query, display and transform models based on meta-object facility (MOF) 2.0 [11].
The QVT language is supported by Eclipse, so on the technical side, there will be no
obstacle to ensure the passage between BPMN and UML that are already supported
by Eclipse.
4 Our Proposal Approach
In our paper, we present an approach in which we realize a transformation of the

independent model of computation CIM to a model independent of platforms PIM;
the two models are represented by two different languages of modeling; we begin
to treat the study of cases (online purchase) by a process diagram in the first place
using BPMN; subsequently transforming it into a UML class diagram using the
QVT transformation language, the tool used for the representation of diagrams and
the development of the transformation is the ECLIPSE software, so our task has
allowed us to have a complete modeling cycle starting with the CIM represented by
a process diagram arriving at the code step or even the desired software the PSM.
We focused on an important and relevant transformation; from the BPMN diagram
as an input, to the UML class diagram as an output, we rethink the transformation
methods in the following papers:
Fig. 1 Transformations of the models in MDA by projecting our study of cases
In [3, 4], the authors propose a transformation of BPMN to the use case of UML
arriving at the UML sequence diagram; the authors in [5, 6] describes a transformation
going from BPMN diagram arriving at the UML activity diagram; in [7], the authors
made a CIM to PIM transformation in a single modeling language, namely UML.
We opted for a direct transformation between two different modeling languages;
using the BPMN model as a use case and since there is a very large intersection
between the BPMN model and the UML activity diagram to get directly to the class
diagram from the BPMN instead of the use case diagram of UML or the UML activity
diagram, then we choose the platform and we build the code (Fig. 1).
4.1 BPMN to Class Diagram Transformation QVT Rules
To have a transformation done correctly by the QVT, we have based on transformation

rules that express each element in one language to what concerns it in the other
language. Practically, the class diagram is more sophisticated, especially to design
the case study unlike the BPMN which is very simple and clear so to properly
determine each transformation rule, we have used conditions in the transformation
rules that we will find the example in “4.3 extracted code by QVT” to distinguish
each case to ensure the passage.
4.2 Case Study: BPMN Diagram by Eclipse
We have chosen a simple and clear case study that is the online purchase; a customer
placing orders online on a website, as shown in the BPMN diagram, uses elements
represented by simple symbols like “Start” that means the start of the process, tasks
like simple, sent, received, for example, “the task choose product.” In the middle of
the process, we observe the collaboration between the client and the site that we call
them participants and this is why we call our example a diagram of collaboration. In
the future as a next step, we can choose a more complicated case study to use more
elements thus more BPMN symbols (Fig. 2).
Fig. 2 BPMN diagram for our case study
4.3 Extracted Code by QVT Language Used by Eclipse
In this piece of the QVT code (view transformation language), we first import the
sources of UML and BPMN and how to write the transformation rules in QVT, for
example, the transformation between the task and the class or the operation, and to put
the difference, we used the condition in the ID of the task following the conception
of course [12, 13].
Figure 3 presents the principle part of the M2M transformation with QVT
language.
As we showed in Fig. 3, we distinguished the transformed elements by a modi-
fication of the id of the task; we notice in Fig. 4 “properties of the ‘receive order’
task,” for example, that each element of the BPMN model has an id, a source and a
destination according to its position in the diagram.
4.4 The Result of the Transformation
The execution of the previous QVT code allows a transformation of the source
components of the BPMN toward the destination components of the UML while
respecting the transformation rules mentioned in Table 1 (Fig. 5).
Fig. 3 Result of the transformation by eclipse
Fig. 4 Properties of the “receive order” task
Our result is the class diagram produced by the eclipse papyrus tool based on
the structure of the class diagram of Fig. 4. This is the last step of our work of
which we can observe the transformations mentioned in the table of transformation
rules of Table 1. For example, the participant “buyer” in BPMN is transformed into
a class “buyer” on the UML class diagram, the sequence flow “send_paiment” is
transformed into an association “A_pay”, and the received task “receive_order” into
Table 1 BPMN to class diagram transformation QVT rules

Transformation rules Source model element Target model element
Participant (pool) to class Participant (pool) Class
Lane to class Lane Class
Task “Activity” to class or method (Task) Class or method condition
“with condition”| Receive task
Send task
Sequence flow/message flow to Sequence flow Association (condition)
association
Exclusive gateway (OR XOR …) to Exclusive gateway Association or (generalization)
association
Fig. 5 Result of the

transformation by eclipse
an operation “receive order” in the class “order,” etc. Therefore, we arrive at our
result (Fig. 6).
The transformation between two models is an essential step in a modeling project,

and the MDA approach helps to have a perfect conception especially to properly
represent each model by other modeling languages.
In our article, we have, based on the transformation between the BPMN diagram
and the class diagram of the UML, given the importance of the latter for the high-
lighting of the software, and this transformation can be seen as a bridge between the
simple users and the IT developers.
Fig. 6 Result in class diagram
This result led us to think about finding other results with other transformations
using other modeling languages, (diagrams), other languages of transformations or
to use the ADM architecture-driven modernization standard, practically to carry out
an inverse transformation with which we use in our paper, i.e., to transform the PIM
in our case, it is the class diagram of UML into a business process model, namely
the BPMN which represents the CIM.
References
1. Object Management Group (OMG): https://www.omg.org/bpmn/index.htm

2. MDA en action Ingénierie logicielle guidée par les modèles, Xavier Blanc, édition Eyrolles,
2005
3. Addamssiri, N., Kriouile, A., Balouki Y., Taoufiq, G.: J. Comput. Sci. Inf. Technol. 2(3 & 4)
(2014)
4. Rhazali, Y., Hadi, Y., Mouloudi, A.: Third World Conference on Complex Systems (WCCS),
2015
5. Baĳ o, N.Q.: Vietnamese-German University—BIS (2010)
6. Macek, O., Richta, K.: Proceedings of the Dateso 2009 Annual International Workshop on
DAtabases, TExts, Specifications and Objects, Spindleruv Mlyn, Czech Republic, April 15–17,
2009
7. Oualid, B., Saida, F., Amine, A., Mohamed, B.: Applying a model driven architecture approach:
transforming CIM to PIM using UML. Int. J. Online Biomed. Eng. (iJOE) 14(9), 170–181
(2018)
8. Debnath, N., Martinez, C.A., Zorzan, F., Riesco, D.: IEEE Trans. Ind. Informat., December
2013
9. BPM Modeling FAQ: http://www.bpmodeling.com/faq/.
10. Gotti, S., Mbarki, S.: IFVM bridge: a model driven IFML execution. Int. J. Online Biomed.
Eng. (iJOE) 15(4), 111–126 (2019)
11. Sbai, R., Arrassen, I., Meziane, A., Erramdani, M.: QVT transformation by modeling. (IJACSA)
Int. J. Adv. Comput. Sci. Appl. 2(5), (2011)
12. Argañaraz, M., Funes, A.: An MDA Approach to Business Process Model Trans-formations,
université national de SAN LUIS, SADIO electronic journal of informatics operations research,
janvier, 2010.
13. Esbai, R., Elotmani, F., Belkadi, F.Z.: Model-driven transformations: toward automatic gener-
ation of column-oriented NoSQL databases in Big Data context. Int. J. Online Biomed. Eng.
(iJOE). 15(09), 2019 (2019)
14. Blanc, X.: MDA en action, Ingénierie logicielle guidée par les modèles, EYROLLES. Paris,
1st edition (2005)
15. Abdelhedi, F., Ait Brahim, A., Atigui, F., Zurfluh, G.: Processus de transformation MDA d’un
schéma conceptuel de données en un schéma logique NOSQL. In: Congrès INFORSID, 34ème
édition, Grenoble, 31 mai–3 juin 2016
16. Braun, R., Esswein, W.: Classification of Domain-Specific BPMN Extensions. In: 7th IFIP
Working Conference on The Practice of Enterprise Modeling (PoEM), Nov 2014, Manchester,
United Kingdom
17. OMG: UML Infrastructure Final Adopted Specification, version 2.0, September 2003
18. Radoslava, S.K., Velin, S.K., Nina, S., Petia, K., Nadejda, B.: Design and Analysis of a Rela-
tional Database for Behavioral Experiments Data Processing. International Journal of Online
and Biomedical Engineering (iJOE) 14(02) (2018), 117–132 (2019)
Endorsing Energy Efficiency Through
Accurate Appliance-Level Power
Monitoring, Automation and Data
Visualization
Aya Sayed , Abdullah Alsalemi , Yassine Himeur , Faycal Bensaali ,

and Abbes Amira
Abstract Accrediting the fast economic growth and the enhancement of people’s
live standards, the overall household’s energy consumption is becoming more and
more substantial. Thus, the need of conserving energy is becoming a critical task
to help preserve energy resources and slow down climate change, which in turn,
protects the environment. The development of an Internet of Things (IoT) system
to monitor the consumer’s power consumption behavior and provides energy sav-
ing recommendation at a timely manner can be advantageous to shape the user’s
energy saving habits. In this paper, we integrate the (EM)3 framework into a local
IoT platform named Home-Assistant to help centralize all the connected sensors.
Additionally, two smart plug systems are proposed to be part of the (EM)3 ecosys-
tem. The plugs are employed to collect appliances energy consumption data as well
as having home automation capabilities. Through Home-Assistant User Interface
(UI), end-users can visualize their consumption trends together with ambient envi-
ronmental data. The comparative analysis performed demonstrates great potential
and highlights areas of future work focusing on integrating more sensing systems
into the developed platform for the sake of enriching the existing database.
A. Sayed (B) · A. Alsalemi · Y. Himeur · F. Bensaali

Department of Electrical Engineering, Qatar University, Doha, Qatar
e-mail: as1516645@qu.edu.qa
A. Alsalemi
e-mail: a.alsalemi@qu.edu.qa
Y. Himeur
e-mail: yassine.himeur@qu.edu.qa
F. Bensaali
e-mail: f.bensaali@qu.edu.qa
A. Amira
Institute of Artificial Intelligence, De Montfort University, Leicester, UK
e-mail: abbes.amira@dmu.ac.uk
https://doi.org/10.1007/978-981-16-3637-0_43
604 A. Sayed et al.
1 Introduction
The consumer’s electricity demand is projected to rise to 30% by 2040 (compared to

2017) [1, 2]. This is due to a multitude of factors including the increase of popula-
tion, growth in devices and urbanization, thus, creating a strain over current electricity
infrastructure [3, 4]. Moreover, in order to satisfy this amount of demand, the envi-
ronment is drastically and perpetually affected, causing pollution and global warming
[5, 6]. Not to mention, the energy crisis caused by the overwhelming production of
CO2 which has been cited as a serious issue in the energy market [7, 8].
Further, across the world, COVID-19 social distancing measures have resulted in
a significant segment of the population having to stay at home [9]. In certain regions,
this move has resulted in a temporary transfer of energy demand away from business
districts and urban centers to residential areas, resulting in increased utilization and
higher bills, creating an economic effect on customers. Data from the US, Europe,
and India indicates that during the most restrictive containment stage (covering early-
mid-April), energy sector activity decreased by 15% on average, while residential
electricity usage increased by 5% [9].
In this regard, a range of solutions are presented to converse energy. Internet of
energy (IoE) is being one of the top energy conserving solutions[10–12]. Follow-
ing smart grids, IoE has emerged as a renowned technology in the energy sector.
This technology employs the Internet to collect, arrange, optimize, and control the
networks energy data obtained from different edge devices to develop a distributed
smart energy infrastructure [13–15]. By using sensor and communication technolo-
gies, data is collected and used to predict consumer’s demand and supply to further
polish their energy consumption behavior [12, 16].
Since the consumption is a human product, and through taking advantage of the
aforementioned technology, collected domestic energy consumption data can be used
to dissect the factors responsible of shaping behavior and then construct technology-
based techniques to motive a more energy efficient habits [6, 17, 18]. In this con-
text, because of the widespread use of the IoT devices, it becomes at hand to collect
appliance-level energy consumption data, which helps in accurately analyzing power
consumption traces and effectively detecting abnormal consumption behaviors [19].
This is possible either by installing individual IoT smart sensors/meters [16, 20] or
through deploying the energy disaggregation technology that allows the extraction
of the appliances’ signatures from the main supply only by using artificial intelli-
gence tools [21–23]. By consequence, sub-metering plays an essential role in order
produce more accurate and personalized recommendations, which enable end-users
in performing the best energy-saving decision-making at the right time [24].
1.1 Our Contributions
In this paper, we investigate a new approach of integrating the consumer engagement

toward energy saving behavior by means of exploiting micro-moments and mobile
Endorsing Energy Efficiency Through Accurate Appliance-Level … 605
recommendation systems (EM)3 ecosystem into an IoT platform called Home-

Assistant. Then, we examine commercial smart plug systems. After, we include
a comparative analysis between the (EM)3 smart plugs. Finally, we discuss the faced
challenges while integrating (EM)3 framework with the Home-Assistant platform.
1.2 Paper Organization
The rest of this article is organized as follows. Section 2 highlights related work on
energy efficiency systems . Section 3 gives an overview and the characteristics of
the (EM)3 framework. Section 4 describes the system design along with declaring
the different components it comprises. Section 5 provides an overview of Home-
Assistant platform and different applications. Section 6 goes in details over the (EM)3
smart plugs. Section 7 presents the evaluation of results in real-time and discuss
further limitations. Finally, the paper is concluded with future work in Sect. 8.
2 Related Work
Adopting a more accountable, efficient, and environmentally-aware power consump-

tion is becoming essential to maintain the longevity of the modern electrical grid [25,
26]. According to the authors of [27], smart meters are being implemented in mil-
lions of houses around the world, allowing a bi-directional communication between
the utility companies and the households, and as a result, generating a sheer high
volume of data. By using big data technologies capable of taking this hefty amount
of data and extracting consumer energy usage patterns, predicting demand, and opti-
mizing usage, the utility providers’ capabilities can be revolutionized. Primarily, the
paper presents an intelligent data mining model to investigate, forecast, and visualize
energy time series to reveal numerous energy usage habits.
Owing to the accessibility of electricity and its integral role, it can be quite the chal-
lenge to model the consumer demeanor toward electricity saving [28]. Utilizing smart
home technologies can help bridge the gap between understanding the consumer’s
behavior and integrating that understanding into advancing smart technology. Smart
homes technologies are not only intended to improve the energy usage, but also in fact
provides other benefits such as enhanced lifestyle, security, and safety. Some of the
many technologies used in smart homes to change the consumer’s power consump-
tion behavior includes smart meters, smart plugs, and home automation devices. In
[29], a practical case study of IoT dedicated to home automation is illustrated. The
paper discusses the implementation of three hardware modules, involving a smart
home, smart building, and smart city. Similar to this work, an IoT platform named
Home-Assistant was used with the smart home implementation.
The authors in [30] highlight the significance of collecting and analyzing power
consumption data in buildings, in order to recognize the moments when the occupants
606 A. Sayed et al.
could adjust their energy patterns. Additionally, it has been described that the change
in the residence’s behavior can aid reaching the net-zero energy goal for a building,
creating an opportunity for overall energy saving. This occurs once the total power
consumption of the building is equal to the power generated from the renewable
energy systems operating in that property.
Moreover, in [31], the authors present a consumer-oriented energy monitor built
on the Raspberry Pi that allow smart energy monitoring and services in resi-
dences. Named YoMoPie, the system monitors both active and apparent power, saves
recorded data on the board as well as enabling the energy sensor access via a Python
API. The API allows executing user-designed services to increase energy quality in
buildings and households.
Also, an energy management system (EMS) that regulates HVAC appliances is
developed for fostering energy conservation in houses [32]. Based on the plan-do-
check-act (PDCA) cycle, the EMS has to manage data collection, data processing,
and execution. In this article, we are introducing a real-world EMS deployed in a
real-world building. Powered by a Home-Assistant-based platform, micro-controller
enabled sensing units integrate those units with actuators and the database. The units
are inter-connected via a mesh network.
Another Home-Assistant contribution is the VOLTTRON-enabled EMS for homes
[33]. The intention of this initiative is to develop open-source, easy-to-use EMS
accessible on the market so that anyone can profit from it. The system is a mixture
between a home automation system and a distributed control and sensing software
(i.e., VOLTTRON). As a result, the EMS is to comply with government requirements
such as system monitoring and regulation, smooth connectivity between devices, DR,
intelligence, data processing, and protection.
3 Overview of the (EM)3 Framework
The (EM)3 framework has been developed to promote behavioral improvement for
customers through improving understanding of energy use in domestic households
and buildings. It involves four key steps described as: collection of data (i.e., con-
sumption footprints and atmospheric conditions) from various appliances in institu-
tional buildings [34, 35], processing of consumption footprints in order to abstract
energy micro-moments that help identify anomalies [36, 37], implementation of
consumer preference details to detect correlations between them, and generation of
customized feedback to minimize abnormalities [38, 39] and visualize consumption
footprints [40].
Sensing components play a crucial role in collecting and securely preserving data
in the platform store. By using micro-controller unit (MCU), data from different
sensors is extracted and wirelessly transfered from various cubicles to the (EM)3
storage server housed in the Qatar University (QU) lab. Figure 1 illustrates the overall
design of the (EM)3 energy efficiency framework.
Fig. 1 Overview of (EM)3 energy efficiency framework
In addition, a recommendation engine focused on an algorithm that considers

consumer needs, energy priorities, and availability in order to optimize approval of
the suggested action and improve the performance of the [41] recommended method.
The algorithm is based on the user’s extracted patterns involving the repetitive usage
of devices at some times during the day [20]. It is derived from data on energy usage
and room occupancy reported in the recent background of activities of the customer
(or office) [42].
In summary, the framework is comprised of the following functions:
1. Data Collection: collect data for power usage and environmental monitoring [43].
2. Micro-Moment Classification: analyzes abnormal energy patterns [44].
3. Suggestions and Automations: provides personalized guidance to end-users to
endorse energy-saving management activities coupled with recommended actions
for environmental change [45] .
4. Visualization: upload results, observations, and feedback in an accessible and
engaging manner via a mobile application.
4 Proposed Platform
The system encompasses several sensing components to capture different types of

data such as power consumption, occupancy, light level, and air information (i.e.,
temperature and humidity). The collected data is sent over Wi-Fi to the local server
608 A. Sayed et al.
Fig. 2 System block diagram
to be viewed via a personal computer (PC) or any mobile device. The data is also
stored on the local server so it can later be feed into the micro-moment classifier and
the recommender system. Figure 2 depicts a block diagram of the proposed system
with the predefined stages. The main devices that make up the system are mentioned
next.
4.1 Home-Assistant Platform
A Raspberry Pi 4 B+ runs the latest version of Home-Assistant platform with

ESPHome integration. ESPHome is used to establish a connection between ESP32
and the platform. The library also allows full control over the ESP32 via the Home-
Assistant UI.
4.2 Sensing Components
All the selected sensors with their operating ranges are shown in Table 1 including:
1. The DHT22 is a digital sensor used to measure relative humidity in percent and
temperature in Celsius;
2. The TSL2591 measures the intensity of light in Lux units;
3. The AM312 is a passive infra-red (PIR) motion sensor;
4. The HLW8012 is used to measure the real power of a connected appliance in
Watt;
5. A 5V relay is used as a switch to control the appliances remotely; and
6. The Sonoff POW is a wirelessly connected device used to monitor electricity
usage and can also be utilized as a smart switch.
Table 1 List of components

Component name Operating range
DHT22 Temperature: −40◦ C−480◦ C and Humidity:
0–100%
TSL2591 188µLux-88,000Lux
AM312 3–5 m
HLW8012 Voltage: 90-240VAC, max current: 15A and
max power: 3500W
5V Relay 10A and 250VAC
Sonoff POW Voltage: 90-250VAC, max current: 16A and
max power: 3500W
5 Overview of Home-Assistant
When investing in a home automation ecosystem, there are a lot of factors to consider.
Currently, most of the devices available for purchase are linked to some kind of
cloud service, offering to the end-user a level of convenience [46]. Simultaneously,
numerous problems arise when using such cloud services, an obvious one being,
providing the host company a full access to the user’s personal data. There are several
alternatives to explore when planning to regain control over your smart devices and
achieving local control [47]. Home-Assistant being one of these options. Home-
Assistant is a platform that helps to centralize all the sensors and gadgets available
at home. The platform utilizes message queuing telemetry transport (MQTT) to
communicate with other devices [48]. MQTT is a lightweight messaging protocol
utilized to exchange messages between devices.
5.1 Connected Nodes
Through ESPHome, the connection between Home-Assistant and the different hard-
ware modules is made possible. ESPHome is an add-on available through Home-
Assistant to allow control over a verity of MCU, in this case ESP32. This is feasible
by simply editing a configuration file, and then the nodes can be controlled remotely
through ESPHome dashboard.
Two nodes are created into ESPHome to represent the two physically available
units, the environmental module, and the power module, in the lab as displayed in
Fig. 3. The former setup gathers contextual information on temperature, humidity,
occupancy, and level of light. The latter setup provides data on power usage and
also involves a relay used as a switch over, however, many distance to control any
appliance.
Furthermore, in Fig. 4, the nodes included in Home-Assistant through ESPHome
are indicated. The “powertest” is connected with the power unit, and the “testnode”
610 A. Sayed et al.
(a) Environmental module (b) Power module
Fig. 3 Hardware configuration
Fig. 4 The nodes in ESPHome dashboard interface
is linked with the environmental unit. In the configuration file corresponding to each
node, the sensors comprising each setup are included.
5.2 Data Visualization Options
Through Home-Assistant history component, data is tracked and saved as an SQLite

database. The end-user has access to all stored information. The end-user additionally
can select a specific time to view data at. Power consumption data is presented in
Figs. 5 and 6. The figures display the power consumption patterns of a 60W lamp.
This is also true for other sensors connected with Home-Assistant, such as data
collected from the environmental unit can be viewed in the same manner as Fig. 5
and Fig. 6. Data can be also viewed instantaneously through the overview similar to
what illustrated in Fig. 7.
Fig. 5 Power consumption of a 60W lamp measured by HLW8012 sensor
Fig. 6 Power consumption of a 60W lamp measured by the Sonoff POW device
(a) Mobile UI (b) Web UI
Fig. 7 End-user interface
6 (EM)3 Smart Plug
To determine the power consumption for a given appliance, a proper sensing device
is required. To perform that task, two smart plug alternatives are proposed, which
are the HLW8012 an power sensor and the Sonoff POW device.
Figure 8 provides illustration on how the connection is done between the Sonoff
POW smart plug, the power source, and the load. The power plug should be connected
to the grid to provide the needed power to any appliance connected to the socket.
612 A. Sayed et al.
Fig. 8 Overview of the Sonoff POW smart plug connection
Fig. 9 Overview of the HLW8012 connection
Over the Home-Assistant UI, the end-user can view the power consumption data and
additionally can turn on/off the coupled appliance.
As for the HLW8012 power sensor, the connection with the source and load was
performed similar to the Sonoff POW device; however, in the case of this sensor,
additional pins must be connected with an ESP32 in order to power the sensor and also
acquire the power consumption reading. The sensor is powered using 5V and Ground,
and the other pins are connected with digital pins on the ESP32. The connection of
the HLW8012 sensor is demonstrated in Fig. 9.
To evaluate the new integration and the proposed smart plugs, the performance of
the smart plugs is reported along with a discussion of the current limitations. The
test-bed overview is demonstrated next.
Fig. 10 PX110 watt-meter
7.1 Test-Bed Overview
Figure 2 depicts the test-bed configuration. It includes either the web or the mobile
application UI, local server with Home-Assistant, the various sensors to provide
contextual, and power consumption information. The test-bed was designed mainly
to measure the communication latency and concurrently evaluate the performance
of the smart plug alternatives against a reliable watt-meter. The remaining sensors
were evaluated versus a reference measurement apparatus in [49].
7.2 Comparative Analysis Between Smart Plugs
The reference measurement tool used is the PX110 watt-meter shown in Fig. 10. The
meter has an accuracy of ±1.5% for real power [50]. Table 2 depicts the reference used
and accuracy of the smart plug alternatives connected with different home appliances.
The test was conceded with a duration of 45–60 s depending on the appliance. For
example, the kettle operating time (i.e., time needed to boil water to 100◦ C) was 45
s; therefore, the readings were collected for that time period.
When assessing the efficiency of the smart plugs, it is clear from Table 2, the
overall superiority of the (EM)3 smart plug against the Sonoff POW. This due to the
fact that the (EM)3 smart plug was exactly calibrated to match the reference, while
Sonoff POW unit was not.
614 A. Sayed et al.
Table 2 Smart plugs accuracy results

Appliance (EM)3 Smart plug accuracy Sonoff POW accuracy (%)
(%)
60W lamp 99.28 97.58
Air cooler 98.87 96.51
Monitor 96.55 95.95
Kettle 99.12 96.93
Laptop 96.96 97.44
Overall 98.16 96.88
7.3 Discussion
The deployed Home-Assistant platform displays the environmental data with an

update interval of 1 sec, as for the power plugs the interval is set to be 5 sec. Hence,
all the data can be viewed on the UI in real-time, either from the web interface or
from the open-source smartphone app (on iOS and Android). The performance of
the smart plugs is satisfactory, with the (EM)3 in the lead with an accuracy of 98.16%
and Sonoff POW finishes last with accuracy of 96.88%.
With that being said, the integration is still limited, since the micro-moment clas-
sifier and the recommender system are still not implemented into the platform.
Another constraint is, data collected is solely stored on the Raspberry Pi with no
backup, preventing the risk of losing valuable data. Moreover, the current setup is
only implemented for one room (i.e., research cubicle) and not been verified for
several end-users. However, since the current setup connection is displaying good
stability with a high measured accuracy, it can be extended to support more end-users
and appliances.
8 Conclusion
In this paper, the integration of (EM)3 framework with Home-Assistant is presented.

The framework aims to develop more environmentally conscious habits through
providing energy saving recommendation at the right micro-moment when the end-
user is more receptive to change. The framework is supported by a system of sensors
to collect rich habitual and environmental data. The additional smart plugs proposed
to be added to the system are tested against a trustworthy reference to measure their
performance. High stability and accuracy are reached with future work, focusing
on expanding the size of the sensing system linked with Home-Assistant with the
intention of increasing the database scope and diversity.
Acknowledgements This paper was made possible by National Priorities Research Program
(NPRP) grant No. 10-0130-170288 from the Qatar National Research Fund (a member of Qatar
Foundation). The statements made herein are solely the responsibility of the authors.
References
1. Miglani, A., Kumar, N., Chamola, V., Zeadally, S.: Blockchain for internet of energy manage-
ment: review, solutions, and challenges. Comput. Commun. 151, 395–418 (2020)
2. Himeur, Y., Alsalemi, A., Bensaali, F., Amira, A.: A novel approach for detecting anomalous
energy consumption based on micro-moments and deep neural networks. Cogn. Comput. 12(6),
1381–1401 (2020)
3. Cao, X., Dai, X., Liu, J.: Building energy-consumption status worldwide and the state-of-the-
art technologies for zero-energy buildings during the past decade. Energy Build. 128, 198–213
(2016)
4. Alsalemi, A., Himeur, Y., Bensaali, F., Amira, A., Sardianos, C., Varlamis, I., Dimitrakopoulos,
G.: Achieving domestic energy efficiency using micro-moments and intelligent recommenda-
tions. IEEE Access 8, 15047–15055 (2020)
5. Keho, Y.: What drives energy consumption in developing countries? The experience of selected
African countries. Energy Policy 91, 233–246 (2016)
6. Himeur, Y., Alsalemi, A., Al-Kababji, A., Bensaali, F., Amira, A.: Data fusion strategies for
energy efficiency in buildings: overview, challenges and novel orientations. Inf. Fusion 64,
99–120 (2020)
7. Sardianos, C., Varlamis, I., Chronis, C., Dimitrakopoulos, G., Alsalemi, A., Himeur, Y., Ben-
saali, F., Amira, A.: The emergence of explainability of intelligent systems: Delivering explain-
able and personalized recommendations for energy efficiency. Int. J. Intell. Syst. 36(2), 656–680
(2021)
8. Sardianos, C., Varlamis, I., Chronis, C., Dimitrakopoulos, G., Himeur, Y., Alsalemi, A., Ben-
saali, F., Amira, A.: A model for predicting room occupancy based on motion sensor data. In:
2020 IEEE International Conference on Informatics, IoT, and Enabling Technologies (ICIoT),
IEEE, pp. 394–399 (2020)
9. Snow, S., Bean, R., Glencross, M., Horrocks, N.: Drivers behind residential electricity demand
fluctuations due to covid-19 restrictions. Energies 13(21), 5738 (2020)
10. Alsalemi, A., Himeur, Y., Bensaali, F., Amira, A.: An innovative edge-based internet of energy
solution for promoting energy saving in buildings. Sustain. Cities Soc. 1–20 (2021)
11. Al-Ali, A.-R., Zualkernan, I.A., Rashid, M., Gupta, R., Alikarar, M.: A smart home energy
management system using IoT and big data analytics approach. IEEE Trans. Consum. Electron.
63(4), 426–434 (2017)
12. Shahzad, Y., Javed, H., Farman, H., Ahmad, J., Jan, B., Zubair, M.: Internet of energy: oppor-
tunities, applications, architectures and challenges in smart industries. Comput. Electr. Eng.
86, 106739 (2020)
13. Sardianos, C., Varlamis, I., Chronis, C., Dimitrakopoulos, G., Himeur, Y., Alsalemi, A., Ben-
saali, F., Amira, A.: Data analytics, automations, and micro-moment based recommendations
for energy efficiency. In: 2020 IEEE Sixth International Conference on Big Data Computing
Service and Applications (BigDataService), IEEE, pp. 96–103 (2020)
14. Kabalci, E., Kabalci, Y.: From Smart Grid to Internet of Energy. Academic Press (2020)
15. Alsalemi, A., Himeur, Y., Bensaali, F., Amira, A., Sardianos, C., Chronis, C., Varlamis, I.,
Dimitrakopoulos, G.: A micro-moment system for domestic energy efficiency analysis. IEEE
Syst. J. 1–8 (2020)
16. Himeur, Y., Alsalemi, A., Bensaali, F., Amira, A.: An intelligent non-intrusive load monitoring
scheme based on 2d phase encoding of power signals. Int. J. Intell. Syst. 36(1), 72–93 (2021)
616 A. Sayed et al.
17. Zhang, C.-Y., Yu, B., Wang, J.-W., Wei, Y.-M.: Impact factors of household energy-saving
behavior: an empirical study of Shandong Province in China. J. Cleaner Prod. 185, 285–298
(2018)
18. Azizi, Z.M., Azizi, N.S.M., Abidin, N.Z., Mannakkara, S.: Making sense of energy-saving
behaviour: a theoretical framework on strategies for behaviour change intervention. Procedia
Comput. Sci. 158, 725–734 (2019)
19. Himeur, Y., Elsalemi, A., Bensaali, F., Amira, A.: Smart power consumption abnormality
detection in buildings using micro-moments and improved k-nearest neighbors. Int. J. Intell.
Syst. 1–25 (2021)
20. Elsalemi, A., Himeur, Y., Bensaali, F., Amira, A.: Appliance-level monitoring with micro-
moment smart plugs. In: The Fifth International Conference on Smart City Applications (SCA),
pp. 1–5 (2020)
21. Himeur, Y., Alsalemi, A., Bensaali, F., Amira, A.: Efficient multi-descriptor fusion for non-
intrusive appliance recognition. In: IEEE International Symposium on Circuits and Systems
(ISCAS). IEEE, vol. 2020, 1–5 (2020)
22. Himeur, Y., Alsalemi, A., Bensaali, F., Amira, A.: Improving in-home appliance identification
using fuzzy-neighbors-preserving analysis based qr-decomposition. In: International Congress
on Information and Communication Technology. Springer, Berlin, pp. 303–311 (2020)
23. Himeur, Y., Alsalemi, A., F.ensaali, Amira, A., Sardianos, C., Varlamis, I., Dimitrakopoulos,
G.: On the applicability of 2d local binary patterns for identifying electrical appliances in non-
intrusive load monitoring. In: Proceedings of SAI Intelligent Systems Conference. Springer,
Berlin, pp. 188–205 (2020)
24. Himeur, Y., Alsalemi, A., Bensaali, F., Amira, A.: Building power consumption datasets: survey,
taxonomy and future directions. Energy Build. 227, 110404 (2020)
25. Al-Kababji, A., Alsalemi, A., Himeur, Y., Bensaali, F., Amira, A., Fernandez, R., Fetais, N.:
Energy data visualizations on smartphones for triggering behavioral change: Novel vs. con-
ventional. In : 2nd Global Power, Energy and Communication Conference (GPECOM). IEEE,
vol. 2020, pp. 312–317 (2020)
26. Sardianos, C., Chronis, C., Varlamis, I., Dimitrakopoulos, G., Himeur, Y., Alsalemi, A., Ben-
saali, F., Amira, A.: Real-time personalised energy saving recommendations. In: The 16th
IEEE International Conference on Green Computing and Communications (GreenCom), pp.
1–6 (2020)
27. Singh, S., Yassine, A.: Big data mining of energy time series for behavioral analytics and energy
consumption forecasting. Energies 11(2), 452 (2018)
28. Bhati, A., Hansen, M., Chan, C.M.: Energy conservation through smart homes in a smart city:
a lesson for Singapore households. Energy Policy 104, 230–239 (2017)
29. Debauche, O., Mahmoudi, S., Moussaoui, Y.: Internet of things learning: a practical case for
smart building automation. In: 2020 5th International Conference on Cloud Computing and
Artificial Intelligence: Technologies and Applications (CloudTech). IEEE, pp. 1–8 (2020)
30. Chou, C.-C., Chiang, C.-T., Wu, P.-Y., Chu, C.-P., Lin, C.-Y.: Spatiotemporal analysis and visu-
alization of power consumption data integrated with building information models for energy
savings. Resour. Conserv. Recycl. 123, 219–229 (2017)
31. Klemenjak, C., Jost, S., Elmenreich, W., Yomopie: a user-oriented energy monitor to enhance
energy efficiency in households. In: 2018 IEEE Conference on Technologies for Sustainability
(SusTech), IEEE, pp. 1–7 (2018)
32. Najem, N., Haddou, D.B., Abid, M.R., Darhmaoui, H., Krami, N., Zytoune, O.: Context-
aware wireless sensors for IoT-centeric energy-efficient campuses. In: 2017 IEEE International
Conference on Smart Computing (SMARTCOMP), IEEE, pp. 1–6 (2017)
33. Zandi, H., Kuruganti, T., Fugate, D., Vineyard, E.A.: Volttron-enabled home energy manage-
ment system, Tech. rep., Oak Ridge National Lab.(ORNL), Oak Ridge, TN (United States)
(2019)
34. Himeur, Y., Alsalemi, A., Bensaali, F., Amira, A.: Effective non-intrusive load monitoring of
buildings based on a novel multi-descriptor fusion with dimensionality reduction. Appl. Energy
279, 115872 (2020)
35. Himeur, Y., Elsalemi, A., Bensaali, F., Amira, A.: Recent trends of smart non-intrusive load
monitoring in buildings: a review, open challenges and future directions. Int. J. Intell. Syst.
1–28 (2020)
36. Sardianos, C., Chronis, C., Varlamis, I., Dimitrakopoulos, G., Himeur, Y., Alsalemi, A., Ben-
saali, F., Amira, A.: Smart fusion of sensor data and human feedback for personalised energy-
saving recommendations. Int. J. Intell. Syst. 1–20 (2021)
37. Himeur, Y., Alsalemi, A., Bensaali, F., Amira, A., Varlamis, I., Bravos, G., Sardianos, C.:
Dimitrakopoulos, Techno-economic analysis of building energy efficiency systems based on
behavioral change: a case study of a novel micro-moments based solution. Appl. Energy 1–25
(2021)
38. Himeur, Y., Ghanem, K., Alsalemi, A., Bensaali, F., Amira, A.: Artificial intelligence based
anomaly detection of energy consumption in buildings: a review, current trends and new per-
spectives. Appl. Energy 287, 116601 (2021)
39. Himeur, Y., Elsalemi, A., Bensaali, F., Amira, A.: The emergence of hybrid edge-cloud comput-
ing for energy efficiency in buildings. In: Proceedings of SAI Intelligent Systems Conference,
pp. 1–12 (2021)
40. Al-Kababji, A., Alsalemi, A., Himeur, Y., Bensaali, F., Amira, A., Fernandez, R., Fetais, N.:
Interactive visual analytics for residential energy big data. Inf. Vis. 1–20 (2021)
41. Himeur, Y., Elsalemi, A., Bensaali, F., Amira, A: Appliance identification using a histogram
post-processing of 2d local binary patterns for smart grid applications. In: Proceedings of 25th
International Conference on Pattern Recognition (ICPR), pp. 1–8 (2020)
42. Varlamis, I., Sardianos, C., Dimitrakopoulos, G., Alsalemi, A., Himeur, Y., Bensaali, F., Amira,
A.: Reshaping consumption habits by exploiting energy-related micro-moment recommen-
dations: a case study. In: Communications in Computer and Information Science, Springer
International Publishing, Cham, pp. 1–22 (2020)
43. Alsalemi, A., Ramadan, M., Bensaali, F., Amira, A., Sardianos, C., Varlamis, I., Dimitrakopou-
los, G.: Endorsing domestic energy saving behavior using micro-moment classification. Appl.
Energy 250, 1302–1311 (2019). https://doi.org/10.1016/j.apenergy.2019.05.089 https://doi.
org/10.1016/j.apenergy.2019.05.089
44. Himeur, Y., Alsalemi, A., Bensaali, F., Amira, A.: Robust event-based non-intrusive appliance
recognition using multi-scale wavelet packet tree and ensemble bagging tree. Appl. Energy
267, 114887 (2020)
45. Sardianos, C., Varlamis, I., Dimitrakopoulos, G., Anagnostopoulos, D., Alsalemi, A., Bensaali,
F., Himeur, Y., Amira, A.: Rehab-c: recommendations for energy habits change. Future Gener.
Comput. Syst. 112, 394–407 (2020)
46. Himeur, Y., Alsalemi, A., Bensaali, F., Amira, , A., Sardianos, C., Dimitrakopoulos, G., Var-
lamis, I.: A survey of recommender systems for energy efficiency in buildings: Principles,
challenges and prospects. Inf. Fusion 1–33 (2020)
47. Alsalemi, A., Al-Kababji, A., Himeur, Y., Bensaali, F., Amira, A.: Cloud energy micro-moment
data classification: a platform study. In: 2020 IEEE/ACM 13th International Conference on
Utility and Cloud Computing (UCC), IEEE, pp. 420–425 (2020)
48. Home Assistant. Available online https://www.home-assistant.io/. Accessed 30-12-2020
49. Alsalemi, A., Ramadan, M., Bensaali, F., Amira, A., Sardianos, C., Varlamis, I., Dimitrakopou-
los, G.: Boosting domestic energy efficiency through accurate consumption data collection. In:
IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing,
Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People
and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI). IEEE,
pp. 1468–1472 (2019)
50. TRMS three- and single phase digital wattmeters. Available online: http://www.farnell.com/
datasheets/3649.pdf. Accessed 30-12-2020
Towards a Smart City Approach:
A Comparative Study
Zineb Korachi and Bouchaib Bounabat
Abstract There are conflicts and ambiguity regarding smart city strategies. Some
works present smart city processes with vague and inconsistent steps. While others
propose smart city elements and dimensions, rather than providing a clear and holistic
approach. The most of smart city strategy works overlap one another, which creates
ambiguity for smart city leaders. To fill this gap and reduce this ambiguity, the current
paper presents and describes the main components of a smart city strategy framework,
which are: strategic vision, action plan, and management strategy. To evaluate the
relevance of these elements, the present paper conducts a comparative study.
1 Introduction
To be a smart city, it is first necessary to have specific goals and strategies and to be
committed to fulfilling those goals [1, 2]. Despite the importance of smart cities, still
few studies investigate how developing a clear and consistent smart city approach to
help cities become smarter [3]. There is no agreement about the smart city definition,
domains, and indicators [4]. In addition to the challenges surrounding the smart city
definition and indicators, there is also an ambiguity regarding the definition of the
smart city strategy [5].
Many of the smart city strategy efforts are fragmented, stressing only some aspects
of the smart city, rather than approaching them in an integrated way [6–8]. Some of
these works treat city objectives and indicators, whereas others emphasize solution
architectures and technical details [9]. This enhances the misunderstanding and ambi-
guity regarding the smart city strategy, rather than resolving it and enabling action-
able smart city planning [8]. Evidence of this ambiguity is presented by recent works
Z. Korachi (B) · B. Bounabat

ADMIR Laboratory, Rabat Information Technology Center, ALQUALSADI Team,
ENSIAS/Mohamed V University in Rabat, Rabat, Morocco
e-mail: zineb.korachi@um5s.net.ma
B. Bounabat
e-mail: b.bounabat@um5s.net.ma
https://doi.org/10.1007/978-981-16-3637-0_44
620 Z. Korachi and B. Bounabat
illustrating the existence of overlapping smart city development processes, which

generate conflict and ambiguity regarding smart city development paths [7, 10–13].
Even though there is no ‘one-best-way’ of designing a smart city strategy, there is
a need for an integrated and comprehensive framework that highlights the different
components of a smart city and which provides guidelines for smart city strategy
design [5]. The current work fills this gap by presenting a comparative study illus-
trating and describing the main smart city strategy components which are: smart city
strategic vision, action plan, and management strategy.
The paper is structured as follows: the next section presents literature review,
followed by methodology, problem of statement, results, discussion, and conclusion.
2 Literature Review
This section analyzes existing smart city frameworks and strategies. It shows the
relevant blocks, strengths, and weaknesses of these frameworks. This analysis helps
to identify and factorize the relevant components of these frameworks and models
in a unified solution.
• Agbali et al. (2019) present a comparative analysis between three cities: Boston,
Manchester, and San Diego. The comparison is conducted using a smart city
ecosystem composed of the following domains: Smart infrastructure, smart
institution, smart people [3].
• Oliveira et al. (2020) highlight the concepts surrounding cities as follow: mobility,
health care, governance industry, and services [14]. These concepts are important
elements that should be incorporated into the smart city strategy. Oliveira et al.
(2020) cite that the citizen is the core center of the smart city system.
• Darmawan et al. (2019) identify crucial factors in the process of readiness and
application of the smart city concept to regional governments in Indonesia [15].
These factors are presented as follow: Perceived of use, service quality, system
quality, information quality, and intention use [15]. These factors provide city
leaders with characteristics that can help to identify a successful strategy.
• Dabeedooal et al. (2019) propose the smart tourism dimensions as follow: smart
infrastructure, smart business, governance, and urban metabolism. It presents
a framework for smart tourism composed of the following components: Tech-
nology applications, leadership, human capital, entrepreneurs, innovation, social
capital, tourism experience, tourism competitiveness [16]. The smart tourism
framework should achieve the following characteristics: Attractions, accessibility,
amenities, available packages, activities, ancillary services [16]. This framework
concentrates only on some elements that should be highlighted by the smart
tourism strategy, rather than providing a clear and integrative smart tourism
implementation process.
• Gokozan et al. (2017) present the following smart city components and their
definitions: smart care, smart energy, smart society, smart office, smart mobility,
Towards a Smart City Approach: A Comparative Study 621
and smart space [2]. It defines the smart city management center concept, which
works like the brain and central nervous system of smart cities, which connect
and integrate information and processes [2].
• Rotuna et al. (2019) show that the blockchain is a relevant solution for a wide
range of challenges faced by the smart city, but the implementation depends on the
city characteristics and context [17]. A blockchain-based smart city infrastructure
has the advantages of increased efficiency due to the automated interactions with
its citizens, optimized distribution of resources, and fraud reduction [17].
• Saba et al. (2019) provide smart city definitions, characteristics (Sustainability,
smartness, life quality, and urbanization), trends and needs, the architecture of the
smart city (sensing layer, transmission layer, data management layer, application
layer), smart city main components (territory, infrastructure, people, government),
and smart city pillars (sustainability, technology, flexibility, citizen involvement).
This paper presents that open data is a crucial element in the development of smart
cities [18].
• Einola et al. (2019) present the advantages of the open strategy in a smart city
[19]. An open strategy that includes the participation of external and internal
stakeholders has many undeniable benefits: increasing collective commitment
and, through commitment, enabling more effective strategic actions and joint
sensemaking [19]. Open strategizing can improve creativity by capturing more
diverse views [19]. Einola et al. (2019) provide a smart city process that involves
citizens in the definition of the strategy through crowdsourcing [19].
• Afonso et al. (2015) propose a Brazilian smart city maturity model composed of
five levels namely: Simplified, managed, applied, measured, and turned [4]. The
model is based on the following domains: Water, education, energy, governance,
housing, environment, health, security, technology, transport [4].
• Aljowder et al. (2019) provide a systematic literature review on maturity models
that assess the level of maturity for smart city projects [20]. It provides an analysis
and classification of these models based on their components and perspectives.
This study can help to identify the list of elements that should be highlighted
through the smart city strategy (ex: education, health, energy…).
• Komninos et al. (2015) present an overall ontology for the smart city and defines
its building blocks [21]. They defined a set of smart city indicators, namely: appli-
cation ontology size, the maximal length of nodes, the number of the object, data
properties and super classes, the position of the ontology within the overall smart
city ontology, the type of the digital space, the knowledge generation processes,
the highest level of innovation to be achieved [21]. The definition of the smart
city ontology simplifies the smart city strategy definition and implementation.
• Kuyper (2016) defines the concept of the smart city strategy [5]. This theoret-
ical debate was then applied to two practical examples of smart cities: Barcelona
and Amsterdam, to present how have they approached the smart city implemen-
tation [5]. The comparison between Barcelona and Amsterdam is done based
on the following characteristics: Direction of strategy, main focus, planning
horizon, strategic choices, SMART framework, smart city reference model, citizen
empowerment and inclusion, smart city pilot projects & upscaling [5].
Smart city strategy components provided by the above studies overlap one another,
this creates ambiguity and misunderstanding regarding how to implement smart
cities. Hence the need for a clear and consistent smart city approach. Few frameworks
are addressing clear guidelines for smart city strategy development. The majority
focus on providing smart city components, and dimensions, rather than integrating
them in a clear and coherent smart city strategy to frame and facilitate smart city
implementation.
3 Methodology
This work aims to identify a clear and consistent smart city approach, by selecting
recent smart city frameworks and presents them comparatively using the smart city
approach proposed by Korachi and Bounabat in [1] and [22]. This approach consists
of three main processes: strategic vision definition, action plan elaboration, and
management strategy definition [1]. Figure 1 illustrates the process of how developing
a smart city vision. Figure 2 presents the process for the action plan definition.
Figure 3 illustrates the management strategy definition process. The components of
these processes are presented in Table 1. This current paper compares these processes
with recent smart city frameworks to evaluate their originality and completeness.
Fig. 1 Strategic vision process [22]
Fig. 2 Action plan process [22]

Fig. 3 Management strategy process [22]
Table 1 Korachi and Bounabat smart city approach components [1, 22, 23]
Processes Activities [1, 22]
Strategic vision (1) Why the city need smart transformation
(2) Gather information on the internal and external environment of the
city
(3) Identify stakeholders and their engagements
(4) Identify and describe strategic goals
(5) Identify challenges
(6) List smart city trends
(7) Lead benchmarking
(8) Determine city strengths and weaknesses
(9) Identify the main components that will be highlighted through the
Transformation Strategy
(10) Define desired outcomes, changes, and impact of the smart
transformation (SV10)
(11) Identify the required components and resources for achieving
desired goals and outcomes
(12) Identify gaps
(13) Identify opportunities
Action plan (1) Determine existing success potentials
(2) Establish a list of city departments and business processes
(3) Identify the engagements (activities) of each city department for
achieving strategic goals
(4) Establish the list of activities, define their input and output, and
determine dependencies between activities
(5) Identify the programs list
(6) Identify the projects list
(7) Identify required resources to achieve smart city projects
(8) Elaborate a timesheet for the smart strategy implementation
Management strategy (1) Define appropriate KPIs
(2) Evaluate the digital transformation maturity level
(3) Smart city dashboard
(4) Control the smart city evolution and rank
4 Problem of Statement
This section summarizes the above studies by comparing them according to the
components illustrated in Table 1. The comparison analysis is illustrated in Table
2. It illustrates that the cited works focus mainly on describing the requirements,
components, technologies, and dimensions of smart cities (Strategic Vision 9),
without proposing a comprehensive approach providing mechanisms to simplify
their development.
Table 2 Comparative analysis of smart city works
624
Strategy Smart City

Source Strategic Strategic Strategic Strategic Strategic Strategic Strategic Strategic Strategic Strategic Strategic Strategic Strategic
Vision 1 Vision 2 Vision 3 Vision 4 Vision 5 Vision 6 Vision 7 Vision 8 Vision 9 Vision Vision Vision Vision
10 11 12 13
[3] ˛ ˛
[14] ˛ ˛
[15] ˛
[16] ˛
[2] ˛
[17] ˛ ˛
[18] ˛ ˛ ˛ ˛
[19] ˛ ˛
[4] ˛
[20] ˛
[21] ˛ ˛ ˛
[5] ˛ ˛ ˛ ˛ ˛
(continued)
Z. Korachi and B. Bounabat
Table 2 (continued)
Strategy Smart City
Source Action Action Action Action Action Action Action Action Management Management Management Management
Plan1 Plan 2 Plan 3 Plan 4 Plan 5 Plan 6 Plan 7 Plan 8 Strategy 1 Strategy 2 Strategy 3 Strategy 4
[3]
[14]
[15] ˛
[16]
[2]
[17]
[18]
[19] ˛
[4] ˛
[20] ˛
Towards a Smart City Approach: A Comparative Study
[21] ˛ ˛
[5]
625
The lack of a comprehensive and practical smart city approach, integrating busi-
ness and technological processes and ensuring their strategic alignment, confirms the
need for a new framework, which organizes all the smart city aspects and concerns
in a unified solution.
5 Results
The smart city strategy is not a global approach, which must be implemented in the
same way and with the same processes all over the world [24]. However, there are
some standard aspects that are common to smart cities all over the world, and this
section aims to explore them.
5.1 Smart City Objectives
Several works describe the smart city strategic objectives and this section factors
them into the list below [6, 24–26]:
• Natural resources management and protection
• Creating a competitive economy
• Networking development
• Digitization of public and private services
• The central role of technology
• Advancing human and social capital
• Reducing CO2 emissions
• Creation of an urban infrastructure that meets the expectations and needs of current
and future generations
• Improving the quality of life
• Development of a clean and sustainable environment.
5.2 Smart City Challenges
Cloud computing, the Internet of Things, open data, semantic web, and future Internet
technologies are the leading technologies for the development of the smart city.
All these technologies have their challenges and limitations and they together form
a complex system that increases the challenges [25] that can be grouped into the
following two categories:
• Managerial challenges [26]: managerial attitudes and behavior, resistance to
change, a blurred vision of IT management.
• Technical challenges [27, 28]: interoperability, processing of huge amounts of

data in real-time, mashup of heterogeneous resources, high-energy consumption,
privacy, and cyber-attacks.
5.3 Smart City Main Components
The current section identifies the main components contributing to the design and
implementation of the smart city strategy. Previous studies deal with these compo-
nents in an unstructured and unorganized way which makes it difficult to iden-
tify and understand them. This section aims to structure them according to the
following categories: requirements, dimensions, risks, policy, recommendations, and
architecture.
5.3.1 Strategic requirements of the smart city
The smart cities requirements can be divided into the following two categories:
• Managerial requirements [26, 29]: vision, strategy, leadership, collaboration,
management, organization, governance, political context, people and communi-
ties, and culture.
• Technical requirements [24–30]: technology, sustainable infrastructure, envi-
ronment, integration of services and applications, broadband, communication
channels, and sensors.
5.3.2 Smart city dimensions
Different overlapping smart city dimensions are presented in the literature. This
section aims to simplify their understanding by grouping them in Table 3.
According to Table 3, each study defines the concept of the “smart city dimension”
in its own way. For example, the dimensions proposed by [31] and [27] represent the
areas of the city in need of digital transformation, such as the economy, governance,
citizens, life, transport, and the environment. While [37–39] suggest a set of dimen-
sions which include city domains and the techniques required for their development
such as technology, infrastructure, and innovation.
5.3.3 Smart City Risks
Smart city projects have many advantages, but the security risks for data and services
cannot be avoided. To address this, Guntur & Ibrahim (2019) propose a cybersecurity
strategy based on three dimensions: citizens, technologies, and institutions [40].
628
Table 3 Smart city dimensions

Source Economy Governance People Life Mobility Environment Stakeholders Technologies Infrastructure Strategy legal Data Sustainability Innovation Context Security
[31] ˛ ˛ ˛ ˛ ˛ ˛
[32] ˛ ˛ ˛ ˛
[33] ˛ ˛ ˛ ˛
[27] ˛ ˛ ˛ ˛ ˛ ˛
[34] ˛ ˛ ˛ ˛ ˛ ˛
[35] ˛ ˛ ˛ ˛
[36] ˛ ˛ ˛ ˛
[37] ˛ ˛ ˛ ˛ ˛
[30] ˛ ˛ ˛
Z. Korachi and B. Bounabat
Fig. 4 Smart City layers [39]
5.3.4 Smart City Implementation Policy
The various existing smart city implementation policies are [6]:

• National versus local strategies
• Stage of development: new projects or continuation of existing projects
• Hardware-oriented strategies or software-oriented strategies
• Economically or geographically oriented strategies.
5.3.5 Smart city architecture
Smart city architecture layers [33, 39] are data collection, data processing, data
analysis and integration, production, and use of information. They interact with each
other as illustrated in Fig. 4.
5.3.6 Recommendations
Among the recommendations that can improve the development of smart cities are
the following [6]:
– The study of what is already in place and how it can be improved
– Prioritization of areas that need to be improved urgently
– Selectivity, synergies, and prioritization are three standard core processes in the
planning of a smart city.
– Stakeholders engagement.
5.4 Smart City Stakeholders
Smart city stakeholders involve those responsible for the development of the smart
city and those affected by its outcome [27]. They include citizens, educational
institutions, health care, public safety providers, and government organizations.
5.5 Smart City Resources
Smart city development requires a set of resources such as [25, 27]:

• Network infrastructure: the smart city needs a highly interconnected, cost-
effective, energy-efficient, and reliable infrastructure.
• Technology services: consisting of the acquisition of applications, services, data
management systems, IoT platforms, and data sources.
• Devices: sensors, smartphones, RFID, and NFC.
• Financial resources: the smart city project requires the acquisition of a powerful
technological infrastructure, millions of sensors, thousands of network equipment,
and computing devices are needed. This means the need for a huge budget.
6 Conclusion
There is still a conflict among the smart city definitions and frameworks. Various
studies in the literature addressing smart city frameworks and strategies. The literature
analysis shows that these studies overlap one another and create ambiguity around
smart city strategies. To fill this gap, the present paper conducts a comparative analysis
that allows to identify the smart city main components and processes. This study aims
to reduce the ambiguity and misunderstanding that surrounds smart city approaches
and frameworks, by presenting a relevant smart city approach composed of three
processes namely: strategic vision definition, action plan development, management
strategy identification. The approach was evaluated comparatively with recent smart
city approaches. The result of the comparative analysis shows that the approach is
holistic and original. Future studies can investigate and analyze the elements of the
proposed smart city approach to collect more information about them from literature
or smart city cases.
References
1. Korachi, Z., Bounabat, B.: Towards a platform for defining and evaluating digital strategies
for building smart cities. 2019 3rd International Conference On Smart Grid And Smart Cities
(ICSGSC) (2019). https://doi.org/10.1109/icsgsc.2019.00-22
2. Gokozan, H., Tastan, M., Sari, A.: Smart cities and management strategies. Chapter 8 In Book:
2017 Socio-Economic Strategies. ISBN: 978–3–330–06982–4 (2017)
3. Agbali, M., Trillo, C., Ibrahim, I., Arayici, Y., Fernando, T.: Are smart innovation ecosystems
really seeking to meet citizens’ needs? insights from the stakeholders’ vision on smart city
strategy implementation. Smart Cities 2(2), 307–327 (2019). https://doi.org/10.3390/smartciti
es2020019
4. Afonso, R.A., dos Santos Brito, K., do Nascimento, C.H., Garcia, V.C., Álvaro, A: Brazilian
smart cities. Proceedings of the 16th Annual International Conference on Digital Government
Research—Dg.o ’15 (2015). https://doi.org/10.1145/2757401.2757426.
5. Kuyper, T.: Smart City Strategy & Upscaling: Comparing Barcelona and Amsterdam. https://
doi.org/10.13140/RG.2.2.24999.14242. Master Thesis, Msc. IT & Strategic Management
(2016)
6. Angelidou, M.: Smart city policies: a spatial approach. Cities 41(2014), S3–S11 (2014). https://
doi.org/10.1016/j.cities.2014.06.007
7. Mora, L., Deakin, M., Aina, Y., Appio, F.: Smart City Development: ICT Innovation for Urban
Sustainability. Encyclopedia of the UN Sustainable Development Goals, pp. 589–605 (2020).
https://doi.org/10.1007/978-3-319-95717-3_27
8. Angelidou, M.: Smart cities: a conjuncture of four forces. www.elsevier.com/locate/cities.
Cities 47 (2015) 95–106 (2015a). http://dx.doi.org/https://doi.org/10.1016/j.cities.2015.05.004
9. Bastidas, V., Bezbradica, M., Helfert, M.: Cities as enterprises: a comparison of smart city
frameworks based on enterprise architecture requirements. Smart Cities, pp. 20–28 (2017).
https://doi.org/10.1007/978-3-319-59513-9_3
10. Korachi, Z., Bounabat, B. (2018). Data driven maturity model for assessing smart cities. ACM
International Conference Proceeding Series, pp. 140–147, 2nd International Conference on
Smart Digital Environment, ICSDE’18, October 18–20, 2018, Rabat, Morocco; © 2018 Asso-
ciation for Computing Machinery; ACM ISBN 978–1–4503–6507–9/18/10; ENSIAS Rabat;
Morocco. https://doi.org/10.1145/3289100.3289123
11. Korachi, Z., Bounabat, B.: Integrated methodological framework for digital transformation
strategy building (IMFDS). Int. J. Adv. Comput. Sci. Appl. 10(12) (2019). https://doi.org/10.
14569/ijacsa.2019.0101234
12. Korachi, Z., Bounabat, B.: Towards a maturity model for digital strategy assessment. Adv.
Intell. Syst. Comput. 1105, 456–470 (2020). Springer, Cham. https://doi.org/10.1007/978-3-
030-36674-2_47
13. Korachi, Z., Bounabat, B.: Towards a frame of reference for smart city strategy development
and governance. J. Comput. Sci. 16(10), 1451–1464 (2020). https://doi.org/10.3844/jcssp.2020.
1451.1464
14. Oliveira, T., Oliver, M., Ramalhinho, H.: Challenges for connecting citizens and smart cities:
ICT. E-Governance Blockchain. Sustain. 12(7), 2926 (2020). https://doi.org/10.3390/su1207
2926
15. Darmawan, A., Siahaan, D., Susanto, T., Hoiriyah, Umam, B.: Identifying success factors in
smart city readiness using a structure equation modelling approach. 2019 International Confer-
ence On Computer Science, Information Technology, And Electrical Engineering (ICOMITEE)
(2019). https://doi.org/10.1109/icomitee.2019.8921312.
16. Dabeedooal, Y., Dindoyal, V., Allam, Z., Jones, D.: Smart tourism as a pillar for sustainable
urban development: an alternate smart city strategy from mauritius. Smart Cities 2(2), 153–162
(2019). https://doi.org/10.3390/smartcities2020011
17. Rotuna, C., Gheorghita, A., Zamfiroiu, A., Smada, D.: Smart city ecosystem using blockchain
technology. Informatica Economica, 23(4/2019), 41–50 (2019). https://doi.org/10.12948/iss
n14531305/23.4.2019.04.
18. Saba, D., Sahli, Y., Berbaoui, B., Maouedj, R.: Towards smart cities: challenges, components,
and architectures. Toward Social Internet of Things (Siot): Enabling Technologies, Archi-
tectures And Applications, pp. 249–286 (2019). https://doi.org/10.1007/978-3-030-24513-
9_15
19. Einola, S., Kohtamäki, M., Hietikko, H.: Open Strategy in a Smart City. Tchnology Innovation
Management Review, September 2019 (Volume 9, Issue 9) (2019)
20. Aljowder, T., Ali, M., Kurnia, S.: Systematic literature review of the smart city maturity model.
2019 International Conference on Innovation and Intelligence for Informatics, Computing, and
Technologies (3ICT) (2019). https://doi.org/10.1109/3ict.2019.8910321
21. Komninos, N., Bratsas, C., Kakderi, C., & Tsarchopoulos, P. (2015). Smart City Ontologies:
Improving the effectiveness of smart city applications. Journal Of Smart Cities, 1(1). https://
doi.org/10.18063/jsc.2015.01.001.
22. Korachi, Z., Bounabat, B.: Integrated methodological framework for smart city development.
Proceedings of the International Conferences ICT, Society, and Human Beings 2019; Connected
Smart Cities 2019; and Web Based Communities and Social Media (2019). https://doi.org/10.
33965/csc2019_201908l030
23. Korachi, Z., Bounabat, B.: Towards a frame of reference for smart city strategy development
and governance. J. Comput. Sci. 16(10), 1451–1464 (2020). https://doi.org/10.3844/jcssp.2020.
1451.1464
24. Dameri, R.P., Benevolo, C., Veglianti, E., Li, Y.: Understanding smart cities as a glocal strategy:
a comparison between Italy and China. Technol. Forecast. Soc. Chang. (2018). https://doi.org/
10.1016/j.techfore.2018.07.025
25. Kadhim, W.: Case study of Dubai as a Smart City. Int. J. Comput. Appl. 178(40), 35–37 (2019).
https://doi.org/10.5120/ijca2019919291
26. Chourabi, H., Nam, T., Walker, S., Gil-Garcia, J. R., Mellouli, S., Nahon, K., Scholl, H.J.,
et al.: Understanding smart cities: an integrative framework. 2012 45th Hawaii International
Conference on System Sciences (2012). https://doi.org/10.1109/hicss.2012.615.
27. Petrolo, R., Loscrì, V., Mitton, N.: Towards a smart city based on cloud of things, a survey
on the smart city vision and paradigms. Trans. Emerging Telecommun. Technol. 28(1), e2931
(2015). https://doi.org/10.1002/ett.2931
28. Khatoun, R., Zeadally, S.: Smart cities: concepts, architectures, research oppor-tunities.
Commun. ACM 59(8), 46–57 (2016). https://doi.org/10.1145/2858789
29. Allam, Z., Newman, P.: Redefining the smart city: culture, metabolism and governance. Smart
Cities 1(1), 4–25 (2018). https://doi.org/10.3390/smartcities1010002
30. Kesswani, N., Kumar, S.: The smart-X model for smart cities. 2018 IEEE 42nd Annual
Computer Software and Applications Conference (COMPSAC) (2018). https://doi.org/10.
1109/compsac.2018.00112.
31. Giffinger, R.,Fertner, C.,Kramar, H., Kalasek, R., Pichler-Milanović, N., Meijers, E.: Smart
cities—ranking of European medium-sized cities. [Online] Centre of Regional Science (SRF),
Vienna University of Technology in October 2007 (2007). Available at: http://www.smart-cit
ies.eu/download/smart_cities_final_report.pdf [Accessed 17 Jun 2019]
32. Dameri, R.P., Rosenthal-Sabroux, C.: Smart city and value creation. Progress IS, pp. 1–12
(2014). https://doi.org/10.1007/978-3-319-06160-3_1
33. Asri, N.A.M., Ibrahim, R., Jamel, S.: Designing a model for smart city through digital transfor-
mation. Int. J. Adv. Trends Comput. Sci. Eng. (2019). https://doi.org/10.30534/ijatcse/2019/
6281.32019
34. Joshi, S., Saxena, S., Godbole, T., Shreya: Developing smart cities: an integrated framework.
Procedia Comput. Sci. 93, 902–909 (2016). https://doi.org/10.1016/j.procs.2016.07.258
35. Hämäläinen, M.: A framework for a smart city design: digital transformation in the Helsinki
Smart City. Contribut. Manage. Sci. 63–86,(2019). https://doi.org/10.1007/978-3-030-236
04-5_5
36. Haller, S., Neuroni, A., Fraefel, M., Sakamura, K.: Perspectives on smart cities strategies.
Proceedings of the 19th Annual International Conference on Digital Government Research
Governance in the Data Age—Dgo ’18 (2018). https://doi.org/10.1145/3209281.3209310
37. Maestre-Gongora, G., Bernal, W.: Conceptual model of information technology management
for smart cities. J. Glob. Inf. Manag. 27(2), 159–175 (2019). https://doi.org/10.4018/jgim.201
9040109
38. Hämäläinen, M., Tyrväinen, P.: Improving smart city design: a conceptual model for governing
complex smart city ecosystems. 31st Bled Econference: Digital Transformation: Meeting the
Challenges (2018). https://doi.org/10.18690/978-961-286-170-4.17
39. Kumar, M.: Building Agile Data Driven Smart Cities (IDC: October 2015), White Paper,
(Sponsored by EMC) (2015). http://docplayer.net/8696930-Building-agile-data-driven-smart-
cities.html [Last Access 21/06/2020]
40. Guntur Alam, R., Ibrahim, H.: Cybersecurity strategy for smart city implementation. The Inter-
national Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences,
vol. XLII-4/W17, 2019. 4th International Conference on Smart Data and Smart Cities, 1–3
October 2019, Kuala Lumpur, Malaysia (2019). https://doi.org/10.5194/isprs-archives-XLII-
4-W17-3-2019
Hyperspectral Data Preprocessing
of the Northwestern Algeria Region
Zoulikha Mehalli, Ehlem Zigh, Abdelhamid Loukil, and Adda Ali Pacha
Abstract Hyperspectral images provide a rich source of information and allow us to

broaden the field of applications like material identification, agriculture, and environ-
mental studies. However, these images commonly suffer from atmospheric effects,
sensor noise, etc. Preprocessing is a crucial task to reduce/remove these effects
for better extraction of the desired information. The present paper aims to apply
some preprocessing techniques on Djebel Meni zone, situated in the Northwestern
of Algeria, to prepare reliable images for the next processing stages. The database
used in this work is the Hyperion level L1R dataset. The preprocessing approach
adopted in this work involves: bad band removal, i.e., removing the bands with no
information, radiometric calibration, and atmospheric correction in order to handle
the atmospheric effects, correcting the continuous vertical stripes and the abnormal
pixels in bands. Also, we have applied dimensionality and noise reduction to deal
with the huge volume of spectral data and the noise present in the Hyperion hyper-
spectral dataset. These preprocessing steps are an indispensable prerequisite for any
further processing and accurate interpretation of spectra of different surface objects.
Also this study recommends that QUAC atmospheric correction has a capability to
compensate the atmospheric effects than IARR method for mineralogy studies
1 Introduction
Remote sensing measurement of the earth’s surface is influenced by the atmosphere

due to many causes like gases, sensor noise, and other illumination effects. For that,
Z. Mehalli (B) · A. Ali Pacha

Coding and Information Security Laboratory, University of Science and Technology of Oran -
Mohamed Boudiaf, Oran, Algeria
E. Zigh
LaRATIC laboratory. National Institute of Telecommunications and ICT of Oran, Oran, Algeria
A. Loukil
Department of Electronics, Faculty of Electrical Engineering, University of Science and
Technology Oran, Oran, Algeria
https://doi.org/10.1007/978-981-16-3637-0_45
636 Z. Mehalli et al.
it is crucial to reduce these effects in order to improve the quality of hyperspectral

images. The preprocessing of these images could be considered as the first important
step to ensure better further processing and to allow more accurate interpretation of
objects spectra in images. Many scientific researches have been done in the field of
hyperspectral images preprocessing [1–5].The preprocessing steps include generally
bad band removal, geometric correction, radiometric calibration, atmospheric correc-
tion, and noise reduction. The geometric correction aims to correct the distortion in
the acquired image caused by the earth rotation and curvature. Atmospheric correc-
tions aim to match the image endmember spectra with the reference spectral libraries
or ground data. They also aim to minimize the effect of atmospheric agents that alter
radiance data. There are different atmospheric correction techniques divided into
empirical approaches (Statistical Based) and atmospheric models (Physics Based).
The empirical approaches include quick atmospheric correction (QUAC) [6],
empirical line correction (ELC) [7], internal average relative reflectance (IARR) [8],
and flat field (FF) [9]. The atmospheric models are: ATmospheric REMoval Program
(ATREM) [10], Atmospheric CORrection Now (ACORN) [11], Fast Line-of-sight
Atmospheric Analysis of Spec-tral Hypercube (FLAASH) [6], and ATmospheric
CORrection (ATCOR) [12].
To preprocess the Hyperion dataset of Djebel Meni (Northwestern Algeria),
we have been inspired by [13] to apply an empirical approach quick atmo-
spheric correction that determines atmospheric compensation parameters directly
from the observed pixel spectra, without ancillary information. It generally
produces reflectance spectra within the range of roughly+-15% of the physics-based
approaches [14].
The next preprocessing operation is a vertical stripes removal to enhance the
clarity of the images. After that, we have applied MNF technique to decrease data
dimension while conserving the important information. This final pre-processing
stage is recommended to ensure the low cost for downstream processing [15].
2 Study Area and Data Used
Djebel Meni is located in Mostaganem (Northwestern of Algeria). The area falls

between latitudes 36°04’28.06” and 36°03’45.11” N and longitudes 0°23’15.02”
and 0°31’08.59” E (Fig. 1). For this paper, data from EO-1 Hyperion satellite have
been used. The EO-1 space-borne launched on November 21, 2000 by the National
Aeronautics and Space Administration (NASA) for earth observation studies, and it
orbits the earth in a sun-synchronous orbit altitude of 705 Km [16]. The Hyperion
sensor has two spectrometers, the first operates in a visible near infrared region
(VNIR), 0.355–1 um, having 70 bands and the second operates in short wave infrared
region (SWIR), 0.9 um to 2.5 um, having 172 bands. So there are 242 wavebands with
a sampling interval of 10 nm. The EO-1 hyperspectral image has a spatial resolution
of 30m. The Hyperion hyperspectral data characteristics used in this paper are given
in Table 1.
Hyperspectral Data Preprocessing of the Northwestern Algeria … 637
Fig. 1 Study area

localization
Table 1 Summary of characteristics of the Hyperion hyperspectral data used in this study
Hyperion hyperspectral image of the region of Djebel Meni
Acquisition date 17–12–2010
Spatial resolution 30 m
Spectral resolution 10 nm
Number of bands 242
3 Methodology
In this section, we described the steps of the adopted method for hyperspectral image
preprocessing illustrated and resumed in flowdiagram (Fig. 2).
3.1 Bad Bands Removal
Hyperion hyperspectral data has 242 bands. The bands which do not have any pixel
information are called zero bands, so we need to remove them. Concerning our image,
bands 1 to 7 and 225 to 242 are not illuminated, bands 58 to 78 fall in the overlap
region of the two spectrometers (VNIR, SWIR), i.e., bands 56, 57 and 77, 78. We
need to remove the water vapor absorption bands too. These last are identified as
bands 120 to 132,165 to 182 and 221 to 242 [17].The summary of removed bands is
given in Table 2.
Fig. 2 Flowdiagram of the

methodology
Table 2 Summary of
Bands removed Reason
removed bands
1–7 Not illuminated
58–78 Overlap region
120–132 Water vapor absorption band
185–187 Identified by Hyperion bad band list
225–242 Not illuminated
3.2 Radiometric Calibration and Atmospheric Correction
The electromagnetic signal captured by the Hyperion hyperspectral sensor is modi-

fied by the atmosphere between sensor and the ground. The atmospheric effects are
scattering and absorption by gases such as carbon dioxide, ozone, water vapor, and
other gases. The radiance recorded by the hyperspectral sensor is influenced by the
atmosphere in two ways, the first way is by attenuating the energy illuminating the
object of the earth and the second way is by adding the path radiance to the signal
captured by the sensor. These two ways effects are represented mathematically as
follows:
P.E.T
Rtot = + RP (1)
2π
where Rtot is the total spectral radiance measured by the sensor, R p is the path
radiance, p is the reflectance of the object, T is the Transmitted energy, and E is the
irradiance on object caused by directly reflected sunlight and diffused skylight.
The radiometric calibration is performed in order to remove the path radiance,
and we need also to compensate the atmospheric attenuation effect. For that, we
have used quick atmospheric correction method (QUAC) in order to get the true
reflectance energy of the ground object. Quick atmospheric correction method deter-
mines the atmospheric compensation parameters directly from the observed pixel
spectra in hyperspectral image. This method is based on the empirical finding that
the mean spectrum of a collection of diverse material spectra, such as the endmember
spectra in a scene, is not dependent on each scene. Quick atmospheric correc-
tion method is suitable for real-time applications than the first-principles methods
because his faster computational speed. Furthermore, it performs a more approxi-
mate atmospheric correction than fast line-of-sight atmospheric analysis of spectral
Hypercube (FLAASH) or other physics-based first-principles methods, generally
producing reflectance spectra within the range of approximately 10 percent of the
ground truth [13, 18]. QUAC also allows for any view angle or solar elevation angle,
if a sensor does not have proper radiometric or wavelength calibration, or if the solar
illumination intensity is unknown (with cloud decks, for example).
3.3 Destriping of Band
Among the problems associated with Hyperion hyperspectral images is the vertical
stripes caused by calibration differences in Hyperion sensor array and temporal
variations in the sensor response[19]. These stripes contain corrupted pixels that make
the image unclear and will give a negative impact on further processing results[20].
We have proposed to remove these stripes by calculating the mean of every nth line
and normalizing each line to its respective mean.
3.4 Minimum Noise Fraction (MNF) Transform
The minimum noise fraction transform is used to determine the inherent dimen-
sionality of image data, to maximize the signal to noise ratio (SNR) of image data,
and to minimize the computational requirements for subsequent processing [21]. The
MNF can be processed as two consecutive data principal component transformations.
The first is the conversion of the noise covariance matrix to an identity matrix, and
the second is the principal component transformation of the noise whitened dataset
maximizing the signal to noise ratio (SNR) and removing the noise from the acquired
signal [22, 23]. The noise statistics are calculated using the shift difference method.
In the proposed methodology, Hyperion hyperspectral image of the region of Djebel

Meni is of MxNxB size, where B is the number of bands in the image and MxN is
the size of each band as shown in Table 3.
The true color image used as an input image is displayed in Fig. 3a and Fig. 4
depicts a Hyperspectral data cub
Only 158 of the 242 total hyperspectral bands were used because many of bands
do not have any pixel information which are called zero bands. Accepted bands and
noisy ones are shown in Fig. 5. Next radiometric calibration is applied to remove the
path radiance effect from the acquired signal, and the result is shown in Fig. 3b. The
quick atmospheric correction should be performed to handle the atmospheric effect
(Fig. 3c).
A comparison of the reflectance spectral profiles of a randomly selected pixel in:
the original image, after bad bands removal, after radiometric calibration, and after
atmospheric corrections is shown in Fig. 6.
The operation of destriping should be performed on the Hyperion hyperspectral
image of Djebel Meni because many bands suffer from this striping issue. The pixels
at stripes can be replaced with either the mean values of neighboring columns, and
the result of destriping for bands 2,3,4 is shown in Fig. 7
After applying destriping method on Hyperion data, minimum noise fraction
(MNF) is applied to segregate the random noise and reliable estimation of data
dimensionality from the signal information. The MNF eigenvalue graph is shown
Table 3 Data used

Hyperion hyperspectral image of the region of Djebel Meni
characteristics
Type HDF EO-1 Hyperion
Columns (N) 256
Rows (M) 3296
Bands (B) 242
Fig. 3 Hyperspectral image, a original image, b image after radiometric calibration, c image after
atmospheric correction
Fig. 4 Hyperion image cube
in Fig. 8. It represented each band with its corresponding eigenvalue. Bands with
eigenvalues close to 1 are mostly noise.
Table 4 listed the first ten MNF bands selected with their corresponding
eigenvalue, these bands contain the higher percentage of information.
Fig. 5 Gray scale displaying Hyperion bands, a bands 179,180,123 removed, b bands 13,14.16
accepted
The results showed that the maximum noise fraction of an image has been reduced
without losing information. Therefore, the overall proposed methodology is able to
guarantee interesting preprocessed data for further processing or analysis.
5 Comparison of Efficient Techniques of Atmospheric

Corrections on Hyperspectral Data for Mineralogy
Studies
5.1 Methodology
In this section, we described the steps of the adopted method for a comparative
analysis between internal average relative reflectance (IARR) atmospheric correction
Fig. 6 Spectral profile of a randomly selected pixel, a original image, b after bad bands removal,
c after radiometric calibration, d after atmospheric correction
and quick atmospheric correction (QUAC), illustrated and resumed in flowdiagram

(Fig. 9).
5.1.1 Bad Bands Removal and Radiometric Calibration (Previously

Explained in Sect. 3.1 and 3.2)
5.1.2 Internal Average Relative Reflectance (IARR).
IARR method divided each pixel in the image into a reflectance spectrum to
generate relative reflectance. However, the reflectance spectrum for IARR is the
mean spectrum of the complete image. It works best for arid areas with no vegetation
[24, 25].
Fig. 7 Destriping of bands 2, 3, 4, a bands before destriping, b bands after destriping

Fig. 8 MNF eigenvalue graph
Table 4 First ten MNF bands

MNF Eigenvalue
with their corresponding
eigenvalue 1 59.3072
2 24.3788
3 16.9522
4 12.2546
5 10.7573
6 7.8286
7 6.1890
8 5.5210
9 4.8821
10 4.4532
5.1.3 Quick Atmospheric Correction (QUAC) (Previously Explained

in Sect. 3.2)
5.1.4 Spectral Angle Mapper (SAM)
After applying IARR and QUAC atmospheric corrections, the spectral angle mapper
algorithm is applied for IARR corrected image end QUAC corrected image.
The spectral angle mapper algorithm calculates the spectral similarity between the
spectral signature of each pixel of the image and the spectral signatures of 21 clays
minerals which are represented in Fig. 10 (5 spectral signatures of Illite, 8 spectral
signatures of kaolinite and 8 spectral signatures of montmorillonite) introduced via
the United States Geological Survey (USGS) spectral library [26]. These spectral
Fig. 9 Flowdiagram of the methodology
signatures of clay minerals were chosen due to the geological nature of the study
area of Djebel Meni which is usually covered with this clay minerals[13].
SAM determines the similarity of an unknown spectrum t to a reference spectrum
r by applying [27, 28]:
Fig. 10 Spectral signature of clay minerals from USGS spectral library
⎛ ⎞
nb
⎜ i=1 ti ri ⎟
α = cos−1 ⎝ 0.5 0.5 ⎠ (2)
nb 2 nb 2
i=1 ti i=1 ri
where nb is the number of bands, ti is the test spectrum, and ri is the reference
spectrum.
The SAM classification result is represented by a color-coded image that show
the best SAM match at each pixel.
5.2 Results and Discussion
After the image bands have been resized to 158 bands (see Sect. 3.1), radiometric
calibration is applied, and the result is shown previously in Fig. 3b.
Next the IARR and QUAC atmospheric corrections methods are performed, and
the result of original image, QUAC, and IARR corrected images is shown in Fig. 11.
Visual analysis of the result (Fig. 11) shows that there is no significant difference in
the three images (original image, QUAC corrected image, IARR corrected image),
so to compare and evaluate the differences between QUAC and IARR atmospheric
corrections methods, we used the spectral angle mapper method that permits a rapid
mapping by calculating the similarity between spectral signature of each pixel on
the hyperspectral image and the spectral signatures of 21 clay minerals (illite, kaoli-
nite, montmorillonite) of the USGS spectral library (this clay minerals were chosen
because the Djebel Meni area is covered with it). The atmospheric correction method
with which we identify more types of clay minerals is considered as the best.
Fig. 11 a Original image, b IARR corrected image, c QUAC corrected image
The result of SAM method applied to QUAC corrected image and IARR corrected
image is illustrated in Fig. 12, and the histogram of Fig. 13 shows the clay minerals
types identified with their number of pixel covered.
After analyzing the results of Figs. 12 and 13, QUAC atmospheric correction
permits to identify 13 spectral signatures profils of illite, kaolinite, and montmo-
rillonite clay minerals; but with IARR atmospheric correction, we identified only
10 spectral profils of illite et montmorillonite. Kaolinite is not identified, also with
QUAC corrected image more pixels are classified than IARR corrected image as
shown in Tables 5 and 6. So QUAC is rigorous method and shows better correction
results, it has a capability to compensate the atmospheric effects than IARR method,
and it performs a good approximate atmospheric correction to IARR.
Fig. 12 a SAM applied to IARR corrected image, b SAM applied to QUAC corrected image
6 Conclusion
This scientific research aimed to propose a preprocessing scheme for the Hyperion
dataset of the region of Djebel Meni (Northwestern Algeria). It includes four main
steps, well-chosen to overcome input image drawbacks like geometric distortions,
striping, low signal to noise ratio, and high dimensionality. Therefore, bad bands are
Fig. 13 Histogram of clay minerals identified. a SAM applied to IARR corrected image, b SAM
applied to QUAC corrected image
Table 5 Capability of clay

Situation
minerals identification
SAM applied to SAM applied to
QUAC corrected IARR corrected
image image
Illite Identified Identified
Kaolinite Identified Not identified
Montmorillonite Identified Identified
Table 6 Comparison
Number of clay Number of pixels
between SAM applied to
minerals profiles classified
QUAC and IARR corrected
identified
images
SAM applied to 13 25,601
QUAC corrected
image
SAM applied to 10 17,513
IARR corrected
image
firstly removed, only 158 of the 242 total Hyperion bands were used, than, radiometric
calibration was performed to eliminate the path radiance effect from the acquired
signal, after that, quick atmospheric correction (QUAC) is applied to compensate
the effects of atmospheric absorption. This atmospheric absorption effect can lead to
wrong interpretation and identification of objects because it influences the reflectance
spectra .After removing atmospheric effects, destriping method is performed in order
to correct the abnormal pixels of vertical stripes. Finally, to process and analyze the
hyperspectral imagery with low computational cost, we have applied the minimum
noise fraction transform to decrease the dimensions of the data; while conserving the
important information, this method chooses the new components to maximize the
signal to noise ratio (SNR) and orders them according to increasing image quality or
decreasing noise. The obtained results show the high contribution of preprocessing
proposed stages to enhance the quality of input image.
A methododology for a comparative analysis between internal average rela-
tive reflectance (IARR) atmospheric correction and quick atmospheric correc-
tion (QUAC) for mineralogy studies is also proposed. QUAC compensates the
atmospheric effects more than IARR method in the field of mineral identification.
Comparison of other atmospheric correction will be an interesting perspective for
this research, and it is highly recommended for the future.
References
1. Amigoa, J.M., Santosb, C.: Preprocessing of hyperspectral and multispectral images. Elsevier,
pp. 37–53 (2020)
2. Jia, B., Wang, W., Ni, X., Lawrence, K.C., Zhuang, H., Yoon, S.C.,Gao, Z.: Essential processing
methods of hyperspectral images of agricultural and food products. Elsevier, pp. 1–11 (2020)
3. Kale, K.V., Solankar, M.M., Nalawade, D.B., Dhumal, R.K., Gite, H.R.: A Research Review
on Hyperspectral Data Processing and Analysis Algorithms, pp. 541–555, Springer (2017)
4. Tripathi, M.K., Govil, H.: Evaluation of Aviris-NG Hyperspectral Images for Mineral
Identification and Mapping. Elsevier, pp. 1–10 (2019)
5. Gore, R., Mishra, A., Deshmukh, R.: Mineral mapping at lonar crater using remote sensing. J.
Sci. Res. pp. 359–365 (2020)
6. Rani, N., Mandla, V.R., Singh, T.: Evaluation of atmospheric corrections on hyperspectral data
with special reference to mineral mapping. Elsevier, –12 (2016)
7. Karpouzli, E., Malthus, T.: The empirical line method for the atmospheric correction of
IKONOS imagery. Int. J. Remote Sens. pp. 1143–1150 (2003)
8. Tuominen, J, Lipping, T.: Atmospheric correction of hyperspectral data using combined empir-
ical and model based method. In: Proceedings of the 7th European Association of Remote
Sensing Laboratories Sig-imaging Spectroscopy Workshop (2011)
9. Kumar, M.V., Yarrakula, K.:Comparison of efficient tech-niques of hyper-spectral image
preprocessing for mineralogy and vegetation studies (2017)
10. Thompson, D.R., Gao, B.C., Green, R.O., Roberts, D.A., Dennison, P.E.: Lundeen SR Atmo-
spheric correction for global mapping spectroscopy: ATREM advances for the HyspIRI
preparatory campaign. Remote Sens. Environ. 167, 64–77 (2015)
11. Gao, B.C., Montes, M.J., Davis, C.O., Goetz, A.F.: Atmospheric correction algorithms for
hyperspectral remote sensing data of land and ocean. Remote Sens. Environ. 113, S17–S24
(2009)
12. Pflug, B., Main-Knorn, M.: Validation of atmospheric correction algorithm ATCOR. SPIE
Proc. Lidar Radar Passive Atmos. Measure. II, 9242(92420W), 1–8 (2014)
13. Zazi, L., Boutaleb, A., Guettouche, M.S.: Identification and mapping of clay minerals in the
region of Djebel Meni (Northwestern Algeria) using hyperspectral imaging, EO-1 Hyperion
sensor. Springer, 2–10 (2017)
14. Vignesh Kumar, M., Yarrakula, K.: Comparison of efficient techniques of hyper-spectral image
preprocessing for mineralogy and vegetation studies. Indian J. Geo Marine Sci. pp. 1008–1021
(2017)
15. Wang, J., Chang, C.I.: Independent component analysis-based dimensionality reduction with
applications in hyperspectral image analysis. IEEE Trans. Geosci. Remote. Sens. 44(6), 1586–
1600 (2006)
16. Pearlman, J., Carman, S., Lee, P., Liao, L., Segal, C.: Hyperion imaging spectrometer on the
new millennium program Earth Orbiter-1 system. In Proceedings, International Symposium on
Spectral Sensing Research (ISSSR), Systems and Sensors for the New Millennium, published
on CD-ROM, International Society for Photogrammetry and Remote Sensing (ISPRS) (1999)
17. Datt, B., McVicar, T.R., Van Niel, T.G., Jupp, D.L.B., Pearlman, J.S.: Preprocessing eo-1
hyperion hyperspectral data to support the application of agricultural indexes. IEEE Trans.
Geosci. Remote Sens. 41(6), 1246–1259 (2003)
18. Bernstein, L.S., Adler-Golden, S.M., Jin, X., Gregor, B., Sundberg, R.L.: Quick atmospheric
correction (QUAC) code for VNIR-SWIR spectral imagery: algorithm details. In Hyperspectral
Image and Signal Processing (WHISPERS), 2012 4th Workshop on (pp. 1–4). IEEE (2012)
19. Acito, N., Diani, M., Corsini, G.: Subspace-based striping noise reduction in hyperspectral
images. IEEE Trans. Geosci. Remote Sens. (2010)
20. Han, T., Goodenough, D.G., Dyk, A., Love, J.: “Detection and correction of abnormal pixels
in Hyperion images,” In IEEE International Geoscience and Remote Sensing Symposium,
Toronto, Ont.,Canada, pp. 1327–1330
21. Shirmard, H., Farahbakhsh, E., Pour, A.B., Muslim, A.M., Müller, R.D., Chandra, R.: Inte-
gration of selective dimensionality reduction techniques for mineral exploration using ASTER
satellite data. MDPI, pp. 1–29 (2020)
22. Phillips, R.D., Watson, L.T., Blinn, C.E., Wynne, R.H.: An adaptive noise reduction technique
for improving the utility of hyperspectral data. In: Proceedings of the 17th William T. Pecora
Memorial Remote Sensing Symposium, pp. 16–20 (2008)
23. Islam, M.R., Hossain, M.A., Ahmed, B.: Improved Subspace Detection Based on Minimum
Noise Fraction and Mutual Information for Hyperspectral Image Classification. Springer,
pp. 631–641 (2020)
24. Chakouri, M., Lhissou, R., El Harti, A., Maimouni, S., Adiri, Z.: Assessment of the image-
based atmospheric correction of multispectral satellite images for geological mapping in arid
and semi-arid regions. J. Preproof, pp. 1–33 (2020)
25. Merzah, Z.F., Jaber, H.S.: Assessment of Atmospheric Correction Methods for Hyperspectral
Remote Sensing Imagery Using Geospatial Techniques. IOP Publishing, 1–7 (2020)
26. Ren, Z., Sun, L., Zhai, Q.: Improved k-means and spectral matching for hyperspectral mineral
mapping. Elsevier, pp. 1–12 (2020)
27. Gopinath, G., Sasidharan, N., Surendran, U.: Landuse classification of hyperspectral data by
spectral angle mapper and support vector machine in humid tropical region of India. Springer,
pp. 1–9 (2020)
28. Govil, H., Mishra, G., Gill, N., Taloor, A., Diwan, P.: Mapping Hydrothermally Altered
Minerals and Gossans using Hyperspectraldata in Eastern Kumaon Himalaya, India. Elsevier,
pp. 1–7 (2021)
Smart Agriculture Solution Based on IoT
and TVWS for Arid Regions
of the Central African Republic
Edgard Ndassimba, Nadege Gladys Ndassimba, Ghislain Mervyl Kossingou,

and Samuel Ouya
Abstract The Central African Republic is threatened by desertification. Its economy

is mainly dependent on agriculture. However, in arid zones where rainfall is low,
environmental factors have made it difficult to develop sustainable agriculture. The
collapse of the level of agricultural production makes populations vulnerable to food
insecurity. Today, there are several innovative solutions to monitor and automate
field irrigation in these arid and semi-arid regions. This paper proposes a system for
monitoring and automatic irrigation of fields in these areas. The proposed solution
is based on the Internet of Things (IoT) using a raspberry pi 3 b + with temperature
and soil moisture sensors, a submersible water pump and a relay. The system uses a
TVWS network that offers broadband Internet to farmers allowing them to connect
their smartphones via Wi-Fi. These farmers use a web interface to access the field
temperature and soil moisture monitoring screen and then remotely activate or deac-
tivate the irrigation system, although this can also be done automatically. The results
of this experiment allow the implementation of an automatic irrigation system in arid
and semi-arid regions with low rainfall. This solution would strengthen the Central
African Republic’s economy and help develop sustainable agriculture to meet the
urgent needs of the population threatened by famine.
1 Introduction
Over the last ten years, several regions of Central African Republic, covering an
area of about 623,000 km2 , have been threatened by desertification. The localities
in the North East are practically devoid of water. Locally, the climate has changed:
temperature has increased and rainfall has decreased [1]. Due to a long dry season,
food is also no longer produced because the climatic conditions are not favorable
for agriculture. One of the solutions to promote agriculture is the use of intelligent
agriculture.
E. Ndassimba (B) · N. G. Ndassimba · G. M. Kossingou · S. Ouya

Laboratory of Computing Networks and Telecommunications (LIRT) of Higher Polytechnic
School At the University Cheikh Anta Diop (UCAD), Dakar, Senegal
https://doi.org/10.1007/978-981-16-3637-0_46
654 E. Ndassimba et al.
Several research works have been carried out on the use of the Internet of Things
based on the Raspberry pi in the field of agriculture. The authors in [2] presented a
soil quality monitoring system using wireless sensor nodes. Others in [3] developed
intelligent monitoring and security devices based on IoT for agriculture. The results
of this work have shown that with the Internet of Things (IoT), we can predict
and analyze the parameters of the greenhouse effect to improve crop quality and
productivity, and then ensure the safety of farms against rodents.
Our approach differs from existing uses. Firstly, because of the use of TV White
Space to bring broadband internet to these arid areas, and secondly on how to monitor
and control the irrigation system which can be done automatically by the system and
manually remotely from a smartphone of the farmer connected to the TV White
Space Wi-Fi network.
The rest of this paper is organized as follows. Section 2 is reserved for the state of
the art. Section 3 presents the materials and methodology used. Section 4 presents
the results and discussions about our solution, and finally Section 5 provides the
conclusion.
2 State of the Art
2.1 Impact of Desertification on Agriculture in the Central

African Republic
The concept of desertification is defined as land degradation in arid, semi-arid and

dry sub-humid areas, often referred to simply as “drylands." It is believed to result
from a combination of factors, including climate change and human activity. Arid
zones occupy 41.3% of the Earth’s surface and are characterized by water scarcity
[4, 5].
In the Central African Republic, climate change is perceptible over the entire
national territory and deserves special attention from the main land users whose
ecosystem functions need to be preserved in order to have an impact on productivity
and farm incomes. Droughts are now frequent in the northern, northeastern and
eastern regions, once renowned for their agricultural production. There is growing
evidence that groundwater reserves are being depleted, leading to a sharp reduction
in productivity in these areas [6].
Desertification is increasing every year, contributing to the disappearance of
several million hectares and affecting farmers. This creates food insecurity, poverty
and high-economic cost in the Central African Republic.
Smart Agriculture Solution Based on IoT and TVWS for Arid … 655
2.2 Climate-Smart Agriculture
Climate-smart agriculture (CSA) is an approach that helps guide the actions needed
to transform and reorient agricultural systems to effectively support development
and ensure food security in a changing climate. Climate-smart agriculture is one of
the techniques that maximize agricultural yields through good management of inputs
according to climate conditions [7]. The CSA has three main objectives: to sustain-
ably increase agricultural productivity and incomes; to adapt and build resilience
to climate change; and to reduce and/or eliminate greenhouse gas emissions wher-
ever possible [8, 9]. Intelligent agriculture uses new technologies, such as satellite
imagery and computers, satellite positioning systems such as GPS, also through the
use of sensors that will collect useful information on soil condition, moisture content,
mineral salt content, etc. and send this information to the farmer to take the neces-
sary measures to ensure good production. Generally, irrigation is used to improve
agricultural production in the face of climate change. Irrigation is the supply of water
to crops by artificial means to enable agriculture in arid areas and to compensate for
the effects of drought in semi-arid areas. Where traditional rain-fed agriculture is at
high risk, irrigation can help to ensure stable production.
The Central African Republic has a hot and humid equatorial climate, charac-
terized by two seasons: a rainy season that lasts from April to October, and a dry
season, between November and March. Annual rainfall is higher in the Ubangi Valley
(1780 mm) than in the central part (1300 m) and in the semi-arid northeastern and
eastern areas (760 mm). The development of sustainable agriculture in the Central
African Republic can help mitigate the secondary effects of the country’s dire situa-
tion. Smart agriculture can lead to a growing economy at the micro and macro levels
by increasing production [10].
Vulnerability to climate change in the Central African Republic and low capacity
to adapt to its adverse effects pose serious threats to the management of ecosystems
and agricultural resources and to sustainable development, hence the importance of
using smart agriculture in desertification-affected regions.
2.3 The Internet of Things in Agriculture
The Internet of Things (IoT), refers to the set of infrastructures and technologies
set up to make various objects work through an Internet connection. We then speak
of connected objects. These objects can be controlled remotely, most often using a
computer, a smartphone or a tablet.
Numerous research works have shown that the Internet of Things (IoT) can be used
in several fields such as transport, health, home automation, agriculture, etc. In [11],
the authors proposed an intelligent agricultural system (AgriSys) that can analyze an
agricultural environment and intervene to maintain its suitability. The authors in [12]
focused their work on the introduction of a Smart Drone for crop management where
real-time data from the UAV, combined with IoT and Cloud Computing technologies,
help to build a sustainable intelligent agriculture. The authors in [13] presented an
intelligent solution, gCrop, to monitor the growth and development of leafy crops and
to update the status in real time using IoT, image processing and machine learning
technologies. In [14], the authors proposed adapted good practices to reduce the
water footprint in agricultural crop fields with traditional methods. The combination
of biochemistry and the Internet of Things contributes to improve the competitiveness
of agricultural economic activities near cities and at the same time to avoid water
crisis.
The results of this work have proved that the challenges of IoT are numerous.
Among these challenges, agriculture is undergoing its digital transformation. Farmers
can accurately control environmental parameters (air and soil humidity, temperature,
etc.) recorded by sensors, and remotely control the irrigation of their fields for better
productivity and profitability.
2.4 The TVWS in Agriculture
The TV White Space is a technology that allows the use of free frequencies from tele-
vision to provide a broadband network in a given region. The free UHF frequency
bands used are 470–790 MHz in Europe and 54–698 MHz in the United States.
The use of White Space is based on secondary unlicensed dynamic spectrum
access (DSA—Dynamic Spectrum Alliance) under the principle of non-detrimental
interference for television operators operating in the area.
By revolutionizing traditional wireless broadband connectivity, the TV White
Space is typically used to bring broadband Internet to rural areas with difficult access.
Several researchers have worked on the relevance of using TVWS. In [15], the
authors showed the opportunity of vehicular communications on TV White Space
in the presence of secondary users. The results show that there are opportunities for
vehicular access even when a White-Fi network occupies the TVWS. In [16], the
authors studied and adopted TV White Space technology as a rural telecommuni-
cation solution in Indonesia in relation to its performance. They concluded from a
simulation that TV White Space is an appropriate technological alternative for rural
conditions.
The results of this work proved the potential that TV White Space has offered for
a wide range of innovative applications. For example, White Space TV can establish
high-bandwidth links between a farmer’s home Internet connections and an on-farm
IoT base station with sensors for intelligent agriculture in arid rural areas.
3 Materials and Methodology
3.1 Architecture of Our Solution
The architecture of our solution includes:

• A Raspberry Pi 3 b + equipped with a 40-pin GPIO extension board and a GPIO
cable mounted on a breadboard, and deployed on a RAB platform;
• A Relay 4 Module (VCC: 5 V positive supply voltage, GND: ground, IN1: relay
control port);
• A DHT11 Temperature/Humidity Sensor (VCC: 5 V supply voltage, GND:
ground, DATA: serial data port, single bus);
• Soil Moisture Sensor (VCC: 5 V supply voltage, GND: ground, GIS);
• A mini submersible pump (Motor Submersible Water Pump);
• Silicone tubing;
• 5 V power supply for the pump and Raspberry Pi;
• A TV White Space network with a Base Station (BS) and customers called
Customer Premises Equipment (CPE);
• Internet connectivity.
The following figure shows the interconnection of the intelligent agriculture
system test bed components (Fig. 1).
The following figures show the different elements of the Intelligent farming system
test bed (Fig. 2).
4 relay module is a 4-channel 5 V relay interface card and each channel requires a
driver current of 15–20 mA. It can be used to control various devices and equipment
with high current. It is equipped with high current relays operating under AC250V
Fig. 1 Proposed solution architecture

Fig. 2 Relay module
10A or DC30V 10A. It has a standard interface that can be directly controlled by a
microcontroller (Fig. 3).
The DHT11 digital temperature and humidity sensor are a composite sensor
that contains a calibrated digital signal output of temperature and humidity. The
application of dedicated digital module collection technology and temperature and
humidity sensing technology ensures that the product has high reliability and excel-
lent long-term stability. The sensor consists of a resistive wet component sensor and
NTC temperature measuring device, and is connected to a high-performance 8-bit
microcontroller (Fig. 4).
Fig. 3 DHT11
temperature/humidity sensor
Fig. 4 Soil moisture sensor

Fig. 5 Mini DC 3-6 V

120L/H brushless motor
submersible water pump
The SparkFun soil moisture sensor is a simple escape to measure the moisture
content of soil and similar materials. The two large exposed studs act as probes for
the sensor, acting together as a variable resistance. The more water in the soil, the
better will be the conductivity between the pads, resulting in lower resistance and
higher SIG output (Fig. 5).
The mini water pump is made of plastic material and electronic components. It
operates with a DC voltage of 2.5-6 V and can provide a flow rate of 80-120L/H with
a power of 0.4–1.5 W (Fig. 6).
The raspberry pi is a motherboard of a mini-computer that can be connected to
any device (mouse, keyboard…). This card is made to help study computers and to
represent a means of learning computer programming in several languages (python,
scratch…).
The raspberry pi 3 model B + includes a Broadcom BCM2837B0 64-bit quad-
core ARM Cortex-A53 processor running at 1.4 GHz, a new CYW43455 supporting
Dual-band 802.11ac Wi-Fi and Bluetooth version 4.2, and the support for power over
Ethernet through an additional element.
The following figure shows the cabling of our test environment (Fig. 7).
Fig. 6 Raspberry Pi 3 model

B+
Fig. 7 Testing environment
3.2 How the Solution Works
To operate, the mini submersible water pump must be completely submerged in

a water tank in order to pump water. This pump is powered by a 5 V charger
and connected to the relay. The pump is connected to the GPIO4 as a LED. The
DHT11 sensor is connected to the GPIO17. The soil moisture sensor is connected
to the GPIO14 as a button. For the SparkFun soil moisture sensor to work, we have
connected the VCC and GND pins to the raspberry pi. We receive a SIG, which
depends on the amount of water in the soil. When the soil moisture sensor gives
the raspberry pi an input voltage of 1 or 0 depending on whether it is wet or dry, it
behaves like a button pressed when wet, and a button released when dry. The pump
will receive from the python written program of the raspberry pi via the relay an
output voltage activated or deactivated like a LED that triggers watering and admin-
isters irrigation periods. The soil moisture sensor detects the condition of the soil (its
dryness rate) in order to adjust the water supply accordingly to efficiently water the
fields. The DHT11 sensor allows you to stop watering immediately when it begins
to rain.
The system can operate in automatic mode, manual mode, and factory water mode.
In order to access our raspberry remotely via the TV White Space Wi-Fi broadband
network and control the GPIOs, we need to make our card a Nginx web server capable
of hosting the website and thus ensure communication with customers and respond
to their requests using the HTTP network protocol. This allows them to consult in
real time on a smartphone or computer via a dynamic web page of the web server
that displays the ground parameters.
Fig. 8 Execution of the Planty.py python code
The results of the test bench are based on the libraries and dependencies of dedicated
peripherals for the temperature/humidity sensor and the soil moisture sensor and
GPIO raspberry pi that we downloaded and installed. We created our application
with PubNub functions.
4.1 Irrigation System Test Without Wi-Fi Connection
To start the system, we executed the Planty.py python code by connecting to the
Raspberry via SSH as shown in the following Fig. 8:
Figure 9 showed the values of temperature and humidity obtained on the computer
connected by SSH to the raspberry showing the values of the two variables: Temp
and humidity.
Figure 10 showed the automatic irrigation triggered by the Raspberry Pi when we
remove the soil moisture sensor from the soil. This sensor emitted a voltage when
wet and none when dry. Removing this sensor sent a signal that the soil is dry, which
caused the pump that started the irrigation to start.
4.2 Remote Control of the Irrigation System by Simulating

the Connection to the TVWS Network
To demonstrate the relevance of using TVWS technology for broadband Internet

connectivity in rural areas for remote monitoring and control of the irrigation system,
we used a Wi-Fi network card (LB-LINK) installed on the raspberry pi and connected
to a Wi-Fi access point.
We used the web interface: http://192.168.43.140/IoT_Plant/plant.html for remote
monitoring of field soil parameters and control of the irrigation system.
Figures 11 and 12 showed the graphs of the temperature and humidity data
obtained in real time from the Nginx web server installed on the raspberry pi.
On the web interface, there are three buttons for remote control of the system.
Fig. 9 Publication of
temperature and humidity
results
Fig. 10 Automatic start of

the irrigation system
Fig. 11 Graph of data on a computer
Fig. 12 Data graph on a smartphone

Fig. 13 Remote control of the irrigation system
• Auto ON button: when the automatic watering mode is activated.

• Auto OFF button: when the automatic watering mode is off.
• Water Plant Button: to remotely control watering.
Figure 13 showed the remote control of the irrigation system. Pressing the water
plant button will remotely control the irrigation.
Contrary to the different types of networks used in Internet of Objects solutions
(standard Wi-Fi IEEE 802.11, LoRa, GSM, etc.), TVWS technology is best suited
to bring high-speed Internet to rural areas in relation to its technical and economic
advantages. Its technical advantages are its extended range of 10 to 30 km; broad-
casting over uneven areas, through obstacles, forests or buildings. The economic
advantages concern the deployment of the technology at a lower cost.
In Central African Republic, unused television White Space (TVWS) would solve
the problem of Internet connectivity to establish a farmer’s home broadband Internet
connection and the Internet of Things (IoT) deployed in the field through inter-
connected TVWS customer premium equipment (CPE) on a TVWS base station
(BS).
5 Conclusion
In this work, we proposed a smart agriculture solution for the arid regions of the
Central African Republic. This system was designed based on a raspberry pi 3 b +
and a set of sensor networks connected to a TV White Space broadband network.
Intelligent farming practices have proven effective in several countries that have
experienced droughts, and could be implemented in agricultural systems in countries
that have experienced food crises due to droughts.
This solution will help farmers to remotely control the soil parameters of their
fields and then trigger the irrigation system, although this can also be done auto-
matically. The impact will be to increase production and solve the problems of food
insecurity and climate change.
In future work, we will propose an independent mobile network infrastructure
based on TVWS in rural areas of the Central African Republic using the Internet
of Things with D-GSM in Osmocom and Freeswitch to send SMS/MMS alerts to
farmers on the critical state of soil parameters (If the soil moisture level is below the
normal threshold, the farmer receives an SMS notification and remote control of the
automatic irrigation).
References
1. Forest Management and Deforestation in Central African Republic. http://www.ajer.org/pap

ers/v5(04)/I0504079090.pdf, last accessed 2020/11/21
2. Shinde, D., Siddiqui, N.: “IOT Based environment change monitoring & controlling in
greenhouse using WSN,” 2018 International Conference on Information , Communication,
Engineering and Technology (ICICET), pp. 1–5, Pune (2018)
3. Baranwal, T., Nitika and P.K. Pateriya, “Development of IoT based smart security and moni-
toring devices for agriculture,” 2016 6th International Conference - Cloud System and Big
Data Engineering (Confluence), pp. 597–602, Noida (2016)
4. United Nations Decade for Deserts and the Fight Against Desertification. https://www.un.org/
en/events/desertification_decade/whynow.shtml, last accessed 2020/11/21
5. Trees, forests and land use in drylands The first global assessment. http://www.fao.org/3/a-i59
05e.pdf, last accessed 2020/11/21
6. SOCIAL WATCH poverty eradication and gender justice. https://www.socialwatch.org/node/
13981, last accessed 2020/11/22
7. Tenzin, S., Siyang, S., Pobkrut, T., Kerdcharoen, T.: “Low cost weather station for climate-
smart agriculture,” 2017 9th International Conference on Knowledge and Smart Technology
(KST), pp. 172–177. Chonburi (2017)
8. Climate-Smart Agriculture. http://www.fao.org/climate-smart-agriculture/en/, last accessed
2020/11/24
9. Climate-smart agriculture for food security. https://www.researchgate.net/publication/273448
307_Climate-smart_agriculture_for_food_security, last accessed 2020/11/24
10. Sustainable Agriculture in the Central African Republic. https://www.borgenmagazine.com/
sustainable-agriculture-in-the-central-african-republic/, last accessed 2020/11/25
11. Abdullah, A., Al Enazi, S., Damaj, I.: “AgriSys: a smart and ubiquitous controlled-environment
agriculture system,” 2016 3rd MEC International Conference on Big Data and Smart City
(ICBDSC), pp. 1–6, Muscat (2016)
12. Namani, S., Gonen, B.: “Smart agriculture based on iot and cloud computing,” 2020 3rd
International Conference on Information and Computer Technologies (ICICT), pp. 553–556,
San Jose, CA, USA (2020)
13. Kumar, S., Chowdhary, G., Udutalapally, V., Das, D., Mohanty, S.P.: “gCrop: Internet-of-
Leaf-Things (IoLT) for monitoring of the growth of crops in smart agriculture,” 2019 IEEE
International Symposium on Smart Electronic Systems (iSES) (Formerly iNiS), pp. 53–56,
Rourkela, India (2019)
14. Larios, V.M., Michaelson, R., Virtanen, A., Talola, J., Maciel, R., Beltran, J.R.: “Best practices
to develop smart agriculture to support food demand with the rapid urbanization trends in Latin
America,” 2019 IEEE International Smart Cities Conference (ISC2), pp. 555–558, Casablanca,
Morocco (2019)
15. Arteaga, A., Céspedes, S., Azurdia-Meza, C.: Vehicular communications over TV white spaces
in the presence of secondary users. IEEE Access 7, 53496–53508 (2019)
16. Aji, L.S., Wibisono, G., Gunawan, D.: “The adoption of TV white space technology as a rural
telecommunication solution in Indonesia,” 2017 15th International Conference on Quality in
Research (QiR) : International Symposium on Electrical and Computer Engineering, pp. 479–
484 Nusa Dua (2017)
Model-Driven Engineering: From SQL
Relational Database
to Column—Oriented Database in Big
Data Context
Fatima Zahra Belkadi and Redouane Esbai
Abstract The growth of application architectures in all areas (e.g., astrology, mete-
orology, E-commerce, social network, etc.) has resulted in an exponential increase
in data volumes, now measured in Petabytes. Managing these volumes of data has
become a problem that relational databases are no longer able to handle because
of the acidity properties. In response to this scaling up, new concepts have emerged
such as NoSQL. In this paper, we show how to design and apply transformation rules
to migrate from an SQL relational database to a big data solution within NoSQL. For
this, we use the model driven architecture (MDA) and the transformation languages
like as MOF 2.0 QVT (Meta-Object Facility 2.0 Query-View-Transformation) and
Acceleo which define the meta-models for the development of transformation model.
The transformation rules defined in this work can generate, from the class diagram,
a CQL code for creation column-oriented NoSQL database.
1 Introduction
In recent years, the world of data storage is changing rapidly. New technologies and
new actors are settling when the old ones make the move. This scientific revolu-
tion that has invaded the world of information, and the Internet has imposed new
challenges on researchers in recent years and has led them to design new tools for
specific storage and manipulation. The development of these tools is generating a
growing interest among scientific and economic actors to offer them the possibility of
managing all these masses of data with reasonable response times. Big data is corre-
lated between four notions generally grouped under the acronym “4 V,” namely:
Volume, variety, velocity and variability [1].
Our focus in this paper is only on big data storage. Using relational databases
prove to be inadequate for all applications, particularly ones involving large volumes
of data. In this context, NoSQL databases offer new storage solutions in large-scale
environments, replacing many traditional database management systems [2]. The
F. Z. Belkadi · R. Esbai (B)

Mohammed First University, Oujda, Morocco
https://doi.org/10.1007/978-981-16-3637-0_47
668 F. Z. Belkadi and R. Esbai
key feature of NoSQL databases is that they are schema-less, meaning that data can
be inserted in the database without upfront schema definition. Nevertheless, there
is still a need for a semantic data model to define how data will be structured and
related in the database [3]; it is generally accepted that UML meets this requirement
[4].
This paper aims to rethink the work presented in [5]. However, we develop the
transformation rules using the MOF 2.0 QVT standard to generate a file which
contains a code for creation a column-oriented NoSQL model [6]. Our approach
includes UML modeling and automatic code generation using Acceleo with the aim
to facilitate and accelerate the creation of column-oriented NoSQL database.
This paper is organized as follows: related works are presented in the second
section, the third section defines the MDA approach, and the fourth section presents
the NoSQL and its implementation as a database, column-oriented in this case. In the
fifth section, we present the source and target meta-models. In the sixth section, we
present the transformation process M2M and M2T from UML class diagram model
to the column-oriented NoSQL database. The last section concludes this paper and
presents some perspectives.
2 Related Works
Many researches on MDA and the process of transforming relational databases into
a NoSQL model have been conducted in recent years. The most relevant are [3,
5–10]: Chevalier et al. [7] defined rules to transform a multidimensional model into
NoSQL column-oriented and document-oriented models. The links between facts
and dimensions have been converted using imbrications. Although the transformation
process proposed by authors start from a multidimensional model, it contains facts,
dimensions and one type of links only. Gwendal et al. [3] describe the transformation
from an UML conceptual model into a graph databases via an intermediate graph
meta-model. These transformation rules are specific to graph databases used as a
framework for storing, managing and querying complex data with many connections.
Li et al. [8] propose a MDA approach to transform UML class diagram into HBase.
After building the meta-models of UML class diagram and HBase, the authors have
proposed mapping rules to realize the transformation from the conceptual level to the
physical level. These rules are applicable to HBase only. Another works followed the
same logic and have been the subject of a work Vajk et al. [9]. The authors propose
a mapping from a relational model to document-oriented model using MongoDB.
The purpose of the work [10] presented by Abdelhedi et al. is to implement a
conceptual model describing big data into NoSQL database and they choose to focus
on column-oriented NoSQL model.
This paper aims to rethink and to complete the work presented by Abdelhedi
et al. [5, 10], by applying the standard MOF 2.0 QVT and Acceleo to develop
the transformation rules aiming at automatically generating the creation code of
column-oriented NoSQL database. It is actually the only work for reaching this goal.
Model-Driven Engineering: From SQL Relational Database … 669
3 Model Driven Architecture (MDA) Approach
3.1 The Transformations of MDA Model
The MDA identifies several transformations during the development cycle [11]. It is
possible to make three different types of transformations: CIM to PIM, PIM to PSM
and PSM to code.
In this paper, we chose two types of transformation, we start with the transforma-
tion PIM to PSM using the approach by modeling. This type of transformation will
allow us to automatically generate a column-oriented NoSQL model from an UML
model. The second transformation is of type PSM to Code using the approach by
template with Acceleo to develop the transformation rules aiming at automatically
generating the creation code of column-oriented NoSQL database [12].
3.2 The Elaborationist Approach
The elaborationist approach is the one used in the present paper. The main advantage
of MDA in the development of column-oriented NoSQL databases is the automation.
This way, to demonstrate the automation support provides by our MDA approach,
we are using the “Elaborationist approach” (see Fig. 1). With the elaborationist
approach, the definition of the application is built up progressively as you progress
through from PIM to PSM to code. When the PIM has created, the tool generates
a skeleton or first-section PSM which the developer can then “elaborate” by adding
more detail. Similarly, the final code is generated from PSM, and this can also be
elaborated.
Fig. 1 Elaborationist
approach [13]
4 Column-Oriented NoSQL Database
There are four basic types of NoSQL databases: key-value, document-oriented,

column-oriented and graph-oriented [2]. In this paper, we choose to focus on
column-oriented NoSQL model.
The column-oriented databases were originally created by Facebook to store
messages (non-instant) between users [14]. It is a key-value database extension,
because the column model is more evolved, it is called super-column or column-
family that a line identifier can store a structured set of data. A column-family has
the following characteristics: the data is sorted, associated, and can contain an array
of columns of unlimited size.
The storage of column-oriented databases is by column and not by row. These
bases can evolve over time, either in number of rows or in number of columns. In
other words, and unlike a relational database where columns are static and present
for each row, the column-oriented databases are dynamic and present only when
needed.
In column-oriented databases such as Cassandra [15] or HBase [16], there are
some additional concepts that are the column-family, which are a logical grouping
of rows. In the relational world, this would be equivalent to a table. Cassandra offers
an extension to the base model by adding an extra dimension called “Super Column”
which itself contains other columns.
The concept of column-oriented databases is created by the big web actors, to meet
the processing needs of large volumes of data precisely to manage large volumes of
structured data. Often, these databases integrate a minimalist query system close to
SQL called CQL [2].
In this paper, we choose the principle actor of column-oriented database such as
Cassandra.
5 Source and Target Meta-Models
In our MDA approach, we opted for the modeling and template approaches
to generate the column-oriented NoSQL database. As mentioned above, these
approaches require a source meta-model and a target meta-model. We present in this
section, the various meta-classes forming the UML class diagram source meta-model
and the column-oriented NoSQL target meta-model.
5.1 UML Source Meta-Model
Figure 2 illustrates the simplified UML source meta-model based on packages

including data types and classes. Those classes contain typed properties and they
Fig. 2 Simplified UML source meta-model
are characterized by multiplicities (lower and upper). The classes are composed of
operations with typed parameters.
UmlPackage: is the concept of UML package. This meta-class is connected to
the meta-class Classifier.
Classifier: This is an abstract meta-class representing both the concept of UML
class and the concept of data type.
Class: is the concept of UML class.
DataType: represents UML data type.
Operation: is used to express the concept of operations of a UML class.
Parameter: expresses the concept of parameters of an operation. These are of
two types, Class or DataType. It explains the link between parameter meta-class and
classifier meta-class.
Property: expresses the concept of properties of a UML class. These properties
are represented by the multiplicity and meta-attributes upper and lower.
The works [17, 18] contain more details related to this section topic.
5.2 Column-Oriented Target Meta-Model
To fully understand the data model used by Cassandra [19], it is important to define
a number of concepts used:
Keyspace: Appears as a namespace, this is usually the name given to the
application.
Column: Represents a value, and it has three fields (see Fig. 3): its name, its value
and a timestamp representing the date on which this value was inserted.
Super-Column: it is a list of columns (see Fig. 4), if you want to compare them
with an SQL database, it is a row. It contains the key-value correspondence; the key
identifies the super column, while the value is the list of columns that compose it.
Column-Family: it is a container of several columns or super-columns. Its notion
is closer to the SQL table (see Fig. 5).
Figure 6 presents these concepts through the target meta-model.
By default, we store the database in a single Keyspace. This Keyspace is comprised
of a set of column-families [20]. Each Column-family is identified by a unique
identifier called “PrimaryKey” and contains several columns or super-columns that
must be declared up front at schema creation time.
Fig. 3 Structure of the

column element

super-column element

Column-Family element
Fig. 6 Simplified
column-oriented target
meta-model
6 The Process of Transforming UML Source Model

to Column—Oriented Target Code
We first developed ECORE models corresponding to our source and target meta-
models. The development of many meta-models requires multiple model transfor-
mations. From these developed meta-models, M2M (Model to Model) and M2T
(Model to Text) transformations are needed, to generate the code needed to create
the column-oriented database. We have implemented the M2M transformation algo-
rithm (see Sect. 6.1) using the QVT Operational Mappings language [21], and then the
second M2T transformation is done with the Acceleo language [22] (see Sect. 6.2).
6.1 The Transformation Rules M2M
This transformation uses, in entry, a model of the UML type, and in output a model
of column-oriented database. The first transformation rule establishes the correspon-
dence between all the elements of the UML package and the element of the Keyspace
type of the column-oriented database. The purpose of the second rule is to transform
each UML class and association into a family of columns by creating the columns
and references for each column-family. It is a question of transforming each property
of these classes in column, without forgetting to give names and types to the various
columns.
Figure 7 presents the principle part of the M2M transformation with QVT
language.
Fig. 7 M2M transformation with QVT from UML to NoSQL model

Fig. 8 M2T transformation with Acceleo to generate a CQL code
6.2 The Transformation Rules M2M
The transformation M2T toward the creation code of column-oriented database in

Cassandra is realized with Acceleo transformation language, and the writing of the
transformation rules itself does not present any problems in practice. It simply boils
down to creating a text file where the transformation rules are written.
Figure 8 presents the transformation rules with Acceleo to generate a CQL file.
6.3 Result
To validate our transformation rules, we conducted several tests.

For example, we considered the class diagram composed by the classes depart-
ment, employee and city (see Fig. 9).
After applying the transformation on the UML source model, we generated the
column-oriented PSM target model (see Fig. 10).
Figure 10 shows the result after applying the transformation rules M2M.
Figure 11 illustrates the results of our M2T transformation. Our application
generates a CQL code for a department management database on Cassandra platform.
Fig. 9 UML source model: Class diagram EMF model and Class diagram instance model
Fig. 10 Column-oriented cassandra PSM: resource set and their properties
In this paper, we have proposed an MDA approach to migrate UML class diagram
representing a relational database to a column-oriented database. The transforma-
tions rules were developed using QVT to transform the class diagram into column-
oriented model and then the automatic code generation using Acceleo with the goal
to accelerate and makes easy the creation of NoSQL databases in Cassandra plat-
form. In future, this work should be extended to allow the generation of other NoSQL
Fig. 11 CQL file generated
solutions such as document-oriented and graph-oriented. Afterward we can consider

integrating other big data platforms like HBase, Redis, Neo4j and others.
References
1. Chen, C.L.P., Zhang, C.: Data-intensive applications, challenges, techniques and technologies:
a survey on big data. Inf. Sci. 275, 314–347 (2014)
2. Cattell, R.: Scalable SQL and NoSQL data stores. ACM SIGMOD Rec. 39(4), 12–27 (2011)
3. Gwendal, D., Gerson, S., Jordi, C.: UMLtoGraphDB: mapping conceptual schemas to graph
databases. In: The 35th International Conference on Conceptual Modeling (ER) (2016)
4. Abello, A.: Big data design. In: Proc. of the ACM Eighteenth International Workshop on Data
Warehousing and OLAP, Australia (2015)
5. Abdelhedi, F., Brahim„ A.A., Faten, A., Zurfluh, G.: MDA-based approach for NoSQL
Databases Modelling, In: International Conference on Big Data Analytics and Knowledge
Discovery (DaWaK 2017), Lyon, France, (28–31 Aug 2017)
6. OMG, XML Metadata Interchange (XMI), version 2.1.1, OMG (2007)
7. Chevalier,M., El Malki, M., Kopliku, A., Teste, O., Tournier, R. : Implementing multidimen-
sional data warehouses into NoSQL. In: International Conference on Enterprise Information
Systems (ICEIS 2015), Barcelona, Spain (2015)
8. Li, Y., Gu, P., Zhang, C.: Transforming UML Class Diagrams into HBase Based on Meta-model.
Information Science, Electronics and Electrical Engineering (ISEEE) (2014)
9. Vajk, T., Feher, P., Fekete, K., Charaf, H.: Denormalizing data into schema-free databases. In:
4th International Conference CogInfoCom. pp. 747–752 (2013)
10. Abdelhedi, F., Brahim, A.A., Atigui, F., Zurfluh, G.: Big Data and knowledge management:
how to implement conceptual models in NoSQL systems?. In: 8th International Conference on
Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2016),

Porto, Portugal, (9–11 Nov 2016)
11. Gotti, S., Mbarki, S.: IFVM bridge: a model driven IFML execution. Int. J. Online Biomed.
Eng. (iJOE). 15(4), 111–126 (2019)
12. Miller, J., Mukerji, J.: MDA Guide Version 1.0.1, OMG, (2003)
13. Papajorgjin, P., Pardalos, P.M.: Towards a model-centric approach for developing enterprise
information systems. Enterprise Information Systems and Implementing It Infrastructures:
Challenges and Issues. IGI Global; 1st edn. pp. 140–158 (2010)
14. Radoslava, S.K., , S.K., Nina, S., Petia, K., Nadejda, B.: Design and analysis of a relational
database for behavioral experiments data processing. Int. J. Online Biomed. Eng. (iJOE). 14(02)
(2018), 117–132 (2019)
15. Apache Cassandra, http://cassandra.apache.org/
16. Apache HBase, https://hbase.apache.org/
17. Oualid, B., Saida, F., Amine, A., Mohamed, B.: Applying a model driven architecture approach:
transforming CIM to PIM using UML. Int. J. Online Biomed. Eng. (iJOE). 14(9), 170–181
(2018)
18. Arrhioui, K., Mbarki, S., Erramdani, M.: Applying CIM-to-PIM model transformation for
development of emotional intelligence tests platform. Int. J. Online Biomed. Eng. (iJOE).
14(8), 160–168 (2018)
19. Abadi, D., Boncz, P., Harizopoulos, S.: The design and implementation of modern column-
oriented database systems”. Found. Trends Databases 5(3), 197–280 (2012)
20. Angadi, A.B., Angadi, A.B., Gull, K.C.: Growth of New databases & analysis of NOSQL
datastores. Int. J. Adv. Res. Comput. Sci. Softw. Eng. 3(6) (June 2013)
21. OMG, Meta Object Facility (MOF) 2.0 Query/View/Transformation, V1.1 (2011)
22. Acceleo, http://www.eclipse.org/acceleo
23. OMG, UML Infrastructure Final Adopted Specification, version 2.0, September 2003
24. OMG, Meta Object Facility (MOF), version 2.0, OMG (2006)
Data Lake Management Based on DLDS
Approach
Mohamed Cherradi, Anass EL Haddadi, and Hayat Routaib
Abstract Over the past few years, big data is at the center of the concerns of actors
in all fields of activity. The rapid growth of this massive data requires the question
of its storage. Data lakes meet these storage needs, offering data storage without
a predefined schema. In this context, a strategy for building a clear data catalog is
fundamental for any organization that stores big data, helping to ensure the effective
and efficient use of information. Setting up a data catalog in a data lake remains a
complicated task and presents a major issue for data managers. However, the data
catalog is still essential. This article presents the use of XML and JAXB technologies
in the modeling of the data catalog by proposing an approach called DLDS (stands
for Data Lake Description Service) and enables to build a central catalog file that
allows the users to search, locate, understand and query different data sources stored
in the lake.
1 Introduction
The term “data lake” was established by James Dixon founder and former CTO of
Pentaho. According to Dixon, Data lake is very efficient compared to data marts,
offering as an efficient solution to the problem of data silos linked to data marts:
“If you think of a data mart as a store of bottled water, one packaged for easy
consumption, the data lake is a great source of water in its natural state.” [1] The
emergence of the data lake concept in the last five decides has seen increasing interest
compared to the data warehouse as shown in Fig. 1, which represents the number of
times the terms “data lake” and “data warehouse” searched in the last five years on
Google trends.
Data lake is a concept linked to the big data movement, refers to a centralized
storage space of the new generation, which allows to store large amounts of data,
whatever their format, without time limit and strict schema, this model described as
M. Cherradi (B) · A. EL Haddadi · H. Routaib

Data Science and Competitive Intelligence Team (DSCI), Applied Science Laboratory
ENSAH/UAE, AL Hoceima, Morocco
https://doi.org/10.1007/978-981-16-3637-0_48
680 M. Cherradi et al.
Fig. 1 “Data lake” queries in google trends
“schema on-read” [2], and the on-read model schema does not impose any structuring
on the data, thus maintaining their original appearance. This flexibility thus ensures
that the data is used either for analysis purposes in order to make effective decisions.
The data lake can easily turn into a “data swamp” [3], and due to the absence of any
enforced schema, the fact of not imposing a well-defined schema for the data during
their ingestion presents an obvious risk of quality, reliability, trust, etc. In this context,
data governance therefore appears to be one of the major challenges for ensuring the
proper functioning of a data lake. He corresponds to a set of processes, rules, standards
to ensure the effective and efficient use of information in the data lake. It defines the
responsibilities ensuring the quality and security of data within an organization; data
governance ingests in their process data catalog who is combined with governance
and also ensures reliability of the data. A data lake governance provides assurance
that the data is accessible, reliable and of high quality. In contrast, the data catalog
authenticates the data stored in the lake using structured workflows.
The content of this paper is organized like this as follows: Identification of the
related work is described in Sect. 2. Section 3 presents the challenges of data gover-
nance in data lake. Section 4 explains the main factors to consider in the design of
the data lake to solve the issue of data swamp. Section 5 presents a formalization
of the data lake and our proposition about the architecture of data lake adopted in
our approach DLDS. Section 6 gives an overview about technical details associated
with our approach. Section 7 presents the results and a critical study. Finally, Sect. 6
concludes the fruit of our work.
2 Related Work
Some people mistakenly think that a data lake is only version 2 of a data warehouse.
Although in reality, the last two storage techniques are totally different. In fact, in
Data Lake Management Based on DLDS Approach 681
the literature, there is an extremely large agreement in defining a data lake concept,
in which Ref [4] defines data lake as “big data repositories which store raw data and
provide functionality for on-demand integration with the help of metadata descrip-
tions.” On other hand, Ref [5] resumes a data lake like a “massive scalable storage
repository that holds a vast amount of raw data in its native format (as is).” Then, it
is clear that data lake uses a flat architecture that stores data in their native format,
following Ref [6], “Each data entity in the lake is associated with a unique identifier
and a set of extended metadata.”
In the absence of metadata management, data lake can easily be changed into
data swamps [3]. It is important to note that the management of data lakes to a large
extent is based on metadata management systems. Indeed, metadata management is
a crucial element and key component in the architecture of data lakes. In [7], the
authors propose a generic model named MEDAL for managing the metadata of a
data lake. This model adopts a modeling of the metadata system based on graphs.
Any data lake design must integrate a metadata storage strategy [8] to allow
users to search, locate and understand the datasets that are available in the lake. In
this context, our paper aims to propose a comprehensive data catalog that contains
metadata about all assets that have been ingested into our data lake. The compre-
hensive data catalog created based on Java Architecture Xml Binding (JAXB) API,
which makes it possible to match an XML document to a set of classes and vice
versa via serialization/deserialization operations called marshaling/unmarshaling.
The data catalog does not contain the data itself but rather metadata about the data,
like its source, owner and other metadata if available.
3 Challenges of Data Catalog in Data Lake
It is no secret that the amount of data is growing exponentially, we often talk about
big data. If we are talking about big data today, tomorrow we are going to talk about
what we call “Huge Data.” That creates the issue of storage. In this context, data
lake appeared as an efficient and powerful solution for the storage of big data [9].
A data lake acts as a flexible repository that stores different types of data in their
native format, exactly the way it comes in (as is) without any schema defined. One
of the keys to this flexibility is the absence of a strict schema imposed on incoming
flows [10]. Beyond the storage stage, one of the challenges of the data lake is to be
able to facilitate access to data with the objective of carrying out advanced analyzes
and meeting business needs. Seeing this observation, a data catalog appeared as one
of the best ways to facilitate the management of data lakes in order to avoid their
transformation into a data swamp.
One of the key components that must be taken into consideration to effectively
manage a data lake is the incorporation of a data catalog [11], to enable business users
to search, locate and understand the datasets contained in the lake. This is where the
data catalog comes in. A data catalog is a descriptor center where users come to
find the data they need to solve a business or technical problem at hand. The catalog
contains only the descriptive metadata like the source, authors and title.
The data catalog provides the ability to query all assets stored in the lake. The
catalog was also designed to provide a single source of information about the contents
of the data lake. It presents an overview of the content contained in the data lake. It is
interesting to note that there are several tools for building a data catalog [12–15]. Since
most of these tools are paid solutions or are limited in terms of functionality, we have
developed our own data catalog which ensures accessibility and an understanding
of the different sources of the lakes, and it is indeed a centralized repository. We
identify in our paper five major functionalities that must be provided in any data
management system that favors the data catalog that ensures governance, such as
Data enrichment (DE) is one of the most important characteristics which has
the role of supplementing the data, improving it and structuring it, so that it provides
valuable information. Data enrichment is more than correcting errors it also improves
data accuracy [16].
Data indexing (DI) consists of organizing a collection of data sources so that
we can later easily find the one that interests us according to specific keywords.
It is thanks to this functionality that we can simplify and speed up the operations
of searching, sorting, joining, etc. It is very useful for structured and unstructured
textual data [17].
Data versioning (DV) is a method of managing versions of the same data source.
It consists of working directly on the lake’s data source, keeping all the previous
versions. And therefore, it also supports a continuous evolution of the data, in
particular in their schema [18].
Number of descriptive variables (NDV) these variables consist of describing
the data stored in the lakes. These variables are varied; the more we have a large
number of variables, the more we have a clear vision of the content of the lake. It is
very useful to organize the data in a synthetic way [19].
Data accessibility (DA) allows all users to access the different data sources that
exist in the lakes and easily navigate through the location of the different data sources.
Data accessibility is defined as the range in which the different data sources are
available or easily and quickly retrieved [20].
Table 1 presents a comparative and synthetic study of some data management
solution that ensures data governance based on data catalog. The main objective of
this study is to give a global overview of the functionalities provided by our model
compared to others. It appears clearly that our system is complete in terms of these
six features that we have proposed (Table 2).
In our opinion, this lack of functionalities clearly demonstrates the complexity
of design and implementation. But to design an effective system, it turns out that
there are other features that can give great added value. In our approach, we have to
concentrate only on the characteristics which seem necessary to us.
Table 1 Features provided by data catalog tools
Systems DE DI DV NDV DA
Ckan [12]
Collibra [21]
Data Lake Management Based on DLDS Approach
Erwin [14]
CoreDB [15]
DLDS
683
Table 2 Overview of the

Documents ID Metadata
data lake resources
Document 1 Identifiant 1 Metadata 1
… … …
Document n Identifiant n Metadata n
4 What to Consider When Designing a Data Lake
Data lake design is a complex task, but necessary process. It involves a functional
information system capable of managing all the data (structured, semi-structured and
unstructured) of a company in one place, called a “data lake repository.”
When designing a data lake, there are a lot of factors to consider in order to ensure
that it can do what is required of it. Among these factors to be considered during the
database design process, we cite:
1. Metadata storage: Metadata provides information about the structures that
contain the data lake (it is data about data). A data lake design must include a
metadata storage functionality to allow users to locate, search and understand
the datasets in the lake. According to [22], the most significant keyword to
express data lake is “metadata.”
2. Independent of fixed schema: With the problems of increasing data volumes
and the insufficiency of traditional methods, another approach was born, known
by what is called “schema on–read,” which allows for data to be inserted without
applying any schema (upload data as it arrives without any transformations).
With this type of approach, we do not talk about the extract transformation load
(ETL) process, and on the other hand, we talk about extract load transformation
(ELT). With the absence of a fixed schema (schema on-read), data lake can easily
adapt to change, contrariwise schema on write. According to [22], an important
keyword used to exprime data lake also is schema “on-read” (or “on-demand”).
3. Support for different data: The main objective of a data lake is to create a
centralized repository that supports a large amount of data sources in different
types, whatever the format (structured, semi-structured or unstructured), acces-
sible to a variety of end users like data scientist and data analyst. This flexibility
enables organizations to store anything in raw format [23].
4. Scalability and durability: A data lake architecture designed to store different
data sources for long periods of time. That makes data scalability and durability
very important keys to design data lakes effectively, often traditional rdbms we
face the limitations of scalability and durability due to their design [24].
A data lake offers other key factors, but in our article, we looked at the necessary
elements that must be present in any data lake architecture in order to provide faster
query results with low-cost storage.
5 Data Lake Architecture
5.1 Formalization of Our Approach
Before moving to the architecture of a data lake, it is essential to give it a formaliza-

tion. A data lake has a flat architecture [25] for storing heterogeneous and voluminous
data. Each data has a unique identifier and is informed by metadata, as you can see
in Table 1.
In our approach, data lake DL is designed as a set of documents, like this:
DL = { document1, document2, …, documentn / document € Data Sources }

With: document = {(id, metadata) / id € ID and metadata € Metadata)}
It is interesting to note that the formalization of our data lake focuses only on
description metadata, which allows you to identify, select the document and finally
access it. They meet bibliographic objectives by adapting them according to our
needs via our data lake descriptor, and they obviously act from our data catalog file.
To keep things simple and clear, for companies to effectively maximize the data
stored in data lakes, they need to add context to their data based on policies that
classify and identify information in the lake, in order to give an overview of the
contents of the lakes. This cannot be achieved unless we have a global descriptor
named in our paper DLDS (stands for Data Lake Descriptor Service) that can describe
and index data. It also gives the ability to trust the data which it needs for its business
value or to gain a competitive advantage. Figure 2 shows the different features of the
data catalog.
In related research to data quality, most researchers stress the importance of data
quality for the efficient construction of the data catalog. In addition to this, they agree
that every company needs a data catalog to improve its use of its data.
Fig. 2 Main features of the data catalog

Fig. 3 Data lake architecture
5.2 Proposed Architecture
In this paper, we present a new architecture for management different data sources
that exists in the lake via data catalog; as you see in Fig, 3, this architecture contains
the main steps proposed by our approach, like (1) metadata extractor spyder, which
is based on the extraction of the various descriptive attributes of each data source,
while using specific API for each data source in order to guarantee the powerful
extraction of metadata, (2) data catalog, which allows users to explore different data
sources and understand their content via descriptive metadata and (3) catalog query,
which allows to query the data stored in the data catalog according to the requested
need.
6 Technical Details
In this article, we present a new methodology to build a data catalog to provide users
with a guide to discover and understand the content of the different data sources
in the lake, with the main objective of preventing data lake from turning into data
swamp.
The proposed service presents an interface to users and shows how to use the data
lake and how to interact with it. DLDS is based on XML and allows you to precisely
describe the metadata of each document in the lakes such as location, title, authors
and keywords.
DLDS ensures the messaging part in the data lake architecture, and it is used to
provide a description of the data lake resources to enable its use. Indeed, to allow
a customer to consume and perform analyzes according to their needs on the data
sources that exist in the lake, and the latter needs a detailed description of the service
before being able to interact with it. A DLDS provides this description in an XML
document. DLDS plays an important role in the data lake architecture by providing
Fig. 4 Overview of the

DL- Document
resources that make up the
Repository
data lake
Metadata
the description part: It contains all the information necessary to invoke the service it
describes.
As described in Sect. 4.2, the data lake has two essential parts (data and metadata),
architectured as shown in Fig. 4.
In order to efficiently organize all these documents stored in the lake and benefit
from the wealth of these data, DLDS has been designed to meet this need. It is an
XML document which describes the lake documents independently of any language.
This shows the flexibility of our approach. Using DLDS as an xml-based data catalog,
it allows tools from different systems, platforms and languages to use the content of
a WSDL to generate code to consume different source data in the lake.
To respond to the approach we have designed, we based on the jaxb specification,
as shown in Fig. 5.
The major objective of relying on this architecture is to facilitate the construction
of the data catalog, by converting an object into an XML document. This document
is linked to an XML schema which gives an overview of our data container, and it
describes all the elements necessary to interact with our data lake. The description
of this schema file is described below.
Fig. 5 Jaxb specification

architecture
In this section, we present the result of our DLDS approach, which gives birth to a
new approach for structuring the different data sources existing in the lake via data
catalog. Figure 6 shows the data catalog we designed, menu with its associated XML
schema to standardize it.
In fact, this schema defines the existing contract between users and our data lake.
To interact with data lake, we will use the catalog that we have generated based on
the descriptive metadata of each data source. Figure 7 shows an extract of the data
catalog that we built.
This section also aims to present the critical aspect of our approach in the form
of a discussion. It is true that our approach presents an original idea and makes a
Fig. 6 Generated XML schema associated with the DLDS
Fig. 7 Extract from the data catalog that we built

strong contribution to the scientific literature, especially in terms of the structuring of

heterogeneous data stored in the lake via our approach called DLDS. But this does not
prevent us from saying that our article is complete, on the other hand perhaps missing
the following points, which can be considered as the subject of future research:
• Demonstration of metastore link to an available XML standards and semantic
vocabularies.
• Connectors to storage space, such as: HDFS, Hive, Cassandra, MongoDB or
others as needed, instead of an XML descriptor.
• Leveraging the power of machine learning to enrich metadata extraction process.
• etc.
8 Conclusion
Improving the quality of data organization via data catalog is becoming a brilliant
technique in the world of heterogeneous data management. Indeed, the construction
of the data catalog with a contract in the form of an XML schema considered to be a
reference architecture to fully understand and interact with the different data sources
exists in the lake.
This paper should not be interpreted as finalized. But rather it is the starting point
or the birth of a new approach that deserves to be complemented by other work whose
objective is to effectively manage data lake.
References
1. Dixon, J.: Pentaho, Hadoop, and Data Lakes | James Dixon’s Blog (2010). https://jamesdixon.
wordpress.com/2010/10/14/pentaho-hadoop-and-data-lakes/. Accessed 10 Feb 2021
2. Mathis, C.: Data lakes. Datenbank-Spektrum 17, 1–5 (2017). https://doi.org/10.1007/s13222-
017-0272-7
3. Suriarachchi, I., Plale, B.: Crossing Analytics Systems: A Case for Integrated Provenance in
Data Lakes (2016)
4. Hai, R., Geisler, S., Quix, C.: Constance: an intelligent data lake system. In: Proceedings of the
2016 International Conference on Management of Data. Association for Computing Machinery,
USA, pp. 2097–2100, New York, NY, (2016)
5. Miloslavskaya, N., Tolstoy, A.: Big data, fast data and data lake concepts. Procedia Comput.
Sci. 88, 300–305 (2016). https://doi.org/10.1016/j.procs.2016.07.439
6. Rangarajan, S., Liu, H., Wang, H., Wang, C.-L.: Scalable architecture for personalized health-
care service recommendation using big data lake. In: Beheshti, A., Hashmi, M., Dong,
H., Zhang, W.E. (eds.) Service Research and Innovation, pp. 65–79. Springer International
7. Scholly, E., Sawadogo, P.N., Favre, C., Ferey, E., Loudcher, S., Darmont, J.: Système de
métadonnées d’un lac de données : modélisation et fonctionnalités (2019)
8. Sawadogo, P.N., Darmont, J.: On data lake architectures and metadata management. J. Intell.
Inf. Syst. 56, 1–24 (2021). https://doi.org/10.1007/s10844-020-00608-7
9. Khine, P., Wang, Z.: Data lake: a new ideology in big data era. ITM Web Conf. 17, 03025
(2018). https://doi.org/10.1051/itmconf/20181703025
10. Sawadogo, P.N., Scholly, E., Favre, C., Ferey, E., Loudcher, S., Darmont, J.: Metadata Systems
for Data Lakes: Models and Features (2019)
11. Chen, M.: Why Data Lakes Need a Data Catalog (2019). https://blogs.oracle.com/bigdata/why-
data-lakes-need-a-data-catalog. Accessed 15 Feb 2021
12. ckan ckan. In: Data Cat. https://ckan.org/. Accessed 15 Feb 2021
13. Collibra Data Catalog on-demand demo. In: Data Manag. Data Cat. https://www.collibra.com/
download/data-catalog-demo. Accessed 15 Feb 2021
14. Erwin Data Catalog Free Demo. In: Erwin Inc. https://erwin.com/erwin-data-catalog-free-
demo/. Accessed 15 Feb 2021
15. Beheshti, A., Benatallah, B., Nouri, R., , V., Xiong, H., Zhao, X.: CoreDB: a Data Lake Service,
pp. 2451–2454 (2017)
16. Azad, S., Wasimi, S., Ali, A.B.M.: Business Data Enrichment: Issues and Challenges, pp. 98–
102 (2018)
17. Singh, K., Paneri, K., Pandey, A., Gupta, G., Sharma, G., Agarwal, P., Shroff, G.: Visual
Bayesian Fusion to Navigate a Data Lake (2016)
18. Hellerstein, J.M., Sreekanti, V., Gonzalez, J.E., Dalton, J., Dey, A., Nag, S., Ramachandran,
K., Arora, S., Bhattacharyya, A., Das, S., Donsky, M., Fierro, G., She, C., Steinbach, C.,
Subramanian, V., Sun, E.: Ground: A Data Context Service. 12
19. Yellapu, V.: Descriptive statistics. Int J. Acad Med 4, 60 (2018). https://doi.org/10.4103/IJAM.
IJAM_7_18
20. Pipino, L.L., Lee, Y.W., Wang, R.Y.: Data quality assessment. Commun ACM 45, 211–218
(2002). https://doi.org/10.1145/505248.506010
21. Collibra Trusted data for your entire organization. In: Collibra. https://www.collibra.com/.
Accessed 16 Feb 2021
22. Chihoub, H., Madera, C., Quix, C., Hai, R.: Architecture of Data Lakes. pp. 21–39 (2020)
23. Anne Laurent Dominique Laurent Cédrine Madera (2020) Data Lakes | Wiley Online Books.
book/https://onlinelibrary.wiley.com/doi/10.1002/9781119720430. Accessed 16 Feb 2021
24. Bhawkar, A.: A Comparative Study to Analyze Scalability, Availability and Reliability of
HBase and MongoDB (2018). In: ResearchGate. https://www.researchgate.net/publication/
330675690_A_comparative_study_to_analyze_Scalability_Availability_and_Reliability_of_
HBase_and_MongoDB. Accessed 16 Feb 2021
25. Alrehamy, H., Walker, C.: Personal Data Lake With Data Gravity Pull (2015)
Evaluation of Similarity Measures
in Semantic Web Service Discovery
Mourad Fariss, Naoufal El Allali, Hakima Asaidi, and Mohamed Bellouki
Abstract The semantic Web service discovery is the process of finding services
that could potentially meet the consumer requirements by choosing between several
services. The matchmaking between the consumer request and the semantic Web
services is the main task of any semantic Web service discovery mechanism. To
make this, many works use similarity measures to choose a similar semantic Web
service and the consumer’s request. In this paper, we assess similarity measures
to help us define their problems and determine which measure is most appropriate
for Web service discovery approaches. For this matter, we used a test collection of
semantic Web services with different domains. The results of our evaluation show
the weakness of similarity measures to have an interesting efficiency.
1 Introduction
Service-oriented architecture (SOA) is an architectural style for designing distributed

applications using functionality implemented by third-party providers. In a SOA, the
service consumer satisfies its specific needs by using services offered by service
providers. One concrete technology used for implementing SOA is Web service.
According to the W3C, a Web service is defined as “a software system designed to
support interoperable machine-to-machine interaction over a network” [1]. Its inter-
face can be described as a Web service description language (WSDL) document that
contains structured information about the Web service’s location, its offered oper-
ations and the input/output parameters. Interface descriptions (WSDL documents)
enable Web services to be discovered, used by applications or other Web services
and composed into new more complex Web services.
The Web service architecture consists of three entities: the service provider, the
service registry and the service consumer. The service provider simply creates or
offers the Web service. The service provider describes the Web service in a standard
M. Fariss (B) · N. El Allali · H. Asaidi · M. Bellouki

Laboratory MASI, FPN, UMP, Nador, Morocco
e-mail: m.fariss@ump.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_49
692 M. Fariss et al.
Fig. 1 Web service standard

model
format, often in XML, and publishes it in a central service registry. The service
registry contains additional information about the service provider, such as the
address and contact details of the company providing the services, as well as tech-
nical details about the service. The service consumer extracts the information from
the registry and uses the resulting service description to bind and call the Web service
(Fig. 1).
The Universal Description Discovery and Integration (UDDI) registry was
proposed for the publication of services. The service consumer can access this registry
in order to find the best service that meets their needs. Since UDDI uses syntactic
information and does not use semantic information, the search for the most appro-
priate service is limited in the sense that consumers cannot make requests asking
for specific desirable properties such as quality of service parameters related to
reliability, performance, security, response time, etc.
As long as several researches have adopted the semantic description of Web
services [2, 3] [4], a new problem has appeared to measure the degree of similarity
between the service consumer and the Web services registered in the database. In this
paper, we propose an evaluation of the most used similarity measures in Web service
discovery. Section 2 presents the problem of Web service discovery. Section 3 cites
the related works for the similarity measures on the Web service discovery. Session 4
describes the similarity measures studied. The evaluation of the similarity measures
studied is represented in Session 5. Finally, we present the conclusion and future
works.
2 Web Service Discovery
The Web service discovery process is carried in three stages. Firstly, the service
provider publishes the Web service in public repositories by registering the Web
service description file written in WSDL [5]. In the second step, the service consumer
sends a request with the requirements in predefined format to the Web service registry.
Evaluation of Similarity Measures in Semantic Web Service … 693
Fig. 2 Web service discovery. Since it has been observed that the search capabilities of the
current discovery mechanisms are limited because they are mostly based on keyword matching, the
service consumer searches the Web service in the UDDI register and submits the requirements with
keywords. This requires a different mechanism, which includes the location of Web services based
on the features they provide. Semantic technologies in Web service play an important role in the
seamless integration of different services based on different terminologies
Web service matcher finds the corresponding Web service candidates to the consumer
request. Finally, the selection and invocation of one of the retrieved Web services
(Fig. 2).
The combination of the theory of semantic Web and Web services gives rise to what
is known as semantic Web services. There are several approaches to add semantic
information to services such as OWL-S [6], WSDL-S [7] and WSMO [8]. The term
“ontology” is used in many fields such as artificial intelligence, software engineering,
information architecture and many others. Ontology is a structural framework for
organizing the representation of information or knowledge about the world or part of
a world, which describes a set of concepts, its properties and the relationship between
them. Ontology is a “formal specification of a shared conceptualization,” providing
a shared vocabulary, a taxonomy for a particular domain that defines objects, classes,
its attributes and their relationships.
An ontology provides semantics to Web services and its resources and can greatly
improve discovery mechanisms. To implement the vision of the semantic Web, the
researchers proposed several languages, algorithms and architectures. Semantic Web
services are part of the semantic Web because it uses markings that make the data
machine readable [9]. Semantic Web services use standards such as OWL-S, WSDL-
S, WSMO, OWLS-LR and others.
3 Related Works
Several works are concentrated on semantic Web service discovery using similarity
measures [10–13]. In this section, we give some studies that use the three evaluated
similarity measures.
The Web service discovery method based on semantic and clustering was proposed
[14], and the similarity measure used a WordNet-based semantic similarity measure-
ment. It uses the cosine similarity to compute the score of similarity on this method.
The cosine measure was used to compute similarities and retrieve relevant WSDL
service descriptions [15]. To do this, the authors create the service request vector
according to the domain ontology and then project the description vectors and the
request vector. Some researchers were present an approach for automated service
discovery based on ontology [16]; this approach adds a filtering method based on
logical reasoning before using the cosine similarity in the matching algorithm for the
Web services filtered. In [17], Wu and Palmer’s similarity was used to compute the
similarity between the document word vectors obtained by analyzing the OWL-S
Web service documents, before the LDA clustering in the Web service discovery
process. The same measure of similarity was used to compute the similarity between
words related to the requested query and service parameters on the proposed approach
[18]. The Web service discovery approach combines the LDA clustering and the k-
Medoids to reduces search space for Web service discovery. The Jaccard coefficient
was used to calculate the similarity between Web services in a method to improve the
Web service discovery using the user’s experiences with similar queries [19]. Based
on the experimental results of measuring the performance of similarity metrics for
text information retrieval provided by Cohen et al. [20], the authors of [21] selected
the top-performing ones to build the OWLS-MX variants on the proposed approach
of Web service discovery. These symmetric token-based string similarity measures
are the cosine similarity, the extended Jacquard similarity, the intentional loss of
information and the Jensen–Shannon information divergence.
Researchers are still finding a difficulty to choose the most appropriate similarity
measure to solve the Web service discovery problem. These measures aim to measure
the similarity between the consumer request and the Web service available in the Web
services registry. This paper provides a tool to evaluate and compare the similarity
measures most used in the literature, to help researchers choose the most satisfactory
of them for Web services defined in OWLS.
4 Similarity Measures
Semantic similarity between concepts is a method of measuring semantic similarity

or the semantic distance between two concepts according to a given ontology. In other
words, semantic similarity is used to identify concepts that have common “charac-
teristics”. Although a human does not know the formal definition of the connection
between concepts, he can judge the relationship between them. For example, a young
child may say that “apple” and “peach” are more related than “apple” and “toma-
toes”. These few concepts are interrelated, and their definition of structure is formally
called the “is-a” hierarchy. Semantic similarity methods are used intensively for most
applications of intelligent systems with knowledge-based and semantic information
retrieval (identify an optimal match between query terms and documents) [22] [23],
the ambiguity of the senses [24] and bioinformatics [10]. Semantic similarity and
semantic relatedness [25] are two related words, but the semantic similarity is more
specific than relatedness and can be considered as a type of semantic relatedness.
For example, “Student” and “Teacher” are related terms, which are not similar. All
similar concepts are related, but the reverse is not always true.
Semantic similarity and semantic distance are defined in the reverse direction. Let
C 1 and C 2 be two concepts belonging to two different nodes n1 and n2 in a given
ontology, and the distance between the nodes determines the similarity between these
two concepts. Both n1 and n2 can be considered as an ontology that contains a set of
synonymous concepts and consequently. Two terms are synonymous if they are in
the same node, and their semantic similarity is maximized.
The use of ontologies to represent the concepts or terms (humans or computers)
that characterize different communication sources is useful in making knowledge
comprehensible. Furthermore, it is possible to use different ontologies to represent
the concepts for each source of knowledge. Then, the mapping or comparing concepts
based on the same or different ontologies ensures the exchange of knowledge between
concepts. The mapping must find the similarity between terms or concepts based
on domain-specific ontologies. The similarity between concepts or entities can be
identified if they share common attributes or if they are related to other semantically
related entities in an ontology [26].
The algorithm used to compute the similarity between two concepts, for as the
Web service and the consumer request, is presented as follows:
Algorithm: Semantic Similarity Score Computation

Input: WSs and CR // WSs is the Web services, and CR is the consumer request
Output: Semantic Similarity Score
2. Convert CR to vector Vr.
3. for i = 1 to n do
//n is the number of web services
4. compute Measure_Similarity(V[i], Vr).
//Measure_Similarity is replaced by the measure tested
5. end
To calculate the similarity score between Web services and the consumer request,
first, we convert the Web services and the customer request to a vector, and then, we
calculate the similarity score according to the wanted similarity measure for each
Web service.
In the following, the definition of the similarity measures evaluated in this paper.
4.1 The Cosine Similarity
In the process of semantic Web service discovery, it is very important to find out
the semantic Web service which can meet the consumer’s needs in specific contexts
through the specific keyword query based on the keyword of the text. The cosine simi-
larity algorithm uses the cosine of angle between two different vectors in the vector
space to determine the difference on content between two vectors, it is mainly based
on consumer’s own preference and the level of difference between provided Web
services, determine a semantic Web service that ultimately conforms to consumer’s
context, and then, feed the semantic Web service back to the consumer to meet the
different needs of the consumer in different contexts (Fig. 3).
In two dimensions, the angle cosine of the vector a and the vector b is calculated
as follows:
a·b
cos(α) =
a × b
(x1 , y1 ) · (x2 , y2 )
=
x12 + y12 × x22 + y22
x1 x2 + y1 y2
= (1)
x12 + y12 × x22 + y22
In the case of multidimensional, the angle cosine of the vector a and the vector b
is calculated as follows:
n
i=1 (x i × yi )
cos(α) = n n (2)
i=1 (x i ) × i=1 (yi )
2 2
Fig. 3 Angle cosine of the

vector a and the vector b
Since the cosine similarity algorithm focuses on the difference of vectors’ direc-
tions, it is not sensitive to their size, so it is mainly used by consumers to determine
whether the content of the Web service is interested.
4.2 Wu and Palmer’s Similarity Measure
Path-based similarity measure usually utilizes the information of the shortest path
between two concepts, of the generality or specificity of both concepts in ontology
hierarchy and of their relationships with other concepts. Wu and Palmer [27] present
a similarity measure finding the most specific common concept that subsumes both of
the concepts being measured. The path length from the most specific shared concept
is scaled by the sum of IS-A links from it to the compared two concepts.
SW &P (C1 , C2 ) = 2N /(N1 + N2 + 2N ) (3)
where N 1 and N 2 are the distance that separates, respectively, the concept C 1 and C 2
from the specific common concept, and N is the distance which separates the closest
common ancestor of C 1 and C 2 from the root node.
4.3 Jaccard Index
The Jaccard index, also known as the Jaccard similarity coefficient, is a statistic used
for comparing the similarity and diversity of sample sets. The Jaccard coefficient
measures similarity between sample sets and is defined as the number of common
objects divided by the total number of objects, minus the number of common objects.
g (J ) (xa , xb ) = xaT xb /(xa 22 + xb 22 − xaT xb ) (4)
5 Experimentation and Results
In this section, we will present the experimentation results done on the similarity
measures evaluation, using the well-known service repository of OWLS-TC v4.0
test collection. There are 1083 semantic Web services written in OWL-S 1.1 and 42
requests. There are nine service domains: education, medical, food, travel, commu-
nication, economy, weapon, geography and simulation. Table 1 shows the details of
OWLS-TC v4.0.
Table 1 Details of
Domains rainfall Number of services Number of requests
OWLS-TC v4.0
data
Education 286 6
Medical care 73 1
Food 34 1
Travel 197 6
Communication 59 2
Economy 395 12
Weapon 40 1
Geography 60 10
Simulation 16 3
Some services appear in more than one category. Therefore, the number of services
is 1083 if we consider just the first occurrence of each service and 1140 if we consider
repetitions across different categories.
To evaluate the three measures of similarity cited below, we used five consumer’s
requests and across about 1000 Web services. For each consumer, we calculated
the degree of similarity with different similarity measures, and the execution time
between each request of the Web services returned during the search. In order to
judge the effectiveness of similarity measures, we compared the similarity measures
with the number of Web service of each domain. The following Table 2 shows the
samples chosen for the consumer request:
The following Table 3 represents some results of similarity measures for the
consumer 1 request “car_price_service” with different services from all domains of
dataset.
From the results of the similarity measurements tested, we can conclude the
following points:
• The time to calculate similarity measures remains acceptable for all the similarity
measures tested.
• We cannot say that there is an agreement between the different similarity measures,
which will influence the Web service discovery process based on their similarity
measures (Fig. 4).
Table 2 Consumer’s request

Consumer request Request name Domains rainfall data
Consumer 1 car_price_service Travel
Consumer 2 hospital_investigating_service Medical care
Consumer 3 book_price_service Education
Consumer 4 getLocationOfUSCity Geography
Consumer 5 preparedfood_price_service Food
Table 3 Similarity result for consumer 1
Web services Domains rainfall data Similarity measure
Cosine Time (µs) Wu and Time (µs) Jaccard Time (µs)
similarity Palmer’s index
similarity
AcademicBookNumberOrISBNSearch Education 30 1 15.7 3 13.9 1
HealthInsurance_service Medical care 73.7 1 10.4 3 56.3 2
corporation_apple_service Food 11 3 30.8 1 66.4 1
_internalchange_EarthSystemservice Travel 100 0 30.1 2 100 2
_filmvideomediaDiscoveryChannel_service Communication 33.5 1 89.3 2 43.4 3
4wheeledcar_price_service Economy 88.6 1 78.8 2 93.2 1
Evaluation of Similarity Measures in Semantic Web Service …
governmentweapon_funding_service Weapon 23.4 2 12.4 4 28.4 1

GetCoordinatesOfAddress Geography 100 1 88.4 2 23.4 0
greenLight_to_off Simulation 34.5 2 45.7 3 67.7 1
699
Fig. 4 Similarity measures of web services for consumer 1
– For example:
– For Consumer 1, the degree of similarity of the service
“HealthInsurance_service” is 73.7% for the cosine similarity and 56.3%
for the Jaccard index, but for the Wu & Palmer’s Similarity, it is 10.4%
– For Consumer 1, the degree of similarity of the “GetCoordinatesOfAddress”
service is 100% for cosine similarity and 88.4% for the Wu & Palmer’s
Similarity, but for the Jaccard index, it is 23.4%,
• The number of Web services discovered varies depending on the measure of simi-
larity used, because of the degree of similarity of each service and the consumer’s
request (Fig. 5).
Fig. 5 Number of
discovered services for
consumer 1
Fig. 6 Number of web services discovered for each consumer
• The results obtained with Consumer 1 for the discovered Web services based on
similarity measures remain reliable for five different consumer’s requests tested
in different domains (Fig. 6).
• The results show that all the similarity measures tested do not give effective
results. The similarity values obtained do not always meet the consumer’s request.
Furthermore, the Web services discovered using the tested similarity measures
stay far away from the number of Web services in each domain.
To measure the accuracy of each similarity measure, we calculate the precision
and the recall. Precision is the number of correct results divided by the number of
all returned results. Recall is the number of correct results divided by the number of
results that should have been returned as shown in the following equations:
Precicion(P) = Correct Relevant Services Found/Total Relevant Services Found

(5)
Recall(R) = Correct Relevant Services Found/Relevant Services should be Found

(6)
• The recall and the precision presented in Fig. 7 show the weakness of the three
studied similarity measures, namely the cosine similarity, the Wu & Palmer’s
Similarity and the Jaccard index.
In this paper, we proposed an evaluation of the three most used similarity measures
in the semantic Web service discovery. Indeed, our objective was to give the mean
to compare and evaluate similarity measures with existing measures. The studied
Fig. 7 Comparison of recall

and precision of similarity
measures
similarity measures show a weakness of decision either at the level of degree of

similarity and also at the level of precision and recall in the Web service discovery.
In addition, we aim to
• Suggest solutions to solve the problem of similarity measures.
• Study other measures of similarity that exist and integrate them into this
evaluation.
• Introduce the concept of quality of service at the task of the Web service discovery.
References
1. Moreau, J.J., Chinnici, R., Ryman, A., Weerawarana, S.: Web services description language
(WSDL) version 2.0 part 1: core language. Candidate Recomm. W3C, 7 (2006)
2. Fethallah, H., Ismail, S.M., Mohamed, M., Zeyneb, T.: “An outranking model for web service
discovery.” Int. Conf. Math. Inf. Technol. (ICMIT) 2017, 162–167 (2017)
3. Fariss, M., El Allali, N., Asaidi, H., Bellouki, M.: Review of Ontology Based Approaches for
Web Service Discovery. Springer International Publishing (2019)
4. Malburg, L., Klein, P., Bergmann, R.: “Using Semantic Web Services for AI-Based Research
in Industry 4.0,” arXiv Prepr. arXiv2007.03580 (2020)
5. Christensen, E., Curbera, F., Meredith, G., Weerawarana, S. et al.: “Web Services Description
Language (WSDL) 1.1.” Citeseer (2001)
6. Martin, D., et al.: “OWL-S: Semantic markup for web services,” W3C Memb. Submiss. 22(4)
(2004)
7. Akkiraju, R., Farrell, J., Miller, J.A., Nagarajan, M., Sheth, A.P., Verma, K.: “Web Service
Semantics-wsdl-s,” (2005)
8. Roman, D., et al.: Web service modeling ontology. Appl. Ontol. 1(1), 77–106 (2005)
9. Malaimalavathani, M., Gowri, R.: “A survey on semantic web service discovery.” Int. Conf.
Inf. Commun. Embed. Syst. ICICES 2013, 222–225 (2013)
10. Ehsani, R., Drabløs, F.: TopoICSim: A new semantic similarity measure based on gene ontology.
BMC Bioinf. 17(1), 1–14 (2016)
11. Wu, J., Chen, L., Zheng, Z., Lyu, M.R., Wu, Z.: Clustering web services to facilitate service
discovery. Knowl. Inf. Syst. 38(1), 207–229 (2014)
12. Alwasouf, A.A., Deepak, K.: “Research challenges of web service composition, software
engineering,” Adv. Intell. Syst. Comput., vol. 731, 2019.
13. El Allali, N., Fariss, M., Asaidi, H., Bellouki, M.: “Semantic web services composition model
using ant colony optimization,” 4th Int. Conf. Intell. Comput. Data Sci. ICDS 2020 (2020)
14. Wen, T., Sheng, G., Li, Y., Guo, Q.: “Research on web service discovery with semantics and
clustering,” Proc.—2011 6th IEEE Jt. Int. Inf. Technol. Artif. Intell. Conf. ITAIC 2011, 1, 62–67
(2011)
15. Paliwal, A.V., Adam, N.R., Bornhövd, C.: “Web service discovery: adding semantics through
service request expansion and latent semantic indexing,” Proc.—2007 IEEE Int. Conf. Serv.
Comput. SCC 2007, no. Scc, pp. 106–113 (2007)
16. Fang, M., Wang, D., Mi, Z., Obaidat, M.S.: Web service discovery utilizing logical reasoning
and semantic similarity. Int. J. Commun. Syst. 31(10), 1–13 (2018)
17. Zhao, H., Chen, J., Xu, L.: “Semantic web service discovery based on LDA clustering,” In Web
Information Systems and Applications, pp. 239–250 (2019)
18. Jalal, S., Yadav, D.K., Negi, C.S.: “Web service discovery with incorporation of web services
clustering,” Int. J. Comput. Appl., 0(0), 1–12 (2019)
19. Nayak, R.: Data mining in Web services discovery and monitoring. Int. J. Web Serv. Res. 5(1),
63–81 (2008)
20. Cohen, W.W., Ravikumar, P., Fienberg, S.E.: “A comparison of string distance metrics for
name-matching tasks,” Proc. IJCAI-2003 Work. Inf. Integr. Web (2003)
21. Klusch, M., Fries, B., Sycara, K.: “Automated semantic web service discovery with OWLS-
MX,” in Proceedings of the Fifth International Joint Conference on Autonomous Agents and
Multiagent Systems—AAMAS ’06, p. 915 (2006)
22. Budan, I.A., Graeme, H.: Evaluating wordnet-based measures of semantic distance. Comuta-
tional Linguist. 32(1), 13–47 (2006)
23. Sim, K.M., Wong, P.T.: “Toward agency and ontology for web-based information retrieval,”
IEEE Trans. Syst. Man, Cybern. Part C (Applications Rev., vol. 34, no. 3, pp. 257–269, 2004.
24. Patwardhan, S.: Incorporating Dictionary and Corpus Information into a Context Vector
Measure of Semantic Relatedness. University of Minnesota, Duluth (2003)
25. Gracia, J., Mena, E.: “Web-based measure of semantic relatedness,” in International Conference
on Web Information Systems Engineering, pp. 136–150 (2008)
26. Li, Y., Bandar, Z.A., McLean, D.: An approach for measuring semantic similarity between
words using multiple information sources. IEEE Trans. Knowl. Data Eng. 15(4), 871–882
(2003)
27. Palmer, M.: “Verb semantics and lexical zhibiao W u,” 32nd Annu. Meet. Assoc. Comput.
Linguist., pp. 133–138 (1994)
Knowledge Discovery for Sustainability
Enhancement Through Design
for Relevance
Abla Chaouni Benabdellah, Asmaa Benghabrit, Imane Bouhaddou,
and Kamar Zekhnini
Abstract Data is collected and accumulated at a dramatic pace from many different
resources and services across a wide variety of fields, particularly for industrial
companies. Hence, to capture long-term revenue, sustainable assessment and help
humans extract valuable knowledge from the rapidly increasing amounts of digital
data, companies must adopt a new generation of analytical theories and methods. A
well-known fundamental task is the use of knowledge discovery in databases (KDD).
In this respect, the aim of this paper is to adopt the KDD process to extract informa-
tion from data that are generated through the use of different design for X techniques
named Design for Relevance. Since we are looking to find a structure for sustain-
ability enhancement in an unlabeled dataset related to collaborative product devel-
opment, clustering is the most appropriate data mining (DM) task in our context.
However, with the modified applications for various domains, several clustering
algorithms have been provided. This multiplicity makes it difficult for researchers
to define both the appropriate algorithms and the appropriate measures of validity.
To deal with these issues, a related work focusing on comparing various clustering
algorithms for real industrial datasets is presented. After that, for Design for Rele-
vance dataset and by following the KDD process, several clustering algorithms were
implemented and compared using several internal validity indices. In addition, we
highlighted the best performing clustering algorithm that gives us successful clusters
for the dataset to achieve improvement in sustainability.
1 Introduction
Over the past decade, scientific consensus on the importance of sustainability has
tended to exist, both at the organizational level [1] and at the national and global level
A. C. Benabdellah (B) · I. Bouhaddou · K. Zekhnini

L2M3S Laboratory, ENSAM, Moulay Ismail University, 50500 Meknes, Morocco
A. Benghabrit
LMAID Laboratory, ENSMR, Mohamed V University, 10070 Rabat, Morocco
https://doi.org/10.1007/978-981-16-3637-0_50
706 A. C. Benabdellah et al.
[2, 3]. Organizations are not only supposed to be sustainable within this framework,
but it is also in their interest to do so [4]. Hence, companies are currently in a
state of change. In fact, additional criteria have been imposed on new products
regarding consumer concerns about total cost of ownership, perceived efficiency, cost
savings, long-term product support, and environmental effects. Different markets,
digital demand, changing business environment, uncertainty and cost pressure, labor
costs, are considered as the new challenges that companies have to face in order
to stay competitive [5]. Thus, companies must commit to bringing new goods to
market on a recurring basis to gain long-term sales and sustainable competitive
advantage by knowing what consumers want [6] and considering a new generation
of computational theories and tools to assist human in extracting useful information
from the rapidly growing volumes of digital data.
Data produced by machines and computers, product lifecycle management (PLM)
solutions, design for X techniques (DFX), supply chain processes, product creation,
production planning systems, or quality and inventory management systems have
reached an enormous volume of more than 1,000 Exabytes per day in line with the
massive advancement and development of internet and online world technologies [7].
As a consequence, a well-known fundamental task for discovering useful knowledge
from data and involves the evaluation and interpretation of the patterns to make
decision is the use of knowledge discovery in databases (KDD).
The main objective of KDD is to extract high-level knowledge from low-level
information, or in other words, to automatically process large quantities of raw data,
identify the most significant and meaningful patterns, and present these as knowledge
appropriate for achieving the user goals [8]. This process is mainly based on data
mining step that is responsible for extracting patterns or generating models from the
output of the data processing step and then feeding them into the decision-making
step, which takes care of transforming its input into useful knowledge.
Traditional DM approaches have worked successfully in modeling variables of
interest, and their core technologies have been identified and categorized into mining
technologies for clustering, classification, and frequent patterns [9]. In this regard,
three main considerations are target, data characteristics, and mining algorithm in
choosing the applicable mining technologies for the problem to be solved by the
KDD technology. However, since we are looking to find a structure for sustainability
enhancement in an unlabeled dataset related to collaborative product development
while considering different virtues X, clustering is the most appropriate method
in our context. Indeed, without a priori knowledge of the structure of a database,
only unsupervised classification can automatically detect the presence of relevant
subgroups (or clusters).
Clustering is considered to be one of the most challenging tasks due to its unsu-
pervised nature [10]. The numerous algorithms developed by the researchers over
the years lead to different data clusters, including for the same algorithm, the selec-
tion of different parameters or the order of presentation of data objects may have
a major impact on the final clustering partitions. However, with the vast number of
surveys and comparative studies concerning the clustering algorithms, exploring the
algorithm that cluster industrial sparse dataset still remains an open issue. Therefore,
Knowledge Discovery for Sustainability Enhancement … 707
to deal with these issues, a related work focusing on comparing various clustering
algorithms for real industrial datasets is presented. After that, by considering the
Design for Relevance dataset which implements the sustainability concerns while
considering different virtues in the product development and by following the KDD
process, several clustering algorithms were implemented and compared using several
internal validity indices. In addition, we highlighted the best performing clustering
algorithm that gives us successful clusters for the dataset to achieve improvement in
sustainability.
This paper is structured as follows; Sect. 2 presents the basic terminologies.
Section 3 presents a related work over the past twenty years on research and review
articles with a focus on comparative research. By providing a categorized framework,
Sect. 4 provides a description of the considered clustering algorithms which will be
evaluated properly. Section 5 detailed the KDD process for Design for Relevance
dataset. Section 6 concludes the paper and discusses future research.
2 Preliminaries
The concepts of “Data Mining” (DM) and “Knowledge Discovery in Databases”

(KDD) are confused by many researchers. More clearly, in addition to being used as
one of the phases of the KDD system, some consider the DM as a synonym for KDD.
In this section, key terms are defined to prepare the groundwork for the subsequent
study.
2.1 Knowledge Discovery in Databases (KDD)
KDD has evolved, and continues to evolve, from the intersection of research in such
fields as databases, machine learning, pattern recognition, statistics, artificial intel-
ligence, and reasoning with uncertainty, knowledge acquisition for expert systems,
data visualization, machine discovery [8], scientific discovery, information retrieval,
and high-performance computing. Theories, algorithms, and methods from both of
these areas are integrated into KDD software systems.
The literature analysis shows that there is no unified definition of KDD [8, 9, 12–
14]. But generally, there is a common consensus that KDD is essentially an iterative
and interactive process of discovering useful knowledge from a collection of data.
More clearly, the goal of the KDD process is to transform data (large, multifaceted,
stored on media that can be transmitted in different formats) into information. It is
possible to convey this information in the form of general concepts that enrich the
user’s semantic field in relation to a query that concerns him. For decision making,
they can be represented as a mathematical or logical model. Based on this, Fayyad
et al. [11] define KDD as a process based on nine steps such as (Fig. 1):
Fig. 1 Overview of the steps constituting the KDD process [11]
• Developing an understanding of the application domain and the related prior

knowledge and, define the purpose of the KDD process from the customers point
of view.
• Creating a target dataset.
• Data cleaning and preprocessing.
• Data reduction and projection
• Matching the goals of the KDD process to a particular data mining method.
• Exploratory analysis and model and hypothesis selection.
• Data mining.
• Interpreting mined patterns.
• Acting on the discovered knowledge.
The KDD process requires crucial steps. In [15], Yoong and Kerschberg claim that
the discovery of information critically depends on how well a database is described
and how the current and discovered knowledge is constantly generated. In this way,
the phase description of the KDD process channel has a strong effect on the final
mining results. For example, for mining, not all the attributes of the data are useful.
For example, for mining, not all the attributes of the data are useful. The result is that
Date Mining algorithms can find it difficult to find useful information if the selected
attributes are unable to fully reflect the data characteristics [9]. However, the question
arises, what is exactly data mining?
2.2 Data Mining (DM)
Fayad et al. in [11] define DM as a step in the KDD process involving the application
of data analysis and the discovery of algorithms that generate a specific enumeration
of patterns (or models) over the data. But any DM process needs a previous phase in
data processing, also known as data warehousing (DW) [11]. DW refers to collecting
and cleaning transactional data to make them available for online analysis and deci-
sion support. DW helps set the stage for KDD in two important ways: data cleaning
and data access.
Data mining is the application, under human control, of low-level mining methods,
which are in turn defined as algorithms designed to analyze data or to extract specific
patterns (or patterns) from a large amount of data [16]. More clearly, DM focuses
mostly on discovering knowledge in association with six basic tasks:
• Clustering or segmentation aims at grouping dataset (objects) into clusters such

that similar objects are aggregated together in the same cluster while dissimilar
ones should be belonging to different clusters. Furthermore, from an optimization
perspective, the main goal of clustering is to maximize both the (internal) homo-
geneity within a cluster and the (external) heterogeneity among different clusters
[17].
• Classification includes evaluating and assigning to a given class the attributes
of a specific object. For example, to discern suspicious characters at an airport
security check, identify a fraudulent transaction, or identify potential for a new
service, classification may be used to split a customer base into best, average, and
low-value customers.
• Estimation assigns an object some numeric value that is continually valued.
Credit risk evaluation, for instance, is not simply a yes/no question; it may be
some kind of scoring that assesses a propensity to default on a loan. As part of
the classification process, estimation may be used (such as using an estimation
model as part of a market segmentation process to guess the annual salary of a
person).
• Prediction attempts to identify objects according to any potential actions
predicted. By applying historical data where the classification is already defined
to construct a model, classification and estimation can be used for prediction (this
is called training). To predict future behavior, the model can then be applied to
new data.
• Affinity grouping evaluates relationships or associations between data elements
that demonstrate some kind of affinity between objects.
• Definition tries to explain what has been discovered or attempting to explain the
results of the process of data mining.
On the basis of KDD, the landscape of business decision support is evolving
with an increasing body of emerging applications, such as risk analysis, targeted
marketing, customer retention, portfolio management, and brand loyalty [18]. In
modeling variables of interest, traditional DM methods have proven to be successful,
so that these variables can be projected in future scenarios and effective decisions
made based on that forecast can be taken. More clearly, three key considerations
in choosing the applicable mining technologies for the problem to be solved by the
KDD technology: objective, characteristics of data, and mining algorithm. However,
how to test a candidate model is one of the most critical aspects of DM problems,
and obviously, this issue depends on the type of DM task at hand. Therefore, most
of the DM issues can be considered as optimization issues, where the goal is to
create a candidate model that optimizes those performance parameters which is a
multi-objective concern in nature.
Since we are looking to find a structure for sustainability enhancement in an
unlabeled dataset related to collaborative product development while considering
different virtues X, clustering is the most appropriate method in our context [7].
Indeed, without a priori knowledge of the structure of a database, only unsuper-

vised classification can automatically detect the presence of relevant subgroups (or
clusters).
3 Related Works
The clustering publications have been rising over the past years, showing that
researchers are paying more and more attention to this issue. Some researchers have
improved clustering algorithms for a particular domain, while others have imple-
mented new algorithms, or have researched and compared clustering algorithms.
By using different databases such as Scopus, Elsevier, Emerald, Taylor & Francis,
Springer, IEE, and Google Scholar, we have grouped the literature into four key cate-
gories: review and surveys, comparative studies; clustering methods dealing with a
new algorithm, and finally clustering applications. According to Fig. 2 and without
limitations constraints, twenty-three percent of publications focused on comparing
various algorithms, while thirty-eight percent of publications applied clustering to
several domains such as image processing, speech processing, information retrieval,
web application processing.
Using the Newbert methodology [5, 19], this systematic methodology ensured
that the analysis was thorough, leading to a large number of comparative studies
being reviewed concerning to clustering algorithms for different domains. Among
the 87 papers chosen in the literature review, 28 were eventually used for comparative
studies to research these works in detail, while 10 were used for the application of the
clustering approach in the industry field. As a result, some of the selected studies are
presented in Table 1 using both the comparative studies carried out in the literature
for all fields and the clustering algorithms used in the industry.
140
120 127
100
102
80
80
78
60
40
20
0
Review and surveys Comparave studies Clustering approaches Clustering applicaons
Fig. 2 Classification of clustering publications

Table 1 The literature review of clustering algorithms with respect to the considered classification
References Description
[22] Comparative studies A detailed survey of current clustering algorithms in data
mining with making a comparison among them according to
their scores (merits). The authors identified the problems to
be solved and the new directions in the future according to
the application requirements in multimedia domain
[20] A survey of clustering algorithms that require less amount of
knowledge about the domain being clustered
[23] The performance of several clustering algorithms is been
computed with respect to several indices such as
homogeneity, separation scores, silhouette width, redundant
score, and WADP
[21] A comparison of performance of various clustering
algorithms based on size of dataset, time taken to form
clusters, and the number of clusters formed
[24] Both A survey of different clustering algorithm (DBSCAN and
K-means) for analyzing different financial datasets for a
variety of applications: credit cards fraud detection,
investment transactions, stock market, etc
[25] A comparative study with three algorithms such as K-means,
hierarchical agglomerative clustering, and SOM was used to
modularize a packaging system by reducing the variety of
packaging sizes
[17] A comparative study of different clustering algorithms such
as K-means, DBSCAN, hierarchical agglomerative
algorithm, and SOM for analyzing four different real
industrial dataset
[26] Applications Applied SOM to visualize and analyze large data bases of
industrial systems such as forest industry. The goal was to
cluster the pulp and paper mills of the world
[29] The identification of high-profit, high-value and low-risk
customers by the customer clustering which is a one of the
data mining technique
[27] A K-means clustering approach to determine and categorize
the environmental risk zoning. The clustering result with the
optimal clustering number is then used for the environmental
risk zoning, and the zoning result is mapped using the
geographic information system
[28] The relationship between perceptions of clusters and
innovation for firms operating through technicities that are
accepted as innovative clusters. To test the propositions, a
field survey using questionnaires was conducted
[30] A survey on data mining in steel industries. Steel consists of
alloy iron, carbon, and manganese with small amount of
silicon, phosphorous, and Sulfur. Steel production stages:
Heating, cooling, melting, solidification
With an in-depth study of these papers, we may state that several studies [20,
21] have extensively studied common and known algorithms such as K-means,
DBSCAN, DENCLUE, K-NN, fuzzy k-means, and SOM to discuss their advantages
and disadvantages, taking into account several factors that may affect the criterion
when selecting a suitable clustering algorithm. While other studies [22, 23] have
examined the clustering algorithm surveys based on various parameters, such as
their score (merits), their problems solved, their applicability, their domain aware-
ness, and also on the size of the dataset, the number of clusters, the type of dataset,
the software used, the complexity of time, stability, etc. Researchers [24–30] have
examined various algorithms such as K-means, DBSCAN, agglomerative hierar-
chical clustering, and SOM algorithm for cluster packaging and environmental risk,
financial, female employees, consumer preferences, industrial hygiene, and forest
industry datasets in further research and in particular in the field of industry.
However, there are some shortcomings of all these surveys and comparisons found
in the literature, such as the characteristics of the algorithms are not well studied, no
systematic empirical study was carried out to assess the value of one algorithm for a
particular type of dataset over another. Thus, the only paper that deals with such issue
is one of our previous work [10] which deals with various algorithms with different
real industrial datasets. Therefore, overviewing and exploring the algorithms that
determine the best clusters for sparse industrial dataset remains an open issue.
4 Methodology
Algorithms for clustering have a strong relationship with many fields, especially
statistics and science. Different starting points and parameters typically lead to
various clustering algorithm taxonomies. As a result, a large number of algo-
rithms have been proposed in the literature. Depending on the approach consid-
ered in terms of data processing, clustering techniques can be divided into five
categories [17]: Partitioning-based algorithms; hierarchy-based algorithms; density-
based algorithms; grid-based algorithms; and model-based algorithms. However,
based on the comparison realized in the paper [17] and due to the popularity, versa-
tility, applicability to industrial datasets, we can claim that the chosen algorithms
are: k-means algorithm, ward-distance agglomerative hierarchical algorithm, and
self-organization map (SOM).
4.1 K-means
The K-means algorithm is the best-known approach, used and extended in the
different communities dedicated to clustering. The principle is “natural,” given
the distribution of individuals X in the description space and a fixed number of
groups, the objective is to minimize the dispersion of individuals relative to a set
of prototypes representative of these groups. In other words, the first step is to

randomly select k, where k equals the number of clusters that you chose. Data points
representing the center of a cluster are called centroids [31]. By a two-step method
called expectation–maximization, the key elements of the algorithm operate. Each
data point is assigned to its closest centroid by the expectation stage. Then, for each
cluster, the maximization step computes the average of all the points and sets the
new centroid. The pseudo-code of the K-means algorithm is as follows:
Algorithm: K-means
1: Specify the number of k of clusters to assign
2: Randomly initialize k centroids
3: repeat
4: expectation: Assign each point to its closest centroid
5: maximization: Compute the new centroid (mean) of each cluster
6: until The centroid positions do not change
4.2 Hierarchical
By creating a hierarchy, hierarchical clustering decides cluster assignments. Either

a bottom-up or a top-down approach implements this: the bottom-up approach is
agglomerative clustering. Until all points have been grouped into a single cluster,
it merges the two most similar points. In contrast, the top-down strategy is divisive
clustering. It begins as one cluster with all points, and at each stage separates the least
related clusters until only single data points remain [32]. A tree-based hierarchy of
points called a dendrogram is generated by these techniques. The number of clusters
(k) in hierarchical clustering is always predetermined by the user, similar to partitional
clustering. Clusters are allocated by cutting the dendrogram at a given depth, resulting
in smaller dendrograms in k groups. The pseudo-code of the hierarchical algorithm
is as follows:
Algorithm: Hierarchical agglomerative

1: Compute the proximity matrix for data points
2: repeat
3: combine the two closest clusters into one cluster
4: update the proximity matrix
5: until there is only one cluster
4.3 Self-Organizing Maps
A self-organizing map (SOM) is an unsupervised competitive learning algorithm

from an artificial neural network. It is a very popular nonlinear technique for
reducing and visualizing data. When an observation is recognized, the activation of
a neuron in the network, selected by a competition between neurons, has the effect
of strengthening this neuron and inhibiting others (this is the rule of “Winner Takes
All”)[33]. Therefore, each neuron specializes during learning in recognizing a certain
type of observation. The self-organizing map is made up of a set of neurons connected
by topological links that form a two-dimensional grid. Each neuron is connected to n
inputs (corresponding to the n dimensions of the representation space) according to n
weights which form the prototype vector of the neuron. Neurons are also connected to
their neighbors by topological links. The dataset is used to organize the map according
to the topological constraints of the input space. Thus, a configuration between the
input space and the network space is constructed; two nearby observations in entry
space activate two nearby units on the map. The pseudo-code of the SOM algorithm
is as follows:
Algorithm: Self-organizing maps

1: Select random values for the initial weight vector wj
2: repeat
3: Select a sample from the input dataset
4: Find the winning neuron for the sample input
5: Adjust the weights of nearby neurons
5: until the feature map stops changing
In this section, we describe all the phases presented in the generic KDD pipeline
(Fig. 3) to discover and extract useful knowledge from a collection of industrial data.
More clearly, we are going to describe the considered dataset, the data-preprocessing
Fig. 3 KDD pipeline for clustering design for relevance dataset

and transformation tasks, the data mining process which is clustering as well as the
useful knowledge which are the obtained clusters.
5.1 Dataset Collection
Manufacturing industries that rely on a higher degree of value-added manufacturing,

such as the automotive industry, are increasingly trying to innovate in the face of
rising sustainability challenges by bringing new products to market and providing
more complete solutions that incorporate both product and service components
[5]. However, it is a difficult mission to shift from a product-centric perspective
to a solution-oriented perspective. It involves the use of modern sustainable busi-
ness models to ensure design activities and to ensure sustainable assessment, from
creation, through manufacturing, use, until the end-of-life of the product, meets
consumer and societal needs. In this respect, the “design for” or DFX literature is
reviewed by this new product development study to consolidate the existing body of
expertise and to seek the future direction of the industry [7]. In terms of economics
(dominated by supply chain design techniques), ecology (dominated by environ-
mental design techniques), and social equity, DFX techniques can be put under the
heading of sustainability [34].
However by analyzing the different DFX techniques which are most applicable
during the design phase and by considering also the relationships and trade-offs
between design decision across the main phases of product lifecycle while consid-
ering the IATF 16,949 quality attributes, the most relevant DFX in the automotive
sector are Design for manufacture and assembly, Design for quality, Design for
service; Design for safety; Design for supply chain and finally design for environ-
ment [7]. We shall from now, that the six DFX chosen in this paper will be named
by Design for Relevance. The term relevance refers to the relation and interaction
between DFX techniques and attributes of quality in a way that makes it benefi-
cial to consider second when considering first. That implies that the two principles
contribute to unity and cannot be isolated. More clearly, the Design for Relevance
dataset contains 492 design factors (knowledge) and 29 modules.
The aim of clustering in Design for relevance is to assign related objects to the
same cluster and to distinguish them from clusters containing modules that are not
similar. We refer from this point onwards to objects as modules and present a group
as Family Modules (MF). The basic concept is that the MF helps a project team to
understand and design the design factors (knowledge) common to a group of modules
at the same time. The efficacy of the designs importance will be improved by this
technique. The dataset is presented in details in paper [7].
5.2 Data-Preprocessing and Transformation
Data-preprocessing is one of the most important steps in a data mining process. To

improve their consistency, it deals with the preparation and transformation of the
initial dataset. Generally, this phase is divided into four categories: data cleaning,
data incorporation, transformation of data, and data reduction [11]. In our Design
for Relevance dataset, we need the data cleaning and data transformation in order to
improve the quality of our data, and we can start our clustering with no noise features
and with numerical inputs dataset.
Data cleaning task consists of cleaning the information from noisy and inconsis-
tent design factors. In this step, we first begin by unifying concepts that have the same
thoughts but with different names. After that, and since our dataset is small, we define
some noisy concepts with the aid of an expert and compute their neighborhood for
each noisy point to decide whether it is a core point (the volume for each point is the
same and the mass of neighborhood is variable). If yes, we start a cluster somewhere
else around this point, we mark the point as an outlier. These two steps are repeated
until all points are either allocated to a cluster or designated as an outlier [35].
Data transformation task deals with the conversion or aggregation of data into
suitable forms for mining. The set of modules is described in the vector space model
in this phase. This implies that the interactions between modules and design variables
are determined using the tf scores [36] and are expressed in the incidence matrix of the
binary module design factor [aij], where I represent the ith design variables associated
with all modules and j represents the jth module. After doing this transformation;
the only entries of our dataset are 0 and 1. Entry 1 indicates that the design factor of
the ith is part of the jth module, while entry 0 indicates that the design factor of the
ith is not part of the jth module. We note that with the aid of the designer interpreted
as one expert or a team of experts, the binary assignment is achieved.
5.3 Data-Mining: Clustering
In several real-world cluster applications, one of the most important questions is

how to decide a suitable number of clusters K. There is no easy way to know the
number without a priori information, the only way to do it is by computing several
internal indices such as CH index [37], Gamma index [38], DB index [39] and Hubert
index[40] that allow us to choose the best cluster number [17]. The optimal number
of clusters for each index is based on the maximum (or minimum) index value, the
maximum (or minimum) index hierarchy difference, the maximum (or minimum)
index hierarchy value, the maximum (or minimum) second index level difference,
or the use of the critical value [41]. To compute the indices, user can request the
indices one by one, by setting the argument index to the name of the index. Hence
and according to the majority rules, 5 would be the best number of clusters for our
dataset. In addition, if we look at the Hubert index, which is a graphical tool, the
Fig. 4 Hubert index for design for relevance dataset
optimal number of clusters in the plot of index values against the number of clusters
is defined by a significant knee. As the number of clusters ranges from minimum
to maximum, this knee leads to a significant increase or significant decrease in the
index. In other words, the relevant number of clusters is indicated by a large peak
in the second differential value map. Hence, as shown in Fig. 4, the Hubert index
confirms our purpose and propose 5 as the best number of clusters.
We can now run the selected three algorithms after selecting the required number
of clusters in our data to compare, first, the best clustering algorithm in our case
and, secondly, to define the groups of modules. However, we should consider the
following points before beginning our comparative analysis:
• The SOM algorithm differs from other algorithms for clustering, especially from
other artificial neural networks. Indeed, SOM is a common nonlinear dimension
reduction and data visualization technique that does not provide clusters or groups
[26]
• SOM uses a neighborhood function to preserve the topological properties of the
input space.
Thus, an efficient method to grouping problem is based on the learning of a SOM.
We use SOM to measure a set of reference vectors representing local means of data
based on topological properties in the first step of the method. In the second point, the
vectors obtained are clustered using the two standard K-means clustering methods
and the agglomerative hierarchical algorithm to form the final partitioning. This
method is most commonly referred to in the literature as two-level clustering algo-
rithms [42]. Choosing the best clustering algorithm based only on a single measure
for our dataset can lead to misleading conclusions. For this reason, Fig. 5 presents
the results of the clustering of four candidates with regard to five internal validity
measures presented above.
First, it can be seen that, compared to the remaining clustering algorithms except
for the CH index, where k-means performs well, the SOM/K-means algorithm
provides the best clustering performance based on most internal and stabilization
indices. The SOM/HC followed by K-means is the second-best clustering algorithm
in terms of internal validity. The agglomerative hierarchical algorithm is, thus, the
worst. Moreover, it can also be seen that the SOM/K-means algorithm also generates
connected clusters compared to other clustering algorithms [43]. In fact, according
to Fig. 6, the connectivity of SOM clusters to K-means is approximately 15.97,
suggesting that only 16% of related artifacts are not in the same cluster. This result
is normal since in some modules, there are concepts that are partially similar to each
Fig. 5 Internal validation indices for the selected algorithms
Fig. 6 Connectivity index 35

for the selected algorithms
30
25
20
15
10
5
0
Connecvity
K-means SOM-K SOM-HC HC
other but are in the same module. In terms of compactness and separation, it can
be seen that not only compact clusters, but also well-separated clusters can often be
generated by SOM/HC than K-means, which in turn is better than the Hierarchical
algorithm. This is substantiated by indices of tau, Gamma, and DB performance.
Finally, and after analyzing the results of testing the clustering algorithms and
evaluating them under different indices, we can remark that SOM combining with
agglomerative hierarchical algorithm shows more accuracy in clustering most of the
objects into their suitable clusters.
5.4 Knowledge for Sustainability Enhancement of the Design

for Relevance
According to our previous comparison, SOM combined with hierarchical clustering

algorithm indicates that the overall Design for Relevance can be accomplished in
five self-contained modules:
• Module Group 1 addresses logistics issues, it consists of packaging criteria,
transport criteria, warehouse criteria, and Industrial footprint modules.
• Module Group 2 addresses reverse logistics issues, it consists of recycling and
disassembly criteria, reverse logistics criteria, and environmental issues modules.
• Module Group 3 addresses service issues, it consists of product service system,
service control, quality insurance, and service criteria modules.
• Module Group 4 addresses feasibility issues, it consists of assembly features,
assembly requirements, environmental requirements, environmental considera-
tions, materials safety, materials selection, safety criteria, quality performance,
service requirements, and human safety modules. Finally,
• Module Group 5 addresses design issues, it consists of manufacture criteria,
assembly criteria, assembly process, quality control, quality system, safety
requirement, and manufacture planning and control modules. In designing these
modules groups, the project team considers only the necessary and relevant design
factors.
By so doing, the scope and degree of complexity of the design become more
manageable and take a holonic perspective for sustainability assessment while devel-
oping a product from all its life cycle. For example, not only the widely used manu-
facturing design factors are considered by the design module community, but also
across other modules, and other important and required design factors are included
such as controlled parameters and control plan in quality control module; Assembly,
transport and environment requirements and safety requirements, etc.). This simply
signifies that the design module group is not an isolated activity and may not be
designed in a vacuum. As a result, by achieving and integrating other important
and necessary factors such as consistency, protection, climate, operation, assembly,
and logistics, this approach improves the effectiveness as well as the sustainability
consideration of the product design.
6 Conclusion
With the rapid growth in number and dimension of databases and database applica-
tions in business, administrative, industrial, and other fields, the automated extraction
of information from these broad databases needs to be examined. These have become
rich and reliable sources of information generation and verification due to knowl-
edge extraction from databases, and knowledge discovery can be applied in software
management, process querying, decision making, process control, and many other
fields of interest. In addition to that, given the problems facing product manufacturing,
which are becoming increasingly complex, companies need to consider complexity
in both technological and other multidisciplinary fields. Manufacturing firms will
not only need to implement flexible strategies to reap the gains in the future, but
they must also successfully innovate and manage various problems with different
X-factors to reach the sustainability assessment [7].
In this respect, this paper provides a comprehensive survey and intends to compare
popular, flexible, and applicable clustering algorithms in industry field. More clearly,
by considering the Design for Relevance dataset, we facilitate the interface and
the collaboration between all project teams. However, to manage and organize a
large number of design factors involved in the design of integrated DFX in an unla-
beled dataset, we have used a clustering data mining task clustering. However, many
researchers have developed and provided several clustering algorithms with updated
implementations for different domains. As a result, finding suitable algorithms helps
to organize information significantly and extract the correct queries from various
database queries.
Through an exhaustive search, a related work was first presented in four cate-
gories, namely review and surveys; comparative studies; clustering approaches and
clustering applications, according to twenty last year. After that, and based on the
clustering categorization framework presented in paper [17], the most representa-
tive clustering algorithm of each category (excepted for the grid-based) has been
implanted and compared for the Design for Relevance dataset. Following the KDD
process, in the type of hierarchy, we have generated a taxonomy, i.e., a category of
knowledge entities, according to the presumed relationships of the real-world enti-
ties they represent. SOM combined with hierarchical clustering offers an effective
clustering algorithm for solving integrated DFX problems and achieving sustainable
progress after comparing three commonly used algorithms (k-means, agglomera-
tive hierarchical clustering with ward distance, and self-organizing SOM maps).
This algorithm creates self-contained clusters that can be interpreted, changed, and
applied more easily.
Furthermore, we have tried to reveal future directions for the researchers. While
manufacturing processes are among the most controlled and supervised, this trend
is being driven today by IoT growth. Data collected by manufacturing firms is

increasing sharply, but software management tools, on the other hand, are typi-
cally unable to perform a detailed analysis of this information. In other words, KDD
methods are not common, and businesses are still not leveraging the full poten-
tial of their collected data, especially about information acquisition. Thus, recent
developments in innovative architectures and methods, such as IoT and DM, enable
manufacturing management systems to rapidly evolve toward capacity forecasting,
modeling, and optimization.
References
1. Salvioni, D.M., Gennari, F., Bosetti, L.: Sustainability and Convergence: The Future of
Corporate Governance Systems? mdpi.com. https://doi.org/10.3390/su8111203
2. Clayton, T., Radcliffe, N.: Sustainability: A Systems Approach; Routledge, New York, NY,
USA (2015)
3. Drexhage, J., Murphy, D.: Sustainable development: from Brundtland to Rio 2012. Background
paper prepared for consideration by the High-Level Panel on Global Sustainability at its first
meeting 19 September 2010. Adv. Appl. Sociol., 5(12) (2015)
4. Zbuchea, A. Are customers rewarding responsible businesse...—Google Scholar
5. Benabdellah AC, Bouhaddou I, Benghabrit A, Benghabrit O (2019) A systematic review of
design for X techniques from 1980 to 2018: concepts, applications, and perspectives. Int. J.
Adv. Manuf. Technol., pp. 1–30. https://doi.org/10.1007/s00170-019-03418-6
6. Holmes, A., Moore, L., Sundsfjord, A., et al.: Understanding the mechanisms and drivers of
antimicrobial resistance. Elsevier
7. Benabdellah, A.C., Benghabrit, A., Bouhaddou, I., Benghabrit, O.: Design for relevance
concurrent engineering approach: integration of IATF 16949 requirements and design for X
techniques. Res. Eng. Des. 31, 323–351 (2020). https://doi.org/10.1007/s00163-020-00339-4
8. Matheus, C.J., Chan, P.K., Piatetsky-Shapiro, G.: Systems for knowledge discovery in
databases. IEEE Trans Knowl Data Eng 5, 903–913 (1993). https://doi.org/10.1109/69.250073
9. Tsai, C.-W., Lai, C.-F., Chiang, M.-C., Yang, L.T.: Data mining for internet of things: a survey.
IEEE Commun. Surv. Tutorials 16, 77–97 (2014). https://doi.org/10.1109/SURV.2013.103013.
00206
10. Benabdellah, A.C., Benghabrit, A.: Science IB-P computer. A survey of clustering algorithms
for an industrial context. Elsevier (2019)
11. Fayyad, U., Piatetsky-Shapiro, G.: Magazine PS-AI, 1996. From data mining to knowledge
discovery in databases. aaai.org
12. Hilbert, M., Lopez, P.: The world’s technological capacity to store, communicate, and compute
information. Science 80(332), 60–65 (2011). https://doi.org/10.1126/science.1200970
13. Cicardi, M., Aberer, W., Banerji, A., et al.: Classification, diagnosis, and approach to treatment
for angioedema: consensus report from the hereditary angioedema international working group.
Allergy 69, 602–616 (2014)
14. Shen, W., Hao, Q., Yoon, H., et al.: Applications of agent-based systems in intelligent
manufacturing: An updated review. Elsevier
15. Yoon, J.P., Kerschberg, L.: A framework for knowledge discovery and evolution in databases.
IEEE Trans Knowl Data Eng 5, 973–979 (1993). https://doi.org/10.1109/69.250080
16. Klösgen, W., data JŻ-A in knowledge discovery and, 1996. Knowledge discovery in databases
terminology.dl.acm.org
17. Benabdellah, A.C., Benghabrit, A., Bouhaddou I (2019) A survey of clustering algorithms for
an industrial context. In: Procedia Computer Science
18. Gamarra, C., Guerrero, J.: Energy EM-R and S, 2016. A knowledge discovery in databases
approach for industrial microgrid planning. Elsevier
19. Newbert, S.L.: Empirical research on the resource-based view of the firm: an assessment and
suggestions for future research. Strateg. Manag. J. 28, 121–146 (2007)
20. Treshansky, A.: … RM-T for S, 2001. Overview of clustering algorithms. spiedigitallibrary.org
21. Chen, L., Ellis, S., Holsapple, C.: Supplier development: a knowledge management perspective.
Knowl. Process. Manag. 22, 250–269 (2015). https://doi.org/10.1002/kpm.1478
22. He, L., Wu, L.: Computers YC-A research of, 2007. Survey of Clustering Algorithms in Data
Mining. en.cnki.com.cn
23. Chen, G., Jaradat, S., Banerjee, N., et al.: Evaluation and comparison of clustering algorithms
in analyzing ES cell gene expression data. JSTOR
24. Cai, F., Le-Khac, N.-A., Kechadi, T.: Clustering Approaches for Financial Data Analysis: a
Survey (2016)
25. Zhao, C., Johnsson, M., He, M.: Data mining with clustering algorithms to reduce packaging
costs: a case study. Packag. Technol. Sci. 30, 173–193 (2017). https://doi.org/10.1002/pts.2286
26. Simula, O., Vasara, P.: JV-IA of, 1999. The self-organizing map in industry analysis.
books.google.com
27. Shi, W., Zeng, W.: Application of k-means clustering to environmental risk zoning of the
chemical industrial area. Front. Environ. Sci. Eng. 8, 117–127 (2014). https://doi.org/10.1007/
s11783-013-0581-5
28. Yıldız, T.: Sciences ZA-P-S and B. Clustering and innovation concepts and innovative clusters:
an application on technoparks in Turkey. Elsevier (2015)
29. Saraee, M., Moghimi, M.: AB the FIW on. Modeling batch annealing process using data mining
techniques for cold rolled steel sheets.dl.acm.org (2011)
30. Umeshini, S., PSumathi, C.: A Survey on Data Mining in Steel Industries.
pdfs.semanticscholar.org
31. MacQueen, J.: Some methods for classification and analysis of multivariate observations. Proc.
Fifth Berkeley Symp. Math. Stat. Probab. 1, 281–297 (1967)
32. Johnson, S.C.: Hierarchical clustering schemes. Psychometrika 32, 241–254 (1967). https://
doi.org/10.1007/BF02289588
33. Kohonen, T.: 1 The Self-Organizing Map (SOM) (2001)
34. Arnette, A.N., Brewer, B.L., Choal, T.: Design for sustainability (DFS): the intersection of
supply chain and environment. J. Clean. Prod. 83, 374–390 (2014)
35. Campello, R.J.G.B., Moulavi, D., Sander, J.: Density-Based Clustering Based on Hierarchical
Density Estimates, pp. 160–172 (2013)
36. Aizawa, A.: An information-theoretic perspective of tf–idf measures. Inf. Process. Manage.
39(1), 45–65 (2003)
37. Caliński, T.: JH-C in S. A dendrite method for cluster analysis. Taylor Fr (1974)
38. Baker, F.B., Hubert, L.J.: Measuring the power of hierarchical cluster analysis. J. Am. Stat.
Assoc. 70, 31–38 (1975). https://doi.org/10.1080/01621459.1975.10480256
39. Davies, D.: Analysis DB-I transactions on pattern. A cluster separation measure. ieeex-
plore.ieee.org (1979)
40. Statistical HL-J of the A. On the Kolmogorov-Smirnov test for normality with mean and
variance unknown. amstat.tandfonline.com (1967)
41. Sajana, T., Rani, C.M.S.: KVN-I Journal of S. A survey on clustering techniques for big data
mining. researchgate.net (2016)
42. Cabanes, G.: Maps YB-S-O. Learning the number of clusters in Self Organizing Map.
intechopen.com (2010)
43. Xing, G., Wang, X., Zhang, Y., Lu, C., Pless, R., Gill, C.: Integrated coverage and connectivity
configuration for energy conservation in sensor networks. ACM Trans, Sens. Netw. (TOSN)
1(1), 36–72 (2005)
Location Finder Mobile Application
Using Android and Google SpreadSheets
Adeosun Nehemiah Olufemi and Melike Sah
Abstract In this paper, we introduce a mobile application (app) called Location

Finder that is built with Java using Android Studio and the Google Spreadsheets. This
application works semantically as different locations can be automatically detected. It
is a complete native Android application to find different locations such as restaurant,
hotel, shops, bus stops, ATM, universities, hospitals, gas station and many more. It
provides advanced mapping systems that will show different places of interest for
app users. This paper discusses the system architecture and details of the design. One
of the advantages of the proposed mobile app is that it is easy to develop and based
on interoperable Google Spreadsheets platform.
1 Introduction
With the recent developments in the cellular world, the high-end mobile phones and
PDAs are becoming pervasive and are being used in different application domain.
Integration of the Web services and cellular domains leads to the new application
domain, mobile Web services [1].
In this paper, we propose Location Finder which is a mobile Web application
that allows people to explore the world around them by leveraging contexts that are
meaningful to them. Someone that finds himself/herself in a city or town that he/she
is not familiar with, say North Cyprus, for example, may decide to go to a restaurant
and have no clue of how to find one. He/she may decide to open the mobile app and
search for query like “Restaurant”. The Location Finder mobile application (Fig. 1)
combines an innovative interface and architecture to support ready exploration of
rich information. The mobile application was built with Java using Android Studio,
A. N. Olufemi
Software Engineering Department, Near East University, Nicosia, North Cyprus
e-mail: 20195289@std.neu.edu.tr
M. Sah (B)
Computer Engineering Department, Near East University, Nicosia, North Cyprus
e-mail: melike.sah@neu.edu.tr
https://doi.org/10.1007/978-981-16-3637-0_51
724 A. N. Olufemi and M. Sah
Fig. 1 Interface of Location Finder mobile application; different features are supported by the app
such as map, favorites and categories
and the Google Spreadsheets has been used for the backend. Further, the application
supports the publication of new information in the backend: Public toilet, accident-
prone area, etc., can be semantically published which will dynamically reflect on the
mobile application. In these respects, Location Finder application uses technologies
that provides an effective information exploration experience.
Another related work is mSpace Mobile [2, 3]. It is a semantic mobile application.
The main goal of this semantic mobile app is to provide information about topics of
chosen interest, based upon the location, as determined by an optional GPS receiver.
It queries the mSpace Query Service (MQ), which is connected to RDF triplestore
knowledge interfaces (MK). Simple Object Access Protocol (SOAP) and HTTP
are used for the communication between the mSpace Mobile application and the
mSpace Query Service. mSpace Mobile uses semantic Web technologies to explore
information. The mSpace Mobile interface is designed to let users of small screen
devices run complex queries through direct manipulation, and no typing is required.
To this end, the application utilizes the primary features of the mSpace interaction
model [4]. mSpace Mobile is an application that was developed for Windows phone
users. Whereas in our work, Location Finder has been built and developed with
Android Studio which will run on Android devices and based on interoperable Google
SpreadSheets.
Other related works are AWARE [5], Momento [6], Device Analyser [7],
MetricWire [8] and Google Timelines [9]. AWARE [5] is an open source platform for
catching, deducing and creating context on mobile devices. It is an Android applica-
tion which collects data from several sensors such as emails and messages which can
be used to create context-aware software. Momento [6] is also another context-aware
mobile application which explicitly asks the user his/her location, nearby persons
Location Finder Mobile Application Using Android and Google … 725
and other information in order to provide context-aware data. Device Analyser [7]
is a free data collection tool for Android devices, which can be used for ubiquitous
computing. MetricWire [8] is an Android mobile application that is used for data
collection from the user’s mobile phone. Google Timelines [9] mobile application
can also be utilized to track the location of the user.
2 System Architecture of Location Finder
Location Finder mobile application has been built with Java using Android Studio
and Online Google Spreadsheet. The project is as simple as possible for better under-
standing. It contains few JAVA and XML files. The app has been built on Google
standard Google map to ensure highest accuracy and portability and has a dynamic
backend based on Online Google Spreadsheet, and this will help the administrator
to update and maintain backend data without any additional server or database. In
the following sub-sections, we explain the details of system components.
2.1 Google Spreadsheets for Storing Information of Places
First, information will be inserted to the Google Sheets as shown in Fig. 2. Google
Sheets is a part of free, Web-based office software offered by Google Drive service.
The service also includes Google Docs and Google Slides, a word processor and
presentation program, respectively. It makes data to be created charts and graphs.
Fig. 2 Information is inserted into Google SpreadSheets

Built-in formulas, pivot tables and conditional formatting options save time and
simplify common spreadsheet tasks.
2.2 Material Design and Optimization
Then, we design the UI/UX of the Location Finder. Material Design [10] is a visual
language that synthesizes the classic principles of good design with the innovation
of technology and science. It is very important to have a presentable user interface
by using a well-pleasing design. To identify a set of UI component categories that
frequently occur in Android apps, we referenced popular design tools and languages
that expose component libraries such as Balsamiq [11] and Google’s Material Design
[12]. The Material Design has been maintained for Location Finder mobile appli-
cation to decorate the user interface and to ensure the app is mobile friendly and
fully optimized. All basic components are nicely decorated with unique looking and
gorgeous color combination.
Optimization of code is essential to make the app run smoothly without lagging
and that is why app has been developed with the code fully optimized. Every single
implementation has been optimized for highest performance. Also, the code has been
beautifully crafted and modularized to enable other developers to easily understand
the code. Comments are used where necessary to describe certain line of code.
2.3 Android Studio Setup
This software is built purposely to develop Android apps. After installing this soft-
ware, the next thing is to download the necessary plugins including Java IDE and
SDK. In Android studio, the project folder is named. Once that has been selected,
the project will be synced by Gradle, and the screenshot below will be displayed. We
have created different categories for certain places such as hotel, restaurant, ATM
point, bus stops and so on. Category takes the name of this places and displays it
on the app in real-time mode. Subcategories are needed in order to put the values
of specific places under each category, e.g., Ezic restaurant is a subcategory under a
category called restaurant. All these information has been recorded manually in our
backend (on Google Sheet) which will be automatically displayed on the app for app
users.
Backend Customization. This is where we add the data that will appear on app
dynamically. We start by creating a new Excel sheet. This can be done by visiting
the link [13] with a Gmail account. After doing this, we open the Google Drive and
then create a new sheet by going to New > Google Sheets. Three sheets were created,
namely category, subcategory and items. The files in the folder contain the Java and
XML files.
Fig. 3 Creating categories
Category Sheet Section. In this section, we have three columns, namely cate-
gory_id (number must be unique), category_title (the name of category) and
image_url (link at which image that will be displayed for each category is located)
(Fig. 3).
Subcategory Sheet Section. This section consists of four columns, namely
subcategory_id (a unique number for this column), category_id (the id of cate-
gory in which subcategory will fall), sub_category_title (the title of subcategory)
and image_url (link at which image that will be displayed for each subcategory is
located) as shown in Fig. 4.
Fig. 4 Creating subcategories

Fig. 5 Item page sheet page
Item Sheet Section. Lastly, we created a sheet named item, and this section is
having eleven columns which are item_id (unique name for item), category_id (the
column at which item will fall), sub_category_id (the subcategory id at which item
will fall), item_title (name of item), image_url (item URL), address (business address
of a particular location), longitude (distance from current location), latitude (distance
from current location), contact_number (business contact of a particular location) and
description (business description) shown in Fig. 5.
After creating the Google Sheet, then we save it and copy the sheet id from the
URL as shown Fig. 6.
The marked id will be copied, and this is what will be used in Android Studio
to display information on the app. A Java class has been created in Android Studio
called HttpParams.java. This class takes the parameters of the Google Spreadsheet
id, category, subcategory and items. After adding this project id to Android Studio,
the APK file is generated by selecting Build > Generate Signed Bundle/APK, as
illustrated in Fig. 7.
After generating the APK file, the app can be either installed on any Android
device or published on Google Play Store for people to install. Categories are added
to the spreadsheet, and more categories can still be added (the amount of categories
that can be added are unlimited). For the demonstrations of our mobile app, categories
presented in Fig. 8 are created. Once categories are added to the spreadsheet, it will
be reflected automatically on the app as it works dynamically. Having explained the
architecture design of Location Finder application, we can now move to the user
interface.
Fig. 6 Project URL

Fig. 7 Generating the .apk file
Fig. 8 Created subcategory

hierarchy for demonstrations
3 User Interface of Location Finder
The Location Finder app interface is designed to allow users of small screen devices
run a search for a particular query like “Restaurant, Hotel and so on”. To this end, the
application utilizes the primary features of the app interaction model. The features
of Location Finder app are as follows:
• App will automatically detect user current location.
• The contact number of each business can be easily dialled for more information
about a particular location.
• It allows user to easily navigate to desired destination
• List of locations are already created.
• Users can save an item or a particular location of interest for further exploration
in future.
• Categories of different items (i.e., names of available restaurants).
These user interactions are illustrated in Fig. 9. The user just opens up Location
Finder app (Fig. 9a). The first screen that the user will see is the splash screen activity
with a beautifully designed logo and a progress bar. User gives permission to app to
access the user’s current location (Fig. 9b). Then, user’s location is detected (Fig. 9c).
User then navigates using the navigation bar (Fig. 9d). User can set a radius setting in
km, where the app looks for items of interest in this area (Fig. 9e). User just selected
“Restaurant” from the list since s/he wants to find nearby restaurants (Fig. 9f). Ekor
Premier has been chosen by the user (Fig. 9g). The user is trying to locate Ekor
Premier in Famagusta using the map (Fig. 9h). A particular location can also be
saved to easily access it next time as shown in Fig. 9i.
To summarize, Location Finder is a complete Android application to find different
categories such as restaurant, hotel, popular places, shopping mall, ATM, hospital,
fuel station, popular food item place, public toilet, accident-prone area and many
more. Every categories also have a subcategory. The app has been built on standard
Google Map to ensure highest accuracy and portability. The backend has been built
dynamically based on Google Sheet. It will help us to update and maintain data
without additional server or database. It provides advanced mapping systems that
will show different places for a specific category. It would be easy to search nearby
place and navigate in a user-friendly way. Initially, it will find the nearest points based
on current user location. User will be able to find a particular location from a different
place using built-in custom search facility. Our code is also reusable since it has been
built with JAVA and Google SpreadSheets. In the next section, we also compare
performance of Location Finder with other similar work in terms of execution time.
Fig. 9 Nine screens that explain how the app works. The user just opens up Location Finder app
and is about to find a restaurant. a The first screen that user will see is the splash screen activity
with a beautifully designed logo and a progress bar. b Gives permission to app to access the user’s
current location. c User’s location has been detected. d User navigates using the navigation bar.
e Settings screen. f User just selected “Restaurant” menu. g Ekor Premier has been chosen by the
user. h The user is trying to locate Ekor Premier in Famagusta. i This shows that a particular location
can be saved to easily access it next time
Fig. 9 (continued)
Fig. 9 (continued)
4 Evaluation and Comparison of Mobile Application

Performances
To assess performance of the proposed mobile application, Location Finder, we test

the response time using Android Profiler on Android Studio. Android Profiler is
part of Android Studio 3.0 that provides real-time graphical display of CPU usage,
memory and network usage for the running a mobile application. Furthermore, we
compare the results of the Location Finder, with other mobile applications, namely
AWARE [5], Device Analyser [6], Momento [7], MetricWire [8] and Google Time-
lines [9]. Experiments are tested in a mobile phone with the following technical
details: 4 GB RAM, Android version 6 and CPU Octa-core.
For the experiment, we follow the guidelines from [14]. In particular, four test
cases are defined: (1) Accessing an element from an array, (2) Iterate through an
array, (3) Iterate through a matrix and (4) Calculation of Fibonacci number [14].
The test cases require input parameters, and three ranges are defined such as small,
medium and large. For the small range, an arbitrary and small enough value is taken
that contributes to changing the utilization of the memory and CPU. For the highest
range, a rough upper boundary value is taken, after which the device has insufficient
Table 1 Comparative results

Mobile applications Platform Response time
for response time
Wi-Fi (s) 4G (s)
AWARE [5] Android/iOS 7.09 9.03
Device Analyser [6] Android 15.04 16.11
Momento [7] Android 8.91 9.02
MetricWire [8] Android/iOS 10.19 10.34
Google Timeline [9] Android/iOS 7.23 7.31
Proposed approach Android 5.28 5.71
resources to perform the action. For the details of the experimental ranges and the
settings, please refer to [14].
We compare performances of all mobile applications with the proposed mobile
application. Results are shown in Table 1. Among the different mobile applications,
Device Analyser has the highest response time for Wi-Fi and 4G network with
15.04 s and 16.11 s, respectively. Comparing to other applications, AWARE and
Google Timelines achieve quicker response times. But among all mobile applications,
Location Finder has the best response times with 5.28 s and 5.71 s for Wi-Fi and 4G
network, respectively.
The good performance of Location Finder is attained for a number of reasons.
Firstly, the mobile application has been developed on Android platform. Secondly,
the proposed mobile application is fully optimized by getting rid of codes that are
not adding any value, checking the application’s efficiency, trying and testing the
code, use profiling tools for monitoring, emphasizing on increasing app usability
and focusing on the user interface. Having putting these into consideration, we are
able to have a better response time which makes our application faster than the
other ones. In our work, beyond identifying the location, the processing is kept to
minimum. Thus, it can save battery life. On the other hand, many of the compared
popular mobile applications run numerous background processes, data collection
and sharing tools, which add extra processing time and increase the response time
of the mobile applications.
5 Conclusions
We explained the proposed Location Finder mobile application that is based on

Google SpreadSheets and Android devices. This application works semantically
as different locations can be automatically detected. It is a complete native Android
application to find different locations such as restaurant, hotel, shops, bus stops, ATM,
universities, hospitals, gas station and many more. It provides advanced mapping
systems that will show different places of interest for application users. Users can
also adjust the search settings (categories, searching area radius) and can save a place
for finding it next time. As a result, we provided an easy to use and reusable mobile
application. Experiments on the performance of Location Finder also shows that the
proposed approach provides much faster response times and better than other popular
mobile applications in the domain.
References
1. Farley, P., Capp, M.: Mobile web services. BT Technol. J. 23(2), 202–213 (2005)
2. Harris, C., et al.: mSpace: exploring the semantic web. In: A Technical Report in Support of the
mSpace Software Framework, p. 98. IAM Group, University of Southampton, Southampton
(2004)
3. Max, W., Alistair R., Daniel A., Alisdair O.: mSpace Mobile: A Mobile Application for the
Semantic Web. IAM Research Group School of Electronics and Computer Science University
of Southampton, SO17 1BJ. http://mspace.fm/
4. Schraefel, M.C., Karam, M., Zhao, S.: mSpace: interaction design for userdetermined, adapt-
able domain exploration in hypermedia. In: International Workshop on Adaptive Hypermedia.
2003. Nottingham, UK
5. Ferreira, D., Kostakos, V., Dey, A.K.: AWARE: mobile context instrumentation framework.
Front. ICT, 20 April 2015. https://doi.org/10.3389/fict.2015.00006
6. Carter, S., Mankoff, J., Heer, J.: “/”. In: Proceedings of the 2007 Conference on Human Factors
in Computing Systems, 2007. https://doi.org/10.1145/1240624.1240644
7. Wagner, D.T., Rice, A., Beresford, A.R.: Device analyzer: understanding smartphone usage.
In: International Conference on Mobile and Ubiquitous Systems: Computing, Networking and
Services, 2014
8. MetricWire Mobile Application. Last accessed at 16th of February 2021. https://play.google.
com/store/apps/details?id=com.metricwire.android3
9. Rodriguez, A.M., Tiberius, C., van Bree, R., Geradts, Z.: Google timeline accuracy assess-
ment and error prediction, Forensic Sci. Res. 3, 3, 240–255 (2018). https://doi.org/10.1080/
20961790.2018.1509187
10. Material Design. https://material.io/design/
11. Balsamiq Studios: 2018. basalmiq. (2018). https://balsamiq.com/
12. Call-Em-All: 2018. Material-UI. (2018). https://material-ui-next.com/
13. Google Account: https://accounts.google.com
14. Andonoska, A., Jakimoski, K.: Performance Evaluation of Mobile Applications (2018). https://
www.researchgate.net/publication/337437805
Sign Language Recognition with
Quaternion Moment Invariants:
A Comparative Study
Ilham El Ouariachi, Rachid Benouini, Khalid Zenkouar, Arsalane Zarghili,

and Hakim El Fadili
Abstract In this paper, we aim to carry out a brief study of a new sets of Quater-
nion Discrete Moment Invariants (QDMI) in uniform lattice named: Quaternion
Tchebichef Moment Invariants (QTMI), Quaternion Krawtchouk Moment Invariants
(QKMI), and Quaternion Hahn Moment Invariants (QHMI), for hand gesture-based
sign language recognition. For this purpose, we present briefly the principles of those
moment invariants. Then, we test them on several datasets that contain challenging
attributions, with regard to invariability and recognition under noise-free and noisy
conditions. We conclude our paper by discussing the obtained results and the future
works in this field.
1 Introduction
Sign language is a challenging task of research due to the complexity involved by

gestures of signs. The basic component of sign language is hand gestures, aimed
at communicating a message, by combining it with other body postures and facial
expressions. Hand gesture-based sign language recognition requires solving a num-
ber of computer vision issues, including (i) hand gesture detection (ii) hand gesture
feature extraction (iii) hand gesture recognition[1, 2].
In general, feature extraction algorithms for hand gestures-based sign language
can be classified in two main types: dynamic and static algorithms [3]. In the case
of static systems, different methods and algorithms have been proposed to extract
features from hand gestures, including adaptive skin color model switching [4],
hierarchical elastic graph matching algorithm [5], histogram of oriented gradient
I. El Ouariachi (B) · R. Benouini · K. Zenkouar · A. Zarghili

Laboratory of Intelligent Systems and Application (LSIA), Faculty of Sciences and Technology,
University Sidi Mohamed Ben Abdellah, Fez, Morocco
e-mail: ilham.elouariachi@usmba.ac.ma
H. El Fadili
Ecole Nationale des Sciences Appliquees of Fez, University Sidi Mohamed Ben Abdellah, Fez,
Morocco
https://doi.org/10.1007/978-981-16-3637-0_52
738 I. El Ouariachi et al.
[6], Gabor filters [4], scale invariant feature transform [7], and local binary pattern
[8].
Despite the huge work that have been done in this area[3–8], feature extraction
from hand gesture-based sign language remains unresolved issue due to appearance
variations of the moving hand gesture, the complex background, and rotation scale
and translation (RST) deformations.
Recently, with the availability of 3D sensors, such as Microsoft Kinect, and Asus
Xtion [3], there has been an increasing concern in the research area related to hand
gesture-based sign language. In fact, the RGB-D sensors allow to capture red, green,
blue, and depth (RGB-D) informations from a scene at the same time [3].
Currently, image moments and moment invariants are one of the most active
research areas in the fields of pattern recognition and computer vision [1, 9], and this
is due to their properties that represent the information with minimum redundancy,
robustness to different kinds of noise, and the invariability property against geometric
deformations [10].
In that respect, various methods have been proposed for hand gesture-based sign
language feature extraction using image moments and moment invariants [11–13].
However, the effectiveness of the moment invariants in hand gesture-based sign
language recognition by using RGB-D images has not been amply examined, and
only very few researches have been done with this interest.
Motivated by the search in moments and moment invariants area related to hand
gesture-based sign language, we present a comparison between three Quaternion Dis-
crete Moment Invariants(QDMI) in uniform lattice: Quaternion Tchebichef Moment
Invariants(QTMI), Quaternion Krawtchouk Moment Invariants(QKMI), and Quater-
nion Hahn Moment Invariants(QHMI).
In this paper, we propose a new sets of moment invariants named: Quaternion hahn
moment invariants(QHMI), we present a brief study of QDMI in the uniform lattice:
QTMI, QKMI, and QHMI, also we evaluate and compare them using challenging
hand gesture-based sign language databases.
The remainder of this paper is organized as follows: Sect. 2 we present a brief
introduction of Quaternion algebra. Then, in Sect. 3, we introduce the principles of
the three evaluated methods: QTMI, QKMI including the proposed QHMI. Section 4
reports the evaluation of the experiments result. Finally, Sect. 5 concludes this paper
and projects some future works.
2 Quaternion Algebra
Quaternions have been introduced by Hamilton in 1843 [14]. The quaternion number
has four parts, one real part and three imaginary parts. The formula of a quaternion
q is defined as follows:
q = qr + qi i + q j j + qk k, (1)
Sign Language Recognition with Quaternion Moment Invariants … 739
where qr , qi , q j , qk are real numbers and i, j, k are complex operators, which obey
the following rules:
i 2 = j 2 = k 2 = −1, i j = − ji = k, jk = −k j = i, ki = −ik = j. (2)
Let f (x, y) be an RGB-D image function defined in cartesian coordinate system,

using the quaternion representation, each pixel can be represented as a quaternion
[1] as following:
f (x, y) = f D (x, y) + f R (x, y)i + f G (x, y) j + f B (x, y)k, (3)
with f D (x, y), f R (x, y), f G (x, y) and f B (x, y) are, respectively, the depth, red,
green, and blue components of the pixel (x, y).
3 Evaluated Quaternion Moment Invariants Methods
3.1 Quaternion Tchebichef Moment Invariants
Elouariachi et al. [1] proposed a robust hand gesture recognition system using a new
set of Quaternion Tchebichef Moment Invariants. The authors derived directly the
proposed invariants from their orthogonal moments, based on the algebraic properties
of the discrete Tchebichef polynomials. The proposed method based on quaternion
algebra is suggested to process the four components holistically, for a robust and
efficient hand gesture recognition system.
Let f d (x, y) be the deformed version of the original image f (x, y), defined by:
f d (x, y) = f (a11 x + a12 y, a21 x + a22 y). (4)
R ST
The Quaternion Tchebichef Moment Invariants QTMIn,m of a deformed image
f (x, y), which is rotation, scaling, and translation invariance is defined as:
d
n m i
j i+ j−s−t s+t
i j
R ST
QTMIn,m = (−1) j−t × An,i Am, j
i=0 j=0 s=0 t=0 u=0 v=0
s t
i+ j+2
× Bi+ j−s−t,u Bs+t,v × (λ f )− 2 (cos θ f )i+t−s (sin θ f ) j−t+s QT M tu,v .
(5)
Along this paper, it will be noted as QTMI.
3.2 Quaternion Krawtchouk Moment Invariants
The Krawtchouk polynomials were introduced by Mikhail Kravchuk in [15]. The

classical Krawtchouk polynomials of the nth order are defined in terms of the hyper-
geometric function 2 F1 (.) as follows:
kn (x; p, N ) = 2 F1 (−n, −x; −N ; 1/ p). (6)
Similarly, to the Quaternion Tchebichef Moment Invariants, we can define the

following QKMIRST n,m which is rotation, scaling, and translation invariance as follows:
n m i
j i+ j−s−t s+t
j−t i j
QKMIRST
n,m = (−1) × Cn,i Cm, j
i=0 j=0 s=0 t=0 u=0 v=0
s t
i+ j+2
× Di+ j−s−t,u Ds+t,v × (λ f )− 2 (cos θ f )i+t−s (sin θ f ) j−t+s Q K M tu,v ,
(7)
Along this paper it will be noted as QKMI.
3.3 Quaternion Hahn Moment Invariants
The Hahn polynomials have been firstly introduced in the field of image analysis
by Zhu et al. in [16]. Similarly to Tchebichef and Krawtchouk polynomials, we can
define the n − th order Hahn polynimials in terms of the hypergeometric function
3 F2 (.) as follows:
(−1)n (b + 1)n (N − n)n

h (a,b)
n (x; N ) =
n!
3 F2 (−n, −x, n + 1 + a + b; b + 1, 1 − N ; 1), (8)
R ST
We can define the following Q H M I n,m which is rotation, scaling, and translation
invariance as follows:
n m i
j i+ j−s−t s+t
i j
n,m =
QHMIRST (−1) j−t × E n,i E m, j
i=0 j=0 s=0 t=0 u=0 v=0
s t
i+ j+2
× Fi+ j−s−t,u Fs+t,v × (λ f )− 2 (cos θ f )i+t−s (sin θ f ) j−t+s Q H M tu,v ,
(9)
Along this paper it will be noted as QHMI.
4 Experimental Results
In this experimental study, three real hand gesture-based sign language databases,
namely HKU [17], NTU [3], and ASL[18] are used to evaluate the three sets of
discrete quaternion moment invariants. The first one contains 10 gestures with 20
different poses from 5 subjects. Therefore, there are a total of 1000 cases, each of
which consists of a pair of color and depth information. The NTU contains 1000
cases of 10 signs from 10 subjects. The third database captures about 65000 samples
of 24 signs (English letters expect j and z) from 5 subjects. The hands are located and
segmented using the hand-wrist belt for the second dataset and depth thresholding
for the others.
All the experiments are conducted on a personal computer equipped with an
Intel(R) Core i7 2.1 GHz and 4 GB of memory. And all the algorithms are coded in
Matlab 8.5.
4.1 Invariability Property
The task of sign language recognition is highly challenging, due to the geometric
deformations performed by the hand, during the execution of the sign. To solve this
issue, an analysis of the three moment invariants with respect to the variation of
image geometric deformation, translation, scale, and rotation is presented.
This section is intended to test the performance of the proposed QTMI, QKMI,
and QHMI quaternion discrete orthogonal moment invariants using RGB-D images.
We evaluate the invariability property under various geometric deformations and
noise conditions. For this, we use a sign language image, whose size is 150 × 150
pixels [17].
Let R E be the relative error between two sets of quaternion discrete orthogonal
moment invariants corresponding to the original and transformed image as:
||QDMI( f ) − QDMI( f t )||

RE( f, f t ) = ,
||QDMI( f )||
where f, f t , QDMI( f ) and Q D M I ( f t ) represent, respectively, the original and the

transformed images, as well as the quaternion discrete orthogonal moment invariants
of the original image and the transformed one. The testing protocol is then constructed
as follows: the test image is scaled with the factor λ ∈ 0.7, 1.2, and rotated from 0 to
360, with an increment 10; then the test image was translated by (Δx, Δy) = (4, 4).
Accordingly, the obtained relative error for the rotation, scaling, and translation
transformations are shown, respectively, in Fig. 1a–c.
In the second part of this experiments, we will investigate the invariability property
of the proposed QDMI in the presence of noisy effects. In fact, the previous hand
(a) QTMI QKMI QHMI (b) QTMI QKMI QHMI (c) QTMI QKMI QHMI
6,0E-09 2,0E-04 5,0E-11
5,0E-09 4,0E-11
Relative Error
Relative Error
Relative Error
1,5E-04
4,0E-09
3,0E-11
3,0E-09 1,0E-04
2,0E-11
2,0E-09
5,0E-05 1,0E-11
1,0E-09
0,0E+00 0,0E+00 0,0E+00
0 40 80 120 160 200 240 280 320 360 0,7 0,8 0,9 1 1,1 1,2 1,3 -24 -20 -16 -12 -6 0 6 12 16 20 24
Rotation angles Scaling factors Translation values
Fig. 1 Relative errors of the proposed QTMI, QKMI, and QHMI against: a Rotation, b Scaling
and c Translation transformations
gesture test image is corrupted with different kinds of noise. First, the test image
was corrupted by Gaussian noise with standard deviations varying from 0 to 5.
Then, affected by salt-and-pepper noise with a density in the range 0–5%. Finally,
degraded by speckle noise with standard deviation starting from 0 to 5. Consequently,
the obtained results for noise invariance are summarized in Fig. 2. It is important
note that the parameters of Krawtchouk and Hahn polynomials are restricted to
p = 0.5, a = 0 and b = 0. In addition, moment invariants up to the order (n, m ≤ 3)
are used to construct the feature vector.
As it can be seen from the Figs. 1 and 2, QKMI perform better than the QTMI
and QHMI, not only in the case of geometric transformations (scale, rotation and
translation), but also in the presence of noisy effects (salt-and-pepper, Gaussian and
speckle).
According to the obtained results in the two figures, it is clear that the relative error
is very low indicating that the proposed set of QDMI remain unchangeable under
geometric transformations and in the presence of the noisy conditions. Therefore,
we can conclude that these new sets of QDMI could be useful in pattern recognition
tasks, especially in hand gesture-based sign language area, which suffer from various
challenging transformations and noisy effects.
4.2 Recognition Results Under Noise-Free Conditions
The goal of this experiment is to evaluate the performance of the proposed descriptors
without noisy effects in the task of sign language recognition. We carry out a series
of experiments on the three databases as follow:
(a) QTMI QKMI QHMI (b) QTMI QKMI QHMI (c) QTMI QKMI QHMI
2,0E-09 4,0E-08 2,5E-07
2,0E-07
Relative Error
Relative Error
Relative Error
1,5E-09 3,0E-08
1,5E-07
1,0E-09 2,0E-08
1,0E-07
5,0E-10 1,0E-08 5,0E-08
0,0E+00
0,0E+00 0,0E+00
0,2
0,4
0
0,04
0,08
0,12
0,16
0,24
0,28
0,32
0,36
0,44
0,48
0 0,5 1 1,5 2 2,5 3 3,5 4 4,5 5 0 0,08 0,14 0,2 0,26 0,32 0,38 0,44 0,5
Noise densities Noise standar deviation values Noise variance
Fig. 2 Relative errors of the proposed QTMI, QKMI, and QHMI for different types of noise: a
salt-and-pepper, b Gaussian and c speckle
Table 1 Comparative analysis of recognition accuracy (%) on HKU, NTU and ASL, by using the
studied methods: QTMI, QKMI and QHMI, with a varying number of neighbors k
Different distance metric k
Databases Methods k = 1 k=2 k=3 k=4 k=5 k=6 Average
QTMI 71 71.5 72 72,9 73 72.8 72.2
QKMI 70.2 70.9 71.3 72 72.7 72.3 71.57
NTU QHMI 54.2 57.9 63.1 62.1 62.2 62.4 60.32
QTMI 74.3 75 76 76.8 77,1 77 76
QKMI 70.9 71.5 72,3 73.9 75 74.8 73.1
HKU QHMI 55.9 56.6 56 56.6 58.8 59.4 57.27
QTMI 80,6 81 81.9 82 82.7 82.1 81.72
QKMI 75.7 76.5 77.3 78.6 79.1 78.2 77.6
ASL QHMI 57.8 61.7 64.5 66.7 68.2 69.5 64.7
Experiment 1: We first evaluate the recognition accuracy of the studied QDMI,

with a chosen number of coefficients in the feature vector (16 coefficients), and using
k-NN as classifier with different values of the neighborhood parameter k, with 5-folds
cross validation.
Experiment 2: We evaluate the recognition accuracy using different number of
coefficients in the feature vector, with k =5 for the k-NN classifier.
The results of experiment 1 can be found in Table 1, where each row in Table 1
present the recognition accuracies of our moment invariants, for different values of
the neighborhood parameter k; columns from three to ten present the recognition
accuracies of different parameters k. The last column is the average recognition
accuracies of each methods for all the chosen values of k.
As one can see from Table 1, the proposed QTMI demonstrates a superiority in
terms of image classification accuracy for all databases compared with QKMI and
QHMI. Therefore, this experiment approves the discrimination capability of our
methods even though with the simplest classifier.
Generally, the utilization of moment invariants up to a higher order leads to a higher
image description power. However, it has been proved that only a small number
of moment invariants coefficients can produce accurate classification [10]. In this
connection, we will investigate the influence of the number of moment invariants
coefficient in the feature vector. The results of experiment 2 are shown in Table 2.
Each row in Table 2 presents the recognition accuracies with variation in the number
of coefficients in the feature vector, of the used methods; each column represents
the recognition accuracies for a specific number of coefficients in the feature vector.
The last column is the average accuracy of each methods for all the numbers of
coefficients in the feature vector up to 64 coefficients.
Obviously, we can notice from the Table 2 that the augmentation in the number
of coefficients has a good impact on the recognition accuracy for all the methods. In
addition, as remarked in the experiment 1, the QTMI gives the best results among
Table 2 Comparative analysis of recognition accuracy (%) on HKU, NTU, and ASL, by using the
studied methods: QTMI, QKMI, and QHMI with a varying number coefficient in the feature vector
Databases Methods N of coefficients in the feature vector
9 16 25 36 49 64 Average
NTU QTMI 72.4 73 74.4 76.8 78.4 79.5 75.75
QKMI 72 72.7 73.2 75.7 77.9 78.4 74.98
QHMI 63.8 62.2 59.9 62.2 62.8 61.3 62.03
HKU QTMI 69.3 77.1 85.1 84.1 85.3 91.4 82.05
QKMI 73.4 75 77.9 80.8 82.6 88.7 79.73
QHMI 59.5 58.8 58.4 55.6 52.2 57.8 57.05
ASL QTMI 75 82,7 89.7 91.4 91.3 91.2 86.88
QKMI 77.6 79.1 80.7 82.9 85.2 88.4 82.32
QHMI 61.42 68.2 60.3 57.3 52.5 58.9 59.77
the proposed methods. This obtained results are justified by the fact that the QTMI
are global shape descriptors [19], and they extract features from the whole sign
gesture image. In the contrary, QKMI is a local descriptor, since it is computed with
emphasis on a specific region of an image [15]. The proposed QHMI gives the lowest
results compared with both QTMI and QKMI, due to its computational complexity
and numerical instability in the calculation of polynomial values [20]. Finally, in this
work, we choose to use for the rest of experiments, a feature vector of 16 elements [1].
4.3 Recognition Results Under Noisy Conditions
It is important to inspect the capability of our descriptors to remain robust against

different sources of noise. In fact, the three databases used in this paper provide a real
examples of the problems that can be faced in sign language recognition applications
using depth cameras, such as depth noise and the presence of the arm and other parts
of the body [3]. Moreover, we test our QDMI in the presence of other kinds of noise,
salt-and-pepper and Gaussian noise.
Each image of the three datasets is corrupted by salt-and-pepper noise with noise
densities varying from 1% to 5% with 1% increments, as well as the other type of
noise: Gaussian with densities changing from 0.1 to 0.5 with 0.1 increments.
The performance of those moment invariants on the datasets is presented on
Table 3, for the two types of noises, where each row of a dataset corresponds to the
recognition accuracies for the different densities, and each column corresponds to
recognition accuracies of all methods for a specific density. The last column is the
average recognition accuracies of each method for all densities.
(a) (b) (c) (d) (e) (f)
Fig. 3 Examples of hands gesture image, a original depth image, b segmented color image and
different background images from: c Vistex, d Brodatz, e Outex and Amsterdam database f
Considering the presented results in Tables 3, it is obvious that the best aver-
age recognition rate for the three datasets was obtained by QTMI, and it is closely
followed by the proposed QKMI.
4.4 Recognition Results Under Complex Background
In this experiment, we have considered four color texture image databases, namely
VisTex [21], Brodatz [22], Outex [23], and Amsterdam [24], to generate four varia-
tions of color background on each of the testing hand gesture datasets HKU, NTU, and
ASL. In fact, we have selected 50 different images from the three texture databases.
These selected images have been randomly used as complex background for each
sign image. And therefore, we have created four additional testing cases for every
hand gesture-based sign language database (Fig. 3).
Table 4 summarizes, respectively, the obtained recognition results on the three
testing datasets, HKU, NTU, and ASL, for different texture background, by using
the studied quaternion moment invariant, QTMI, QKMI, and QHMI. According to
the presented results in Table 4, we can see that only the proposed QTMI and QKMI
are appropriate to be used to classify the sign gestures with complex background,
with an average recognition rate above 70%. However, the proposed QHMI cannot
obtain a higher average recognition rate, more than 60%. It is important to note that
the QTMI and QKMI obtain quite similar recognition rate for many testing cases,
showing high robustness against complex backgrounds.
Finally, we can deduce the importance incorporating depth information in feature
extraction process, and we can conclude that QTMI and QKMI could be very effective
for the applications involving depth and color information.
746
Table 3 Comparative results of recognition accuracy (%) on HKU, NTU, and ASL affected by different salt-and-pepper and Gaussian noise standard deviations,
using the studied methods: QTMI, QKMI, and QHMI
Databases Methods Noise Salt-and-Pepper Average Gaussian Average
free
1% 2% 3% 4% 5% 0.01 0.02 0.03 0.04 0.05
NTU QTMI 73 72.1 71.8 70.9 70.2 69.5 71.25 71.9 70.8 70.4 69.9 69.2 70.86
QKMI 72.7 71.9 70.5 70.1 70.6 66.8 70.43 71.1 70.4 70.5 70.1 69.8 70.76
QHMI 62.2 61.9 60.8 60.3 59.7 59 60.65 61.9 60.8 60.3 59.7 59 60.65
HKU QTMI 77.1 75.4 74.7 73.3 70.8 70 73.55 76.8 75.7 73 72.3 70.5 74.23
QKMI 75 73.1 72.8 71.2 70.7 70.3 72.18 71 69.8 69.5 68.4 67.6 70.21
QHMI 58.8 57.6 57.3 56.9 55.5 54.7 56.8 57.6 57.3 56.9 55.5 54.7 56.8
ASL QTMI 82.7 80.4 78.8 76.7 74.2 73 77.63 81.2 80.7 78.6 77 76.4 79.43
QKMI 79.1 78.2 76.3 75.8 74.2 73.7 76.21 77.6 76.8 75.5 75.1 75 76.51
QHMI 68.2 67.7 65.3 64.1 63 62.4 65.11 67.7 65.3 64.1 63 62.4 65.11
I. El Ouariachi et al.
Table 4 Influence of different complex backgrounds on the hand gesture-based sign language
recognition accuracy (%) for HKU, NTU and ASL, by using the studied methods: QTMI, QKMI,
and QHMI
Databases Methods Different types of complex background Average
Uniforme background VisTex Brodatz Outex Amsterdam
NTU QTMI 73 71.3 74.3 70.5 73.9 72.6
QKMI 72.7 70.5 73.2 69.8 71.6 71.56
QHMI 62.2 56.1 54.5 64.4 65.8 60.6
HKU QTMI 77.1 76 74.3 75.7 72.1 75.04
QKMI 75 74.8 73.4 74 72.6 73.96
QHMI 58.8 46.1 44.9 63.6 60.5 54.78
ASL QTMI 82.7 80.5 81.7 79.4 78.3 80.52
QKMI 79.1 77.6 78.4 76.3 75 77.28
QHMI 68.2 51.6 49.2 64.6 64.3 59.58
5 Conclusion
In this paper, we carried out a comparative study between a three Quaternion Moment
Invariants in uniform lattice for hand gesture-based sign language recognition. The
experimental results with the sets of moment invariants show that the QTMI performs
well in terms of accuracy and robustness followed by the QKMI. There are many
interesting ways to extend this work in the future. First, the KNN classifier is simple
and could be replaced with a more sophisticated classifiers. Second, it would be
interesting to automate the system of hand gesture-based sign language and use it
for dynamic hand. Finally, we are interested to adopt the three proposed QDMI on
more challenging situations such as luminosity changes, occultation of the target,
and many others issues.
References
1. Elouariachi, I., Benouini, R., Zenkouar, K., Zarghili, A.: Robust hand gesture recognition
system based on a new set of quaternion Tchebichef moment invariants. Pattern Anal. Appl.
1–17 (2020)
2. Elouariachi, I., Benouini, R., Zenkouar, K., Zarghili, A., El Fadili, H.: Explicit quaternion
krawtchouk moment invariants for finger-spelling sign language recognition. In: 2020 28th
European Signal Processing Conference (EUSIPCO), pp. 620–624. IEEE (2021, January)
3. Ren, Z., Yuan, J., Meng, J., Zhang, Z.: Robust part-based hand gesture recognition using kinect
sensor. IEEE Trans. Multimed. 15, 1110–1120 (2013)
4. Huang, D.Y., Hu, W.C., Chang, S.H.: Gabor filter-based hand-pose angle estimation for hand
gesture recognition under varying illumination. Expert Syst. Appl. 38(5), 6031–6042 (2011)
5. Li, Y.T., Wachs, J.P.: HEGM: a hierarchical elastic graph matching for hand gesture recognition.
Pattern Recognit. 47(1), 80–88 (2014)
6. Lin, J., Ding, Y.: A temporal hand gesture recognition system based on hog and motion trajec-
tory. Optik 124(24), 6795–6798 (2013)
7. Patil, S.B., Sinha, G.R.: Distinctive feature extraction for Indian Sign Language (ISL) gesture
using scale invariant feature Transform (SIFT). J. Inst. Eng. (India): Ser. B 98(1), 19–26 (2017)
8. Zhang, F., Liu, Y., Zou, C., & Wang, Y.: Hand gesture recognition based on HOG-LBP fea-
ture. In: 2018 IEEE International Instrumentation and Measurement Technology Conference
(I2MTC), pp. 1–6. IEEE (2018, May)
9. Benouini, R., Batioua, I., Elouariachi, I., Zenkouar, K., Zarghili, A.: Explicit separable two
dimensional moment invariants for object recognition. Procedia Comput. Sci. 148, 409–417
(2019)
10. Flusser, J., Suk, T., Zitov, B.: 2D and 3D image analysis by moments. Wiley, Hoboken (2016)
11. Jadooki, S., Mohamad, D., Saba, T., Almazyad, A.S., Rehman, A.: Fused features mining for
depth-based hand gesture recognition to classify blind human communication. Neural Comput.
Appl. 28(11), 3285–3294 (2017)
12. Hu, Y.: Finger spelling recognition using depth information and support vector machine. Mul-
timedia Tools Appl. 77(21), 29043–29057 (2018)
13. Gallo, L., Placitelli, A.P.: View-independent hand posture recognition from single depth images
using PCA and Flusser moments. In: 2012 eighth international conference on signal image
technology and internet based systems, pp. 898–904. IEEE (2012, November)
14. Hamilton, W.R.: Elements of quaternions. Longmans, Green, & Company (1866)
15. Krawtchouk, M.: On interpolation by means of orthogonal polynomials. Memoirs Agric. Inst.
Kyiv 4, 21–28 (1929)
16. Zhou, J., Shu, H., Zhu, H., Toumoulin, C., & Luo, L.: Image analysis by discrete orthogonal
Hahn moments. In International Conference Image Analysis and Recognition, pp. 524–531.
Springer, Berlin, Heidelberg (2005, September)
17. Wang, C., Liu, Z., Chan, S.C.: Superpixel-based hand gesture recognition with kinect depth
camera. IEEE Trans. Multimedia 17(1), 29–39 (2014)
18. Pugeault, N., Bowden, R.: Spelling it out: Real-time ASL fingerspelling recognition. In: 2011
IEEE International conference on computer vision workshops (ICCV workshops), pp. 1114–
1119. IEEE (2011, November)
19. Mukundan, R., Ong, S.H., Lee, P.A.: Image analysis by Tchebichef moments. IEEE Trans.
Image Process. 10(9), 1357–1364 (2001)
20. Karakasis, E.G., Papakostas, G.A., Koulouriotis, D.E., Tourassis, V.D.: Generalized dual Hahn
moment invariants. Pattern Recogn. 46(7), 1998–2014 (2013)
21. VisionTexture (VisTex). http://vismod.media.mit.edu/vismod/imagery/VisionTexture/Images/
Reference/
22. Colored Brodatz (CBT). http://multibandtexture.recherche.usherbrooke.ca/colored%20_
brodatz.html
23. Outex texture (Outex). http://lagis-vi.univ-lille1.fr/datasets/outex.html
24. Amsterdam Library of Textures (Amsterdam). http://aloi.science.uva.nl/public_alot/
Virtual Spider for Real-Time Finding
Things Close to Pedestrians
Souhail Elkaissi and Azedine Boulmakoul
Abstract The active growth of technology gives a big push to the way we collect
the spatial data either by surveying which means collect the data by surveyors and
then use it to create highly precise maps, they calculate the precise position of points,
distances, and angles through geometry. In this paper, we consider the collection of
timely spatial data by a method based on spiders’ behavior. The data can be collected
by remote sensing which uses satellites orbiting the Earth to capture information
of the surface and atmosphere. One of the major challenges in collecting spatial
data is to get it as fast as possible and to be able to scale contingent on the amount
and the size of the data. In this matter, we see that messages brokers like Kafka
can be very useful, thus it gives the ability to provide a real-time architecture for
a real-time streaming data, it is scalable, durable, and a fault-tolerant published-
subscribe messaging system. The other major challenge in collecting spatial data is
to be able to manipulate this data with a smoother and faster way. Yet geography is a
natural data domain for graphs and graph databases. Geometries and topologies could
simply be drawn on graph databases like Neo4j. With their ability of expressing geo
data intuitive way, graph databases could be used from calculating routes between
locations in an abstract network such as a road or rail network, airspace network, or
logistical network to spatial operations such as find all points of interest in a bounded
area.
1 Introduction
With the growth of technology, the location awareness became more and more fluid,
this is due to the usage of connected objects around us, like GPS devices, smartphones,
sensors, etc. These devices collect a lot of information about individual persons,
communities, and the eco-system that we live in. It collects what each individual does,
or where he goes, when he eats his breakfast, his heartbeat rating, etc. These collected
data can be very massive in terms of field diversity and in term of size, so it should
S. Elkaissi · A. Boulmakoul (B)

LIM/IOS, FSTM, Hassan II University of Casablanca, Casablanca, Morocco
https://doi.org/10.1007/978-981-16-3637-0_53
750 S. Elkaissi and A. Boulmakoul
be stored in a scalable format where the data can be manipulated and transformed
smoothly [16–18]. That said graph databases provide an excellent infrastructure to
link diverse data. With easy expression of entities and relationships between data,
graph databases make it easier for programmers, users, and machines to understand
the data and find insights. This deeper level of understanding is vital for successful
machine learning initiatives, where context-based machine learning is becoming
important for feature engineering, machine-based reasoning, and inferencing [1, 2,
10]. These collected data can be used to enhance the quality and comfort of the
pedestrian on the pedestrian walkaway. We can help to increase the safety of the
pedestrian by minimizing the number of eventual accidents that could happen, and
also, it will help to increase the walkability of the surrounding. We could find out ways
or keys of amelioration of pedestrian walkway. The data allows to understand how the
pedestrian walks. To sum up, the pedestrian study helps to increase the sustainability
of the walking area. In the paper, we have created a spatial data pipeline based on a
virtual spider. Let us imagine that it exists an intelligent virtual spider that can travel
on streets or avenues or even cities and it is conscious of its position, and it is aware
of all the elements that are in its circular area of radius r such as stores, avenues,
highways, institutions. The spider will go through long straight streets and try to
collect all the possible information around. All the collected information is spatial
data, which is a geometry like points, lines or polygons that have a meaningful
semantic match of a real and a physical object. At the end of its journey, the spider
could give us the point of interest based on a criterion. This kind of systems (like
the spider) is extremely useful nowadays, and it could be such a relief in situations
of finding extremely fast some points of interest like, for example, the time when it
is safe to go on the pedestrian walkway [14, 15]. Trajectory-based human mobility
data analysis research has largely focused on the trajectories of people and vehicles,
driven by the fact that geographic information science has traditionally supported
spatial information from moving objects [5–9, 11–13]. Currently, the interest of
spatiotemporal data analysis for road safety is a real challenge for modern cities
[3, 4].
The rest of this article is organized as follows; the following section describes
the main components of the proposed system, namely the graph database and the
messaging broker. Section 3 details the implementation of the proposed architecture.
The article ends with a conclusion and offers further work in Sect. 4.
2 System Components
2.1 Graph Database
Graph databases are purpose-built to store and navigate relationships. Relationships

are first-class citizens in graph databases, and most of the value of graph databases
Virtual Spider for Real-Time Finding Things Close to Pedestrians 751
Fig. 1 Graph of different nodes and edges (Neo4j (https://neo4j.com/))
is derived from these relationships. Graph databases use nodes to store data enti-
ties and edges to store relationships between entities. An edge always has a start
node, end node, type, and direction, and an edge can describe parent–child relation-
ships, actions, ownership, and the like. There is no limit to the number and kind of
relationships a node can have (Fig. 1).
A graph in a graph database can be traversed along specific edge types or across
the entire graph. In graph databases, traversing the joins or relationships is amazingly
fast because the relationships between nodes are not calculated at query times but
are persisted in the database. Graph databases have advantages for use cases such as
social networking, recommendation engines, and fraud detection, when you need to
create relationships between data and quickly query these relationships.
2.1.1 Neo4j Graph Database
Neo4j is an open-source, NoSQL, native graph database that provides an ACID-

compliant transactional backend for your applications. The source code, written
in Java and Scala, is available for free on GitHub or as a user-friendly desktop
application download. Neo4j is referred to as a native graph database because it
efficiently implements the property graph model down to the storage level. This
means that the data is stored exactly as your whiteboard is, and the database uses
pointers to navigate and traverse the graph. In contrast to graph processing or in-
memory libraries, Neo4j also provides full database characteristics, including ACID
transaction compliance, cluster support, and runtime failover—making it suitable to
use graphs for data in production scenarios. Some of the following features make
Neo4j extremely popular among developers, architects, and DBAs:
• Cipher, a declarative query language similar to SQL, but optimized for graphs.
• Constant time traversals in big graphs for both depth and breadth due to efficient
representation of nodes and relationships. Enables scale-up to billions of nodes
on moderate hardware.
• Flexible property graph schema that can adapt over time, making it possible to
materialize and add new relationships later to shortcut and speed up the domain
data when the business needs change.
• Drivers for popular programming languages, including Java, JavaScript,.NET,
Python, and many more.
Neo4j has a variety of plugins that can extend its capacity and can add powerful
capabilities to Neo4j. Plugins are meant to extend the capabilities of the database,
nodes, or relationships.
2.1.2 Neo4j Spatial Plugin
Neo4j spatial is a plugin utility for Neo4j that facilitates the enabling of spatial oper-
ations on data. In particular, you can add spatial indexes to already located data and
perform spatial operations on the data like searching for data within specified regions
or within a specified distance of a point of interest. In addition, classes are provided
to expose the data to GeoTools and thereby to GeoTools-enabled applications like
GeoServer and uDig.
2.2 Kafka Message Broker
Kafka1 technically speaking is an event streaming platform which is the practice

of capturing data in real time from event sources like databases, sensors, mobile
devices, cloud services, and software applications in the form of streams of events;
storing these event streams durably for later retrieval; manipulating, processing, and
reacting to the event streams in real time as well as retrospectively; routing the event
streams to different destination technologies as needed. Event streaming thus ensures
a continuous flow and interpretation of data so that the right information is at the
right place, at the right time.
1 https://kafka.apache.org/.
2.3 Geographical Information System
A geographic information system (GIS) is a conceptualized framework that provides

the ability to capture and analyze spatial and geographic data. GIS applications
(or GIS apps) are computer-based tools that allow the user to create interactive
queries (user-created searches), store and edit spatial and non-spatial data, analyze
spatial information output, and visually share the results of these operations by
presenting them as maps. Modern GIS technologies use digital information for
which various digitized data creation methods are used. The most common method
of data creation is digitization, where a hard copy map or survey plan is trans-
ferred into a digital medium through the use of a CAD program and geo-referencing
capabilities. With the wide availability of ortho-rectified imagery (from satellites,
aircraft, and UAVs), heads-up digitizing is becoming the main avenue through
which geographic data is extracted. Heads-up digitizing involves the tracing of
geographic data directly on top of the aerial imagery instead of by the traditional
method of tracing the geographic form on a separate digitizing tablet (heads-down
digitizing). Geoprocessing is a GIS operation used to manipulate spatial data. A
typical geoprocessing operation takes an input dataset, performs an operation on that
dataset, and returns the result of the operation as an output dataset. Common geopro-
cessing operations include geographic feature overlay, feature selection and analysis,
topology processing, raster processing, and data conversion. Geoprocessing allows
for definition, management, and analysis of information used to form decisions.
3 System Design
In this part, we will talk about the design and architecture of our system (Fig. 2).
We will go and explain in details every section in the following sections.
3.1 Collected Data
We are interested into gathering as much data as possible from many sources at once
and bind that data together and store it into one database. The first part of the system
is the data gathering (Fig. 3).
The spider will go around pedestrian walkways to collect all kind of useful infor-
mation. Each of this information above is sent by the spider using a Kafka producer
to a specific topic in the Kafka broker.
Fig. 2 System architecture of spider’s gathering data
Fig. 3 All data sources of the spider
3.1.1 GIS Surrounding Data
We got an intelligent spider that can detect its position and be aware of all the elements
that are in a circular area of radius R. We can expand the radius R to collect a mush
GIS data as possible for a big surrounding area. The spider can move freely over
streets ups and downs. At each point in its path beside or on pedestrian walkway,
the spider collects all the spatial data on a circular area of a radius R, indeed the
spatial data can be any kind of geometries such as points, polygons, or lines. Every
geometry is an abstraction of real and physical places or locations that may be a point
of interest. You can see in red circles on the next images the points that are beside
Ibn Rochd Avenue. Also, in the next images, we can see in purple the polygons close
to Ibn Rochd Avenue (Figs. 4, 5).
For example, if the spider is at the point with the coordinates (33.987181, −
6.852403) on the road of the Ibn Rochd Avenue and the radius is 70 m, the spider
will collect the following points (Fig. 4).
• Burger kings Rabat,
• Laboratoire Ibn Nafiss,
Fig. 4 Example of closest points to spider
Fig. 5 Polygons beside Ibn Rochd Avenue

• Café Sokaina,
• Latiere Anwar,
• Creches maternelle les P’tits Explorateurs.
All these GIS information helps us to measure the sustainability of the walking
area.
3.1.2 Road Accidents Data
While the spider traveling from road and pedestrian walkways, it detects the acci-
dents closer to its radius. The accidents data (thus it can be retrieved from the local
authorities) contains (see Table 1):
• localization of the accident,
• time of the accident,
• type of the accident (accident between two cars, or between a car and a pedestrian)
In the next table, we present a slice of data retrieved by the spider (Table 2).
Traffic flow data. Also, the spider collects the number of cars going on every
road and pedestrian walkway beside going beside it. The collected data contains:
• The name of the road.
• The number of cars going on the road.
• The time of the collection of the data.
In the next table, we present a slice of data retrieved by the spider (Table 3).
Traffic light data. The spider collects the traffic light state on the roads beside the
pedestrian walkway. The state of the traffic light nearby the pedestrian every minute
Table 1 Example table of

Localization Time Accident type
road accidents data
(33.988849, −6.824750) 2019-01-19 22 h Cars
(33.988039, −6.852750) 2019-03-25 04 h Cars
(33.988129, −6.854112) 2019-05-11 14 h Humans
(33.988479, −6.858350) 2019-5-09 13 h Cars
(33.988409, −6.851250) 2019-11-02 20 h Humans
Table 2 Example of traffic

Name Number of cars Time
flow data on Avenue Ibn
Rochd Avenue Ibn Rochd 15 2019-01-10 21 h 10 min
Avenue Ibn Rochd 3 2019-01-10 21 h 12 min
Avenue Ibn Rochd 09 2019-01-10 21h16min
Table 3 Example table of

Road name Time Is red
traffic lights on Avenue Ibn
Rochd crossroads Avenue Ibn Rochd 2019-01-10 19 h 53 min True
Avenue Ibn Rochd 2019-01-10 21 h 10 min True
Avenue Ma El Aynane 2019-01-10 19 h 53 min False
Avenue Ma El Aynane 2019-01-10 19 h 55 min True
Avenue Ibn Rochd 2019-01-10 19 h 55 min False
Fig. 6 Example of pedestrian count
can be red or green. In the next table, we present a slice of data retrieved by the
spider.
Pedestrian information. Here, the spider collects the number of pedestrians
walking on the pedestrian walkway (Fig. 6). The information collects contains:
• The number of pedestrians walking.
• The time the spider retrieves the data.
As you can see in Fig. 6, on Sunday there are more people outside walking on the
pedestrian walkaway at 2 p.m. Also, we can see that there are more people walking
on the pedestrian walkway between 7 a.m. and 5 p.m. over all the days of week.
Weather. The spider collects the information about the weather state on its
localization, beside the pedestrian walkway. The information contains:
• The localization of the spider.
• The time of the collection of the data.
• The degree in Celsius
• Whether it is raining or not.
Fig. 7 Producing spatial data into Kafka broker
Fig. 8 Kafka consumer
3.2 Kafka Producer
While the spider collects the spatial data, there is a data formatter that processes these
data, deletes the unwanted information from the geometry’s attributes, and deletes
the wrong or inappropriate ones. After the steps of formatting and processing have
been finished, the Kafka producers send the data to the Kafka topics using a Geo
serializer (Figs. 7, 8).
3.3 Kafka Consumer
The spatial Neo4j Kafka consumers listen to the brokers on the servers. If there is any
messaging event coming from the tracked partitions, it gets all these messages and
tries to deserialize it is using the Geo deserializer been implemented (Fig. 8). After
getting and deserializing this data, the consumer sends the data to Neo4j instance
graph databases. We use a plugin in the Neo4j database called neo4j spatial, and
it helps to enable geographic capabilities on your data and complement the spatial
functions that come with graph database Neo4j.
Fig. 9 Data representation in the graph
3.4 Neo4j Graph in the Database
All the collected data is sent to neo4j database, and in this section, we will see how
the data it’s been modeled in the graph database (Fig. 9).
Relationship types in our database. As you can see in Fig. 9, the main node is
the geometry it can be a point (stores…) and lines (highway, pedestrian walkway…),
polygons (building…):
• The connection between two geometries is called IS_NEARBY.
• The connection between a geometry node and accidents node is called ACCI-
DENT_HAPPENED, and it contains the property time of the accident.
• The connection between a geometry and car traffic node is called
HAD_CAR_TRAFFIC, and it contains the property time.
• The connection between the node geometry and the traffic light is red node, it is
called HAD_TRAFFIC_LIGHT, and it contains the property time.
• The connection between the node geometry and pedestrian count node, it is called
HAD_PEDESTRIAN, and it contains the property time.
Note that the property time is a representation of the time with days, hours, and
minutes, this choice was made in purpose to be able to know what happens in or
beside every geometry in our database.
3.5 Find Things Close to Pedestrian
Now that the database contains the geo-temporal-spatial information that we want,
we can make specific requests to the database using cipher (Fig. 10).
Fig. 10 Neo4j database after the insertion of the data
Examples of finding things close to pedestrian. To find places close to pedestrian

in Ibn Rochd avenue at 11 h 00 a.m., we could formulate this cipher query:
Match (:Geometry {Name: ‘Ibn Rochd’})-[:IS_NEARBY {time:
’11:00’}]->(place:Geometry) RETURN place.name
To find the number of accidents that occurs at Ibn rochd avenue in a raining
weather, we could use the following cipher query:
Match (accidents:Accidents)<-[:ACCIDENT_HAPPENED]-(:Geometry
{Name: ‘Ibn Rochd’})-[:HAD_WEATHER]->(weather:Weather) Where
weather.State = ‘Rain’ RETURN accidents.number
4 Conclusion
In this work, we presented the first development of a system for collecting oppor-
tunistic spatial data by a method based on spiders’ behavior. The main challenge in
collecting spatial data is to be able to obtain it as quickly as possible depending on
the quantity and size of the data. In this construction, message brokers like Kafka
are extremely useful, giving the possibility of providing the real-time architecture
for real-time data delivery. Kafka is scalable and offers a fault-tolerant subscription
messaging system. The developed system allows the storage of the collected spatial
data in a spider network defined by a Neo4j graph. This persistence will be used
later for analytical purposes. The specialization of the proposed solution to smart
city needs is scheduled in our future work.
Acknowledgments This work was partially funded by Ministry of Equipment, Transport, Logis-
tics and Water−Kingdom of Morocco, The National Road Safety Agency (NARSA), and National
Center for Scientific and Technical Research (CNRST). Road Safety Research Program# An intelli-
gent reactive abductive system and intuitionist fuzzy logical reasoning for dangerousness of driver-
pedestrians interactions analysis: Development of new pedestrians’ exposure to risk of road accident
measures.
References
1. Boulmakoul, A., Fazekas, Z., Karim, L., Gáspár, P., Cherradi, G.: Fuzzy similarities for road
environment-type detection by a connected vehicle from traffic sign probabilistic data. Procedia
Comput. Sci. 170, 59–66 (2020). ISSN 1877-0509
2. Boulmakoul, A., Karim, L., El Bouziri, A., Lbath, A.: A system architecture for heterogeneous
moving-object trajectory metamodel using generic sensors: tracking airport security case study.
IEEE Syst. J. 9(1), 283–291 (2015)
3. Coulton, C.J., Jennings, M.Z., Chan, T.: How big is my neighborhood? Individual and contex-
tual effects on perceptions of neighborhood scale. Am. J. Community Psychol. 51, 140–150
(2013)
4. Erkan, I.: Cognitive analysis of pedestrians walking while using a mobile phone. J. Cogn. Sci.
18(3), 301–319 (2017). https://doi.org/10.17791/jcs.2017.18.3.301
5. Georgiou, H., et al.: Moving Objects Analytics: Survey on Future Location &Trajectory
Prediction Methods, Technical Report. arXiv:1807.04639 (2018)
6. Gómez, L.I., Kuijpers, B., Vaisman, A.A.: Analytical queries on semantic trajectories using
graph databases. Trans. GIS 23(5). (John Wiley & Sons Ltd)
7. Güting, R.H., de Almeida, V.T.: Ding Z Modeling and querying moving objects in networks.
VLDB J 15(2), 165–190 (2006)
8. Güting, R.H., Schneider, M.: Moving objects databases. Morgan Kaufmann, San Francisco,
CA (2005)
9. Gómez, L.I., Kuijpers, B., Vaisman, A.A.: Analytical queries on semantic trajectories using
graph databases. Trans. GIS. 23, 1078–1101. https://doi.org/10.1111/tgis.12556
10. Maguerra, S., Boulmakoul, A., Karim, L., et al.: Towards a reactive system for managing big
trajectory data. J Ambient Intell. Human Comput. 11, 3895–3906 (2020). https://doi.org/10.
1007/s12652-019-01625-3
11. Open Geospatial Consortium, Inc. OGC KML documentation. http://www.opengeospatial.org/
standards/kml/ (2012)
12. Parent, C., Spaccapietra, S., Renso, C. Andrienko, G., Andrienko, N., Bogorny, V., Yan, Z.:
Semantic trajectories modeling and analysis. ACM Comput. Surv. 45(4), 42:1–42:32 (2013)
13. Popa, I.S., Zeitouni, K., Oria, V., Kharrat, A.: Spatiotemporal compression of trajectories in
road networks. GeoInformatica 19(1), 117–145 (2015)
14. Qi, F., Du, F.: Trajectory data analyses for pedestrian space-time activity study. J. Vis. Exp. 72,
50130 (2013)
15. Vecchio, P., Secundo, G., Maruccia, Y., Passiante, G.: A system dynamic approach for the
smart mobility of people: Implications in the age of big data. Technol. Forecast. Soc. Change
149, 119771 (2019)
16. Yoon H., Zheng Y., Xie X., Woo W.: Smart itinerary recommendation based on user-generated
GPS trajectories. In: Yu, Z., Liscano, R., Chen, G., Zhang, D., Zhou, X. (eds) Ubiquitous Intel-
ligence and Computing. UIC 2010. Lecture Notes in Computer Science, vol. 6406. Springer,
Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-16355-5 5 (2010)
17. Zheng, Y., Zhou, X., et al. :Computing with Spatial Trajectories. Springer, Berlin (2011). ISBN
978-1-4614-1629-6
18. Zhong, R.Y., Huang, G.Q., Shulin, L., Dai, Q.Y., Xu, C., Zhang, T.: A big data approach
for logistics trajectory discovery from RFID-enabled production data. Int. J. Prod. Econ. 165,
260–272 (2015)
Evaluating the Impact of Oversampling
on Arabic L1 and L2 Readability
Prediction Performances
Naoual Nassiri, Abdelhak Lakhouaja, and Violetta Cavalli-Sforza
Abstract Most Arabic educational corpora, which are used to elaborate readability
prediction models, suffer from an unbalanced distribution of texts among difficulty
levels. We argue that readability prediction using machine learning (ML) methods
should be addressed through a balanced learning corpus. In this work, we address, in
a first experiment, the problem of imbalanced data by clustering classes, an approach
adopted by several state-of-the art studies. We then present the results of a second
experiment in which we adopted an oversampling technique on a unbalanced corpus
in order to train the models on balanced data. This experiment was carried out on
four corpora, including three dedicated to Arabic as a foreign language (L2) and
one for Arabic as a first language (L1). The results show that balanced data give a
significant improvement on readability prediction model performances.
1 Introduction
Readability represents the complexity of vocabulary and syntax in a text. It plays

an important role in the process of providing students and readers with materials
at an appropriate difficulty level. Research in this field began first with the use of
mathematical formulas, known as traditional approaches. These scores are a simple
way of measuring whether written information is likely to be understood by the
intended reader. They produce an index that can indicate approximately the level of
education a person will need to be able to read a text easily.
N. Nassiri (B) · A. Lakhouaja

Department of Computer Science, Faculty of Sciences, University Mohamed First, Oujda,
Morocco
V. Cavalli-Sforza
School of Science and Engineering, AI Akhawayn University, Ifrane, Morocco
e-mail: v.cavallisforza@aui.ma
https://doi.org/10.1007/978-981-16-3637-0_54
764 N. Nassiri et al.
Later, researchers used machine learning (ML) approaches to automatically pre-

dict the difficulty level of a given text. The approach is a supervised ML process that
consists of applying a classification algorithm to text features in order to generate
models from a set of pre-classified training texts. The models are then used to clas-
sify new texts. These approaches have shown better results compared to traditional
techniques, but this performance is highly correlated to the learning corpus.
Most Arabic readability studies suffer from the problem of unbalanced data.
Unbalanced data means that at least one class is under-represented or over-represented
compared to the others, thus perturbing learning algorithms and making them unable
to recognize and classify minority instances. This phenomenon occurs in several
application areas, such as fraud detection, text categorization, and medical diagnos-
tics, and represents a challenge for the data mining community.
The task of automatically measuring text readability is frequently faced with chal-
lenges such as the lack of data and the imbalance of available data sets. ML algo-
rithms, applied to automatically predict readability, tend to produce unsatisfactory
classifiers when confronted with unbalanced data sets. If the text whose difficulty we
aim to measure belongs to the minority class, it is generally incorrectly classified.
The main question asked when analyzing such a data is: how can we balance the data
set to overcome this phenomenon? In order to address this issue, we will present in
this paper a set of experiments that are based on balancing the learning data in order
to build automatic readability prediction models on a balanced data set. A number
of techniques can be adopted to address this problem such as:
1. Class clustering.
2. Transformation of the learning algorithm.
3. Sampling methods such as oversampling and under-sampling.
In this paper, we are interested in data sampling methods and particularly “over-
sampling." These approaches are widely used and have given encouraging results in
cases where the data are unbalanced.
The rest of this paper is organized as follows. We review in Section 2 a set of studies
on Arabic automatic readability measurement, and we highlight the weaknesses of
these approaches. Details on different unbalanced data classification techniques are
given in Sect. 3. Section 4 presents the tools, the data, and the process that we adopted
for this study. The tests and the results are discussed in Sect. 5. Finally, Sect. 6 presents
a conclusion and some future directions to further improve the obtained results.
2 Related Work
Readability assessment is considered, in the majority of research, as a classification

problem. This process is carried out in three main steps:
1. Assembling a reference corpus: The quality and the quantity of the texts compos-
ing the reference corpus are important.
Evaluating the Impact of Oversampling on Arabic L1 and L2 … 765
2. Transforming corpus texts into feature vectors: The choice of the features to be
extracted is an essential step in the construction of the prediction model. These
must have a correlation with the level to be predicted.
3. Applying a classification algorithm on the data.
For Arabic as a foreign language (L2), most of the studies carried out following
this process are conducted using unbalanced corpora. In 2014, Forsyth described in
his thesis [8], a study is based on a corpus comprising 179 texts. The corpus texts
were annotated with difficulty levels based on the Interagency Language Roundtable
(ILR) organization scale [7]. The scale is used to measure language proficiency by
concerned entities in the U.S. federal government. This scale assesses language skills
on levels ranging from 0 to 5. Levels 0+, 1+, 2+, 3+, or 4+ are used when a person’s
skills far exceed those of a given level but are insufficient to reach the next level.
Forsyth’s corpus is annotated with five difficulty levels (1, 1+, 2, 2+, and 3) using this
scale. The corpus contains an unbalanced distribution of texts between the different
levels as given in Table 1. Based on cross-validation, he reported a maximum F-score
value of 0.52 and 0.78, respectively, for five and three classes (obtained by grouping
levels 1 and 1+ and levels 2+ and 3). The gain in F-score obtained by using three
classes is due to the combination of the levels, and this leads to a relative balance in
the data.
In 2015, Saddiki et al. carried out a study [13] in which they gathered a corpus of
251 texts distributed over five ILR difficulty levels, as given in Table 1. They reported
a maximum accuracy value of 59.76% and a maximum F-score of 0.58 using five
classes, versus a maximum accuracy of 73.31% and a maximum F-score of 0.73 using
three classes. In 2018, Nassiri et al. [12] collected a corpus of 230 texts (Table 1) and
reported a maximum accuracy value of 89.56% using a five-way classification when
testing the models on the training data.
Studies on Arabic as a first language (L1) started in 2010 with Al-Khalifa and Al-
Ajlan [2], who used, for the development of their tool, a corpus collected manually
from reading books of elementary, intermediate and secondary school curricula in
Saudi Arabia. Their corpus was composed of 150 texts distributed in a balanced way
between three classes (50 text per level). They reported an accuracy rate of 77.77%.
In 2014, Al-Tamimi et al. [4] collected a corpus of ten levels containing a total
of 1,196 texts gathered from the Jordanian curriculum. They then re-annotated the
Table 1 State-of-the-art readability corpora

Level Forsyth [8] Saddiki et al. [13] Nassiri et al. [12]
1 20 33 27
1+ 14 28 19
2 80 91 87
2+ 40 63 62
3 25 63 35
Total 179 251 230
corpus with three classes (easy, medium, and difficult) and achieved an average
accuracy of 83.23%.
Finally, in 2018, Saddiki et al. conducted a study [14] to predict the readability
of L1 texts based on a corpus developed by Al Khalil et al. [3]. This corpus is
composed of 27,688 texts divided into four difficulty levels. The first three levels
have been extracted from UAE textbooks, and the fourth level contains novels. Their
best accuracy result is 94.8%.
Most of the studies we presented in this section used class clustering to address
the problem of unbalanced data. Unfortunately, only Saddiki et al. [14] adopted an
approach in which they limited the length of texts by splitting them into approximately
equal sizes, thus increasing the number of texts in a class. In the remainder of this
paper, we will use both class clustering and oversampling techniques to address this
phenomenon.
3 Classification and Class Imbalance
Most existing classification methods are not suitable for use with the minority class
when the class is extremely unbalanced. This problem has become a challenge for
many researchers, as it is present in many real-world applications. To deal with this
challenge, two approaches are widely used:
1. Under-Sampling the learning data, and
2. Oversampling the learning data.
3.1 Under-Sampling
Under-sampling consists in balancing the data set by decreasing the number of

instances of the majority class. Among the best-known algorithms of this category
of approaches, we cite:
• Random Under-sampling [9]: It involves randomly drawing samples from the

majority class, with or without replacement. However, it can increase the variance
of the classifier and can potentially eliminate useful or important samples.
• Tomek Links [15]: It removes unwanted overlap between classes where majority
class links are removed until all pairs of closest neighbors, at minimum distance,
are of the same class.
3.2 Oversampling
Oversampling consists in balancing the data set by artificially increasing the number
of instances of the minority class. Among the best-known oversampling algorithms,
we have:
• Random Oversampling [9]: It involves supplementing the training data with

multiple copies of some instances of the minority class.
• SMOTE (Synthetic Minority Over-sampling Technique) [6]: Rather than repro-
ducing minority observations, SMOTE creates a user-selected number of synthetic
observations on segments between items close to the minority class. SMOTE intro-
duces the synthetic examples randomly on the line connecting the minority class
point concerned and one of its closest K neighbors (the number of neighbors ‘k’
is a user-defined parameter). The principle is based on the difference between a
minority instance and its closest neighbor, this difference is multiplied by a ran-
dom number between 0 and 1, this causes the selection of a random point along the
line segment between two minority instances. For a sufficient number of synthetic
instances, the SMOTE algorithm consists of:
– Selecting a minority instance A,

– Selecting one instance B among the closest neighbors,
– Selecting a random weight between 0 and 1 (W) to create the new synthetic
instance C, and
– Applying, for each attribute, the formula:
att V alue_C = att V alue_A + (att V alue_B − att V alue_ A) ∗ W
4 Resources and Methodology
In this study, we adopted two approaches to develop readability prediction models

based on balanced data sets:
1. The first approach consists of redistributing the data into three classes: easy,
medium, and difficult. These three levels are obtained by grouping adjacent levels
together.
2. The second approach consists of using the SMOTE technique to re-balance the
learning data.
In this section, we will present the data on which we evaluated our approaches,
the tools we used, and the process we adopted in each approach.
4.1 Data
In this section, we will present three of the most used corpora for L2 readability
measurement and one corpus dedicated to L1 called MoSAR [10].
• L2 readability corpora: For Arabic as L2, most of the published research is per-
formed using corpora from the GLOSS1 platform, whose content was developed
by the Defense Language Institute Foreign Language Center (DLIFLC2 ), consid-
ered one of the best foreign language schools. The platform offers thousands of
lessons in dozens of languages for independent learners to strengthen their lis-
tening skills and reading. MSA texts in GLOSS are annotated with five difficulty
levels using the ILR scale, described earlier. We collected two corpora from this
source, namely:
– GLOSS-Reading, which consists of 271 reading texts.

– GLOSS-Listening, which comprises 227 listening texts.
We have also used the texts available online by Aljazeera3 to form the third L2
corpus. The Aljazeera-learning Web site, for learning Arabic, presents texts for
educational purposes. The texts are annotated with five levels of difficulty, namely:
Beginner1, Beginner2, Intermediate1, Intermediate2, and Advanced. The gathered
corpus contains a total of 321 texts.
In order to facilitate the representation of the results, we relabeled the three corpora
levels into Level 1, Level 2, . . ., Level 5. Note that this new labels meaning varies
from one corpus to another. Table 2 presents the distribution of these three corpora
across the five difficulty levels.
The statistics given in Table 2 allow us to appreciate the great imbalance in the dis-
tribution of corpora texts between the five difficulty levels. For Aljazeera-learning,
for example, we have 220 texts in level 2 while we have only 8 texts in level 5.
Similarly for the GLOSS-listening corpus, we have 18 texts for level 2 compared
to 65 texts for level 3. It is this problem of imbalance that we will try to solve in
the remainder of this paper.
• L1 readability corpus: For L1, we used the modern standard Arabic readability
(MoSAR) corpus. MoSAR is composed of a set of texts collected from textbooks
used in the six Moroccan primary school levels and annotated according to these
six levels. MoSAR consists of 602 texts as given in Table 3.
1 https://gloss.dliflc.edu/.
2 https://www.dliflc.edu/global-language-online-support-system-gloss/.
3 https://learning.aljazeera.net.
Table 2 L2 corpora texts distributed by five difficulty levels

GLOSS-reading GLOSS-listening Aljazeera-learning
Level 1 34 43 12
Level 2 29 18 220
Level 3 95 65 54
Level 4 68 62 27
Level 5 45 39 8
Total 271 227 321
Table 3 MoSAR text distribution by levels

Level Number of books Number of texts
1 1 51
2 2 139
3 1 60
4 2 136
5 1 86
6 1 130
Total 8 602
4.2 Morphological Annotation
For the morphological annotation of the four corpora, we used Alkhalil-Toolkit

which is a morphological analysis and disambiguation system composed of several
independent tools. For this task, we used specifically:
1. Alkhalil-Lemmatizer [5], which assigns a unique lemma to each word taking
into account its context, and
2. Alkhalil-PoS-Tagger [1], which offers a very rich set of basic tags providing
syntactic information for each word of the analyzed text.
4.3 Methodology
The first experiment that we will present consists of building predictive models on
five classes (unbalanced corpora) then on three classes data sets distributed as given
in Table 4.
The experimental process consists of three steps, as follows (see Fig. 1):
1. Morphological analysis: The input of this phase is a text from one of the corpora.
The result of this phase is a file (for each text) annotated with information such
as PoS, lemma, and diacritical signs.
Table 4 Corpora distribution over three levels

Level GLOSS-reading GLOSS-listening Aljazeera-learning MoSAR
Easy 63 61 232 190
Medium 95 65 54 196
Difficult 113 101 35 216
Total 271 227 321 602
Fig. 1 Model generation process on unbalanced data
2. Feature extraction: In this step, we extract and calculate a list of features [11].
For each corpus file, we obtain a feature vector that we will use to prepare the
input file for the classification phase.
3. Classification: We apply a classification algorithm on 80% of the generated vec-
tors (training data), randomly selected, in order to build a prediction model. The
obtained model is tested on the remaining 20% of the generated vectors.
The second experiment consists of regenerating the models on balanced learning

vectors. To do this, we proceed as follows (see Fig. 2).
1. Apply the SMOTE technique on the learning vectors (the 80%) to balance them,
2. Use the balanced vectors set to train and build a prediction model, and
3. Test the obtained model on the remaining 20% unseen data.
This second experiment is applied only on texts distributed among five levels,
since the objective is to evaluate the impact of balancing the training data on the
original corpus distribution.
Fig. 2 Model generation process on balanced data
In this section, we will present and discuss the results of the experiments for which
we have already presented the processes in Sect. 4.3.
5.1 L2 Prediction Models
Table 5 presents the results obtained using a five-class classification on the three
corpora. GLOSS-reading achieved an overall accuracy of 60%. This result hides
how accuracy values vary from one class to another; for example, the model was
able to predict level 4 with an accuracy of 25%, while it achieved an accuracy of
83.33% for level 1. For the GLOSS-listening corpus, we obtained a total accuracy
of 56.52%, and for Aljazeera-learning, we obtained 89.23% as accuracy value.
Merging the classes and switching to a three-class classification, allowed us to
obtain the results presented in Table 6. For this second experiment, classification
accuracy for the GLOSS-reading corpus increased from 60 to 70%, for GLOSS-
Table 5 Test results on L2 corpora using five levels

GLOSS-reading (%) GLOSS-listening (%) Aljazeera-learning (%)
Level 1 83.33 77.77 –
Level 2 66.66 16.66 95.65
Level 3 80 75 66.66
Level 4 25 45.45 88.88
Level 5 66.66 50 0
Total 60 56.52 89.23
Table 6 Test results on L2 corpora using three levels

Easy 90 85.71 94.28
Medium 68.75 76.92 72.72
Difficult 61.11 70.58 83.33
Total 70 75 88
Table 7 Test results on L2 corpora after oversampling

Level 1 83.33 88.88 –
Level 2 87.5 33.33 97.82
Level 3 60 75 100
Level 4 31.25 45.45 100
Level 5 77.77 62.5 100
Total 60 63.04 98.46
listening, it increased from 56.52 to 75%, and for the Aljazeera-learning corpus, it
decreased slightly from 89.23 to 88%.
The improvements obtained through the application of this first approach based
on merging classes are very important and are encouraging, but it should be noted
that sometimes we need to classify the texts according to more detailed levels of
difficulty (school grade levels for example), so we should try to improve the results
of the models based on five classes as well. To do so, we present in Table 7 the results
of applying the SMOTE technique on the learning vectors of our three corpora.
The comparison of the results of Tables 5 and 7 leads to the conclusion that the
use of oversampling techniques has improved the classification results both in terms
of total accuracy and in terms of the accuracy of each level independently.
Table 8 MoSAR test results before and after oversampling

Before SMOTE After SMOTE
Six levels (%) Three levels (%) Six levels (%)
1 0 61.76 25
2 84.21 84.21
3 55.55 67.64 66.66
4 77.77 55.55
5 25 75 37.5
6 91.66 83.33
Total 67.21 68 67.21
5.2 L1 Prediction Models
Table 8 presents the results obtained using five classes and three classes classifications
and those obtained when applying the SMOTE technique on MoSAR corpus. For
MoSAR, the total accuracy of 67.21% with five classes was increased to 68% using
three classes. If we compare the five classes results with those obtained using the
SMOTE technique, we can conclude that the use of sampling techniques improves the
classification results at the individual class level, while keeping the same maximum
value of 67.21% overall.
6 Conclusions and Futures Works
This paper proposes a readability prediction approach of MSA texts based on ML

techniques. The main objective of the study is to deal with unbalanced data sets
when training readability prediction models. For these, we experimented with two
balancing approaches: the first one consists of grouping adjacent levels together and
the second one consists of applying an oversampling technique on the training data.
These approaches are evaluated on four corpora, of which three are intended for L2
learners and one for L1 learners.
Using the four corpora with their original unbalanced distributions, we get an
accuracy value of 60, 56.52, 89.23, and 67.21%, while using three classes instead of
five gave corresponding values of 70, 75, 88, and 68%. The application of the SMOTE
oversampling technique enabled us to achieve accuracy values of 60, 63.04, 98.46,
and 67.21% for the same corpora, respectively, thus maintaining or improving the
performance previously achieved with five classes and, in some cases, even with three
classes. This is a promising result when it is important to allow finer discrimination
of text difficulty.
In our future work, we aim to further improve the results of our models by using
alternative data balancing techniques and making changes to the learning algorithms
to support the unbalanced situations. We aim also to perform a n-fold cross-validation
in order the validate the obtained results.
References
1. Ababou, N., Mazroui, A.: A hybrid Arabic POS tagging for simple and compound morphosyn-
tactic tags. Int. J. Speech Technol. 19(2), 289–302 (Jun 2016). http://orcid.org/10.1007/s10772-
015-9302-8
2. Al-Khalifa, H., Al-Ajlan, A.: Automatic readability measurements of the Arabic text: An
exploratory study. Arab. J. Sci. Eng. 35, 103–124 (12 2010)
3. Al Khalil, M., Saddiki, H., Habash, N., Alfalasi, L.: A leveled reading corpus of Modern Stan-
dard Arabic. In: Proceedings of the Eleventh International Conference on Language Resources
and Evaluation (LREC 2018). European Language Resources Association (ELRA), Miyazaki,
Japan (May 2018), https://www.aclweb.org/anthology/L18-1366
4. Al-Tamimi, A.K., Jaradat, M., Aljarrah, N., Ghanim, S.: Aari: Automatic Arabic readability
index. Int. Arab J. Inf. Technol. 11, 370–378 (07 2014)
5. Boudchiche, M., Mazroui, A.: A hybrid approach for Arabic lemmatization. Int. J. Speech
Technol. 22(3), 563–573 (2019). http://orcid.org/10.1007/s10772-018-9528-3
6. Chawla, N., Bowyer, K., Hall, L., Kegelmeyer, W.: Smote: Synthetic minority over-sampling
technique. J. Artif. Intell. Res. (JAIR) 16, 321–357 (06 2002). https://doi.org/10.1613/jair.953
7. Clark, J.L.D., Clifford, R.T.: The fsi/ilr/actfl proficiency scales and testing techniques: devel-
opment, current status, and needed research. Stud. Second Lang. Acquisition 10(2), 129–147
(1988). http://orcid.org/10.1017/S0272263100007270
8. Forsyth, J.N.: Automatic readability detection for modern standard Arabic. Brigham Young
University (2014)
9. Japkowicz, N., Stephen, S.: The class imbalance problem: A systematic study. Intelligent data
analysis 6(5), 429–449 (2002)
10. Nassiri, N., Cavalli-Sforza, V., Lakhouaja, A.: Mosar: Modern standard Arabic readability
corpus for l1 learners. In: Proceedings of the 4th International Conference on Big Data and
Internet of Things. BDIoT’19, Association for Computing Machinery, New York, NY, USA
(2019). https://doi.org/10.1145/3372938.3372961
11. Nassiri, N., Lakhouaja, A., Cavalli-Sforza, V.: Arabic readability assessment for foreign lan-
guage learners. In: Silberztein, M., Atigui, F., Kornyshova, E., Métais, E., Meziane, F. (eds.)
Natural Language Processing and Information Systems. pp. 480–488. Springer International
12. Nassiri, N., Lakhouaja, A., Cavalli-Sforza, V.: Modern standard arabic readability prediction.
In: Lachkar, A., Bouzoubaa, K., Mazroui, A., Hamdani, A., Lekhouaja, A. (eds.) Arabic Lan-
guage Processing: From Theory to Practice, pp. 120–133. Springer International Publishing,
Cham (2018)
13. Saddiki, H., Bouzoubaa, K., Cavalli-Sforza, V.: Text readability for Arabic as a foreign lan-
guage. pp. 1–8 (11 2015). https://doi.org/10.1109/AICCSA.2015.7507232
14. Saddiki, H., Habash, N., Cavalli-Sforza, V., Al Khalil, M.: Feature optimization for predicting
readability of Arabic L1 and L2. In: Proceedings of the 5th Workshop on Natural Language
Processing Techniques for Educational Applications, pp. 20–29. Association for Computational
Linguistics, Melbourne, Australia (Jul 2018). https://doi.org/10.18653/v1/W18-3703, https://
www.aclweb.org/anthology/W18-3703
15. Tomek, I., et al.: An experiment with the edited nearest-nieghbor rule. IEEE Trans. Syst. Man
Cybern. SMC-6(6), 448–452 (1976). https://doi.org/10.1109/TSMC.1976.4309523
An Enhanced Social Spider Colony
Optimization for Global Optimization
Farouq Zitouni, Saad Harous, and Ramdane Maamri
Abstract An improved variant of the social spider optimization algorithm is

introduced. It is inspired by the hunting and mating behaviors of spiders in nature.
We call it enhanced social spider colony optimization (ESSCO). The performance
of ESSCO is evaluated using the benchmark CEC 2020. To validate the proposed
algorithm, the obtained statistical results are compared to eleven recent state-of-the-
art metaheuristic algorithms. The comparative study shows the competitiveness of
ESSCO in finding efficient solutions to the considered test functions.
1 Introduction
Over millions of years of evolution, many intelligent behaviors have arisen in nature.
Biological agents (e.g., insects, birds, mammals, and fishes) have perpetually exhib-
ited adaptability, self-learning, robustness, and efficiency to solve complex tasks
[4, 14, 18, 42]. Researchers started to mimic natural systems during the last decades
to develop several solutions to address difficult problems. Therefore, special atten-
tion has been given to metaheuristic algorithms. Metaheuristics are nature-inspired
F. Zitouni (B)
Department of Computer Science, Kasdi Merbah University, Ouargla, Algeria
LIRE Laboratory, Abdelhamid Mehri University, Constantine, Algeria
e-mail: farouq.zitouni@univ-constantine2.dz
S. Harous
Department of Computer Science and Software Engineering UAE University,
Abu Dhabi, United Arab Emirates
e-mail: harous@uaeu.ac.ae
R. Maamri
Department of Computer Science, Abdelhamid Mehri University, LIRE Laboratory,
Abdelhamid Mehri University, Constantine, Algeria
e-mail: ramdane.maamri@univ-constantine2.dz
https://doi.org/10.1007/978-981-16-3637-0_55
776 F. Zitouni et al.
algorithms: i.e., they mimic biological and physical phenomena. They provide near-
optimal solutions in a reasonable amount of time [46, 57, 62, 84, 89]. We identify
three classes of metaheuristics: (i) algorithms that are based on Darwinian principles,
(ii) algorithms that are based on laws of physics and chemistry, and (iii) algorithms
that are based on swarm intelligence. Table 1 summarizes popular algorithms in each
class.
Metaheuristic algorithms are grouped into two prominent families: (i) algorithms
that use one agent (i.e., individual) and (ii) algorithms that use many agents (i.e.,
population). In the first family, the individual swarms in the search space for a given
number of iterations [70, 79]. Its final position is considered a solution to the opti-
mization problem. In the second family, several agents swarm together in the search
space for a given number of iterations [32, 76]. The position of the best individ-
ual is considered a solution to the optimization problem. Algorithms of the second
family generally outperform the first family’s algorithms concerning the optimality
of obtained solutions [86]. The No-Free-Lunch theorem [35, 88] has declared the
nonexistence of a metaheuristic algorithm that can efficiently solve all optimization
problems. That is to say, some algorithms that show good performance on a given
problem might exhibit poor performance on another one. However, the averaged
performance of all metaheuristics on all optimization problems is equal.
All metaheuristic algorithms share two common steps: exploration and exploita-
tion of the search space. In the exploration step, the algorithm tries to find promising
areas [30]. In the exploitation step, the algorithm attempts to investigate a given
region [11]. In metaheuristics, efficient balancing between these two steps would
lead the algorithm to find near-optimal solutions [12].
The main contribution of the paper is the improvement of the metaheuristic algo-
rithm proposed in [19] for global optimization. The enhancements are summarized
as follows.
• Adding a hunting phase to the basic algorithm;

• Using the blend crossover [25, 26] for generating new offsprings;
• The enhanced algorithm has four parameters to set: number of insects, number
of male spiders, number of female spiders, and an insect’s probability of getting
stuck.
The remainder of the paper is organized as follows. Section 2 presents the biolog-
ical inspiration of ESSCO, its mathematical model, and algorithms. Section 3 sum-
marizes the obtained numerical results and discusses the performance of ESSCO.
Section 4 concludes the paper with a conclusion and some future work.
An Enhanced Social Spider Colony Optimization for Global … 777
Table 1 Popular metaheuristic algorithms for global optimization

Class 1 Class 2 Class 3
Genetic algorithm [33, 63, 78] Simulated annealing [44] Particle swarm optimization
[43, 71]
Evolution strategies [9] Thermodynamic laws [17] Ant colony optimization [20,
21]
Evolutionary programming Gravitation [64, 82] Artificial bee colony [8, 37]
[27, 91]
Genetic programming [45] Big bang–big crunch[24] Grey wolf optimizer [54]
Differential evolution [75] Charged system [41] Bat algorithm [87]
Biogeography-based Central force [28] Whale optimization algorithm
optimization [73] [53]
Quantum-inspired Chemical reaction [2] Dragonfly algorithm [52]
evolutionary algorithm [77]
Black hole [31] Dolphin echolocation [39]
Ray [40] Fruit fly optimization [61]
Small-world [22] Krill herd [29]
Galaxy-based [69] Bird mating optimizer [5]
General relativity theory [55] Hunting search [60]
Spherical search [47] Firefly algorithm [85]
Solar system algorithm [92] Dolphin partner optimization
[72]
Cuckoo search [90]
Bee collecting pollen
algorithm [49]
Marriage in honey bees [1, 83]
Monkey search [58]
Termite [65]
Fish swarm algorithm [48]
Tunicate swarm algorithm [38]
2 Enhanced Social Spider Colony Optimization
2.1 Inspiration
Spiders belong to the class of arachnids. They have eight legs, and the body is
divided into two parts. All spiders are predators. Some species are active hunters,
whereas other species weave webs to capture prey. Some spiders inject venom into
their victims to kill and eat them. Other spiders immobilize the prey by making
silk wrappers around it. According to their social behavior, spiders can be classified
into solitary and social spiders [50]. Solitary spiders weave webs and live there
lonely. Social spiders share the same web and have social relationships (building and
maintaining the web, hunting, and mating).
A colony of spiders is composed of males and females. Usually, the percentage of
females is high (some studies assume that the percentage of males can barely attain
30% of the population size [6]). There are two forms of interactions between social
spiders: direct (i.e., body contact) or indirect (i.e., vibrations transmitted through the
web) [59]. Thus, the shared web is used as a medium of communication. Intensities
of vibrations are used to encode several messages: e.g., size of captured insects,
nature of neighbors. Vibrations’ intensities depend on the weight and the distance
of the source that has initiated them [68]. Social spiders have two main behaviors:
hunting and mating [23, 81].
• The hunting behavior can be summarized as follows. When an insect is trapped in
the web, it tries to escape, generating vibrations through the web. When spiders
sense them, they move toward the source of vibrations.
• The mating behavior can be summarized as follows. The role of males is to fertilize
females. Males are either dominant or non-dominant. Dominant males are attracted
to the closest females. Non-dominant males move toward the center of the male
population to get females. Females exhibit either an attraction or repulsion to
males. This attitude is an answer to the intensity of received vibrations. It depends
on the weight and the distance of males that provoked them.
2.2 Mathematical Model and Optimization Algorithm
We enhance the metaheuristic algorithm proposed in work [19], for global opti-
mization. It mimics the hunting and mating behaviors of social spiders. We name it
enhanced social spider colony optimization (ESSCO). We suppose a search space of
dimension D and an optimization function f to be minimized. We assume three sets
E, M, and F of N E insects, N M male spiders, and N F female spiders, respectively:
⎧
⎨ E = {e1 , e2 , . . . , e N E }
M = {m 1 , m 2 , . . . , m N M }
⎩
F = { f1 , f2 , . . . , f NF }
To describe the behaviors of insects and spiders in ESSCO, we adopt the following
assumptions.
1. An insect moves randomly in the search space according to a Lévy flight distri-
bution [7].
2. An insect can be trapped in the web according to a given probability.
3. When a spider feels a trapped insect, it moves to its location.
4. A dominant male chooses the closest female for mating.
5. A non-dominant male does not mate. It moves in the direction of the center of
males.
6. A female either has attraction or repulsion for a given male.
7. When two spiders mate, a new spider is generated, and the worst spider in the
current population is replaced. The gender of the new spider is the same as the
replaced one.
Algorithms 1, 2, and 3 outline the steps of ESSCO. Their instructions are explained
in Sects. 2.2.1, 2.2.2, and 2.2.3, respectively.
2.2.1 Algorithm 1
• Lines 1 to 3: For each insect ei ∈ E, we generate a random location X ei using

Eq. 1.
• Lines 4 to 6: For each male spider m i ∈ M, we generate a random location X m i
using Eq. 1.
• Lines 7 to 9: For each female spider f i ∈ F, we generate a random location X fi
using Eq. 1.
• Lines 10 to 13: We iterate the algorithm for a given number of iterations. In each
iteration, we run the hunting (Algorithm 2) and mating (Algorithm 3) behaviors.
• Line 14: We get X mbest : i.e., the best solution among male spiders.
• Line 15: We get X best
f : i.e., the best solution among female spiders.
• Line 16: We get X ∗ : i.e., the best solution among male and female spiders.
X y = (α1 × (xmax

1
− xmin
1
) + xmin
1
, . . . , α D × (xmax
D
− xmin
D
) + xmin
D
) (1)
where
y : might be ei , m i , or f i .
α1 , . . . , α D : uniformly distributed random numbers between 0 and 1.
Algorithm 1: The Enhanced Social Spider Colony Optimization.

Input: D is the dimension of the search space.
Input: Δ1 = [xmin1
, xmax
1
], . . . , Δ D = [xminD
, xmax
D
] are the input domains.
Input: f is the objective function to be minimized.
Input: E = {e1 , e2 , . . . , e N E }: i.e., the set of insects (N E = |E|).
Input: M = {m 1 , m 2 , . . . , m N M }: i.e., the set of mal spiders (N M = |M|).
Input: F = { f 1 , f 2 , . . . , f N F }: i.e., the set of female spiders (N F = |F|).
Input: ρ is the probability of an insect to get trapped.
1 foreach ei ∈ E do
e e
2 Compute X ei = (x1 i , . . . , x D i ) using Equation 1;
3 end
4 foreach m i ∈ M do
m i m
5 Compute X m i = (x1 , . . . , x D i ) using Equation 1;
6 end
7 foreach f i ∈ F do
f f
8 Compute X fi = (x1 i , . . . , x D i ) using Equation 1;
9 end
10 for t ← 1 to I termax do
/* Hunting behavior */
11 Run Algorithm 2;
/* Mating behavior */
12 Run Algorithm 3;
13 end
best
14 X m ← argmin { f (X m i )};
i∈{1,...,N M }
15 X best
f ← argmin { f (X fi )};
i∈{1,...,N F }
16 Return X ∗ = argmin{ f (X ibest )} as best solution;
i∈{m, f }
2.2.2 Algorithm 2
• Lines 1 to 14: This loop translates the swarming behavior of insects ei ∈ E in the
search space.
• Line 2: For each insect ei ∈ E, we generate a random number ρi between 0 and 1.
• Lines 3 to 5: If ρi is greater than or equal to ρ, it means that the considered insect
ei is not trapped in the web.
• Line 4: We update location X ei using Eqs. 2 and 3 [51]. If the values of X ei are
out of permitted ranges, they are adjusted.

ei ei u1 uD
X =X + ,..., (2)
(v1 ) 1/β (v D )1/β
⎧ β1
⎨ Γ (1+β) sin( πβ
2 )
u i ∼ N(0, σ 2 ) , σ = β−1 , i ∈ {1, . . . , D} (3)
Γ ( 1+β
2 )×β×2
⎩ 2
2
vi ∼ N(0, σ ) , σ = 1 , i ∈ {1, . . . , D}
• Lines 6 to 13: If ρi is less than ρ, it means that the considered insect ei is trapped
in the web.
• Lines 7 to 9 and 10 to 12: Each male/female spider m i ∈ M/ f i ∈ F moves in the
direction of the trapped insect.
• Lines 8 and 11: We update location X m i /X fi using Eq. 4. If the values of
X m i /X fi are out of permitted ranges, they are adjusted.
X ei −X y
e−δ
2
X y = X y + (X ei − X y ) (4)

f (X ei )
where
Γ : is the gamma function [3].
N(μ, σ ) : is the normal distribution of mean μ and standard deviation σ .
β : is the power law index 1 ≤ β ≤ 2.
y : might be m j or f j .
δ : is a coefficient that defines the attenuation of vibrations.
Algorithm 2: The hunting behavior of ESSCO.
Input: Δ1 = [xmin1
, xmax
1
], . . . , Δ D = [xminD
, xmax
D
1 for i ← 1 to N E do
2 Generate a random number ρi between 0 and 1;
3 if (ρi ≥ ρ) then
4 Update X ei using Equations 2 and 3. Adjust its values;
5 end
6 else
7 for j ← 1 to N M do
8 Update X m j using Equation 4. Adjust its values;
9 end
10 for j ← 1 to N F do
11 Update X f j using Equation 4. Adjust its values;
12 end
13 end
14 end
Algorithm 3: The mating behavior of ESSCO.

Input: Δ1 = [xmin1
, xmax
1
], . . . , Δ D = [xminD
, xmax
D
/* Movement of male spiders */
1 for i ← 1 to N M do
2 Compute the weight w m i of male m i using Equations 5 and 6;
3 end
4 Compute M̃ the median of weights w m i , where i ∈ {1, . . . , N M };
6 if (w m i ≥ M̃) then
7 Update X m i using Equation 7. Adjust its values;
8 end
9 else
10 Update X m i using Equation 8. Adjust its values;
11 end
12 end
/* Movement of female spiders */
13 for i ← 1 to N F do
14 Update X fi using Equation 9. Adjust its values;
15 end
/* Mating of males and females */
17 if (w m i ≥ M̃) then
18 Choose the closest female X f ;
19 Generate a new spider X new using Equations 10, 11 and 12. Adjust its
values;
20 Replace the worst male or female spider;
21 end
22 end
2.2.3 Algorithm 3
• Lines 1 to 3: For each male m i ∈ M, we compute its weight using Eqs. 5 and 6.
F(X m i )
w m i = (5)
NM
j=1 F(X m j )
f (X m i ) − μ − σ
F(X m i ) = (6)
σ
• Line 4: We compute the median M̃ of weights wm i , where i ∈ {1, . . . , N M }.

• Lines 5 to 12: This loop translates the swarming behavior of male spiders m i ∈ M
in the search space.
• Lines 6 to 8: If value wm i is greater than or equal to M̃, it means that male m i is
dominant.
• Line 7: We update location X m i using Eq. 7. If the values of X m i are out of
permitted ranges, they are adjusted.
⎧ m
⎨ X i = X m i + α1 T 1 + α2 T 2
nf m 2
T 1 = w nf e− X −X i (X nf − X m i ) (7)
⎩
T 2 = (α3 − 0.5)
• Lines 9 to 11: If value wm i is less than M̃, it means that male m i is non-dominant.
• Line 10: We update location X m i using Eq. 8. If the values of X m i are out of
permitted ranges, they are adjusted.
m j
m i m i
NM
j=1 (X × w m j ) m i
X =X + α1 −X (8)
NM m j
j=1 w
• Lines 13 to 15: For each female f i ∈ F, we compute its preference (i.e., attraction
or repulsion) to males. In other words, we update location X fi using Eq. 9. If the
values of X fi are out of permitted ranges, they are adjusted.
⎧ fi
⎪
⎪ X = X fi + (−1)b α1 T1 + (−1)b α2 T2 + α3 T3
⎪
⎪ nb m i 2
⎨ T1 = w nb e− X −X (X nb − X m i )
gb − X gb −X m i 2
T2 = w e (X gb − X m i ) (9)
⎪
⎪
⎪
⎪ T3 = (α4 − 0.5)
⎩
b ∼ B(0.5)
• Lines 16 to 22: For each dominant male m i ∈ M, we choose the closest female
and generate a new spider using Eqs. 10, 11, and 12 [25, 26]. Then, we replace
the worst male or female spider in the current population. The gender of the new
spider is the same as the replaced one.
⎡ m f m f m f ⎤
α1 ( f 2 (x1 , x1 ) − f 1 (x1 , x1 )) + f 1 (x1 , x1 )
⎢ ⎥
X new = ⎣ ... ⎦ (10)
m f m f m f
α D ( f 2 (x D , x D ) − f 1 (x D , x D )) + f 1 (x D , x D )
f 1 (x1 , x2 ) = min{x1 , x2 } − α(max{x1 , x2 } − min{x1 , x2 }) (11)

f 2 (x1 , x2 ) = max{x1 , x2 } + α(max{x1 , x2 } − min{x1 , x2 }) (12)
where
μ : is the mean of values f (X m i ), where i ∈ {1, . . . , N M }.
σ : is the standard deviation of values f (X m i ), where i ∈ {1, . . . , N M }.
α1 , . . . , α D : uniformly distributed random numbers between 0 and 1.
X nf : location of the nearest female spider.
B(0.5) : is the Bernoulli distribution that has probability of success 0.5.
X nb : location of the nearest spider that has the best weight.
X gb : location of the best spider in the current population.
3 Experimental Results and Discussion
The benchmark CEC 2020 [66] is used to assess the performance of ESSCO. The
obtained numerical results are compared to eleven state-of-the-art metaheuristic algo-
rithms. All the experiments were performed on a personal computer (16 GB RAM,
CPU of Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz 2.59 GHz, and Windows 10
OS) using the Java programming language. The benchmark CEC 2020 comprises
ten test functions: one is unimodal, three are multimodal, three are hybrid, and three
are composite. All the functions are created employing 14 basic test functions, which
are: High Conditioned Elliptic, Bent Cigar, Discus, Rosenbrock, Ackley, Weierstrass,
Griewank, Rastrigin, Modified Schwefel, Happy Cat, HGBat, Expanded Griewank
plus Rosenbrock, Expanded Schaffer, and Lunacek bi-Rastrigin. More specifica-
tions on these functions can be found in [66]. Tables 2, 3, 4, and 5 summarize the
obtained outcomes of the comparative study for the benchmark CEC 2020. The next
11 state-of-the-art metaheuristic algorithms are studied.
• Improving Cuckoo Search: Incorporating changes for CEC 2017 and CEC 2020
Benchmark Problems (CSsin) [66].
• A Multi-Population Exploration-only Exploitation-only Hybrid on CEC 2020 Sin-
gle Objective Bound Constrained Problems (MP-EEH) [13].
• Ranked Archive Differential Evolution with Selective Pressure for CEC 2020
Numerical Optimization (RASP-SHADE) [74].
• Improved Multi-operator Differential Evolution Algorithm for Solving Uncon-
strained Problems (IMODE) [67].
• DISH XX Solving CEC2020 Single Objective Bound Constrained Numerical Opti-
mization Benchmark (DISH-XX) [80].
• Evaluating the Performance of Adaptive Gaining-Sharing Knowledge Based Algo-
rithm on CEC 2020 Benchmark Problems (AGSK) [56].
• Differential Evolution Algorithm for Single Objective Bound Constrained Opti-
mization: Algorithm j2020 (j2020) [15].
• Eigenvector Crossover in jDE100 Algorithm (jDE100e) [16].
Table 2 Mean, SD, and performance (P) obtained in 30 independent runs by ESSCO, CSsin, MP-EEH, and RASP-SHADE on 20-D CEC 2020 problem suite
Function ESSCO CSsin RASP-SHADE MP-EEH
Mean SD Mean SD P Mean SD P Mean SD P
F1 0.00E+00 0.00E+00 9.33E+09 2.53E+09 + 0.00E+00 0.00E+00 = 0.00E+00 0.00E+00 =
F2 3.00E−01 1.87E−01 9.83E+01 8.33E+01 + 1.70E+02 9.42E+01 + 1.38E−01 4.54E−02 –
F3 1.72E+01 2.55E+00 2.55E+01 2.27E+00 – 2.33E+01 6.13E+00 + 2.05E+01 1.89E−01 –
F4 8.02E−02 6.23E−02 0.00E+00 0.00E+00 – 4.25E−01 1.41E−01 + 4.53E−01 4.18E−02 –
F5 6.57E+00 3.85E+00 1.16E+02 6.34E+01 + 2.36E+02 7.80E+01 + 1.41E+00 1.25E+00 –
F6 1.59E−01 2.83E−02 6.72E−01 8.22E+00 + 4.27E+01 5.23E+01 + 1.71E−01 5.74E−02 +
F7 4.82E−01 1.53E−01 2.62E+00 2.26E+00 + 7.77E+01 6.52E+01 + 7.27E−01 2.40E−01 +
F8 9.07E+01 5.30E+00 9.89E+01 5.59E+00 + 8.00E+01 4.07E+01 + 1.00E+02 0.00E+00 –
An Enhanced Social Spider Colony Optimization for Global …
F9 1.02E+02 3.09E+00 1.03E+02 1.82E+01 + 9.67E+01 1.83E+01 + 3.42E+02 3.51E+01 +

F10 4.00E+02 6.91E−01 3.99E+02 2.44E−01 – 4.01E+02 3.78E+00 + 4.14E+02 1.14E−13 –
+/−/= 7/3/0 9/0/1 3/6/1
785
786
Table 3 Mean, SD, and performance (P) obtained in 30 independent runs by ESSCO, IMODE, DISH-XX, and AGSK on 20-D CEC 2020 problem suite
Function ESSCO IMODE DISH-XX AGSK
F1 0.00E+00 0.00E+00 0.00E+00 0.00E+00 = 0.00E+00 0.00E+00 = 0.00E+00 0.00E+00 =
F2 3.00E−01 1.87E−01 5.13E−01 7.13E−01 + 8.67E+01 1.11E+02 + 2.68E+03 1.60E+02 +
F3 1.72E+01 2.55E+00 2.05E+01 1.26E−01 – 2.13E+01 3.74E+00 + 7.37E+01 5.25E+00 +
F4 8.02E−02 6.23E−02 0.00E+00 0.00E+00 – 2.47E−04 1.35E−03 – 5.37E+00 4.25E−01 +
F5 6.57E+00 3.85E+00 1.09E+01 4.33E+00 + 5.63E+01 6.63E+01 + 2.44E+02 3.97E+01 +
F6 1.59E−01 2.83E−02 3.02E−01 8.17E−02 + 1.50E+01 3.57E+01 + 3.35E+00 2.17E+00 +
F7 4.82E−01 1.53E−01 5.24E−01 1.64E−01 + 5.09E+00 6.42E+00 + 5.86E+01 1.09E+01 +
F8 9.07E+01 5.30E+00 8.40E+01 1.89E+01 + 1.00E+02 1.39E−13 – 1.00E+02 0.00E+00 –
F9 1.02E+02 3.09E+00 9.67E+01 1.83E+01 + 4.05E+02 2.50E+00 – 4.39E+02 2.95E+01 +
F10 4.00E+02 6.91E−01 4.00E+02 6.18E−01 – 4.14E+02 2.54E−02 – 4.14E+02 8.87E−03 –
+/−/= 6/3/1 5/4/1 7/2/1
F. Zitouni et al.
Table 4 Mean, SD, and performance (P) obtained in 30 independent runs by ESSCO, j2020, jDE100e, and OLSHADE on 20-D CEC 2020 problem suite
Function ESSCO j2020 jDE100e OLSHADE
F1 0.00E+00 0.00E+00 0.00E+00 0.00E+00 = 0.00E+00 0.00E+00 = 0.00E+00 0.00E+00 =
F2 3.00E−01 1.87E−01 2.60E−02 2.47E−02 – 1.48E+00 1.50E+00 + 1.15E+02 7.62E+01 +
F3 1.72E+01 2.55E+00 1.44E+01 9.29E+00 + 2.10E+01 4.09E−01 – 2.52E+01 7.63E+00 +
F4 8.02E−02 6.23E−02 1.80E−01 7.84E−02 + 3.47E−01 8.04E−02 + 1.01E+00 1.22E+00 +
F5 6.57E+00 3.85E+00 7.78E+01 5.75E+01 + 2.37E+00 8.49E−01 – 1.78E+01 4.14E+01 +
F6 1.59E−01 2.83E−02 1.92E−01 1.01E−01 + 1.15E−01 3.31E−02 + 5.17E−01 0.00E+00 –
F7 4.82E−01 1.53E−01 1.98E+00 4.02E+00 + 2.14E−01 1.14E−01 – 8.42E−01 1.61E−01 +
F8 9.07E+01 5.30E+00 9.27E+01 2.21E+01 + 1.00E+02 0.00E+00 – 1.00E+02 7.01E−01 –
An Enhanced Social Spider Colony Optimization for Global …
F9 1.02E+02 3.09E+00 3.39E+02 1.28E+02 + 4.05E+02 1.74E+00 – 1.10E+02 3.06E+01 +

F10 4.00E+02 6.91E−01 3.99E+02 4.02E−02 − 4.13E+02 2.78E+00 + 4.10E+02 6.15E+00 +
+/−/= 7/2/1 4/5/1 7/2/1
787
788
Table 5 Mean, SD, and performance (P) obtained in 30 independent runs by ESSCO, mpm-LSHADE, and SOMA-CL on 20-D CEC 2020 problem suite
Function ESSCO mpm-LSHADE SOMA-CL
Mean SD Mean SD P Mean SD P
F1 0.00E+00 0.00E+00 0.00E+00 0.00E+00 = 0.00E+00 0.00E+00 =
F2 3.00E−01 1.87E−01 3.97E−02 2.12E−02 – 7.36E+00 2.15E+01 +
F3 1.72E+01 2.55E+00 2.04E+01 4.67E−13 – 2.14E+01 8.06E−01 –
F4 8.02E−02 6.23E−02 4.97E−01 4.23E−02 – 1.05E+00 1.83E−01 +
F5 6.57E+00 3.85E+00 1.38E+00 1.45E+00 – 1.45E+02 8.01E+01 +
F6 1.59E−01 2.83E−02 2.05E−01 4.71E−02 + 2.71E−01 8.35E−02 +
F7 4.82E−01 1.53E−01 5.10E−01 1.19E−01 – 9.22E+00 8.93E+00 +
F8 9.07E+01 5.30E+00 1.00E+02 7.47E−13 – 9.91E+01 3.75E+00 –
F9 1.02E+02 3.09E+00 4.01E+02 6.68E−01 – 3.99E+02 4.54E+01 +
F10 4.00E+02 6.91E−01 4.14E+02 2.74E−04 – 4.71E+02 3.23E+01 +
+/−/= 1/8/1 7/2/1
F. Zitouni et al.
• Large Initial Population and Neighborhood Search incoporated in LSHADE to

solve CEC2020 Benchmark Problems (OLSHADE) [10].
• Multi-population Modified L-SHADE for Single Objective Bound Constrained
Optimization (mpm-LSHADE) [34].
• SOMA-CL for Competition on Single Objective Bound Constrained Numerical
Optimization Benchmark (SOMA-CL) [36].
The parameters of ESSCO are |N E | = 50, |N M | = 24, |N F | = 56, and ρ = 0.5.

The search space dimension is set to 20. The maximum function evaluation is fixed
to 1,000,000 for 30 different runs. The range of search space is [-100, 100]d . The
same settings and evaluation criteria in [66] are adapted in this work. The default
values for different metaheuristic algorithms’ parameters are used (as reported in the
corresponding published papers). The mean, standard deviation (SD), and obtained
performance (P) of each metaheuristic are given. Signs ’+’, ’−’, and ’=’ designate,
respectively, that the performance of ESSCO is better/worst/equal than/to the con-
sidered algorithm. The statistical results of different metaheuristics are shown in
Tables 2, 3, 4, and 5.
Based on the result given in the last row of Tables 2, 3, 4, and 5, the performance of
ESSCO is better than CSsin, MP-EEH, RASP-SHADE, IMODE, DISH-XX, AGSK,
j2020, jDE100e, OLSHADE, mpm-LSHADE, and SOMA-CL on 7, 9, 3, 6, 5, 7, 7,
4, 7, 1, and 7 functions, respectively. However, the performance of CSsin, RASP-
SHADE, IMODE, DISH-XX, AGSK, j2020, jDE100e, OLSHADE, mpm-LSHADE,
and SOMA-CL is better than ESSCO on 3, 6, 3, 4, 2, 2, 5, 2, 8, and 2 functions,
respectively.
4 Conclusion
This work improves the optimization algorithm proposed in [19]. It is based on

the hunting and mating behaviors of social spiders. Enhanced social spider colony
optimization is evaluated on ten test functions of the benchmark CEC 2020. The
numerical results are compared to the 11 state-of-the-art metaheuristic algorithms
and show the competitiveness of ESSCO in finding efficient solutions. In future work,
we plan to combine ESSCO with quantum genetic algorithms.
Acknowledgements This research work is supported by UAEU Grant: 31T102-UPAR-1-2017.
References
1. Abbass, H.A.: Mbo: Marriage in honey bees optimization-a haplometrosis polygynous swarm-
ing approach. In: Proceedings of the 2001 Congress on Evolutionary Computation (IEEE Cat.
No. 01TH8546). vol. 1, pp. 207–214. IEEE (2001)
2. Alatas, B.: Acroa: artificial chemical reaction optimization algorithm for global optimization.
Expert Syst. Appl. 38(10), 13170–13180 (2011)
3. Artin, E.: The Gamma Function. Courier Dover Publications (2015)
4. Ashby, W.R.: Principles of the self-organizing system. In: Facets of Systems Science, pp.
521–536. Springer (1991)
5. Askarzadeh, A., Rezazadeh, A.: A new heuristic optimization algorithm for modeling of proton
exchange membrane fuel cell: bird mating optimizer. Int. J. Energy Res. 37(10), 1196–1204
(2013)
6. Aviles, L.: Sex-ratio bias and possible group selection in the social spider Anelosimus eximius.
Am. Nat. 128(1), 1–12 (1986)
7. Barthelemy, P., Bertolotti, J., Wiersma, D.S.: A lévy flight for light. Nature 453(7194), 495–498
(2008)
8. Basturk, B.: An artificial bee colony (abc) algorithm for numeric function optimization. In:
IEEE Swarm Intelligence Symposium, Indianapolis, IN, USA, 2006 (2006)
9. Bergmann, H.W.: Optimization: Methods and Applications, Possibilities and Limitations: Pro-
ceedings of an International Seminar Organized by Deutsche Forschungsanstalt Für Luft-und
Raumfahrt (DLR), Bonn, June 1989, vol. 47. Springer Science & Business Media (2012)
10. Biswas, P.P., Suganthan, P.N.: Large initial population and neighborhood search incorporated
in lshade to solve cec2020 benchmark problems. In: 2020 IEEE Congress on Evolutionary
Computation (CEC), pp. 1–7. IEEE (2020)
11. Blum, C., Roli, A.: Metaheuristics in combinatorial optimization: overview and conceptual
comparison. ACM Comput. Surv. (CSUR) 35(3), 268–308 (2003)
12. Blum, C., Roli, A.: Hybrid metaheuristics: an introduction. In: Hybrid Metaheuristics, pp.
1–30. Springer, Berlin (2008)
13. Bolufé-Röhler, A., Chen, S.: A multi-population exploration-only exploitation-only hybrid on
cec-2020 single objective bound constrained problems. In: 2020 IEEE Congress on Evolution-
ary Computation (CEC), pp. 1–8. IEEE (2020)
14. Borenstein, Y., Moraglio, A.: Theory and Principled Methods for the Design of Metaheuristics.
Springer, Berlin (2014)
15. Brest, J., Maučec, M.S., Bošković, B.: Differential evolution algorithm for single objective
bound-constrained optimization: algorithm j2020. In: 2020 IEEE Congress on Evolutionary
16. Bujok, P., Kolenovsky, P., Janisch, V.: Eigenvector crossover in jde100 algorithm. In: 2020
IEEE Congress on Evolutionary Computation (CEC), pp. 1–6. IEEE (2020)
17. Černỳ, V.: Thermodynamical approach to the traveling salesman problem: an efficient simula-
tion algorithm. J. Opt. Theor. Appl. 45(1), 41–51 (1985)
18. Ciarleglio, M.I.: Modular abstract self-learning tabu search (masts): Metaheuristic search the-
ory and practice (2008)
19. Cuevas, E., Cienfuegos, M., ZaldíVar, D., Pérez-Cisneros, M.: A swarm optimization algorithm
inspired in the behavior of the social-spider. Expert Syst. Appl. 40(16), 6374–6384 (2013)
20. Dorigo, M., Birattari, M.: Ant colony optimization. In: Sammut, C., Webb, G. I. (Eds.) Ency-
clopedia of Machine Learning and Data Mining, pp. 56–59. Springer US, Boston, MA (2017),
ISBN:978-1-4899-7687-1. https://doi.org/10.1007/978-1-4899-7687-1_22.
21. Dorigo, M., Di Caro, G.: Ant colony optimization: a new meta-heuristic. In: Proceedings of the
1999 congress on evolutionary computation-CEC99 (Cat. No. 99TH8406). vol. 2, pp. 1470–
1477. IEEE (1999)
22. Du, H., Wu, X., Zhuang, J.: Small-world optimization algorithm for function optimization. In:
International Conference on Natural Computation, pp. 264–273. Springer, Berlin (2006)
23. Elias, D.O., Andrade, M.C., Kasumovic, M.M.: Dynamic population structure and the evolution
of spider mating systems. In: Advances in Insect Physiology, vol. 41, pp. 65–114. Elsevier
(2011)
24. Erol, O.K., Eksin, I.: A new optimization method: big bang-big crunch. Adv. Eng. Softw. 37(2),
106–111 (2006)
25. Eshelman, L.J.: Crossover operator biases: exploiting the population distribution. In: Proceed-
ings of International Conference on Genetic Algorithms, 1997 (1997)
26. Eshelman, L.J., Schaffer, J.D.: Real-coded genetic algorithms and interval-schemata. In: Foun-
dations of Genetic Algorithms, vol. 2, pp. 187–202. Elsevier (1993)
27. Fogel, D.B.: Artificial intelligence through simulated evolution. Wiley-IEEE Press (1998)
28. Formato, R.: Central force optimization: a new metaheuristic with applications in applied
electromagnetics. prog electromagn res 77: 425–491 (2007)
29. Gandomi, A.H., Alavi, A.H.: Krill herd: a new bio-inspired optimization algorithm. Commun.
Nonlinear Sci. Numer. Simul. 17(12), 4831–4845 (2012)
30. Glover, F.W., Kochenberger, G.A.: Handbook of metaheuristics, vol. 57. Springer Science &
Business Media (2006)
31. Hatamlou, A.: Black hole: a new heuristic optimization approach for data clustering. Inf. Sci.
222, 175–184 (2013)
32. Helbig, M., Engelbrecht, A.P.: Population-based metaheuristics for continuous boundary-
constrained dynamic multi-objective optimisation problems. Swarm Evol. Comput. 14, 31–47
(2014)
33. Holland, J.H.: Genetic algorithms. Sci. Am. 267(1), 66–73 (1992)
34. Jou, Y.C., Wang, S.Y., Yeh, J.F., Chiang, T.C.: Multi-population modified l-shade for single
objective bound constrained optimization. In: 2020 IEEE Congress on Evolutionary Compu-
tation (CEC), pp. 1–8. IEEE (2020)
35. Joyce, T., Herrmann, J.M.: A review of no free lunch theorems, and their implications for
metaheuristic optimisation. In: Nature-inspired algorithms and applied optimization, pp. 27–
51. Springer, Berlin (2018)
36. Kadavy, T., Pluhacek, M., Viktorin, A., Senkerik, R.: Soma-cl for competition on single objec-
tive bound constrained numerical optimization benchmark: a competition entry on single objec-
tive bound constrained numerical optimization at the genetic and evolutionary computation
conference (gecco) 2020. In: Proceedings of the 2020 Genetic and Evolutionary Computation
Conference Companion, pp. 9–10 (2020)
37. Karaboga, D., Basturk, B.: A powerful and efficient algorithm for numerical function opti-
mization: artificial bee colony (ABC) algorithm. J. Glob. Optim. 39(3), 459–471 (2007)
38. Kaur, S., Awasthi, L.K., Sangal, A., Dhiman, G.: Tunicate swarm algorithm: a new bio-inspired
based metaheuristic paradigm for global optimization. Eng. Appl. Artif. Intell. 90, 103541
(2020)
39. Kaveh, A., Farhoudi, N.: A new optimization method: Dolphin echolocation. Adv. Eng. Softw.
59, 53–70 (2013)
40. Kaveh, A., Khayatazad, M.: A new meta-heuristic method: ray optimization. Comput. Struct.
112, 283–294 (2012)
41. Kaveh, A., Talatahari, S.: A novel heuristic optimization method: charged system search. Acta
Mechanica 213(3–4), 267–289 (2010)
42. Keller, E.F.: Organisms, machines, and thunderstorms: a history of self-organization, part two.
complexity, emergence, and stable attractors. Hist. Stud. Natural Sci. 39(1), 1–31 (2009)
43. Kennedy, J., et al.: Encyclopedia of machine learning. Particle Swarm Optim. 760–766 (2010)
44. Kirkpatrick, S., Gelatt, C.D., Vecchi, M.P.: Optimization by simulated annealing. Science
220(4598), 671–680 (1983)
45. Koza, J.R., Koza, J.R.: Genetic programming: on the programming of computers by means of
natural selection, vol. 1. MIT press (1992)
46. Koziel, S., Yang, X.S.: Computational optimization, methods and algorithms, vol. 356.
Springer, Berlin (2011)
47. Kumar, A., Misra, R.K., Singh, D., Mishra, S., Das, S.: The spherical search algorithm for
bound-constrained global optimization problems. Appl. Soft Comput. 85, 105734 (2019)
48. Li, X.: A new intelligent optimization-artificial fish swarm algorithm. Doctor thesis, Zhejiang
University of Zhejiang, China (2003)
49. Lu, X., Zhou, Y.: A novel global convergence algorithm: bee collecting pollen algorithm. In:
International Conference on Intelligent Computing, pp. 518–525. Springer, Berlin (2008)
50. Lubin, Y., Bilde, T.: The evolution of sociality in spiders. Adv. Study Behavior 37, 83–145
(2007)
51. Mantegna, R.N.: Fast, accurate algorithm for numerical simulation of levy stable stochastic
processes. Phys. Rev. E 49(5), 4677 (1994)
52. Mirjalili, S.: Dragonfly algorithm: a new meta-heuristic optimization technique for solving
single-objective, discrete, and multi-objective problems. Neural Comput. Appl. 27(4), 1053–
1073 (2016)
53. Mirjalili, S., Lewis, A.: The whale optimization algorithm. Adv. Eng. Softw. 95, 51–67 (2016)
54. Mirjalili, S., Mirjalili, S.M., Lewis, A.: Grey wolf optimizer. Adv. Eng. Softw. 69, 46–61 (2014)
55. Moghaddam, F.F., Moghaddam, R.F., Cheriet, M.: Curved space optimization: a random search
based on general relativity theory. arXiv preprint arXiv:1208.2214 (2012)
56. Mohamed, A.W., Hadi, A.A., Mohamed, A.K., Awad, N.H.: Evaluating the performance of
adaptive gainingsharing knowledge based algorithm on cec 2020 benchmark problems. In:
2020 IEEE Congress on Evolutionary Computation (CEC), pp. 1–8. IEEE (2020)
57. Molina, J., Rudnick, H.: Transmission expansion plan: Ordinal and metaheuristic multiobjec-
tive optimization. In: 2011 IEEE Trondheim PowerTech, pp. 1–6. IEEE (2011)
58. Mucherino, A., Seref, O.: Monkey search: a novel metaheuristic search for global optimization.
In: AIP Conference Proceedings, vol. 953, pp. 162–173. AIP (2007)
59. Murata, K., Tanaka, K.: Spatial interaction between spiders and prey insects: horizontal and
vertical distribution in a paddy field. Acta arachnologica 53(2), 75–86 (2004)
60. Oftadeh, R., Mahjoob, M., Shariatpanahi, M.: A novel meta-heuristic optimization algorithm
inspired by group hunting of animals: hunting search. Comput. Math. Appl. 60(7), 2087–2098
(2010)
61. Pan, W.T.: A new fruit fly optimization algorithm: taking the financial distress model as an
example. Knowl.-Based Syst. 26, 69–74 (2012)
62. Puchinger, J., Raidl, G.R.: Combining metaheuristics and exact algorithms in combinatorial
optimization: a survey and classification. In: International work-conference on the interplay
between natural and artificial computation, pp. 41–53. Springer, Berlin (2005)
63. Rajeev, S., Krishnamoorthy, C.: Discrete optimization of structures using genetic algorithms.
J. Struct. Eng. 118(5), 1233–1250 (1992)
64. Rashedi, E., Nezamabadi-Pour, H., Saryazdi, S.: GSA: a gravitational search algorithm. Inf.
Sci. 179(13), 2232–2248 (2009)
65. Roth, M., Wicker, S.: Termite: A swarm intelligent routing algorithm for mobile wireless ad-hoc
networks. In: Stigmergic Optimization, pp. 155–184. Springer (2006)
66. Salgotra, R., Singh, U., Saha, S., Gandomi, A.H.: Improving cuckoo search: incorporating
changes for CEC 2017 and CEC 2020 benchmark problems. In: 2020 IEEE Congress on
Evolutionary Computation (CEC), pp. 1–7. IEEE (2020)
67. Sallam, K.M., Elsayed, S.M., Chakrabortty, R.K., Ryan, M.J.: Improved multi-operator dif-
ferential evolution algorithm for solving unconstrained problems. In: 2020 IEEE Congress on
Evolutionary Computation (CEC), pp. 1–8. IEEE (2020)
68. Salomon, M., Sponarski, C., Larocque, A., Avilés, L.: Social organization of the colonial
spider leucauge sp. in the neotropics: vertical stratification within colonies. J. Arachnology
38(3), 446–451 (2010)
69. Shah-Hosseini, H.: Principal components analysis by the galaxy-based search algorithm: a
novel metaheuristic for continuous optimisation. Int. J. Comput. Sci. Eng. 6(1–2), 132–140
(2011)
70. Shi, J., Zhang, Q.: A new cooperative framework for parallel trajectory-based metaheuristics.
App. Soft Comput. 65, 374–386 (2018)
71. Shi, Y., Eberhart, R.C.: Empirical study of particle swarm optimization. In: Proceedings of
the 1999 Congress on Evolutionary Computation-CEC99 (Cat. No. 99TH8406), vol. 3, pp.
1945–1950. IEEE (1999)
72. Shiqin, Y., Jianjun, J., Guangxing, Y.: A dolphin partner optimization. In: 2009 WRI Global
Congress on Intelligent Systems, vol. 1, pp. 124–128. IEEE (2009)
73. Simon, D.: Biogeography-based optimization. IEEE Trans. Evol. Comput. 12(6), 702–713
(2008)
74. Stanovov, V., Akhmedova, S., Semenkin, E.: Ranked archive differential evolution with selec-
tive pressure for CEC 2020 numerical optimization. In: 2020 IEEE Congress on Evolutionary
75. Storn, R., Price, K.: Differential evolution-a simple and efficient heuristic for global optimiza-
tion over continuous spaces. J. Glob. Optim. 11(4), 341–359 (1997)
76. Talbi, E.G., Jourdan, L., Garcia-Nieto, J., Alba, E.: Comparison of population based metaheuris-
tics for feature selection: Application to microarray data classification. In: 2008 IEEE/ACS
International Conference on Computer Systems and Applications, pp. 45–52. IEEE (2008)
77. Talbi, H., Draa, A.: A new real-coded quantum-inspired evolutionary algorithm for continuous
optimization. Appl. Soft Comput. 61, 765–791 (2017)
78. Tang, K.S., Man, K.F., Kwong, S., He, Q.: Genetic algorithms and their applications. IEEE
Signal Process. Mag. 13(6), 22–37 (1996)
79. Van Laarhoven, P.J., Aarts, E.H.: Simulated annealing. In: Simulated Annealing: Theory and
Applications, pp. 7–15. Springer, Berlin (1987)
80. Viktorin, A., Senkerik, R., Pluhacek, M., Kadavy, T., Zamuda, A.: Dish-xx solving cec2020
single objective bound constrained numerical optimization benchmark. In: 2020 IEEE Congress
on Evolutionary Computation (CEC), pp. 1–8. IEEE (2020)
81. Vollrath, F., Rohde-Arndt, D.: Prey capture and feeding in the social spider Anelosimus eximius.
Zeitschrift für Tierpsychologie 61(4), 334–340 (1983)
82. Webster, B., Philip, J., Bernhard, A.: Local search optimization algorithm based on natural
principles of gravitation, ike’03, las vegas, Nevada, USA (2003, June)
83. Yang, C., Tu, X., Chen, J.: Algorithm of marriage in honey bees optimization based on the
wolf pack search. In: The 2007 International Conference on Intelligent Pervasive Computing
(IPC 2007), pp. 462–467. IEEE (2007)
84. Yang, X.S.: Engineering Optimization: An Introduction with Metaheuristic Applications.
Wiley, Hoboken (2010)
85. Yang, X.S.: Firefly algorithm, stochastic test functions and design optimisation. arXiv preprint
arXiv:1003.1409 (2010)
86. Yang, X.S.: Nature-Inspired Metaheuristic Algorithms. Luniver Press (2010)
87. Yang, X.S.: A new metaheuristic bat-inspired algorithm. In: Nature Inspired Cooperative Strate-
gies for Optimization (NICSO 2010), pp. 65–74. Springer, Berlin (2010)
88. Yang, X.S.: Swarm-based metaheuristic algorithms and no-free-lunch theorems. Theor. New
Appl. Swarm Intell. 9, 1–16 (2012)
89. Yang, X.S.: Optimization and metaheuristic algorithms in engineering. In Metaheuristics in
Water, Geotechnical and Transport Engineering, pp. 1–23 (2013)
90. Yang, X.S., Deb, S.: Cuckoo search via lévy flights. In: 2009 World Congress on Nature &
Biologically Inspired Computing (NaBIC), pp. 210–214. IEEE (2009)
91. Yao, X., Liu, Y., Lin, G.: Evolutionary programming made faster. IEEE Trans. Evol. Comput.
3(2), 82–102 (1999)
92. Zitouni, F., Harous, S., Maamri, R.: The solar system algorithm: a novel metaheuristic method
for global optimization. IEEE Access (2020)
Data Processing on Distributed Systems
Storage Challenges
Mohamed Eddoujaji, Hassan Samadi, and Mohamed Bohorma
Abstract Hadoop is an open-source software framework for storing data and

launching applications on standard machine clusters. This solution provides massive
storage space for all types of data, tremendous processing power, and the ability
to support virtually any amount of work. Based on Java, this framework is part of
the Apache project, sponsored by Apache Software Foundation [Hadoop official site.
http://hadoop.apache.org/]. Thanks to the MapReduce framework, it can handle huge
amounts of data. Rather than having to move the data to a network for processing,
MapReduce allows you to directly move the processing software to the data. No one
will be able to neglect the results realized by the Hadoop technique nor to question
its power and its performance to manage important volumetrics of the data of any
type: structured, semi-structured, or unstructured [https://www.lebigdata.fr/hadoop
]. But it does not shine by its processing speed when datasets are smaller and even
smaller. This problem, the “small files” problem has been well-defined by the Hadoop
community, and also by the researchers, the majority of the proposed solutions and
concepts only deal with the performance and the pressure exerted on the NameNode
memory; some new approaches have been proposed as the casi-random, grouping of
small heterogeneous files and in different formats, this solves the problem of memory
because the number of metadata has been considerably reduced; the real challenge
companies face is the performance of the Hadoop cluster when processing a very
large number of small files.
1 Introduction and Background
Hadoop was designed to efficiently manage effectively large files, especially when
traditional systems are facing limitations to analyze this new data dimension of data
caused by its exponential growth. However, Hadoop is not deployed to handle only
large files! The heterogeneity and diversity of information from multiple sources
M. Eddoujaji (B) · H. Samadi · M. Bohorma (B)

Abdelmalek Essaâdi University, National School of Applied Sciences of Tangier, Tangier,
Morocco
https://doi.org/10.1007/978-981-16-3637-0_56
796 M. Eddoujaji et al.
Fig. 1 Hadoop 2.0 main

components Map Reduce Other
Data Processing Data Processing
YARN
Cluster Ressource Management
HDFS
Redudant, Reliable Storage
(intelligent devices, IOTs, Internet users, log files, and security events) have become
the normal flow of Hadoop architectures.
In today’s world, most domains permanently and constantly generate a very large
amount of information in the form of small files. Multiple domains store and analyze
millions and millions of small files, such as analysis for multimedia data mining [3],
astronomy [4], meteorology [5], signal recognition [6], climatology [7, 8], energy,
and e-learning. [9], without forgetting the astronomical information processed by
social networks; Facebook stores more than 350 million images every day [10].
In biology, the human genome generates up to 30 million files which do not exceed
on average 190 KB [11] (Fig. 1).
1.1 HDFS—Reliable Storage
Due to its massive capacity and reliability, HDFS is a storage system that is very
suitable for Big Data. In combination with YARN, this system increases the data
management capabilities of the HDFS Hadoop cluster and thus enables the efficient
processing of Big Data. Among its main features, there is the possibility of storing
terabytes, even petabytes of data [12].
The system is capable of handling thousands of nodes without the intervention
of an operator. It allows simultaneous benefits of parallel computing and distributed
computing [33]. After a modification, it allows to easily restore the previous version
of a data.
HDFS can be launched on commodity hardware, which makes it very tolerant of
errors. Each piece of data is stored in several places and can be retrieved under any
circumstances. In the same way, this replication makes it possible to fight against the
potential corruption of the data (Fig. 2).
2 Paper Sections
The following parts in this paper will be organized as follows:

Chapter 3 describes the related work and some of the previous results.
Chapter 3 describes our proposed approach.
Data Processing on Distributed Systems Storage Challenges 797
Fig. 2 Storage and replication of blocks in HDFS
Chapter 4 presents our experimental works and results.

Finally, and to conclude this research,
Chapter 5 for conclusion and expectation.
3 Related Work
3.1 Reminder
In most data-intensive applications, we mainly work on two types of file systems,

the distributed file systems and the parallel file systems [33].
The distributed file system type is broadly used in online services, and one of the
unmistakable examples involves Amazon Simple Storage Service known as (Amazon
S3), Google’s file system, and of course, the open-source Hadoop file system [13].
Parallel file system type is mainly built targeting high compute processing (HPC)
applications that run-on large-scale clusters and require highly extendable and scal-
able storage solutions that can perform simultaneous I/O operations [2]. Some illus-
trations of parallel file systems are covered by Sun’s LustreFS [14], IBM’s Spectrum
Scale based on GPFS [15], and open-source Parallel Virtual FS (PVFS) [14].
3.2 Motivation
HDFS has been designed mainly to be able to manage sizable files, but not small
ones. That is the reason why it may face some issues when asked to manage a big
amount of small files. For example, if we have around 600,000 small files which
have sizes that vary between 1 and 10 KB were stored into HDFS, the process led to
observe the following phenomena [16]:
Unacceptable execution time. It took more than 7 h to store these small-sized files
into HDFS, in the main time in a local file system, for example, “ext3,” the storing
time was about 600 s.
The usage rate of the memory was high. In the process of storing operations, AND
in an idle system, the occupation was at 63%.
In HDFS, each file has its own metadata and is stored with multiple replicas,
three as a default replication value. Managing metadata in HDFS consumes much
time due to the need of the cooperation between at least three nodes. For small file
I/O, so much time is spent in managing metadata while just a little is spent in data
transferring. The big amount of small files raises the overhead of metadata operations
in HDFS. This is why HDFS needs so much time to store these files. The NameNode
manages and stores the metadata, the DataNodes preserve fragmented information,
and all these data are loaded on the physical memory for exploitation.
As a result, the more number of small files the more memory usage rate increases.
We have to adopt an optimized approach to build a middleware on HDFS so as
to satisfy the application demands of small files IO performance, and we should
consider the file access patterns of special applications [17, 18].
4 Existing Solutions
4.1 HAR Files
Hadoop itself provides a tool Archives Hadoop (HAR) [3, 19], and it can conduct
a small file in file processing. It adds an index file to an archive and also provides
convenience for the MapReduce operation, but there are still shortcomings. Due to
the small file in the access to HAR based on the two index files to find which leads to
client relying on the HAR file to access small files, and the merged process takes a
long time, which is less efficient than reading small files from the HDFS. You cannot
modify The HAR file created by this method, instead, you have to recreate the HAR
file if you want to make changes (add or delete content on it) [20] (Fig. 3).
Fig. 3 HAR file layout

5 SequenceFiles
In the case of SequenceFile, we have no solution to list all the keys of a file, except
reading through the entire file. (MapFiles files, which look like SequenceFiles with
sorted keys, keep a partial index, so they cannot list all their keys either—see the
diagram.) [21, 22].
SequenceFile is a binary format program compatible with Hadoop API [3, 23],
whose data structure is composed of a series of key/binary values, the small files
can be stored in a single unit. The principle is simple, and it consists of merging and
grouping a large number of small files into one large file. It provides good support
and high scalability and performance for local MapReduce data management [24]
(Fig. 4).
Unlike the HAR files, the SequenceFile supports compression, and they are more
suitable for MapReduce tasks as they are splittable [3, 23], so mappers can operate
on chunks independently. However, converting into a SequenceFile can be a time-
consuming task, and it has a poor performance during random read access.
To improve the metadata management, Mohd Abdul Ahada, Ranjit Biswasa [10]
merged small files into a single larger file, using the HAR through a MapReduce
task. The small files are referenced with an added index layer (Master index, Index)
delivered with the archive, to retain the separation, and to keep the original structure
of files.
C. Vorapongkitipunet al. [3, 11] proposed an improved approach of the HAR
technique, by introducing a single index instead of the two-level indexes. Their
new indexing mechanism aims to improve the metadata management as well
as the performance during file access without changing the implemented HDFS
architecture.
Rattanaopas and Kaewkeeree. [25], and Mir and Ahmed [26] proposed to combine
files using the SequenceFile method. Their approach reduces memory consumption
on the NameNode but it did not show how much the read and write performances
are impacted.
Fig. 4 SequenceFile layout

Zheng et al. [32] proposed merging related small files according to WebGIS
application, which improved the storage efficiency and HDFS metadata management;
however, the results are limited by the scene.
Mir and Ahmed [7] proposed a modification of the existing HAR. They used a
hashing concept based on the sha256 as a key. This can improve the reliability and
the scalability of the metadata management, also the reading access time is greatly
reduced, but it takes more time to create the NHAR archives compared to the HAR
mechanism.
Niazi and Ronström [27] proposed a scheme for combining small file, merging,
prefetching the related small files which improve the storage and access efficiency
of small files but do not give an appropriate solution for independent small files.
6 Proposed Work
6.1 Little Reminder
Hadoop is designed for powerful management of large files, unlike the management
of large volumes of small files where the technique always requires improvements,
in fact, Hadoop passes each small file to a map () function, which creates a large
number of mappers, and therefore, the solution is not very effective for this kind of
traffic. For example, the 100,000 files smaller than 2Mo will need 10,000 mappers,
which is very inefficient, which could be a problem throughout the system.
The main objective of these improvements and new approaches is on the one
hand to solve this problem, and on the other hand, it is to accelerate the execution
of operations (reading and writing) Hadoop. This solution is to combine small files
into larger files, and this will reduce the number of map () functions that are executed
and thus significantly improve performance [28, 32].
6.2 The Proposed Approach for Small File Management
It is known and even approved by the Hadoop communities that the performance of
Hadoop platforms is greatly impacted by the management of small files. Les solutions
précédentes ou les recherches faites sur ce sujet, surpassent ce souci en empaquetant
de petits fichiers hétérogènes dans de dossiers plus grands. Cette manière de pack-
aging et le facteur principale des améliorations des temps des écritures et des lectures
de la technique Mapreduce, none of the adopted approaches take in consideration
how to organize those small files during the merging phase. The central idea of our
approach is to start the organization and management of files when starting flows
containing small files, combined if relevant, that will be combined with different other
client streams into blocks based on their relevance; moreover, they are arranged in an
efficient way as the files with most fetch probability appear always on top. This effi-
ciency and performance capabilities can be reached utilizing “Hadoop File Server
Analyzer” as shown in “Fig. 6.”
The foundational idea of our approach is to consolidate files from different clients
that contain series of small files, combined by relevance, and unify them through
a merge process, to be stored optimally before the current SFA connection closes.
This realization was carried out in our previous work, see “Fig. 6” as the main task
of the “Small File Analyzer server.” In the present research, we have enhanced the
suggested SFA method, to be able to handle new modules and make it possible for
us to acknowledge other parameters within the merge process [3].
We have introduced a sorting process which can work as an unrelated module in
the SFA server. The compressor module is an extra add-on that has enabled us to apply
a new compression layer on top of the merged files that are no longer in use, or barely
solicited, so that we can have a huge storage capacity advantage. Additionally, we
have used a “prefetch and caching” technique to improve the overall performance
when reading identical small files [3].
6.3 Operational Mode Explanation
The operational mode of our idea can be divided into three major phases:
File combining, File mapping, Prefetching and caching
6.4 File Combining
Conventionally, NameNode stores metadata of files and blocks.

File metadata consists of the following entries:
file name, file permissions, owner and group, access time, modification time, file length,
replication.
Block metadata consists of the following information:

information on all the blocks containing the data of the file, location of these blocks.
Thus, in conventional approaches, storing a large number of small files consumes

a large amount of memory of the “NameNode” machine.
In order to optimize the memory consumption needed to store the NameNode
metadata, several small files will be merged and combined into a single file that we
can call a “combined file.”
As a result, NameNode will store the metadata of a single large file instead of
keeping the metadata of several small files.
In addition, this technique will ensure reading and grouping all similar files
together.
Below is a small description of how MapReduce works for the “combined file”
which will be called the MapReduce combiner:
In the Map phase, the mappers generate key/value pairs.
During the shuffle/sort phase, these pairs are distributed and ordered on one or more nodes
depending on the value of the key.
During the Reduce phase, one or more reducers aggregate the key/value pairs according to
the value of the key.
6.5 File Mapping
The second phase consists of analyzing these files based on their size (how these small
files are used), then put them in appropriate groups in MapFile. As we mentioned
above, the MapFile is another file packaging format developed by Hadoop based on
the indexation of SequenceFile (Figs. 5 and 6).
When a request for a particular file is executed, a request is sent by the HDFS
client to the NameNode machine to obtain the metadata of the requested file.
In the proposed approach, instead of processing file metadata directly, like a
traditional Hadoop processing, the NameNode will opt for metadata processing of
the combined files based on a mapping file.
This mapping file consists of:
the file name, file length, file offset, block number of the combined file.
So we call a “mapping record” each entry in the mapping file.
Fig. 5 Hadoop file server analyzer

Fig. 6 Combiner phase
6.6 Prefetching and Caching
The proposed approach reduces load on NameNode by using prefetching and

caching techniques that are generally used as storage optimization techniques [23].
Prefetching conceals visible I/O cost and improves access time by utilizing corre-
lations between files and fetching data into cache prior to they are requested
[29].
The study of the behavior of data processing is not limited to the study of the
consumption of the memory and the processors, but also depends on storage param-
eters, and we can include here two parameters used in the optimization of storage:
prefetching and caching [30].
Prefetching conceals visible I/O costs and improves file access time by retrieving
data files in the cache before they are requested [31]. This means that access to the
metadata will be acquired directly from the cache instead of the disk volumes.
This greatly improves file access times (Figs. 7 and 8).
Fig. 7 MetaData caching
Fig. 8 Prefetching-based MetaData
7 Experimental
7.1 Experimental Environment
The suggested implementation in this paper can be contrasted to the original HDFS in
terms of NameNode memory usage along with MapReduce job performance, during
sequential, and selective file access. Our simulation was conducted using Hadoop
2.4.0, with a cluster of nodes each node consisting of the following specs:
1. Node x
2. 3.10 GHz clock rate
3. 8 GB of memory
4. 1 Gbit Ethernet—Network interface controller
5. 4 Data Nodes
All nodes offer a 500 GB hard drive and are deployed over Ubuntu server 14.04
distribution. The replication rate will be preserved as the default 3, and the block size
of HDFS is picked as 64 MB. Pilot datasets are mostly auto-generated, and we also
have the ability to use public datasets on [27].
7.2 Phases Algorithms
Input:
Block_Size: Defined block size
CombinedFile: Name of combined file.
smallFile_List: list of small files
smallFile_Dir: Directory containing small files
Output:
File_Index: index of CombinedFile.
Initialize smallFile_Map
if smallFile_List is provided as input
for each file fl in File_List
compute code(fl.name)
add fl to smallFile_Map with key: code,
values:name(fl), length(fl)
end for
else if smallFile_Dir is given as input
for each file f in the directory tree at
smallFile_Dir
compute code(fl.name)
add f to sF_Map with key: code,
values:name(f), length(fl)
end for
end if
if CombinedFile exists
open CombinedFile for append
LocalIndex= LocalIndex(CombinedFile );
blockNum = length(combinedFile)/Block_Size
cur_offset = length(CombinedFile) % Block_Size
else
create CombinedFile; open CombinedFile for write
Initialize File_Index; set cur_offset= 0, blockNum=0
end if
for each file with code fl in smallFile_ap in order
of
increasing length(fl)
name = name(fl), ln=length(fl); combinedIn=CombinedFile
if cur_offset + ln > Block_Size
blockNum++,
start = blockNum*Block_Size, end= start + ln,
curr_offset = ln ++
else if cur_offset + ln = Block_Size
start= blockNum*Block_Size + cur_offset, end= start
+ ln,
curr_offset=0, blockNum++
else
start= blockNum*Block_Size + cur_offset,
end = start + ln,
curr_offset=end ++
endif
append f to CombinedFile
Insert key fl into LocalIndex with name, ln, start,
end, combinedIn as values
end for
close CombinedFile
return File_Index.
7.3 Results
The amount of files generated by size is presented in Fig. 10; the total number of
files is 20,000, and the size of files range is from 1 KB to 4.35 MB (Fig. 9).
The workloads for measuring the time taken for read and write operations are
a subset of the workload used for memory usage experiment, containing the above
datasets (Fig. 11; Table 1).
As shown in the figure, we can conclude that memory consumption using the
approach is 20–30% lower than that consumed by the original HDFS. Indeed, the
NameNode, for original HDFS, stores file and bloc metadata for each file.
This means that by increasing the number of files stored the memory consumption
increases. On the other hand, as for the proposed approach, the NameNode stores
only the metadata of each small file. For the block metadata, the NameNode stores
them as a single combined file and not for every single small file, which explains the
reduction of the memory used by the proposed approach.
7000
7000
6000
NUMBER OF FILES BY RANGE
5000
5000
4000
4000
3000
3000
2000
1000
1000
0
0-128 128-512 512-1028 1024-4096 4096-8192
SMALL FILES RANGE (KB)
Fig. 9 Distribution of file sizes in our experiment
3500
3000 3087.2
MetaData size (KB)
2500
2240.5
2000
1500 1550.9
1000
800
500
353.6
0 1.88 7.3 16.6 25.1 31.9
2500 5000 10000 15000 20000
Number of files
HDFS HFSA approach
Fig. 10 Memory consumption
7.4 Performances Comparison
7.4.1 Writing Test
Results of writing time are shown in this Fig. 12.

In the following, we performed MapReduce jobs on five datasets. As shown in
the diagram, when the number of files is important, the gain in writing performance
becomes more important using our approach. For example, we have a performance
of 9% for 1000 files beside 36% for 20,000 files.
Normal HDFS HFSA Algorithm
2500 2305
2010
TIME CONSUMPTION (S)
2000
1682.65
1410 1467.3
1500
980 916.5
1000
686
500
0
5000 10000 15000 20000
NUMBER OF SMALL FILES
Fig. 11 Memory usage by NodeName
Table 1 Comparison of the

Dataset # Number of Time consumption (s)
NameNode memory usage
small files Normal HDFS HFSA
algorithm
1 5000 980 268
2 10,000 1410 390
3 15,000 2010 530
4 20,000 2305 620
Time Consumpon (s) Normal HDFS Time Consumpon (s) HFSA Algorithm
8000
7000
6000
WRITING TIME (S)
5000
4000
3000
2000
1000
0
2500 5000 10000 15000 20000
Fig. 12 Performance evaluation: writing time

Time Consumpon (s) Normal HDFS Time Consumpon (s) HFSA Algorithm
3500
3000
READING TIME (S)
2500
2000
1500
1000
500
0
2500 5000 10000 15000 20000
Fig. 13 Performance evaluation: reading time
Through the above comparison, it is proved that our approach can correctly
enhance the effectivity of file writing.
7.4.2 Reading Test
The average sequence reading time of HFSA is 788,94 s, and the average read time
of the original HDFS is 1586,58 s. The comparison shows that the average reading
speed of SFS is 1.36 times of HDFS, 13.03 times of HAR.
Applying our approach, we had a performance of around 33% for writing process
and more than 50% for reading (Fig. 13).
8 Conclusion and Future Works
In this paper, we described, in a detailed way, our approach and solution to address
Hadoop technology defects related to distributed storage of large volumes of small
files.
The Hadoop Server File Analyzer supports the combination of a set of files into
MapFile and then categorizes them.
This technique greatly improved the write and read performance of the classic
Hadoop system and also greatly reduced the RAM consumption on the DataNode.
Several researches and several scenarios have been launched to meet the same need
and to improve the technique, such as HAR and NHAR or other technologies such
as SPARC and STORM but each proposed solution, and each approach developed
does not only respond to a very specific need.
If we opted for the solution of the combination of small files, it is to gain in

performance of course; but it is also to maintain a cleaner file system HDFS, because
we do not want to have thousands and millions of checks for each file!
The next phase is to improve the performance of the search on small files in a
huge volume of data; using graph theory techniques in the first phase, especially
the Dijkstra’s and Bellman–Moore algorithms, this first phase will be the initial
basis that will feed our A* algorithm that we will use in our approach to artificial
intelligence.
References
1. Hadoop official site. http://hadoop.apache.org/

2. https://www.lebigdata.fr/hadoop
3. Achandair, O., Elmahouti, M., Khoulji, S., Kerkeb, M.L.: Improving Small File Management
in Hadoop, pp. 1–14 (2017).
4. Bende, S., Shedge, R.: Dealing with Small files problem in hadoop distributed file system.
Procedia Comput. Sci. 79, 1001–1012 (December 2016)
5. Cai, X., Chen, C., Liang, Y.: An optimization strategy of massive small files storage based
on HDFS. In: 2018 Joint International Advanced Engineering and Technology Research
Conference (JIAET 2018) (2018)
6. Niazi, S., Ronström, M., Haridi, S., Dowling, J.: Size Matters: Improving the Performance of
Small Files in Hadoop. Middleware’18. ACM, Rennes, France (2018)
7. Mir, M.A., Ahmed, J.: An Optimal Solution for Small File Problem in Hadoop. Int. J. Adv.
Res. Comput. Sci. (2017)
8. Alange, N., Mathur, A.: Small sized file storage problems in hadoop distributed file system. In:
Second International Conference on Smart Systems and Inventive Technology (ICSSIT 2019),
IEEE Xplore (2019)
9. Archid, A.S., Mangala, C.N.: Improving Hadoop Performance in Handling Small Files. Int. J.
Eng. Res. Technol. (IJERT) (2016)
10. Ahada, M.A., Biswasa, R.: Architecture for Efficiently Storing Small Size Files in Hadoop.
Procedia Comput. Sci. 132, 1626–1635 (2018)
11. Vorapongkitipun, C., Nupairoj, N.: Improving performance of smallfile accessing in hadoop.
In: IEEE International Conference on Computer Science and Software Engineering (JCSSE),
pp. 200–205 (2014)
12. Sheoran, S., Sethia, D., Saran, H.: Optimized MapFile based storage of small files in hadoop.
In: ACM International Symposium on Cluster, Cloud and Grid Computing
13. https://searchstorage.techtarget.com/definition/parallel-file-system
14. Carns, P.H., Ligon III, W.B., Ross, R.B., Thakur, R.: Pvfs: A parallel file system for linux
clusters. In: Proceedings of the 4th Annual Linux Showcase and Conference, pp. 317–327.
USENIX Association.
15. Alange, N., Mathur, A.: Small sized file storage problems in hadoop distributed file system. In:
2019 International Conference on Smart Systems and Inventive Technology (ICSSIT). IEEE
(2019)
16. https://dataottam.com/2016/09/09/3-solutions-for-big-datas-small-files-problem/
17. Bende, S., Shedge, R.: Dealing with small files problem in hadoop distributed file system. In:
7th International Conference on Communication, Computing and Virtualization (2016)
18. Implementing WebGIS on Hadoop: A Case Study of Improving Small File I/O Performance
on HDFS
19. Wang, K., Yang, Y., Qiu, X., Gao, Z.: MOSM: An approach for efficient string massive small
files on Hadoop. In: International Conference on Big Data Analysis (ICBDA), IEEE (2017)
20. Huang, L., Liu, J., Meng, W.: A review of various optimization schemes of small files storage on
Hadoop. In: Joint International Advanced Engineering and Technology Research Conference
(JIAET 2018) (2018)
21. Tchaye-Kondi, J., Zhai, Y., Lin K.J., Tao, W., Yang, K.: Hadoop perfect file: a fast access
container for small files with direct in disc metadata access. IEEE
22. Ciritoglu, H.E., Saber, T., Buda, T.S., Murphy, J., Thorpe, C.: Towardsa better replica manage-
ment for hadoop distributed file system. IEEE Big Data Congress ‘18At: San Francisco
(2018)
23. Cheng, W., Zhou, M., Tong, B., Zhu, J.: Optimizing Small File Storage Process of the HDFS
Which Based on the Indexing Mechanism. In: 2nd IEEE International Conference on Cloud
Computing and Big Data Analysis (2017)
24. Venkataramanachary, V., Reveron, E., Shi, W.: Storage and rack sensitive replica placement
algorithm for distributed platform with data as files. In: 2020 12th International Conference on
Communication Systems & Networks (COMSNETS) (2020)
25. Rattanaopas, K., Kaewkeeree, S.: Improving Hadoop MapReduce Performance with Data
Compression: A Study Using Wordcount Job. IEEE (2017)
26. El-Sayed, T., Badawy, M., El-Sayed, A.: SFSAN approach for solving the problem of small
files in Hadoop. In: 2018 13th International Conference on Computer Engineering and Systems
(ICCES) (2018)
27. Niazi, S., Ronström, M.: Size Matters: Improving the Performance of Small Files in Hadoop.
In: The 19th International Middleware Conference
28. Climate Data Online, available from National Centers for Environmental Information at https://
www.ncdc.noaa.gov/cdo-web/datasets
29. Merla, P.R., Liang, Y.: Data analysis using hadoop MapReduce environment. IEEE
30. Tao, W., Zhai, Y., Tchaye-Kondi, J.: LHF: A New Archive based Approach to Acclerate Massive
Small Files Access Performance in HDFS. EasyChair Preprint n°. 773 (2017)
31. Shah, A., Padole, M.: Optimization of hadoop MapReduce model in cloud computing
environment. IEEE (2019)
32. Zheng, T., Guo, W., Fan, G.: A method to improve the performance for storing massive small
files in Hadoop. In: The 7th International Conference on Computer Engineering and Networks
(CENet2017) Shanghai (2017)
33. https://arxiv.org/ftp/arxiv/papers/1904/1904.03997.pdf
COVID-19 Pandemic
Data-Based Automatic Covid-19 Rumors
Detection in Social Networks
Bolaji Bamiro and Ismail Assayad
Abstract Social media is one of the largest sources of propagating information;

however, it is also a home ground for rumors and misinformation. The recent extraor-
dinary event in 2019, the COVID-19 global pandemic, has spurred a web of misin-
formation due to its sudden rise and global widespread. False rumors can be very
dangerous; therefore, there is a need to tackle the problem of detecting and miti-
gating false rumors. In this paper, we propose a framework to automatically detect
rumor on the individual and network level. We analyzed a large dataset to eval-
uate different machine learning models. We discovered how all our methods used
contributed positively to the precision score but at the expense of higher runtime.
The results contributed greatly to the classification of individual tweets as the dataset
for the classification task was updated continuously, thereby increasing the number
of training examples hourly.
1 Introduction
In our world today, we have economic, technological, and social systems built with
high complexity to help the human society. However, these systems can be highly
unpredictable during extraordinary and unprecedented events. The most recent global
pandemic called COVID-19 started gaining attention from late December 2019 and
has affected the world greatly with more than 45 million cumulative worldwide
cases1 of infection currently [1]. During these shocking periods, cooperation is crucial
to mitigate the impact of the pandemic on the collective well-being of the public.
1 https://coronavirus.jhu.edu/map.html
B. Bamiro (B)
African Institute for Mathematical Sciences, Mbour, Senegal
e-mail: bolaji.r.bamiro@aims-senegal.org
I. Assayad
LIMSAD Faculty of Sciences and ENSEM, University Hassan II of Casablanca, Casablanca,
Morocco
https://doi.org/10.1007/978-981-16-3637-0_57
816 B. Bamiro and I. Assayad
Social media, which is a complex society that aids in global communication and
cooperation, has, however, become one of the major sources of information noise
and fake news. ‘Fake news spreads faster and more easily than this virus, and is
just as dangerous.’ were the words of Dr. Tedros Adhanom Ghebreyesus at the
World Health Organization, Munich Security Conference, February 15, 2020 [2].
The waves of unreliable information being spread may have a hazardous impact on
the global response to slow down the pandemic [3]. Most of the fake news is harmful
and problematic as they reach out to thousands of followers. The possible effects
are widespread fear [4], wrong advice that may lead to the encouragement of risky
behavior, and contribution to the loss of life during the pandemic [5, 6].
There are recognized organizations that have dealt with rumors such as the Inter-
national Fact-Checking Network (IFCN) [7], the World Health Organization (WHO),
and United Nations Office (UN).
This paper aims to achieve the goal by designing a framework that can effectively
detect rumor over time by analyzing tweets on twitter. Twitter is one of the largest
social media platforms [8]; therefore, we obtained the dataset from the platform. The
contributions made in this paper are as follows:
• evaluate methods and models to detect rumors with high precision using a large
dataset obtained from Twitter;
• addition of image analysis to text analysis.
• designing a unified framework that detects rumor effectively and efficiently in
real time.
2 Related Works
Research on rumor detection has been receiving a lot of attention across different
disciplines for a while now [9, 10]. New approaches have been arising to tackle the
problem of fake news specifically on social media using computational methods.
These methods have been shown by [11–13] to be efficient not just in solving the
rumor detection problem but also in identification of such fake news on time [14].
Some of the methods used are machine learning, n-gram analysis, and deep learning
models to develop detection and mitigation tools for classification of news [14, 15].
Some take this further to apply several tools for higher precision [13]. Much previous
research approaches these problems by analyzing a large number of tweets during
the COVID-19 epidemic to analyze the reliability of the news on social media that
poses serious threat amidst the epidemic [10, 16]. Another approach that has been
used to study social media news is fake images, especially on COVID-19 [17]. Few
studies have investigated the reliability of images on social media. Such methods
have been used by [18] by analyzing a large number of tweets to characterize the
images based on its social reputation and influence pattern using machine learning
algorithms.
Most studies make use of response information: agreement, denial, enquiry,
and comment in their rumor detection model, and have shown good performance
Data-Based Automatic Covid-19 Rumors Detection in Social Networks 817
improvement [19]. Text content analysis is also an important method that has been
employed by most previous studies on rumor detection. It includes all the post text and
user responses. Deceptive information usually has a common content style that differs
from that of the truth, and researchers explore the sentiment of the users toward the
candidate rumors such as [19]. It is important to note that although textual content
analysis is quite important in rumor detection, many studies point that just is not
sufficient [20]. Visual features (images or videos) are also an important indicator for
rumor detection and have been shown [17, 21]. Rumors are sometimes propagated
using fake images which usually provokes user responses.
Network-based rumor detection is very useful because it involves construction
of extensible networks to indirectly collect possible rumor propagation information.
Many studies have utilized this method such as [13, 18, 22]. Knowledge base (KB)
has also been shown to be quite important for detecting fake news. This involves
using known truth about a situation. Some studies on employing KB in the past such
as [23]. Very few previous studies have, however, been designed for real-time rumor
detection systems [19, 24]. This paper aims to develop a framework for a practical
rumor detection system that uses available information and models by collectively
involving major factors which are text analysis, knowledge base, deep learning,
natural language processing (NLP), network analysis, and visual context (Images).
3 Background
The definition of rumor is quite ambiguous and inconsistent as various publications

have different definitions for it. Having a solid definition is crucial toward making a
well-informed classification of news by understanding its properties. However, this
paper will emphasize on these two definitions.
3.1 A Rumor is a Statement Whose Truth Value is Either

True, False or Unverified [15]
The definition generally means that the truth value of a rumor is uncertain. The main
problem arises when the rumor is false, and this is often referred to as false or fake
news.
3.2 A Rumor is Defined as Unreliable Information that is

Easily Transmissible, Often Questioned [25, 26]
For further clarity on this definition, we emphasize on the properties of a rumor which
are unreliable information, easily transmissible, often questioned. Rumors cannot be
relied upon because its truth value is uncertain and controversial due to lack of
evidence. Rumors easily transmit from one person/channel to another. Also, study
from [27] shows that false rumor spreads wider and faster than true news. Rumors
cause people to express skepticism or disbelieve, that is, verification, correction, and
enquiry [13].
4 Problem Statement
This paper aims to solve the rumor detection problem: A post p is defined as a set
of i connected news N = {n1 , n2 , …, ni } where n1 is the initial news in which other
posts spanned from. We define a network of posts from a social media platform
where the nodes are the messages while the edges are the similarities between the
two nodes. Each node ni has its unique attributes that represent its composition. This
consists of the user id, post id, user name, followers count, friend count, source, post
creation time, and accompanying visuals (images). Given these attributes, the rumor
detection function takes as input the post p, the set of connected news N together
with the attributes to return an output {True, False} that determines whether that post
is a rumor or not.
5 Real-Time Computational Problem
For each tweet from the tweet stream containing the text and image information, we
extract its attributes as a set {t 1 , t 2 , … t i }. For the rumor detection problem, we aim
to predict whether the tweet is a rumor or not using the attributes of each tweet.
6 Methodology
According to Fig. 1, there are four phases involved in this paper’s framework
for rumor detection. The methods used in this section are a modification and
improvement of the paper [13]. These phases are as follows.
(A) Extraction and storage of tweets;
(B) Classification of tweets;
(C) Clustering of tweets and;
Fig. 1 Framework showing the different phases of rumor detection
(D) Ranking and labeling of tweets.
6.1 Extraction and Storage of Tweets
The goal of this paper is to detect rumor early, and hence, extraction and storage of
tweets are quite important. The tweets are streamed using a python library called
Tweepy and cleaned to remove stop words, links, and special characters, and also
extract mentions and hashtags. The tweets are then stored in a MySQL database.
6.2 Classification of Tweets
The second phase involves classifying the tweets into signal and non-signal tweets.
In this paper, signal tweets are tweets that contain unreliable information, verification
questions, corrections, enquiry, and fake images. Basically, a tweet conveying infor-
mation usually contains a piece of knowledge, ideas, opinions, thoughts, objectives,
preferences, or recommendations.
Verification/confirmation questions have been found to be good signals for rumors
[13] and also visual attributes [17]. Therefore, this phase will explore text and image
Table 1 Method and result comparison with enquiring minds [13]

Text Image NLP Deep Network Knowledge Average Average
analysis analysis learning analysis base precision precision
without with ML
ML
Enquiring Yes No Yes No Yes No 0.246 0.474
minds
[13]
Our Yes Yes Yes Yes Yes Yes 0.602 0.638
method
The machine learning model used in [13] was a decision tree model and was compared with the
random forest model used in this study
analysis. Since this paper is based on COVID-19, we will also use a knowledge-
based method. This involves using known words or phrases that are common with
COVID-19 rumors.
6.2.1 Text Analysis
At this stage, we want to extract the signal tweets based on regular expressions. The
signal tweets will be obtained by identifying the verification and correction tweets
as used in [13]. We will also add known fake websites identified by Wikipedia2 and
WHO Mythbusters as shown in Table 1. We make use of the spaCy python library
to match the tweets to the phases.
6.2.2 Image Analysis
This project aims to use the visual attributes of the tweet as one of the factors to
detect rumor. We approach this by using three stages of analysis for the images. At
the first stage, the image metadata is analyzed to detect software signatures. It is the
fastest and simplest method to classify images. However, image metadata analysis
is unreliable because there are existing programs that can alter it such as Microsoft
Paint. An image without an altered metadata will contain the name of the software
used for the editing, for example, Adobe Photoshop. The second stage makes use of
the error level analysis (ELA) and local binary pattern histogram (LBPH) methods.
ELA basically detects areas in an image where the compression levels are different.
After the ELA, the image is passed into the local binary pattern algorithm. LBPH
algorithm is actually used for face recognition and detection, but, in this paper, it is
useful for generating histograms and comparing them. At the third and final stage,
the image is reshaped to 100px x 100px image. This aspect involves deep learning.
We used the pre-trained model VGG16 and also added a CNN model. Then, these
2 https://en.wikipedia.org/wiki/Fake_news.
10,000 pixels with RGB values will be fed into the input layer of the multilayer
perceptron network. Output layer contains two neurons: one for fake images and one
for real images. Therefore, based on the neuron outputs, we determine whether the
images are real or fake.
6.2.3 Clustering of Tweets
The third phase involves clustering of tweets to highlight the candidate rumors.
Usually, a rumor tweet is retweeted by other users or recreated similarly to the original
tweet. This is why clustering of similar tweets is quite important. However, to reduce
the computational costs, memory, and time used by other clustering algorithms, we
treat the rumor clusters as a network. This method can be quite efficient for this phase
as shown in [13]. We define a network where the nodes represent the tweets while the
edges represent the similarity. Nodes with high similarity are connected. We define
this network as an undirected graph to analyze the connected components, that is, a
path connects every pair of nodes in the graph. We measure the similarity between
two tweets t 1 and t 2 using Jaccard coefficient. The Jaccard coefficient between t 1
and t 2 can be measured using:
|N gram (t1) ∩ N gram (t2)|

J (t1 , t2 ) = (1)
|N gram (t1) ∪ N gram (t2)|
where N is 1-g of tweets t 1 and t 2 .

Jaccard distance is commonly used to measure similarity between two texts [13].
The similarity values range from 0 to 1 and values tending to 1 mean a higher
similarity. However, computing these similarities for each pair of tweets may be time
consuming; therefore, we make use of the MinHash algorithm [28]. The threshold
set for high similarity is at 0.7, that is 70 % similarity between the pair of tweets.
After clustering the signal tweets, we also add the non-signal tweets to the network
also using Jaccard similarity, however, with a threshold of 60 %.
6.2.4 Ranking and Labeling of Tweets
At this phase, each tweet has a degree centrality score. However, tweets with high
degree centrality may not be rumors. Therefore, this phase applies machine learning
to rank the tweets. We extract features from the candidate rumors that may contribute
to predicting whether the candidate is a rumor. Some of these features were used in
[13]. The following are the features used were Twitter_id, follower’s count, location,
source, Is_verified, friends count, retweet count, favorites count, reply, Is_protected,
sentimental analysis, degree centrality score, Tweet length, signal tweet ratio, subjec-
tivity of text, average tweet length ratio, retweet ratio, image reliability, hashtag ratio,
and mentions Ratio.
6.2.5 Experimental Materials and Structure
Dataset
The initial data set used is the COVID-19 tweets selected randomly from February
2020 to October 2020. The total amount of data collected for labeling was 79,856.
The data set used to train the images was obtained from MICC-2000 [29]. It consists
of 2000 images, 700 of which are tampered and 1300 originals.
Ground Truth
The dataset collected was then labeled for training. The labels were assigned
according to the definitions given in Sect. 3, and also, some tweets had to be confirmed
by web search. The reliability achieved a Cohen’s Kappa score of 0.75.
Evaluation Metric
We divided the dataset labeled into train and validation sets. The validation set
contains 13,987 tweets. Then different machine learning models were used to rank
the test set. The evaluation of the model will be based on its top N rumor candidates
where N is varied.
TP
Precision = (2)
TP + TN
The detection time and batch size are also taken into consideration.
Baseline Method
The baseline methods consist of the framework provided without machine learning
and Image analysis.
Text analysis (verification and correction only): This involves using only text
analysis at the classification of tweets into signal tweets. We evaluate the efficiency
of this method without including the visual attributes and knowledge base using the
rank of the output from the machine learning models.
Without machine learning: This involves using only the degree centrality method
to rank the candidate rumor. At phase 4, the different machine learning models are
skipped and evaluated for efficiency. This also includes the omission of the CNN
model at the image analysis stage. Therefore, the rumor detection algorithm based
on this method outputs the rank of the clusters without any machine learning involved
in the process.
Variants
To improve on the baseline methods, we introduce three variants. These variants will
enable us to understand the effectiveness of the method. The variants are as follows:
Text (verification and correction only) and Image Analysis: For this variant, we
use verification and correction, and image analysis to classify the tweets into signal
tweets.
Text analysis (verification and correction, and knowledge base): For this variant,
we use verification and correction, and knowledge base without image analysis to
classify the tweets into signal tweets.
Text (verification and correction, and knowledge base) and Image Analysis: For
this variant, we use verification and correction, knowledge-based method, and image
analysis to classification of tweets into signal tweets. This is our method, and it is
a collation of all the methods in the framework. We evaluate the efficiency of this
method without including the visual attributes using the rank of the output from the
machine learning models.
Machine learning: This variant involves using various machine learning models
to rank the rumor candidates.
Algorithm 1: Ranking Clustered Tweets
Input: Document term, Tweets

Output: Ranked signal tweets
Initialize: N
for text in document term do
| s := signal tweets (get signal tweets)
end
for id, text1 in tweets do
for id, text2 in s do
| J := Jaccard (text1, text2) (calculate Jaccard Coefficient)
| if J > a (set Coefficient threshold)
| | append to dataframe N
| end if
end
end
D := Degree centrality(text)
Rank dataframe N based on D in descending order
Result and Discussions
Precision of Methods
The precision of these methods is evaluated using 10 min of tweets collected on

October 27, 2020. This dataset consists of 13,437 tweets with 213 images in total. The
precision value only takes into account the top 100 tweets ranked by without machine
learning (degree centrality score is used) and with machine learning (CatBoost
model) for the baseline and variants methods, respectively. The results obtained
show that the collation of our methods (text (verification and correction + knowl-
edge base) and image analysis) detected more signal tweets and candidate rumor than
the other method with higher precision with machine learning ranking. The precision
of our method outperformed other methods with a precision of 0.65 with machine
learning. The results also showed that the signal tweets and candidate rumor detected
using our method is much larger than using the baseline method.
Ranking Candidate Rumor
After the clustering of the tweets, we use different ranking methods to rank the candi-
date rumors. The baseline ranking method is ranking based on the degree centrality
score while our method is based on using machine learning models. Among all
machine learning models, we selected a logistic, tree and boosting model-random
forest, logistic regression, and CatBoost model. For the machine learning models,
we use the 20 statistical features described in the methodology section. We trained
the models and tested their performances for comparison. A tenfold cross-validation
was carried out to obtain the average performance. These graphs show a general
reduction in precision as N increases. However, the CatBoost model outperforms
other ranking methods except the Text + Image analysis method where the degree
centrality ranking method performs the best. Logistic regression, however, does not
perform well which may be due to overfitting.
The text + knowledge base method performs best at N = 10 and 20 at with
an average precision value of 0.95 but decreases gradually as N tends to 100. Our
method shows an improvement in most methods especially for the CatBoost model
but its value decreases steadily with increase in N.
Early Detection
It is very useful to detect rumors early; therefore, early detection is a key objective
in this paper. The run time was measured between the baseline method-text analysis
only and variants to determine how early the method can detect the rumor. The results
showed that as the number of tweets increases, the run time increases much faster
in the variant methods as compared to the baseline method. However, the number of
signal tweets detected also increases which improve the precision. This difference
because of the time taken to get each image and classify as rumor or non-rumor. The
higher the number of tweets, the higher the number of images, and hence, the higher
the run time.
Efficiency of Real-Time Detection Framework

Using the above result, we develop an application that detects streaming tweets in
real time while the dataset used for the prediction is appended continuously using the
top 10 candidate rumor detected by the text (verification and correction + knowledge
base) and image analysis hourly. The real-time rumor detection application predicts
an average of 4.04 rumors per second. The web application built using Flask detects
an average of 38 rumors every 8 s.
6.3 Discussion
In this paper, we built a framework to take advantage of the text and visual attributes
of a tweet to classify the tweet as a rumor. It improves the verification and correction
method by including other known regular expressions associated with the problem
and publicly declared fake new websites. We went further to use different ranking
methods to rank clustered rumors using complex network algorithms. We extracted
20 features from the rumors to train models for prediction. We observed that the top
features that have a high impact on the ranking are sentiment analysis, location, and
friends count using the CatBoost model. CatBoost has shown to be very effective in
ranking the candidate rumor because it outperforms other algorithms. This is very
useful because it gives us an idea of the features of tweets needed for real-time
detection and the best model that can deliver the highest precision. The precision can
be improved with higher number of training examples. Our method, however, takes
much longer to run as compared to the baseline method because of the addition of
the image component. Therefore, we have to decide if early detection is a price to
pay for higher precision. The real-time detection component, however, solves this
problem as it predicts the tweet as they stream in.
7 Conclusion
False rumors can be very dangerous especially during a pandemic. Rumor super
spreaders are taking the COVID-19 pandemic to confuse social media users. This
is why it is very important to detect rumors as early as possible. The World Health
Organization is working hard to dispute many false rumors and has provided some
information. Using these details, we built a framework to detect rumors. The approach
used is quite efficient with machine learning because it yields high precision. The
real-time detection model detects 4.04 rumors per second using training examples
appended continuously from the approach. This approach can be improved upon by
reducing the analysis run time for our method.
Acknowledgements Special thanks go out to the African Institute for Mathematical Sciences
(AIMS) and LIMSAD for their support toward this paper.
References
1. W. H. Organization et al.:Coronavirus disease 2019 (covid-19): situation report, 103 (2020)

2. Zarocostas, J.: How to fight an infodemic. The Lancet 395, 676 (2020)
3. Anderson, J., Rainie, L.: The Future of Truth and Misinformation Online, vol. 19. Pew Research
Center (2017)
4. Latif, S., Usman, M., Manzoor, S., Iqbal, W., Qadir, J., Tyson, G., Castro, I., Razi, A., Boulos,
M.N.K., Weller, A., et al.: Leveraging data science to combat covid-19: a comprehensive review
(2020)
5. Tasnim, S., Hossain, M.M., Mazumder, H.: Impact of rumors or misinformation on coronavirus
disease (covid-19) in social media (2020)
6. Hossain, M.S., Muhammad, G., Alamri, A.: Smart healthcare monitoring: a voice pathology
detection paradigm for smart cities. Multimedia Syst. 25(5), 565–575 (2019)
7. Perrin, C.: Climate feedback accredited by the international fact-checking network at poynter.
Clim. Feedback 24 (2017)
8. Kouzy, R., Abi Jaoude, J., Kraitem, A., El Alam, M. B., Karam, B., Adib, E., Zarka, J.. Traboulsi,
C., Akl, E.W., Baddour, K.: Coronavirus goes viral: quantifying the covid-19 misinformation
epidemic on twitter. Cureus 12 (2020)
9. Li, Q., Zhang, Q., Si, L., Liu, Y.: Rumor detection on social media: Datasets, methods and
opportunities. arXiv preprint arXiv:1911.07199 (2019)
10. Shahi, G.K., Dirkson, A., Majchrzak, T.A.: An exploratory study of covid-19 misinformation
on twitter. arXiv preprint arXiv:2005.05710 (2020)
11. Ahmed, H., Traore, I., Saad, S.: Detection of online fake news using n-gram analysis
and machine learning techniques. In: International Conference on Intelligent, Secure, and
Dependable Systems in Distributed and Cloud Environments. Springer, Berlin, pp. 127–138
(2017)
12. Bharadwaj, A., Ashar, B.: Source based fake news classification using machine learning. Int.
J. Innov. Res. Sci. Eng. Technol. 2320–6710 (2020)
13. Zhao, Z., Resnick, P., Mei, Q.: Enquiring minds: early detection of rumors in social media
from enquiry posts. In: Proceedings of the 24th International Conference on World Wide Web,
pp. 1395–1405 (2015)
14. Liu, Y., Wu, Y.-F.B.: Early detection of fake news on social media through propagation path
classification with recurrent and convolutional networks. In: Thirty-second AAAI conference
on artificial intelligence (2018)
15. Qazvinian, V., Rosengren, E., Radev, D., Mei, Q.: Rumor has it: identifying misinformation
in microblogs. In: Proceedings of the 2011 Conference on Empirical 15 Methods in Natural
Language Processing, pp. 1589–1599 (2011)
16. Al-Rakhami, M.S., Al-Amri, A.M.: Lies kill, facts save: detecting covid-19 misinformation in
twitter. IEEE Access 8, 155961–155970 (2020)
17. Jin, Z., Cao, J., Zhang, Y., Zhou, J., Tian, Q.: Novel visual and statistical image features for
microblogs news verification. IEEE Trans. Multimedia 19, 598–608 (2016)
18. Gupta, A., Lamba, H., Kumaraguru, P., Joshi, A.: Faking sandy: characterizing and identi-
fying fake images on twitter during hurricane sandy. In: Proceedings of the 22nd International
Conference on World Wide Web, pp. 729–736 (2013)
19. Liu, X., Nourbakhsh, A., Li, Q., Fang, R., Shah, S.: Real-time rumor debunking on twitter.
In: Proceedings of the 24th ACM International on Conference on Information and Knowledge
Management, pp. 1867–1870 (2015)
20. Chua, A.Y., Banerjee, S.: Linguistic predictors of rumor veracity on the internet. In: Proceedings
of the International MultiConference of Engineers and Computer Scientists, vol. 1, p. 387
(2016)
21. Wang, Y., Ma, F., Jin, Z., Yuan, Y., Xun, G., Jha, K., Su, L., Gao, J.: Eann: event adversarial
neural networks for multi-modal fake news detection. In: Proceedings of the 24th acm sigkdd
International Conference on Knowledge Discovery & Data Mining (2018), pp. 849–857
22. Wu, K., Yang, S., Zhu, K.Q.: False rumors detection on sina weibo by propagation structures.
In: 2015 IEEE 31st International Conference on Data Engineering (IEEE, 2015), pp. 651–662
23. Hassan, N., Arslan, F., Li, C., Tremayne, M.: Toward automated fact-checking: detecting check-
worthy factual claims by claimbuster. In: Proceedings of the 23rd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining (2017), pp. 1803–1812
24. Liu, X., Li, Q., Nourbakhsh, A., Fang, R., Thomas, M., Anderson, K., Kociuba, R., Vedder,
M., Pomerville, S., Wudali, R., et al.: Reuters tracer: a large scale system of detecting &
verifying real-time news events from twitter. In: Proceedings of the 25th ACM International
on Conference on Information and Knowledge Management (2016), pp. 207–216
25. DiFonzo, N., Bordia, P.: Rumor Psychology: Social and Organizational Approaches. American
Psychological Association (2007)
26. Bugge, J.: Rumour has it: a practice guide to working with rumours. Communicating with
Disaster Affected Communities (CDAC) (2017)
27. Vosoughi, S.: Automatic detection and verification of rumors on twitter. Ph.D. thesis,
Massachusetts Institute of Technology (2015)
28. Wu, W., Li, B., Chen, L., Gao, J., Zhang, C.: A review for weighted minhash algorithms. IEEE
Trans. Knowl. Data Eng. (2020)
29. Amerini, I., Ballan, L., Caldelli, R., Del Bimbo, A., Serra, G.: A sift-based forensic method for
copy–move attack detection and transformation recovery. IEEE Trans. Inf. Forensics Secur. 6,
1099–1110 (2011)
Security and Privacy Protection in the
e-Health System: Remote Monitoring
of COVID-19 Patients as a Use Case
Mounira Sassi and Mohamed Abid
Abstract The Internet of Things (IoT) is characterized by heterogeneous technolo-

gies, which contribute to the provision of innovative services in various fields of
application. Among these applications, we find the field of e-Health which provides
a huge amount of data that served the health of patients remotely and in real time but
also medical records, health monitoring and emergency response. e-Health systems
require low latency and delay which is not guaranteed since data are transferred to the
cloud and then back to the application, which can seriously affect performance. Also,
COVID-19 pandemic has accelerated the need of remote monitoring of patients to
reduce chances of infection among physicians and healthcare workers. To this end,
Fog computing has emerged, where cloud computing is extended to the edge of
the network to reduce latency and network congestion. This large amount of data
is downloaded and stored on remote public cloud servers to which users cannot be
fully trusted, especially when we are dealing with sensitive data like health data. In
this scenario, meeting the confidentiality and patient privacy requirements becomes
urgent for a large deployment of cloud systems. In this context, we offer a solution
to secure the personal data of the e-Health system and protect the privacy of patients
in an IoT–Fog–cloud environment while being based on cryptographic techniques
especially CP-ABE and the blockchain paradigm. The results obtained are satisfac-
tory, which allowed us to deduce that the solutions are protected against the most
known attacks in IoT–Fog–cloud systems.
M. Sassi (B)
Laboratory Hatem Bettaher Irescomtah, Faculty of sciences of Gabes,
University of Gabes, Gabes, Tunisia
M. Abid
Laboratory Hatem Bettaher Irescomtah, National School of Engineering of Gabes,
University of Gabes, Gabes, Tunisia
e-mail: mohamed.abid@enig.rnu.tn
https://doi.org/10.1007/978-981-16-3637-0_58
830 M. Sassi and M. Abid
1 Introduction
Countries around the world have been affected by the COVID-19 pandemic since
December 2019, and the health care systems are rapidly adapting to the increasing
demand. E-Health systems offer remote patient monitoring and share of information
between physicians. Hence, it helps facilitate and improve the prevention, diagnosis
and treatment of patients at a distance. Indeed, health data is collected by sensors
and then transmitted through the Internet to the cloud for consultation, evaluation
and recommendations by professionals.
According to the World Health Organization (WHO) [1], “COVID-19 is the dis-
ease caused by a new coronavirus, SARS-CoV-2.” It especially infects the respiratory
system of patients. Some people who have contracted COVID-19, regardless of their
condition, continue to experience symptoms, including fatigue and respiratory or
neurological symptoms. Therefore, doctors use intelligent equipments which collect
the measurements of a patient at home and sends them to the Fog. The latter can
be a treatment center installed in the hospital. Then, the Fog sends this data to the
cloud storage service for consultation by doctors. This process helps professionals
understand the behavior of this pandemic and gives them a hint about its evolution.
Despite the importance of e-Health system and their good results, it is necessary
to protect the confidentiality of the data, to secure the sharing and to protect the
private life of the patients. However, the implementation of treatment and storage in
Fog and cloud to store and process sensitive data raises many security issues (waste,
leakage or theft). Consequently, to use a model based on IoT–Fog–cloud architecture,
a reinforcement of the security measures is mandatory. Thus, the confidentiality,
integrity and access control of stored data are among the major challenges raised
by external storage. To overcome the challenges mentioned above, cryptography
techniques are widely adopted to secure sensitive data.
In this paper, a new solution to secure e-Health applications by exchanging data
confidentially and protecting patient privacy in an IoT–Fog–cloud architecture is
proposed. Our system can offer these basic functionalities:
Achieve a hard authentication and secure key sharing between the oximetry char-
acterized by a limited resource in memory and computation and the Fog. This allows
confidential transfer of data between the two entities.
Apply a public key (one-to-many) encryption scheme for secure cloud storage and
data sharing between a group of physicians. This scheme allows the implementation
of access control according to their attributes.
Combine cryptography technologies and blockchain to strengthen the manage-
ment of decentralized access control, keep the traceability of data traffic, and obtain
a level of anonymity offered by the blockchain.
Our system can effectively resist against the most well-known attacks in IoT and
against tampering of control messages.
The rest of the article is organized as follows. The related works on securing
e-Health systems are discussed in Sect. 2. We present the basic knowledge (prelimi-
naries) in Sect. 3. We describe the secure data sharing e-Health system that protects
Security and Privacy Protection in the e-Health System … 831
patient privacy based on CP-ABE encryption and blockchain in IoT- Fog-cloud archi-
tecture in Sect. 4. We provide security and performance analysis in Sect. 5. Section
6 concludes the article.
2 Related Works
There are many research works focusing in securing IoT application especially for
ambient-assisted living (AAL) application and e-Health system. Some researchers
used public key identity-based security or on lightweight cryptographic primitives
[2], such as one-way hash function and XOR operations. Others concentrated on
securing access to data, and many research used blockchain to secure the e-Health
system.
Chaudhari and Palve [3] developed a system that provides mutual authenti-
cation between all system components (human body sensor, handheld device and
server/cloud). The generated data is encrypted using RSA.
Wang et al. [4] proposed a scheme based on a fully homomorphic design for
the protection of privacy and the treatment of data in the e-Health framework. The
proposed architecture consists in performing a transmission mode for the electronic
health record. This mode ensures diagnostic of the patient based on encrypted records
residing in the cloud by a remote physician without being decrypted.
Bethencourt et al. [5] presented a system to achieve complex access control over
encrypted data based on encryption policy attributes. The access policy is embed-
ded in ciphertext. Attributes are used to describe the credentials of a user. A set S
of descriptive attributes is used to identify the private keys. When a party wishes
to encrypt a message through an access tree structure of a policy, his/her private
keys must satisfy to be able to decrypt. The CP-ABE scheme includes four main
algorithms: initialization, encryption, decryption and generation of the secret key.
Another scheme based on attribute encryption called “CCA” for architectures that
integrate the Fog as an outsourced decryption. It is proposed by Zu et al. [6] in
order to protect data in cloud computing. The main idea, to achieve this schema, is
to allow the decryptor to have the ability to verify the validity of the ciphertext. The
public key used is non-transformable and the type of attribute-based encryption used
is OD-ABE (attribute-based encryption with outsourced decryption).
Wang [7] proposed a secure data sharing scheme to ensure the anonymity and
identity confidentiality of data owners. Symmetric encryption, search-able encryp-
tion and attribute-based encryption techniques are used to keep data outsourced to
the cloud secure. Due to the risk of breach and compromise of patient data, medical
organizations have a hard time adopting cloud-stored services. Moreover, the existing
authorization models follow a patient-centered approach. Guo et al. [4] proposed a
CP-DABKS schema (ciphertext-policy decryptable attribute-based keyword search)
which allows an authorized user to decrypt data in a supposedly completely insecure
network. The architecture of this schema includes the four other components: KDG
(key general center) and the data center are the data owners. The data center lies
down the keywords and the access structure linked to the data. The third data receiver
element is the data consumer. It has a set of attributes, and a generated trap is used to
identify it in order to have the capacity to decrypt the data. Finally, the cloud server
plays the role, in this diagram, of a storage base for the data sent by the data sender
and a verification of the satisfaction of the access structure of a secret key received
by the data receiver.
Blockchain attracts attention in several academic and industrial fields [8]. It is
a technology that was first described in 1991 when researchers Stuart Haber and
W. Scott Stornetta introduced a computer solution, allowing digital documents to be
time-stamped and therefore never backdated or altered [9]. It is based on crypto-
graphic techniques: hash functions and asymmetric encryption. This technology was
at the origin of the Bitcoin “electronic money” paradigm described in the article by
Nakamoto [10] in 2009. Blockchain is an innovation in storage. “It allows informa-
tion to be stored securely (each writing is authenticated, irreversible and replicated)
with decentralized control (there is no central authority which would control the
content of the database)” [11]. This technology has been used in several areas of the
Internet of Objects and often used as a means to secure data as in the paper of Gupta
et al. [12] who proposed a model to guarantee the security of transmitted data and
received by nodes of an Internet of Things network and control access to data [13].
Blockchain is seen as a solution to make secure financial transactions without an
authority. It is also used in several areas with the aim of decentralizing security and
relieving the authorities. For example, the vehicular area integrates this technology to
solve security problems. We mention the article by Yao et al. [14] who have proposed
a BLA (blockchain-assisted lightweight anonymous authentication mechanism) to
achieve inter-center authentication that allows a vehicle to decide to re-authenticate
in another location. At the same time, they used the blockchain to eliminate commu-
nications between vehicles and service managers (SM), which considerably reduces
the communication time.
In recent years, to overcome security issues in e-Health systems, several solu-
tions are based on the blockchain to achieve personal health information (PHI)
sharing with security and privacy preservation due to its advantages of immutability.
C. Nguyen et al. [13] proposed a new framework (architecture) for offloading and
sharing of electronic health records (EHR) that combines blockchain and the decen-
tralized interplanetary file system (IPFS) on a mobile cloud platform. In particular,
they created a reliable access control mechanism using smart contracts to ensure
secure sharing of electronic health records (EDRs) between different patients and
medical providers. In addition, a data sharing protocol is designed to manage user
access to the system. Zhang et al. [15] built two types of blockchain (the private
blockchain and the consortium blockchain) to share PHI on secure health and main-
tain confidentiality. The private blockchain is responsible for storing the PHI, while
the consortium blockchain keeps its records at secure indexes. Next, we give the
basic knowledge that we use in the design of the new solution.
3 Preliminaries
To properly design our solutions, we must have prior knowledge of certain crypto-
graphic tools. Thus, this section is devoted for a generality on mathematical notions
where we refer to the books “Cryptography and Computer Security” [16], “Intro-
duction to Modern Cryptography” by Katz and Lindel [17] and course by Ballet and
Bonecaze [18].
Access Structure
Definition 1 {P1, P2, ..., Pn} is a set of parties. A collection A ⊆ 2{P1,P2,...,Pn} is

monotone if ∀B, C : i f B ∈ A and B ∈ C then C ∈ A. An access structure (respec-
tively, monotone access structure) is a collection (respectively, monotone collection)
A of non-empty subsets of {P1, P2, ..., Pn}, i.e., A ⊆ 2{P1,P2,...,Pn} ∅. The sets in
A are called the authorized sets, and the sets not in A are called the unauthorized sets
[5].
Security Model for CP-ABE
The ABE security model is based on the following functions:

Install: The challenger executes the configuration algorithm and gives the public
parameters, PK to the opponent.
Phase 1: The adversary creates repeated private keys corresponding to the sets of
attributes S1,. . ., Sq1.
Challenge: The opponent submits two messages of equal length M0 and M1.
Moreover, the adversary gives a challenge access structure A * such that none of
the sets S1,. . ., Sq1 satisfy the access structure. The challenger flips a random coin
b and digits Mb under A *. The CT * ciphertext is given to the opponent.
Phase 2: Phase 1 is repeated with the restriction that none of the attribute sets Sq1 +
1,. . ., Sq satisfy the access structure corresponding to the challenge.
Supposition: The opponent makes a supposition b ’of b. The advantage of an adver-
sary A in this game is defined as Pr [b = b]− 21
4 Secure Storage and Sharing of E-Health Data
4.1 An Overview of Our Proposed Schemes
In this section, we present our contribution to secure the sharing and storage of data
and preserve the privacy of patients in the e-Health system. Figure 1 shows the
different components of our architecture.
Connected objects (Oximeter): They generate the data (oxygen level in the blood)
of patients remotely and send them in real time to the Fog (which in turn is responsible
Fig. 1 Global architecture of our contribution
for processing them and sending them to the cloud in order to be stored and consulted
by doctors) in a secure manner using symmetric encryption after an authentication
phase and the exchange of a secret key.
Proxy/Fog Computing: It is an intermediary between health sensors and the cloud.
It offers storage, computation and analysis services. The Fog first decrypts the data
sent by the sensors through an anonymity proxy. He analyzes them. In case of emer-
gency, it sends an alert message to the ambulance. Using attribute-based encryption,
the Fog encrypts data and sends it to the cloud storage server where it will be saved.
Cloud: An infrastructure used in our system to store encrypted data and share it for
legal users. Attributes Authority: The attributes authority manages all attributes and
generates, according to the identity of the users, the set of key pairs and grants them
access privileges to end users by providing them with their secret keys according to
their attributes.
Users: Doctors and caregivers are the consumers of data. They request access
to data according to their attributes from cloud servers. Only users who have the
attributes satisfying the access policies can decrypt the data. Doctors can also add
diagnostics and recommendations to share with colleagues. This data is encrypted
by ABE and stored in the cloud.
Blockchain: It is a decentralized base of trust. It is used to ensure access control
management while ensuring data integrity and traceability of transactions made over
an insecure network.
We need a public key infrastructure (PKI) for entity identity verification for
blockchain operation and digital signatures. Table 1 shows the notations used to
describe the CP-ABE scheme and their meanings.
4.2 Modeling of Our Blockchain
We first present our records on the blockchain in the form of a token presenting a
pseudo transaction.
Table 1 Notation
Notation Description
Pks The set of public attribute keys
SKs The set of secret attribute keys
A List of attributes of a user
CT Ciphertexts (data encrypted by ABE)
The blockchain, in our construction, is used as a distributed, persistent and tamper-

proof book. It manages access control messages. In addition, one of the advantages
of using the blockchain is denying access to data by the cloud. A record (con-
tained in a block) in our distributed database is presented in the form of an access
authorization token (designated authorisation) on blockchain, equivalent to a pseudo
crypto-currency.
• Authorization (idx, @ cl, @ gr, @pr): This is a record in the blockchain. It allows
to specify an authorization by the Fog (the entity that encrypts the data by the ABE
encryption, which signs this token) with a blockchain @pr address so that a group
of doctors with address @gr can access data stored in the @cl cloud (blockchain
address of the storage provider). The cloud identifies data by the idx index. This
idx is calculated by the Proxy/Fog as a sequence idx = HMAC (CT).
We define two transactions carried out in the blockchain:
• GenAutorization (idx, @ cl, @ gr, @pr): It is the source transaction which is

generated by the owners of the data. Once the idx value is calculated, Fog com-
puting broadcasts this transaction to transfer this idx to the group of authorized
consultants.
• RequestAuthorization (authorization (idx, @ cl, @ gr, @pr), @ pr, @ rq): This
transaction is used to load authorization within the blockchain from an actor’s
account to the applicant’s account. The requestor uses their @rq address and sends
a request to the storage provider (@cl). The cloud sends this request to the Fog
(@pr). The token circulation processes to obtain authorization to access data are
illustrated in Fig. 2:
• 1: RequestAuthorization (authorization (idx, @ cl, @ gr, @pr), @ rq, @ cl): The

requestor (@rq) transfers a transaction to the cloud site (@cl) in order to obtain
the permission to download data.
• 2: Authorization request (authorization (idx, @ cl, @ gr, @pr), @ cl, @ pr): The
cloud transfers the transaction from the user request to the Fog (@pr).
• 3: RequestAuthorization (authorization (idx, @ cl, @ gr, @pr), @ pr, @ rq): The
Fog verifies the requester authorization and transfers the transaction to the doctor
site.
Fig. 2 Token circulation

processes in blockchain
4.3 The Proposed ABE Encryption Algorithm
Let G0 and G1 be two bilinear groups of prime order p, g is a generator point of G0,
and also, e is a bilinear map defined by: G0 × G0 → G1.
initialization: Setup()
The algorithm selects the groups G0 and G1 of order p and generator g of G0, and
then, it chooses two random numbers α and β in Zp and produces the keys:
PK = (G0, G1, g, h = g β , Y = e(g, g)α ).

MSK = (α, β)
Message encryptions: Encry (M, T, PK)
The encryption algorithm encrypts a message M under the access structure T. It

first chooses a polynomial qx for each node in the tree T in a descending manner,
from the root node R. Indeed, for each node x of T, we define the degree dx of the
polynomial qx by dx = kx − 1 with kx presents the threshold of this node.
Starting with the root node R, the algorithm chooses a random s ∈ Zp and defines
qR (0) = s. Then, it chooses dR at random in order to define it completely. For any
node x of the structure, this algorithm defines qx (0) = qparent (x) (index (x)), and
also, it chooses the degree dx in a random way.
We put the subset X of the leaf nodes of T, by giving the access structure to the
tree T and building the ciphertext:
CT = {T, E 1 = MY s , E 2 = h s , ∀i ∈ X : E i = gi (0), E = f (attributes(i)qi (0) )}

q
Private key generation: GenerKeyS (AU, MSK)
The algorithm takes an MSK master key and a set of AU attributes as input and
generates a secret key that identifies with that set. We also pose a function:
f : 0, 1 * → G0 maps any attribute described as a binary string to a random group
element. First, the algorithm selects random r ∈ Zp, then a random rj ∈ Zp for each
attribute j ∈ S. Then, it calculates the key as:
(α+r )
SK = (D = g β , D = gr .E2, ∀ j ∈ AU, D j = gr . f ( j)r j , D j = gr j ).
Decryptions
It is a recursive algorithm. To facilitate the calculation, we present the simple form

of decryption and improve it later.
First, we start to define the partial algorithm DecryNode (CT, SK, x) which takes
as input a ciphertext CT, a secret key SK which is associated with a set of attributes
AU, and with a node x of T the access shaft.
If node x is a leaf node, then we set i = attribute (x) and therefore
• If i ∈ AU , then
e(Di ,C x)
DecryNode(CT, SK, x) = e(Di ,C x)
= e(g, g)rq x(0)
• If i is not in AU, then DecryptNode (CT, SK, x) = ⊥.

After that, we move on to recursion. This amounts to saying that when x is a leaf
node the DecryNode (CT, SK, x) algorithm proceeds as follows: For all Z nodes
children of X, it calls DecryNode (CT, SK, z) and stores the output as as long as Fz.
Let Sx be an arbitrary set of child nodes of size kx such that Fz different from ⊥ (if
Fz equal to ⊥, then child node Fz is not satisfied then returns ⊥), so we calculate:

Fx = FZΔi ,S x(0)
, i = index(Z), S x = {index(Z ) : Z ∈ Sx}
r.q p ar ent (z)(index(z)) Δi ,S x(0)
Fx = (e(g, g) )
( Δi ,Sx (0)
Fx = e(g, g) r.qx (i))
Fx = e(g, g)(r.qx (0)) , (polynomialinterpolation)
4.4 Secure Data Storage and Sharing Scheme
In order to ensure effective access control of sensitive recording and protect patient
privacy, we offer a system based specifically on symmetric encrypt, CP-ABE encrypt
and blockchain.
Fig. 3 Authentication phase and key exchange between the device and Fog
Generation of the public key PKs: Initialization
The attributes authority (AA) is responsible for generating the public attribute
keys and transferring them to the Fog and/or doctors for later use if necessary. It
executes the Setup() algorithm.
Data Generation Phase
Figure 3 illustrates the different steps for authenticating and sharing the secure
key in order to obtain a secure channel to transfer data between the data generating
devices and the Fog.
1. First the device selects a secure random ad and calculates the value Rd = ad .G.
2. Then, the device signs an identity idd and encrypts the value Rd and the identity
idd by the public key of Fog P K F og. Then, it sends the information to the Fog:
E P K F og (idd ) E S K d evice (idd ) E P K F og (Rd ) E S K d evice (Rd ).
3. On its part, upon receipt of the message, the Fog decrypts and verifies the mes-
sage in order to obtain the information idd necessary for authentication and Rd
necessary for the calculation of the symmetric key. Then, it performs a signature
verification function.
4. If the signatures received are correct. Then, the device authenticated successfully.
The Fog in turn selects a secure random value a F and calculates R F = a F .G.
Finally, it calculates the symmetric common key SK = Rd .a F .
5. The Fog encrypts and signs the R F value and sends the message to the device:
E P K d evice (R F ) E S K F og (R F ).
6. The device decrypts the E P K d evice (R F ) message and verifies the validity of the
signature. Finally, it calculates the symmetric common key SK = R F .ad .
Note that the public parameters are as follows: - G the generator point. - The public
keys of Fog P K F og and device P K d ivice.
The secure channel is ready to transmit the data generated by the sensors.
Fig. 4 Data logging phase
Data logging phase
Healthcare devices run the E SK (M) algorithm by encrypting the data and sending
it to the Fogs. Once received, the latter executes the algorithm: Encry (M, T, PK) →
CT, it calculates the data identifier “idx” = Hash (CT) and transfers the ciphertext to
the storage provider where it is stored, and simultaneously, the proxy-Fog broadcasts
the transaction:
GenAutorization (idx, @ cl, @ gr, @pr).
The steps of this phase are shown in Fig. 4

1: The Fog calculates the CT = Encry (M, T, PK) and idx = HMAC (CT)
2: The Fog sends CT, idx to the cloud
3: At the same time, the Fog broadcasts
GenAutorization (idx, @ cl, @ gr, @pr).
Access authorization phase
Authorization is given by the data signature (Fog). Indeed, the Fog generates an
authorization (idx, @ cl, @ gr, @pr) which is used to authorize a group to access its
data in the cloud. If a user wants to view data, it defuses a transaction to the site to the
cloud which transmits it to the Fog. Then, the data owner checks the authorization
right of this user and broadcasts the
RequestAutorization (authorization, @ pr, @ rq).

Fig. 5 Authorization and data access phase
Access authorization phase
When a doctor receives authorization to access data, first of all, he/she authen-
ticates to the cloud with his/her professional card which defines his/her attributes.
If the authentication is successful, the attribute authority executes the GenerKeyS
(AU, MSK)→ SK attribute key generation algorithm. The output of this algorithm
is transferred to the requestor in a secure manner. Also, the requestor broadcasts a
DemAutorization transaction (authorization, @ rq, @ cl) to transfer to the cloud.
The storage service sends the requester the encrypted text that is identified by the
idx. It then broadcasts a DemAutorization transaction (authorization, @ cl, @ pr) in
order to inform the Fog that its data has been consulted.
The doctor uses his/her secret ABE key and retrieves the data in clear. Figure 5
summarizes the authorization and data access phase:
1. RequestAuthorization (authorization, @ pr, @ rq).

2. The doctor looks for his/her secret ABE key from the attribute authority.
3. The authority securely sends the secret key to the requester.
4. RequestAuthorization (authorization, @ rq, @ cl)
5. The consultant sends a request to consult the data.
6. The cloud sends the encrypted CT text to the doctor.
7. The cloud broadcasts an Ack D em Autori zation (authorization, @ cl, @ pr).
5 Security And Performance Analysis
In this section, we present the security and performance analysis of our new solution.
5.1 Security Analysis
Unlike traditional communication security and privacy protection, our cryptography

scheme ensures security and privacy. We presented a formal safety analysis and
validated it formally with the AVISPA simulator [19]. The symmetrical scheme
proves through AVISPA that it is safe. The security model is presented as follows:
Suppose we have a polynomial probabilistic adversary A can break our scheme
with a significant advantage AdvA = ε. We will show that we can, then, build a B
simulator that can solve the DBDH problem with a significant advantage.
Simulator B will use A to find the solution to the DBDH problem. For the demon-
stration of the security of the diagram, we estimate the advantage of simulator B.
If μ = 1, the ciphertext gives no information about ϒ. So P [ϒ =ϒ| μ = 1] = 21 .
The decision of B will be based on the result of A, if ϒ different from ϒ, then B
conclude that μ = 1 and if ϒ = ϒ, B will choose μ = 0.
When μ = 0, the advantage of A is ε. By definition P [ϒ = ϒ | μ = 0] = ε + 21 .
B selects μ = 0 and ϒ = ϒ then P [μ = μ | μ = 0] = ε + 21 . Finally, the general
advantage of B: AdvB = ε / 2. And since ε is assumed not to be negligible, then ε /
2 is also not negligible.
5.2 Performance Analysis
To analyze the performance of our solution, we implemented:

an anonymity proxy and attribute-based encryption by Fog computing and secure
data storage in the cloud.
A blockchain to achieve decentralized access control message management and
fine-grained access control is achieved through the encryption scheme based on
attributes conjugated by the Fog.
To check that our scheme achieves its objectives, we analyze the performance
of our implementation by modifying each time the number of attributes N =
{2, 3, 4, 6, 8}, N = {10, 20, 30, 40, 50, 60, 70} which are considered to be repre-
sentative for real-world ranges for attribute-based encryption.
Figure. 6 shows the total execution time (ABE encryption + Symmetric encryp-
tion). The ABE scheme encryption time is a function of the numbers of attributes
in the access structure. The results are considered good since the encryption time is
done slightly as the number of attributes increases, but at a certain level, the execution
time remains stable. On the other hand, the symmetric encryption time is considered
Fig. 6 Total execution time based on number of attributes on PC workstation
to be zero (between the device and the Fog server). Indeed this time has no effect on
the total execution time. This allows us to say that our proposal respects the real-time
constraint.
6 Conclusion
Through this article, we have proposed a solution to secure e-Health applications by

exchanging data confidentially and protecting patient privacy in an IoT–Fog–cloud
architecture. Our solution uses symmetric encryption and asymmetric encryption
(ABE) techniques. It integrates the blockchain in order to strengthen security at the
level of data access control management. In addition, our proposal ensures integrity
and keeps track of data sharing. In order to move to a fully distributed architecture,
we can integrate the smart contract for the execution of the encryption and decryption
algorithms, and we can also strengthen our model by using machine learning to secure
the cloud computing environment and detect the Man-In-The-Middle (MITM) attack
in a network of connected objects for an upcoming job.
References
1. https://apps.who.int/iris/handle/10665/331421
2. Li, X., Niu, J., Karuppiah, M., Kumari, S., Wu, F.: Secure and efficient two-factor user authen-
tication scheme with user anonymity for network based e-health care applications. J. Med.
Syst. 40(12), 268 (2016)
3. Anitha, G., Ismail, M., Lakshmanaprabu, S.K.: Identification and characterisation of choroidal
neovascularisation using e-Health data through an optimal classifier in Electronic Government.
Int. J. 16(1–2) (2020)
4. Wang, X., Bai, L., Yang, Q., Wang, L., Jiang, F.: A dual privacy-preservation scheme for
cloud-based eHealth systems. J. Inf. Secur. Appl. 132–138 (2019)
5. Bethencourt, J., Sahai, A., Waters, B.: Encryption, Ciphertext-Policy Attribute-Based: IEEE
Symposium on Security and Privacy (SP ’07), p. 2007. France, May, Berkeley (2007)
6. Zuo, C., Shao, J., Wei, G., Xie, M., Ji, M.: CCA-secure ABE with outsourced decryption for
fog computing. Future Gener. Comput. Syst. 78(2), 730–738 (January 2018)
7. Wang, H.: Anonymous data sharing scheme in public cloud and its application in E-health
record. In: IEEEaccess May 22, 2018, date of current version June 19 (2018)
8. Liu, Q., Zou, X.: Research on trust mechanism of cooperation innovation with big data pro-
cessing based on blockchain. EURASIP J. Wirel. Commun. Network. 2019, Article number:
26 (2019)
9. https://www.binance.vision/fr/blockchain/history-of-blockchain
10. Nakamoto, S.: Bitcoin : A peer-to-peer electronic cash system (2009). https://doi.org/10.1007/
11823285_121
11. Genestier, P., Letondeur, L., Zouarhi, S., Prola, A., Temerson, J.: Blockchains et smart contracts:
des perspectives pour lInternet des objets (IoT) et pour l’e-santé. Annales des Mines - Réalités
industrielles, août 2017(3), 70–73 (2017). http://orcid.org/10.3917/rindu1.173.0070
12. Gupta, Y., Shorey, R., Kulkarni, D., Tew, J.: The applicability of blockchain in the internet of
things. In: 2018 10th International Conference on Communication Systems Networks (COM-
SNETS), pages 561–564 (2018)
13. Nguyen, C., Pathirana, N., Ding, M., Seneviratne, A.: Blockchain for Secure EHRs Sharing of
Mobile Cloud Based E-Health Systems in ieeeAccess May 17, 2019, date of current version
June 4 (2019)
14. Yao, Y., Chang, X., Misić, J., Misić, V.B., Li, L.: BLA: Blockchain-Assisted Lightweight
Anonymous Authentication for Distributed Vehicular Fog Services. IEEE Internet Things J.
Citation information https://doi.org/10.1109/JIOT.2019.2892009
15. Zhang, A., Li, X.: Towards Secure and Privacy-Preserving Data Sharing in e-Health Systems
via Consortium Blockchain, Springer Science+Business Media, LLC, part of Springer Nature
(2018)
16. Dumont, R.: Cryptographie et Sécurité informatique. http://www.montefiore.ulg.ac.be/
~dumont/pdf/crypto.pdf
17. http://www.enseignement.polytechnique.fr/informatique/INF550/Cours1011/INF550-2010-
7-print.pdf
18. Zhang, P., Chen, Z., Liu, J.K., Kaitai, L., Hongwei, L.: An efficient access control scheme with
outsourcing capability and attribute update for fog computing. Future Gener. Comput. Syst.
78, 753–762 (2018)
19. The Avispa-Project http://www.avispa-project.org/
Forecasting COVID-19 Cases
in Morocco: A Deep Learning Approach
Mustapha Hankar, Marouane Birjali, and Abderrahim Beni-Hssane
Abstract The world is severely affected by the COVID-19 pandemic, caused by

the SARS-CoV-2 virus. So far, more than 108 million confirmed cases have been
recorded, and 2.3 million deaths (according to Statistica data platform). This has
created a calamitous situation around the world and fears that the disease will affect
everyone in future. Deep learning algorithms could be an effective solution to track
COVID-19, predict its growth, and design strategies and policies to manage its spread.
Our work applies a mathematical model to analyze and predict the propagation of
coronavirus in Morocco by using deep learning techniques applied on time series
data. In all tested models, long short-term memory (LSTM) model showed a better
performance on predicting daily confirmed cases. The forecasting is based on history
of daily confirmed cases recorded from March 2, 2020, the day the first case appeared
in Morocco, until February 10, 2020.
1 Introduction
In the last days of December 2019, the novel coronavirus, of an unknown origin, first
time appeared in Wuhan, a province in China. Health officials are still tracing the
exact source of this new virus; early hypotheses thought it may be linked to a seafood
market in Wuhan [1]. After then, it was noticed that some people who visited the
market have developed viral pneumonia caused by the new coronavirus [2]. A study
that came out on January 25, 2020, notes that the individual with the first reported
case became ill on December 1, 2019, and had no link to the seafood market [3].
Investigations are ongoing as to how this virus originated and spread. It appears after
the person has been exposed to the virus for the first time that many symptoms are
showing up within 14 days of the first exposure to the virus, including fever, dry
cough, fatigue, breathing difficulties, and loss of smell and taste.
M. Hankar · M. Birjali (B) · A. Beni-Hssane

LAROSERI Laboratory, Computer Science Department, Faculty of Sciences, University of
Chouaib Doukkali, El Jadida, Morocco
https://doi.org/10.1007/978-981-16-3637-0_59
846 M. Hankar et al.
COVID-19 mainly spreads through the air when people are close to each other
long enough, primarily via small droplets or aerosols, as an infected person breathes,
coughs, sneezes, or speaks [4]. In some cases, people who do not show any symptoms,
or asymptomatic patients, remain infectious to others with a transmission rate equal
to that of symptomatic people [5].
Amid a pandemic that has taken many lives so far, threatens the lives of others
in the world, we are obligated to act, as researchers in machine learning and its
real-world applications, which COVID-19 is one of the biggest actual challenges,
in order to collaborate in the solution process. Machine learning algorithms can be
deployed very effectively to track coronavirus disease and predict epidemic growth.
This could help decision makers to design strategies and policies to manage its spread.
In this work, we built a mathematical model to analyze and predict the growth of this
pandemic. A deep learning model using feedforward LSTM neural network has been
applied to predict COVID-19 cases in Morocco on time series data. The proposed
model based its predictions on the history of daily confirmed cases, as a training
phase, which have been recorded from the start of the pandemic in March 2, 2020,
to February 20, 2020.
After training LSTM model on time series data, we tested it within a period of
60 days to assess the accuracy of the model and compared the obtained results with
other applied models such as auto-regressive integrated moving averages (Auto-
ARIMA), K-nearest neighbor (KNN) regressor, random forest regressor (RFR), and
Prophet.
2 Related Works
Recently, deep learning techniques have been serving the medical industry [6, 7],
bringing with them the new technology and its revolutionary solutions that are
changing the shape of health care. Deep learning provides the healthcare industry
with the ability to analyze large datasets at exceptional speeds and make accurate
model.
Fang et al. [8] investigated the effect of early recommended or mandatory
measures on reducing the crowd infection percentage, using a crowd flow model.
Hu et al. [9] developed a modified stacked auto-encoder for modeling the trans-
mission dynamics of the epidemics. Using this framework, they forecasted the cumu-
lative confirmed cases of COVID-19 across China from January 20, 2020, to April
20, 2020.
Roosa et al. [10] used phenomenological models that have been validated during
previous outbreaks to generate and assess short-term forecasts of the cumulative
number of confirmed reported cases in Hubei Province, the epicenter of the epidemic,
and for the overall trajectory in China, excluding the province of Hubei. They
collected daily report of cumulative confirmed cases for the 2019-nCoV outbreak
for each Chinese province from the National Health Commission of China. They
Forecasting COVID-19 Cases in Morocco: A Deep Learning Approach 847
provided 5, 10, and 15 days forecasts for five consecutive days, with quantified
uncertainty based on a generalized logistic model.
Liu and colleagues [11] used early reported case data and built a model to predict
the cumulative COVID-19 cases in China. The key features of their model are the
timing of implementation of major public policies restricting social movement, the
identification and isolation of unreported cases, and the impact of asymptomatic
infectious cases.
In [12], Peng et al. analyzed the COVID-19 epidemic in China using dynamical
modeling. Using the public data of National Health Commission of China from
January 20th to February 9th, 2020, they estimated key epidemic parameters and
made predictions on the inflection point and possible ending time for 5 different
regions.
In [13], Remuzzi analyzed the COVID-19 situation in Italy and mentioned if the
Italian outbreak follows a similar trend as in Hubei Province, China, the number
of newly infected patients could start to decrease within 3–4 days, departing from
the exponential trend, but stated this cannot currently be predicted because of differ-
ences between social distancing measures and the capacity to quickly build dedicated
facilities in China.
In [14], Ayyoubzadeh et al. implemented linear regression and LSTM models
to predict the number of COVID-19 cases. They used tenfold cross-validation for
evaluation, and root-mean-squared error (RMSE) was used as the performance
metric.
In [15], Canadian researchers developed a forecasting model to predict COVID-19
outbreak using state-of-the-art deep learning models such as LSTM. They evaluated
the key features to predict the trends and possible stopping time of the current COVID-
19 pandemic in Canada and around the world.
3 Data Description
On March 11, 2020, the World Health Organization (WHO) declared COVID-19 as
a pandemic, pointing to over 118,000 confirmed cases of coronavirus in over 110
countries and territories around the world at that time. The data used in this study was
collected by many sources including the World Health Organization, Worldometers,
and Johns Hopkins University, sourced from data delivered by the Moroccan Ministry
of Health. The dataset is in a CSV format taken from the link: https://github.com/dat
asets/covid-19. It is maintained by the team at Johns Hopkins University Center for
Systems Science and Engineering (CSSE) who have been doing a great public service
from an early point by collecting data from around the world. They have cleaned and
normalized data and made it easy for further processing and analysis, arranging dates
and consolidating several files into normalized time series. The dataset is located in
the data folder in a CSV file format. The team has been recording and updating all
the daily cases in the world since January 22, 2020.
Fig. 1 Daily cases over time
The file contains six columns: cumulative confirmed cases, cumulative fatali-
ties, dates of recording these cases, recovered cases, region/country, and finally
province/state. Since we are working on Moroccan data, we filtered it based on
country column to get the cases recorded in Morocco from March 2, 2020, to February
10, 2021. Since we are interested in daily confirmed cases only, which is not found
in the dataset, we had to code a Python script to compute confirmed cases per day
indexed by dates and then feed it to the algorithms.
As mentioned above, we transformed the original given data into univariate time
series format of recorded confirmed cases. The values of the single-columned data
frame are the number of cases per day, indexed by date/time. The plotting of daily
cases in the dataset is shown in Fig. 1.
It is noticeable from the plot above (Fig. 1), COVID-19 daily cases are likely stabi-
lized by the beginning of March with a small margin because of the strict measures
taken by the health authorities. By the end of July, as the authorities start easing these
measures, the cases began to increase exponentially this time due to the increase in
population movements and travels during summer. After November 12, the cases
started obviously to decrease, which could be a result of suppressing the virus by
taken measures, or tested cases have obviously declined.
4 LSTM Model for Forecasting
Hochreiter and Sherstinsky [16, 17] published a theoretical and experimental works
on the subject of LSTM networks and reported astounding results across a wide
variety of application domains, especially on a sequential data. The impact of the
LSTM network has been observable in natural language processing domains, like
speech-to-text transcription, machine translation, and other applications [18].
Fig. 2 LSTM cell architecture
LSTM is the type of recurrent neural networks (RNNs) that have feedback looping,
meaning they are able to maintain information over time. They can process not only
single data points, but also entire sequences of data such as speech or video, much
applicable in time series data [19]. A LSTM unit is composed of a cell, an input gate,
an output gate, and a forget gate (Fig. 2). The cell remembers values over arbitrary
time intervals, and the three gates regulate the flow of information into and out of
the cell [16]. All RNNs keep information in their memory over short period of time,
because the gradient of its loss function fades exponentially [20]. Therefore, it could
be difficult to train standard RNNs to solve problems that require learning long-term
temporal dependencies like time series data. LSTM is a designed RNN architecture
to address the vanishing gradient problem [21]. The reason we chose to implement
this method is that LSTM units include a memory cell that can maintain information
in memory for long period of time. A set of gates is used to control when information
enters the memory, when it is output, and when it is forgotten.
In the equations below, the variables represent vectors. Matrices W q {\displaystyle
W_{q}} and U q {\displaystyle U_{q}} contain, respectively, the weights of the
input and recurrent connections, where the subscript q {\displaystyle _{q}} can
either be the input gate i {\displaystyle i}, output gate o{\displaystyle o}, the forget
gate f {\displaystyle f}, or the memory cell c{\displaystyle c}, depending on the
activation function being calculated. In this section, we are using a “vector notation.”
The equations for the forward pass of a LSTM unit with a forget gate are defined
[22]:

f t = σg W f xt + U f h t−1 + b f (1)
i t = σg (Wi xt + Ui h t−1 + bi ) (2)
ot = σg (Wo xt + Uo h t−1 + bo ) (3)

c̃t = σc (Wc xt + Uc h t−1 + bc ) (4)
ct = f t ◦ ct−1 + i t ◦ c̃t (5)
h t = ot ◦ σh (ct ) (6)
• Activation functions used:
σg : sigmoid function.
σc : hyperbolic tangent function.
σh : hyperbolic tangent function.
• Variables used:
xt ∈ Rd : input vector to the LSTM unit

f t ∈ Rh : forget gates activation vector
i t ∈ Rh : input/update gates activation vector
ot ∈ Rh : output gates activation vector
h t ∈ Rh : hidden state vector also known as output vector of the LSTM unit
c̃t ∈ Rh : cell input activation vector
ct ∈ Rh : cell state vector
W ∈ Rh×d , U ∈ Rh×h and b ∈ Rh : weight matrices and bias
vector parameters which need to be learned during training
RNN architectures using LSTM cells can be trained in a supervised way on

training sequences, using an optimization algorithm, like gradient descent, combined
with backpropagation through time to compute the gradients needed during the opti-
mization process, in order to change each weight of the LSTM network in proportion
to the derivative of the error (at the output layer of the LSTM network) with respect
to corresponding weight [23]. Figure 3 shows the structure of a neural network using
LSTM units.
The reason we proposed to use LSTM neural network goes to the nature of data.
Since we are dealing with COVID-19 cases as time series values, we prioritized
implementing this method over other techniques such as random forest regressor
(RFR), which is rarely applied on times series data. On the other side, LSTM model
showed, among other models, hopeful results and performance in predicting based on
two essential metrics: root-mean-squared-error (RMSE) metric and mean absolute
percentage error (MAPE) (Fig. 4).
Before getting to modeling section, it is a common practice to separate the avail-
able data into two main portions: training and test data (or validation data), where
Fig. 3 Feedforward LSTM network structure
Fig. 4 Data pipeline
the training data is used to estimate the parameters of a forecasting method and the
test data is used to evaluate its accuracy and estimate the loss function. Because the
test data is not used in determining the forecasts, it should provide a reliable indi-
cation of how the model will likely forecast on new data. After splitting the data,
we standardize the values with a MinMax scaler and then reshape the inputs in the
right shape. In time series problem, we predict a future value in a time T based on a
period of time T –N with T is the number time steps to be chosen as hyperparameter.
We obtained good results by taking N = 60 days. Thus, the training inputs have to
be in a three-dimensional shape (training inputs, time steps, and number of features)
before beginning the training.
LSTM network is set to be trained over 300 epochs on more than 80% of the
dataset and tested on a period of 60 days (20% of the dataset). The screenshot below,
taken from the code source, shows the architecture of the trained feedforward LSTM
network (Fig. 5).
The number of hidden layers, the dropout rate, and the optimization method to
minimize the errors are essential hyperparameters to fine-tune in order to achieve
hopeful results and performance of a deep learning model. In our case, the model
contains three LSTM layers with a dropout rate of 0.4 each, dense layer to output
Fig. 5 LSTM model architecture summary
the forecasting results, and the “adam” optimizer given its best results compared to
“rmsprop,” for example.
The part of evaluating the model is a deductive part in our work. Therefore, the choice
of a metric to evaluate the model matters and gives an insight about its performance
on testing data and how the model will perform on new data. The dataset contains 294
records. We left a portion of 80% for training the model and 20% to test it. Since we
used the metrics to compare the performance of an LSTM model with other models,
we evaluated the models by two common methods.
5.1 Root-Mean-Squared Errors
RMSE is a method proposed in 2005 by the statistician Rob J. Hyndman [24] as

the measure of a forecast accuracy. RMSE metric computes the residuals between
predicted values and observed knowing that a forecast error is simply defined by the
equation: et = y − ŷ i , where y is the true value and ( ŷ) is the predicted value.
Accuracy measures that are based only on the error et are, therefore, scale
dependent and cannot be used to make comparisons between series that involve
different units. RMSE method is one of the two most commonly used scale-dependent
measures. It is resulted by the formula:

N
(yi − ŷi )2
RMSE = i=1
N
5.2 Mean Absolute Percentage Error
MAPE is a measure of prediction accuracy of a forecasting method in statistics,

such as time series data in our case, also used as a loss function for regression prob-
lems in machine learning [24]. Percentage errors have the advantage of being scale
independent and frequently used to compare forecast performance across different
datasets. MAPE metric usually expresses the accuracy as a ratio defined by the
formula:
n
1 Yt − Ft
MAPE = ,
n t=1 Yt
where Y t is the actual value and F t is the forecast value. MAPE is also sometimes
reported as a percentage, which is the above equation multiplied by 100 making it
a percentage error: the absolute difference between Y t and F t divided by the actual
value Y t summed for every forecasted point in time and divided by the number of
fitted points n.
Considering the size of the dataset, which is likely small in this case, the model
took an estimated time of 406 s in training over 300 epochs. As we can see in Fig. 6,
the loss function began to minimize the errors in the first fifty epochs, after then the
Fig. 6 Loss function plot over 300 epochs

Table 1 Forecast errors for

Model RMSE MAPE (%)
the tested models
LSTM 357.90 29.31
Prophet 412.022 37.01
Auto-ARIMA 1699.47 215.87
Random forest regressor 977.53 83.41
cost function slowly decreases to the end. As the loss function decreases, the model
on the other way increases its accuracy leading us to get a better outcome.
The results showed a better performance of LSTM model compared to other
models like Prophet (Facebook forecasting algorithm), Auto-ARIMA, and random
forest regressor. Table 1 shows the comparative performance of the four tested models
based on two metrics.
Based on the results above, we chose to forecast COVID-19 daily cases on testing
data using LSTM model (357.90 of RMSE), which outperforms other models by
minimizing the loss function. When compared to bidirectional LSTM neural network
architecture, the results of the latter were much closer to feedforward LSTM model
than other models. In Fig. 6, we plot the whole data frame segmented into training
set (more than 80% of the dataset) and testing set (almost 20%) to see how the
model performs versus actual COVID-19 cases. As the chart shows, LSTM model
accuracy did not reach the best wanted results, but it is very obvious that the model
recognizes the trend within data and learned the overall pattern from the previous
cases of training set. We also noticed that the performance of LSTM model increases
when adding more data to the dataset. Meanwhile, RFR model and Auto-ARIMA
model performances diminish.
To compare the presented forecasting results from the graph above, we tested other
models on the same test set; Fig. 7 illustrates the predictions of Prophet model, RFR
model, and Auto-ARIMA model compared to the performance of LSTM model. It
Fig. 7 LSTM predictions

Fig. 8 Comparing predictions of tested models with the actual cases
is observed that Prophet’s performance is more likely to learn trend from data than
RFR and Auto-ARIMA models. The results of the latter are the worst among all
models which is shown in Table 1 and Fig. 8.
The LSTM model showed a good performance in the training phase, because the
loss function was at its lowest level by increasing the number of epochs to 300. This
may lead to overfitting due to small amount of training data. However, a low training
error might indicate that LSTM model can extract the pattern from data, which is
obvious in our predictions (Fig. 6). Therefore, we certainly assume that the model
could lead to better results if we have more training data. We also notice that the
same proposed model showed good results on Bitcoin time series data predicting the
daily stock prices, since we trained it on a sufficient amount of data. And yet, despite
the small size of the dataset, LSTM model outperformed other models on this task
(Fig. 9).
We assumed before getting into this study that the data provided till date is not
big enough to train the model, meaning that our findings will not be at the very
good level, but remain hopeful showing at least the trending behavior of how the
coronavirus spreads over time, which is a helper factor to anticipate the future of its
growth and give insights to health officials leading them to take actions and slow
down the propagation of the virus, preventing vulnerable people from unbearable
consequences. Due to measures taken during quarantine for more than three months,
the curve of COVID-19 cases was likely stable and the virus propagation was almost
controllable, but shutting down the economy and holding people in their places are
not the ultimate solutions. It could actually be the problem itself.
Fig. 9 Comparing predictions of the models with daily cases [truncated chart]
6 Conclusion
Considering the serious situation of recording thousands of COVID-19 daily cases

in Morocco lately, an early prediction and anticipation of the virus transmission
could help decision makers to take preventive actions to slow down its growth. This
chapter is a contribution to solve this problem by implementing machine learning and
statistical models. The results show that LSTM model yielded a hopeful accuracy
score and a minimum root-mean-squared error.
References
1. Zhu, N., Zhang, D., Wang, W., Li, X., Yang, B., Song, J., Zhao, X., Huang, B., Shi, W., Lu, R.,
Niu, P., Zhan, F., Ma, X., Wang, D., Xu, W., Wu, G., Gao, G.F., Tan, W.: A novel coronavirus
from patients with pneumonia in China, 2019. N. Engl. J. Med. (2020). https://doi.org/10.1056/
nejmoa2001017
2. Zu, Z.Y., Di Jiang, M., Xu, P.P., Chen, W., Ni, Q.Q., Lu, G.M., Zhang, L.J.: Coronavirus disease
2019 (COVID-19): a perspective from China. Radiology (2020). https://doi.org/10.1148/rad
iol.2020200490
3. Huang, C., Wang, Y., Li, X., Ren, L., Zhao, J., Hu, Y., Zhang, L., Fan, G., Xu, J., Gu, X.,
Cheng, Z., Yu, T., Xia, J., Wei, Y., Wu, W., Xie, X., Yin, W., Li, H., Liu, M., Xiao, Y., Gao,
H., Guo, L., Xie, J., Wang, G., Jiang, R., Gao, Z., Jin, Q., Wang, J., Cao, B.: Clinical features
of patients infected with 2019 novel coronavirus in Wuhan, China. Lancet (2020). https://doi.
org/10.1016/S0140-6736(20)30183-5
4. Karia, R., Gupta, I., Khandait, H., Yadav, A., Yadav, A.: COVID-19 and its modes of
transmission. SN Compr. Clin. Med. (2020). https://doi.org/10.1007/s42399-020-00498-4
5. Oran, D.P., Topol, E.J.: Prevalence of asymptomatic SARS-CoV-2 infection: a narrative review.
Ann. Intern. Med. (2020). https://doi.org/10.7326/M20-3012
6. Alhussein, M., Muhammad, G.: Voice pathology detection using deep learning on mobile
healthcare framework. IEEE Access (2018). https://doi.org/10.1109/ACCESS.2018.2856238
7. Yuan, W., Li, C., Guan, D., Han, G., Khattak, A.M.: Socialized healthcare service recommen-
dation using deep learning. Neural Comput. Appl. (2018). https://doi.org/10.1007/s00521-018-
3394-4
8. Fang, Z., Huang, Z., Li, X., Zhang, J., Lv, W., Zhuang, L., Xu, X., Huang, N.: How many infec-
tions of COVID-19 there will be in the “Diamond Princess” predicted by a virus transmission
model based on the simulation of crowd flow. ArXiv (2020)
9. Hu, Z., Ge, Q., Li, S., Jin, L., Xiong, M.: Artificial intelligence forecasting of COVID-19 in
China. ArXiv (2020)
10. Roosa, K., Lee, Y., Luo, R., Kirpich, A., Rothenberg, R., Hyman, J.M., Yan, P., Chowell, G.:
Real-time forecasts of the COVID-19 epidemic in China from February 5th to February 24th,
2020. Infect. Dis. Model. (2020). https://doi.org/10.1016/j.idm.2020.02.002
11. Liu, Z., Magal, P., Seydi, O., Webb, G.: Predicting the cumulative number of cases for the
COVID-19 epidemic in China from early data. Math. Biosci. Eng. (2020). https://doi.org/10.
3934/MBE.2020172
12. Peng, L., Yang, W., Zhang, D., Zhuge, C., Hong, L.: Epidemic analysis of COVID-19 in China
by dynamical modeling. ArXiv (2020). https://doi.org/10.1101/2020.02.16.20023465
13. Remuzzi, A., Remuzzi, G.: COVID-19 and Italy: what next? Lancet (2020). https://doi.org/10.
1016/S0140-6736(20)30627-9
14. Sajadi, M.M., Habibzadeh, P., Vintzileos, A., Shokouhi, S., Miralles-Wilhelm, F., Amoroso, A.:
Temperature and latitude analysis to predict potential spread and seasonality for COVID-19.
SSRN Electron. J. (2020). https://doi.org/10.2139/ssrn.3550308
15. Chimmula, V.K.R., Zhang, L.: Time series forecasting of COVID-19 transmission in Canada
using LSTM networks. Chaos Solitons Fractals (2020). https://doi.org/10.1016/j.chaos.2020.
109864
16. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. (1997). https://doi.
org/10.1162/neco.1997.9.8.1735
17. Sherstinsky, A.: Fundamentals of recurrent neural network (RNN) and long short-term memory
(LSTM) network. Phys. D Nonlinear Phenom. (2020). https://doi.org/10.1016/j.physd.2019.
132306
18. Lin, H.W., Tegmark, M.: Critical behavior in physics and probabilistic formal languages.
Entropy (2017). https://doi.org/10.3390/e19070299
19. Karevan, Z., Suykens, J.A.K.: Transductive LSTM for time-series prediction: an application
to weather forecasting. Neural Netw. (2020). https://doi.org/10.1016/j.neunet.2019.12.030
20. Bai, S., Kolter, J.Z., Koltun, V.: An empirical evaluation of generic convolutional and recurrent
networks for sequence modeling. ArXiv (2018)
21. Graves, A., Schmidhuber, J.: Framewise phoneme classification with bidirectional LSTM and
other neural network architectures. Neural Netw. (2005). https://doi.org/10.1016/j.neunet.2005.
06.042
22. Gers, F.A., Schmidhuber, J., Cummins, F.: Learning to forget: continual prediction with LSTM.
Neural Comput. (2000). https://doi.org/10.1162/089976600300015015
23. Kolen, J.F., Kremer, S.C.: Gradient flow in recurrent nets: the difficulty of learning long term
dependencies. In: A Field Guide to Dynamical Recurrent Networks (2010). https://doi.org/10.
1109/9780470544037.ch14
24. Hyndman, R.J., Koehler, A.B.: Another look at measures of forecast accuracy. Int. J. Forecast.
(2006). https://doi.org/10.1016/j.ijforecast.2006.03.001
The Impact of COVID-19 on Parkinson’s
Disease Patients from Social Networks
Hanane Grissette and El Habib Nfaoui
Abstract Multiple examples of unsafe and incorrect treatment recommendations

shared everyday. The challenge, however, to provide efficient, credible, and quick
relevant access to reliable insight. Providing computational tools for Parkinson’s
Disease (PD) using a set of data-objects that contain medical information is very
desirable for alleviating the symptoms that can help to discover the risk of this
disease at an early stage. In this paper, we propose an automatic CNN-clustering
aspect-based identification method for drug mentions, events, treatments from daily
PD narratives digests. Therefore, a BiLSTM-based Parkinson classifier is developed
regarding both varied emotional states and common senses reasoning, which further
used to seek the impactful COVID-19 insights. The embedding strategy character-
ized polar facts through concept-level distributed biomedical representation associ-
ated with real-world entities, which are operated to quantifying the emotional state
of the speaker context in which aspect are extracted. We conduct comparisons with
neural networks state-of-art algorithms and biomedical distributed systems. Finally,
as a result, the classifier achieves an accuracy of 85.3%, and facets of this study
may used in many health-related concerns such as: Analyzing change in health sta-
tus, unexpected situations or medical conditions, and outcome or effectiveness of a
treatment.
1 Introduction
Having multiple voices who can relate to a similar situation, or who have experienced
similar circumstances, always garner greater persuasion than that of a single brand.1
Understanding emotions is the full-stack study that aims at recognizing, interpret-
ing, processing, and simulating human emotions and affects. Nowadays, affective
1 https://www.pwc.com/us/en/industries/health-industries/library/health-care-social-media.html.
H. Grissette (B) · E. H. Nfaoui

LISAC Laboratory, Faculty of Sciences, Sidi Mohamed Ben Abdellah University, FEZ, Morocco
e-mail: hanane.grissette@usmba.ac.ma
https://doi.org/10.1007/978-981-16-3637-0_60
860 H. Grissette and E. H. Nfaoui
computing (AC) and sentiment analysis (SA) have been considered as significant
emerging approach used to discriminate fine grain information regarding the emo-
tional state of patients. Indeed, instead of just detecting the polarity of given document
[1], they are used to interpret the emotional state of patients, detect misunderstanding
of drug related-information or side effects, and ensure a competitive edge to better
understanding patients’ experiences in a given condition [2].
Parkinson’s Disease (PD) is a condition that can qualify a person for social security
disability benefits. It is second common emotions-related disorders that affect an esti-
mated 7–10 millions people and families worldwide. Few works have been provided
to distil sentiment conveyed towards a drug/treatments on social networks whereby
distinguish impactful facts degrees regarding PD-related drug-aspects. In previous
work [3], authors proved the ability of based-neural network model to probe what
kind of based-treatment target may result in enhanced model performance by detect-
ing genuine sentiment in polar facts. Indeed, numerous of serious factors increase
the failure rate in detecting har m f ul and non − har m f ul patients’ notes regarding
related-medication targets [1]. Noticeably, many of them may fail to retrieve the
correct impact due to the inability to define complex medical components in text.
Each post may cite or refer to a drug reaction or/and misuse, which may lead
to be categorized to harmful impact or beneficial reaction, where beneficial adverse
reactions widely detected as harmful components [4]. This study is an affective
distillation regarding Parkinson’s disease related drug-targets, which further may
used for fine-tuning tasks of many health-related concerns such as defining change
in health status or unexpected situations/medical conditions, and monitoring out-
come or effectiveness of a treatment. In addition, it is a powered neural network
model for detecting polar medical statements, which further investigate what kind
of based-treatment target may result in improved emotion parkinson’s model per-
formance. Technically, it consists of mining and personalizing various changeable
emotional state toward specific objects/subjects regarding various aspect of PD’s
patients, which further used to track the impact of social media messages in daily
PD patients’ lives regarding given aspects and related-medical contexts such the case
of COVID’19 pandemic. At this and (1) Firstly, we investigate a based-phrase tran-
sition method between social media messages and formal description on medical
anthologies such as MedLine, life science journals, and online data systems. Then,
(2) An automatic CNN-clustering-based model regarding PD-aspects for given tar-
gets, e.g., drugs mentions, events, physical treatments, or healthcare organizations
at large. Therefore, (3) a BiLSTM-based emotional parkinson classifier is devel-
oped for the evaluation fine-tuning tasks. The main contributions of this paper can
be summarized as follow: First, a embedding-based conceptualization that relies on
various sub-steps such as medical concept transition-based normalization, the later
is proposed for disambiguating medical-related expressions and concepts. Second,
an affective distinction knowledge and common senses reasoning for improved sen-
timent inference.
The rest of paper is organized as follows: Sect. 2 briefly overviews of senti-
ment analysis and affective computing related works regarding healthcare context.
Section 3 introduces the proposed method and knowledge base that we extended in
The Impact of COVID-19 on Parkinson’s Disease Patients … 861
this paper, and describe the whole architecture. In Sect 4, experimental results are
presented and discussed. Finally, Sect. 5 concludes the paper and presents future
perspectives.
2 Sentiment Analysis and Modelling Language
2.1 Sentiment Analysis and Affective Computing
In a broader scope, sentiment analysis (SA) and affective computing (AC) are allow-
ing the investigation and comprehension of the relation between human emotions
and health services as well as application of assistive and useful technologies in
the medical domain. Recognizing and modeling patient perception by extracting
affective information from related-medical text is of critical importance. Due to the
diversity and complexity of human language, it is been mandatory to prepare tax-
onomy or ontology to capture concepts with various granularities in every domain.
First initiatives in this era was: providing a knowledge-based methods that aim at
building vocabulary and understanding language used to describe medication-related
experiences, drugs issues and other related-therapies topics.
From the literature, efficient methods assume the existing of annotated lexicons
regarding various aspects of analysis. Indeed, many lexicons have been annotated
in term of sentiment for both public and depend-domain. they differ in annotation
schemata: multinomial values (e.g., surprised, fear, joy) or continuous value as sen-
timent quantification that means extract the positiveness or negativeness parameters
of probabilistic model or generative model as well. Existing large scale knowledge
bases including Freebase [5], SenticNet [6], and Probase [7]. Most prior studies
focused on exploring an existing or customized lexicons regarding depend-context,
such as medical and pharmaceutical context.
Typically, neural network brought many success to enhancing these corpora capa-
bilities, for example [6] proposed a SenticNet sub-symbolic and symbolic AI that
automatically discover conceptual primitives from text and link them to common
sense concepts and named entities in a new three-level knowledge representation for
sentiment analysis. Other very related work is [8] that provide for affective common
sense knowledge acquisition for sentiment analysis. Sentiment analysis has been
known as a novel that allows a new form of sentiment-annotation regarding vari-
ous aspects of analysis such as attention, motivation and emotions, namely aspect-
based sentiment analysis (AbSA). Existing AbSA classifiers do not meet medical
requirements. However, various lexicons and vocabularies are defined to identify
very complicated medication concepts, e.g., adverse drug reactions (ADRs) or drug
descriptions. For example, authors in [9] present deep neural network (DNN) model
that utilizes the chemical, biological, and biomedical information of drugs to detect
ADRs. Most of existing models aimed to fulfil two main purposes: (i) identify-
ing the potential ADRs of drugs, (ii) defining online criteria/characteristics of drug
reactions, and (ii) predicting the possible/unknown ADRs of a drug. Traditional

approaches widely used medical concept extraction system such as ADRMine [5]
that uses conditional random fields (CRFs), they used a variety of features, including
a novel feature for modeling words’ semantic similarities.
Parkinson’s Disease Social Analysis Background Reports pour daily into health-
care communities and micro-blogs at a staggering rate range from drug uses, side
effects of some treatments to potential adverse drug reactions. Parkinson’s disease
(PD) is the second most important age-related disorder, after Alzheimer’s disease,
with a prevalence ranging from 41 per 100,000 in the fourth decade of life to over 1900
per 100,000 in people over 80 years of age. Emotional dysregulation is an essential
dimension that may occur in several psychiatric and neurologic disorders. In most
cases, it has been focused on clinical characteristics of emotional state variations
in bipolar disorder in Parkinson’s disease [10]. In both pathologies, the emotional
intensity variability involves important diagnostic and therapeutic issues.
Few data mining and natural language processing techniques have been proposed
in the PD context, and less efficient. Nowadays, advanced machine learning get
researchers and professionals’ attention and result in good results in quantifying
short-term dynamics of Parkinson’s Disease using self-reported symptom data from
social network. Where [11] proved the power of machine learning in a Parkinson’s
Disease digital biomarker dataset construction (NNC) methodology that discrim-
inates patient motor status. Other contributions used ensemble techniques to effi-
ciently enhance models performance. For example, [12] proposed hybrid intelligent
system for the prediction of PD progression using noise removal, clustering and
prediction methods, an adaptive neuro-fuzzy inference system (ANFIS) and support
vector regression (SVR) for prediction of PD progression. This study aims to perform
an efficient neural networks-solution that attempts a public concerns and impact of
COVID-19 pandemic on Parkinson’s disease patients.
2.2 Medical Conceptualization
Patients and health consumers are storming intentionally their medications’ experi-
ences and their related-treatment opinions that describe all the incredibly complex
processes happening in real-time treatment in a given condition. Patients self-reports
on social networks frequently capture varied elements ranging from medical issues,
product accessibility issues to potential side effects. Deep learning based neural net-
works have widely attracted many researchers’ attention on normalization, matching
and classification tasks by exploiting their ability to learn more under distributed
representations.
Embedding approaches are the most accurate methods that used for constructing
vector representations of words and documents [13]. The problem is that these algo-
rithms got in low-medical entities recognition recall which further require interven-
tion of both formal external medical knowledge and real-world examples to learn nat-
ural medical concepts patterns. Recently, researches paid great attention to conquer
these limitations by many alternatives: (i) extending embeddings on large amounts of

data or depend-domain corpora, namely semisupervised feature-based methods [14],
(ii) fusing weighted distributed features, (iii) manifold regularization for enhancing
the performance of sentiment classification of medication-related online examples. In
this study, we aim at imitating and retrieving the corresponding of real-word related
drug entities in the formal medical ontologies by adopting a distributed representa-
tion of two databases PubMed and MIMIC III Clinical notes as described in Table 1.
To summarize, we aim at developing an approach that takes unstructured data as
input, constructing the embedding matrix by incorporating biomedical vocabulary,
and discriminating drug reaction multi-word expressions.
3 Proposed Approach
In this section, we will introduce the proposed methodology that consists of medical
conceptualization and affective-aspect analysis model.
3.1 Text Vectorization and Embedding
The exploitation of such overwhelming unstructured, rich, and high-dimensional data

on social media from patients self-reports is of critical importance. Technically, each
post encoded as vector of the individual document to be input to the neural network
model. Moreover, medication-related text has a wide variety of medical entities and
medical facts, it requires a well-encoding approach to successfully leverage the real
meaning and conveyed sentiment toward a given medical aspects such as drug entity,
treatment, condition...etc.
However, the situation is complex, related-medication text depends on various
aspects. The popular SA approaches ignore many types of information, e.g., sexual
and racial information for medical oriented-analysis that certainly affects the emo-
tion computing process, but using the same information for personalizing patient’s
behavior is not. Moreover, patient self-reports contains various related-medical com-
ponents range from drug/treatment entities, adverse drug events/reactions to potential
unsuspected adverse drug reaction [15]. The existing techniques for both sentiment
analysis and affective computing are not able to efficiently extract those concepts and
topics due to limited resources in this case. Otherwise, probabilistic models could
be used to extract low-dimensional topics from document collections, such models
without any human knowledge often produce topics that are not interpretable. To
address these problems, we approach this step by incorporating real-world knowl-
edge encoded by entities embedding to automatically learn medical representation
from an external biomedical corpora EU-ADR [16] and ADRMINE [5] Lexicon in a
unified model. As depicted in Fig. 2, incorporating depend-domain medical knowl-
edge support two main functionalities in this stage:
• Stage (1): Build an embedding representation: since biomedical knowledge is

multi-relational data, we seek to represent knowledge as triple fact, e.g., (Drug
name, disease, association). We aim at incorporating medical knowledge into
embedding vectors space. Indeed, for this purpose, we combine two annotated
corpora, (1) EU-ADR corpus, The EU-ADR corpus has been annotated for drugs,
disorders, genes and their inter-relationships. Moreover, for each of the drug–
disorder, drug–target, and target–disorder relations three experts have annotated
a set of 100 abstracts, (2) ADRMINE Corpus is results of supervised sequence
labeling CRF classifier is established that extracts mentions of ADR and indica-
tions from inputs, adverse drug reactions (ADRs) training data consisted of 15,717
annotated tweets. Indeed, each annotation includes cluster ID, semantic type (i.e.,
ADR, indication, Drug interaction, Beneficial effect, other), drug name and corre-
sponding UMLS ID that means ADR Lexicon compile an exhaustive list of ADR
concepts and their corresponding UMLS IDs. Moreover, it includes concepts from
SIDER, a subset of CHV (Consumer health vocabulary) and COSTART (The Cod-
ing Symbols for a Thesaurus of Adverse Reaction Terms).
• Stage (2): Define a distance measure for unknown entities. In fact, we want to
rely on a measurement to be internal semantic representation for stream data and
map to some previously sentiment-polarized vectors. We would be able to identify
exact duplicates, making practically online rich information useful for tracking
new issues, thoughts and even probable diseases.
The incorporating knowledge technique is inspired by previous work [17], this latter
aim at defining fact-oriented knowledge graph embedding that automatically captures
relations between entities in knowledge base. However, this knowledge space is
embedded into low-dimensional continuous vector space while new properties have
been created and preserved. Generally, each entity is treated as a new triple with
variable: value format. e.g. [c_1, Drug, aspirin, C0004057]. We combined this latter
with EU-ADR corpus to link drug with probable disease and ADR in the space.
Each time, annotation may have many attributes, while we project them into the
same format, then we copy all the entities triples of each document that shows up in
our training vocabulary. Thus, each term outvocabulary is convolutionally aggregated
to the closest entities in the embedding matrix, we multiply frequencies of the same
term in both text documents. Finally, we defined a distance between two entities by
using soft cosine similarity measure as shown in the formula below Eq. 1.
f i = cosSimilarity(, ci , E) (1)
In such case, we need to consider the semantic meaning that conveyed regarding
medical aspects in texts, where similar entity meanings may attribute varied facets of
sentiments. A similarity metric that gives higher scores forci in documents belonging
to the same topic and lower scores when comparing documents from different topics.
Since, neural network is a method based on nonlinear information processing,
typically we use continuous BOW for building an embedding vector space regard-
ing related-medical concepts. The obtained embedded vectors trained regarding pre-
served ADRMINE parameters and context features where the context is defined with
seven features including the current token ti , the three preceding (ti−3 , ti−2 , ti−1 ), and
three following tokens (ti+3 , ti+2 , ti+1 ), in the input. Moreover, these samples were
passing by a set of normalization and processing steps to be able for our neural
inference model. Indeed, every single tweet including includes spelling, correction,
lemmatization, and tokenization. Our dataset consists of a separate document that
saves the life of correlate entities contained regarding medical and pharma objects.
Convolutional neural network provides an efficient mechanism for aggregating
information at a higher level of abstraction; we exploit convolutional learning to learn
data properties and tackle ambiguities types through common semantics and contex-
tual information.Considering a window of words [wi , wi+1 , . . . , wi+k−1 , wi+k ], the
concatenated vector of the ith window is then:
Si = [wi , wi+1 , . . . , wi+k−1 , wi+k ] ∈ Rk∗d (2)
The convolution filter is applied to each window, resulting in scalar values ri , each
for the ith window:
ri = g(xi ∗ u) ∈ R (3)
In practice one typically applies more filters, u 1 , . . . , u l , which can then be repre-
sented as a vector multiplied by a matrix U and with an addition of a bias term
b:
ri = g(xi ∗ u + b) (4)
with ri ∈ Rl , xi ∈ Rk.d∗l and b ∈ Rl
CNN features are also great at learning relevant features from unlabelled data and
got huge success in many unsupervised learning case study. CNN-based cluster-
ing method use these feature to be input to K-mean clustering and parameterized
manifold learning. It is of extracting the structural representation by polar medical
facts and non-polar facts. This is because of the need to distinct false positives and
negatives usually obtained by baselines.
3.2 Common Sense Detection
Accurate emotions analysis approach relies on the accuracy of vocabulary and the
way we define emotions regarding related-medication concepts (Drugs, ADRs, and
diseases), events, and facts. Indeed, patients self-reports may refer to various con-
cepts in different ways regarding various context. Not only surface analysis of the text
is required, but also common sense analysis based knowledge approach is needed. To
bridge the cognitive and affective gap between word-level natural language data and
the concept-level sentiments conveyed by them, affective common sense knowledge
is needed [17]. For this purpose, a conceptualization technique is involved in discov-
ering the conceptual primitives of each entity by means of contextual embedding.

Each entity may belong to many concepts regarding clusters we preserved from first
learning in previous stage cdrug =treatment, doctor, ADR, indication.
People frequently express their opinions with particular backgrounds and set of
morals aspects, such as ethics, spirituality, and positionality. In this window, it is
widely accepted that before patient or his family make a decision such a straightfor-
wardly information about drugs and adverse drug reactions is learned. Further, we
are willing to put this working hypothesis to the test of rational discourse, believ-
ing that other persons acting on a rational basis will agree. Thus, the weighing and
balancing of potential risks and benefits becomes an essential component of the
reasoning process in applying the principles. Moreover, related-medication mining
involves systematizing and defining concepts-related meanings. Indeed, it seeks to
resolve questions of human perception by defining affective concepts information.
Patient perception is main aspect to define what a person is permitted to do in a spe-
cific situation or a particular domain of action. Technically, we aim at extending the
following assumption regarding these common senses: “a set of affective concepts
correlate with affective words and affective common sense knowledge consists of
information that people usually take for ambiguous status, hence, normally leave
unstated.” Especially, It is concerned the computational treatment representing the
affective meaning of given input that allows a new form. Recognizing emotional
information requires the extraction of meaningful patterns from the gathered data
first otherwise, CNN-based model is calculated.
In this stage, a distinction between means and effect is performed. Meaning, a
distinction between word-level natural language data and the concept-level senti-
ments conveyed by them is also required. Affective common sense, in fact, is not
a kind of knowledge that we can find in formal knowledge such as Wikipedia, but
it consists in all the basic relationships among words, concepts, phrases, emotions,
and thoughts that allow people to communicate with each other and face everyday
life problems and experiences. For this reasons, we chose to use SenticNet [6] as
prior knowledge that is seminal domain-independent knowledge base which is con-
structed for concept-based sentiment analysis through a multidisciplinary approach,
namely sentic computing. Sentics is affective semantics where are operated to extract
the affective concept-based meaning. Thus, a pre-trained embedding from Concept-
Net is basically added to seek the concept-based affective information and collect
such kind of knowledge through label sequential rules (LSR), crowd sourcing, and
GWAP techniques. Indeed, sentic computing is a multidisciplinary approach to opin-
ion mining and sentiment analysis at the crossroads between affective computing and
common sense computing, which exploits both computer and social sciences to better
recognize, interpret and process opinions and sentiments over the Web. It provides
the cognitive and affective information associated to concepts extracted from opin-
ionated text by means of a semantic parser.
4 Experiments and Output Results
The performance of adopting existing SA method to the medical situations and case
studies can be summarized as follow: (1) sentiment analysis systems are able to per-
form sentiment analysis toward a given entity fairly well, but poorly on clarifying
sentiment towards medical targets, (2) they got in low recall in term of distinction
multi-word expressions that may refer to an adverse drug reaction. The paper inves-
tigates the challenges of considering biomedical aspects through sentiment tagging
task. An automatic approach to generate sentimental based-aspect concerning drug
reaction multi-word expressions toward varied related medication contexts, it consid-
ered as domain-specific sentiment lexicon by considering the relationship between
the sentiment of both words features and medical concepts features. From our evalu-
ation on large Twitter data set, we proved the efficiently of our features representation
of drug reaction, which is dedicated to matching expressions from everyday patient
self-reports.
In order to understand the difference of deriving features from various source data,
we choose to utilize a predefined corpora trained on a different corpus. A results from
[7] assumes that is better than using a subselection of the resources and delivered an
unified online corpus for emotion quantification. Technically, deep neural networks
have achieved great success in enhancing model reliability. Authors in [17] provide
a novel mechanism to create medical distributed N-grams by enhancing convolu-
tional representation, which is applied for featuring text regarding medical setting
and clarifying contextual sentiment in a given target. Correlation between Knowl-
edge, experience and common Sense are assessted through this study. Each time,
a sentiment value is contributed to each vector. We use two benchmarks for model
development: (1) lex1: ConceptNet as a representation of commonsense knowledge
2
(2) lex2: SenticNet.3
Since patients perceptions of drug-related knowledge are usually considered
empty of content and untruthful, this application of emotional state comes into focus
of understanding the unique features and polar facts that provide the context for the
case. Therefore, obtaining the relevant and accurate facts is an essential component
of this approach to decision making. Noticeably, we got a great changes and shift in
patient statements on everyday shared conversations on the pandimic period. Table 4
shows a comparison of positives and negatives statements on COVID-19 period
and Before the pandemic period. Where we used a parkisons datasets collected for
previous studies on the year of 2019.
Emotional and common senses detection performance are assessed through exper-
iments on varied online datasets (Facebook, Twitter, Parkinson Forum), as sum-
marized in Table 2. An extensive evaluation of different features, including med-
ical corpora, ML-algorithms, have been performed. As shown in Table 2, a sam-
ple of PD-related posts (dataset can be found in this link4 ) was collected from the
2 https://ttic.uchicago.edu/~kgimpel/commonsense.html.
3 https://sentic.net/downloads/.
4 https://github.com/hananeGrissette/Datasets-for-online-biomedical-WSD.
Table 1 Biomedical corpora and medical ontologies statistics used for biomedical distributed
representation
Sources Documents Sentences Tokens
PubMed 28,714,373 181,634,210 4,354,171,148
MIMIC III Clinical 2,083,180 41,674,775 539,006,967
notes
Table 2 Summarize online datasets from varied platforms used for both training and model devel-
opments
Plateform #Posts Keywords used
source
Twitter 256,703 Parkinson, disorder, seizure, Chloroquine, Corona, Virus,
Remdesivir, Disease, infectious, treatments, COVID-19
Facebook 49,572 COVID-19,Chloroquine, Corona, Virus, Remdesivir, disease,
infectious, Parkinson, disorder, seizure, treatments
PD forum 30,748 Chloroquine, COVID-19 Corona, Virus, Remdesivir, disease,
infectious, treatments
online healthcare community of Parkinson’s Disease and normalized to enrich the

vocabulary. For twitter, we collect more than 25000 tweets in the Par kinson and
C O V I D − 19 contexts that have been prepared to be input to the neural classi-
fier for defining medical concepts and then re-define distributed representation for
unrelated items of natural medical concepts cited in real-life patients narratives. A
based-keywords crawling system is created for the collection of Twitter posts. In this
study, we focused on distracting drug reaction in the COVID19 contexts. We used
a list of related-COVID-19 Keywords, e.g., Cor ona, and Chlor oquine. Thus, we
are interested in getting information attributed by : [’id’,’created_at’, ’source’, ’orig-
inal_text’, ’retweet_count’, ’ADR’, ’original_author’, ’hashtags’, ’user_mentions’,
’place’, ’place_coord_boundaries’]. The Table 2 summarizes statistics of raw data
grouped in terms of some related-drugs keywords in different slice of time. Twitter
Data relies on relies on large volumes of unlabeled data, and thus diminishing the need
for memorable positive and negative statement regarding each based-target class is
assessed through the experiments,whereby, is considered as an automatic supervised
learning problem regarding emotion information. The based CNN-clustering are one
or more axes-matter. For fine-grained analysis, the evaluation may also applied on
varied axes such as: age-axis and gender-axis that allows us to show peaks of posi-
tiveness of PD’s patients.
We conducted many experimentation derived from the application of the pro-
posed method based on hybrid medical corpora based-text and concepts, a minute
conceptualization is released. As illustrated in Table 3, the BiLSTM-based Parkin-
son’s classifier outperforms other neural network algorithms regardless the sentiment
lexicon and the medical knowledge used. The support vector machine (SVM) classi-
Table 3 Experiments results overview on different platforms data using sentiment lexicons dis-
cussed above
Dataset Algorithm Sentiment Medical knowledge/ADRs Accuracy %
Twitter BiLSTM Lex1 PubMED + clinical notes MIMIC 0.71
III + EU-ADR
BiLSTM Lex1 +lex2 PubMED+ EU-ADR+ ADRMINE 0.81
LSTM Lex1+lex2 PubMED + EU-ADR+ ADRMINE 0.73
SVM Lex1+lex2 PubMED + EU-ADR+ ADRMINE 0.61
stacked- Lex1+lex2 PubMED + EU-ADR+ ADRMINE 0.73
LSTM
Facebook BiLSTM Lex1 PubMED + clinical notes MIMIC 0.71
III + EU-ADR
stacked- Lex1+lex2 PubMED + EU-ADR+ ADRMINE 0.71
LSTM
PD forum BiLSTM Lex1 PubMED + clinical notes MIMIC 0.76
III + EU-ADR
Stacked- Lex1+lex2 PubMED + EU-ADR+ ADRMINE 0.80
LSTM
fier has got acceptable results in classifying polar facts and non-polar facts. The based
stacked-LSTM and BiLSTM model consistently improved the sentiment classifica-
tion performance, but is efficient when we exploit proposed configuration on PD post
from forum due to the post’s length(it contains more details and clear drug-related
descriptions). We also conducted an evaluation on different dataset from Facebook,
which is collected in a previous study. The got in low results than other baselines in
term of entity recognition recall, which reflect on model performance (Table 4).
However, it deserves to be noted that if we replace tokens with n-grams and
train on small datasets, which is CNN-clustering based architecture improved the
representations over the obtained biomedical distributed representations on top of
those features, then we may get whopping 1.8 bumps in the accuracy and it boosts
accuracy to over 87%. Thus, we end up by learning some deeper representations and
new multi-word expressions vectors are inserted in the vocabulary each time.
Table 4 Percentage of sentiment terms (positive and negative) extracted before the COVID-19
period and in the pandemic period
Sources Before COVID-19 In COVID-19 period
Positive (%) Negative (%) Positive (%) Negative (%)
PD’s forum 30 10 20 35
Twitter 33 16 28 47
43 13 30 50
51 19 37 43
Facebook 15 17 25 45
32 20 40 52
5 Conclusion
This article is intended to be brief introduction to the use of neural networks to effi-
ciently leverage patient emotions regarding various affective aspects. We proposed
an automatic CNN-clustering aspect-based identification method for drug mentions,
events, treatments from daily PD narratives digests. The experiments proved emo-
tional Parkinson classifier ability to translate varied facets of sentiment and seek
the impactful COVID-19 insights from generated narratives. The study of what is
morally right by patient in given condition and what is not, is our perspectives. We
aim at defining an neural network approach based on set of morals aspects in which
the model rely on variables that can be shown to substitute for morals aspects regard-
ing the emotion quantity. It also involved to provide a proper standard of care that
avoids or minimizes the risk of harm that is supported not only by our commonly
held moral convictions, but by the laws of society as well.
References
1. Grissette, H., Nfaoui, E.H.: Drug reaction discriminator within encoder-decoder neural network
model: Covid-19 pandemic case study. In: 2020 Seventh International Conference on Social
Networks Analysis, Management and Security (SNAMS), pages 1–7 (2020)
2. Grissette, H., Nfaoui, E.H.: A conditional sentiment analysis model for the embedding patient
self-report experiences on social media. In: Advances in Intelligent Systems and Computing
(2019)
3. Grissette, H., Nfaoui, E.H.: The impact of social media messages on parkinson’s disease treat-
ment: detecting genuine sentiment in patient notes. In: Book Series Lecture Notes in Compu-
tational Vision and Biomechanics. SPRINGER International Work Conference on Bioinspired
Intelligence (IWOBI 2020) (2021)
4. Grissette, H., Nfaoui, E.H.: Daily life patients sentiment analysis model based on well-encoded
embedding vocabulary for related-medication text. In: Proceedings of the 2019 IEEE/ACM
International Conference on Advances in Social Networks Analysis and Mining, ASONAM
2019 (2019)
5. Nikfarjam, A., Sarker, A., O’Connor, K., Ginn, R., Gonzalez, G.: Pharmacovigilance from
social media: Mining adverse drug reaction mentions using sequence labeling with word embed-
ding cluster features. J. Am. Med. Inf. Assoc. (2015)
6. Cambria, E., Li, Y., Xing, F.Z., Poria, S., Kwok, K.: SenticNet 6: ensemble application of sym-
bolic and subsymbolic AI for sentiment analysis. In: International Conference on Information
and Knowledge Management, Proceedings (2020)
7. Wu, W., Li, H., Wang, H., Zhu, K.Q.: Probase: a probabilistic taxonomy for text understanding.
In: Proceedings of the ACM SIGMOD International Conference on Management of Data (2012)
8. Cambria, E., Xia, Y., Hussain, A.: Affective common sense knowledge acquisition for sentiment
analysis. In: Proceedings of the Eighth International Conference on Language Resources and
Evaluation (LREC’12), pages 3580–3585, Istanbul, Turkey. European Language Resources
Association (ELRA) (2012)
9. Shiang Wang, C., Ju Lin, P., Lan Cheng, C., Hua Tai, S., Kao Yang, Y.H., Hsien Chiang, J.:
Detecting potential adverse drug reactions using a deep neural network model. J. Med. Internet
Res. (2019)
10. Grover, S., Somaiya, M., Kumar, S., Avasthi, A.: Psychiatric Aspects of Parkinson’s Disease
(2015)
11. Tsoulos, I.G., Mitsi, G., Stavrakoudis, A., Papapetropoulos, S.: Application of machine learning
in a parkinson’s disease digital biomarker dataset using neural network construction (NNC)
methodology discriminates patient motor status. Front, ICT (2019)
12. Nilashi, M., Ibrahim, O., Ahani, A.: Accuracy improvement for predicting Parkinson’s disease
progression. Sci. Rep. (2016)
13. Pennington, J., Socher, R., Manning, C.D.: GloVe: global vectors for word representation. In:
Proceedings of the Conference EMNLP 2014—2014 Conference on Empirical Methods in
Natural Language Processing (2014)
14. van Engelen, J.E., Hoos, H.H.: A survey on semi-supervised learning. Mach. Learn. (2020)
15. Nikfarjam, A.: Health Information Extraction from Social Media. ProQuest Dissertations and
Theses (2016)
16. van Mulligen, E.M., Fourrier-Reglat, A., Gurwitz, D., Molokhia, M., Nieto, A., Trifiro, G.,
Kors, J.A., Furlong, L.I.: The EU-ADR corpus: annotated drugs, diseases, targets, and their
relationships. J. Biomed. Inf. (2012)
17. Grissette, H., Nfaoui, E.H.: Enhancing convolution-based sentiment extractor via dubbed N-
gram embedding-related drug vocabulary. Netw. Model. Anal. Health Inf. Bioinf. 9(1), 42
(2020)
Missing Data Analysis in the Healthcare
Field: COVID-19 Case Study
Hayat Bihri , Sara Hsaini , Rachid Nejjari , Salma Azzouzi ,

Abstract Nowadays, data is becoming incredibly important to manage in a variety of

domains, especially healthcare. A large volume of information is collected through
a lot of ways such as connected objects, datasets, records, or doctor’s notes. This
information can help clinicians in preventing diagnosis errors and reduce treatment
complexity. However, data is not always available and reliable due to missing values
and outliers which lead to a loss of a significant amount of information. In this paper,
we suggest a model to deal with missing values of our system during the diagnosis
of COVID-19 pandemic. The system aims to enhance the physical distancing and
activities limitations related to the outbreak and then providing to the medical staff
the necessary information to make the right decision.
1 Introduction
In the last century, the development of technology and innovation related to the field
of the Internet of Things (IoT) has contributed enormously to improve several sectors
like buildings, transportation, and health [1].
The treatment of collected information from these devices has a very important
role in predicting the future and taking the right decision principally if such data is
complete. However, this is not always the case, as this information is plagued by
missing values and biased data.
In the healthcare domain, preventing diseases and the early detection of a compli-
cation regarding the patient’s situation may save lives and avoid the worst, and we
can find many datasets in the field such as the electronic health records (EHRs)
that contain information about patients (habits, prescriptions of medication, medical
history, the doctors’ diagnosis, nurses’ notes, etc.) [2]. But, one of the common prob-
lems which may affect the validity of the clinical results and decrease the precision
of medical research remains the missing of data [3]. In fact, healthcare data analytics
depend mainly on the availability of information and many others factors such as
H. Bihri (B) · S. Hsaini · R. Nejjari · S. Azzouzi · M. E. H. Charaf

Faculty of Sciences, Ibn Tofail University, Kenitra, Morocco
https://doi.org/10.1007/978-981-16-3637-0_61
874 H. Bihri et al.
disconnection of devices, a deterioration of network connection, or a failure in the

equipment can lead to bias data or loss of precision [4].
In the current circumstance of the COVID-19 pandemic, and due to the severity
of the virus in addition to the risk of transmission and contamination, our idea is to
promote the physical distancing and to minimize the contact of the health staff with
potential patients who could probably be affected by the virus. Thus, we need to deal
with information delivered through the monitoring control, principally if they are not
complete or biased, which can affect the final result and compromise the validity of
the diagnosis.
To avoid this problem, the missing data must be treated and addressed correctly
by the application of the appropriate method that suits the most with the issue of the
study, and then facilitate extraction of useful information from the data collected in
order to assist the decision making.
Furthermore, the use of the IOT technology was through sensors such as
telemedicine tools in order to collect data and to offer then a monitoring control using
smartphones, webcam-enabled computers, or robots of temperature measurement
[5, 6].
In this article, we review the different techniques proposed in the literature to
deal with the missing data treatment in the healthcare domain. Then, we describe our
proposed system to protect both patients and the medical staff from the contamination
by the virus in the context of COVID-19 pandemic, and how to manage missing data,
resulting from data collection through different remote diagnosis tools. The study
is in line with the idea detailed in [7]. The aim is to improve our previous works to
monitor and control the spread of COVID-19 disease.
The paper is organized as follows: Sect. 2 gives some basic concepts related
to missing data and prediction functions. Then, we introduce, respectively, in the
Sects. 3 and 4 the problematic statement and some related works. Afterward, we
present our contribution to manage missing data issues and the workflow of the
proposed approach in the context of COVID-19 in Sect. 5. Finally, we give some
conclusions and future works.
2 Preliminaries
Generally, we distinguish structured data which is an organized collection structured

in rows and columns for storage and unstructured data which is not organized in
predefined manner or model, and the processing of such data requires a specialized
programs and methods to extract and analyze useful information [8, 9].
Furthermore, we can collect data using manually form for entering data, or by
means of measuring features from monitoring systems.
Missing Data Analysis in the Healthcare Field … 875
2.1 Missing Data and Data Collection Mechanisms
The data collected using monitoring systems are usually missed or biased. Indeed,
the type of missing data can indicate the appropriate approach to deal with the issue.
These are three categories [10–12]:
• Missing completely at random (MCAR): can result independently from both the
observed variables and the unobserved variables, and occur entirely in a random
manner. For example, in the case of temperature measurement using an electronic
device. The data can be missed if the device is running out of battery.
• Missing at random (MAR): occurs when the missing data is not unplanned.
However, the missed data can be predictable from variables with complete infor-
mation. For example, the measurement of the temperature failed with children
due to the lack of cooperation of young people.
• Not missing at random (NMAR): In this case, the probability of the variable
that’s missing is directly related to the reason that is looking for or requested by
the study. For example, patients with fever resist to the temperature measurement
for fear of being diagnosed positive.
2.2 Prediction Function
Prediction function or imputation technique is a method used to predict incomplete

information in a dataset. Therefore, several approaches exist to deal with missing
values issue and can be divided into two main categories [13]:
• Deletion-based techniques which ignore incomplete data, we can distinguish:
– List wise deletion that consists of deletion of the incomplete rows completely
which can lead to a loss of a quantity of information;
– Pairwise deletion: This technique maintains the maximum of information and
requires the type MCAR of values that are missing.
• Recovering-based Techniques that aims to recover missing values by estimating
them, through application of a variety of techniques such as:
– Single imputation technique: In this technique, the missing entry is replaced by
a calculated value using an appropriate formula or equation based on variable’s
observed;
– Mean imputation: is one method of the single imputation techniques that
consists of the replacement of the missing value by the mean of the values
observed for this variable;
– Last observation carried forward (LOCF): A kind of deterministic single impu-
tation technique that aims to replace the missing value by the last observed
one;
876 H. Bihri et al.
– Non-response weighting approach is developed to address unit non-response

by creation and application of columns of weights to responses items;
– Multiple Imputations consists of three phases: the imputation phase, the anal-
ysis phase, and the pooling phase. The technique is based on the creation of m
replacement for the missing value using m arbitrary constant;
– Full Information Maximum Likelihood (FIML) is based on the ignorance of
non-response items and considers only the observed data. The results provided
by FIML are optimal in a MAR context.
3 Problematic Statement
Over time, the management of health crises and problems has proven to be very
difficult for both health workers and Nations, especially when it comes to contagious
diseases or pandemics.
In December 2019, a new kind of corona virus (COVID-19) has been discovered
in Wuhan. The disease is then spreading quickly to healthy persons having close
contact with infected ones [14]. The virus causes severe respiratory problems which
increases significantly intensive care unit admission and leads to a high mortality
rate [15].
Therefore, in order to reduce contact between people, which lead to the rapid
propagation of the virus a set of measures have been taken by the government, in
order to break the chains of transmission of the infection, such as physical distancing
measures, school closures, travel restrictions, and activities limitations [16]. Even
more, the medical staff needs to adopt telemedicine as a safe strategy to restrict
contact with infected patients. Indeed, technology can help to exchange information
24/24 and 7/7 using smartphones or IOT components providing thus a real situation
of the patient and allow the health staff to control remotely infected persons [17].
However, the use of sensors and telemedicine tools to collect data can be faced to
missing data issues.
In this paper, we suggest our approach to manage outliers and missing data in
order to help medical staff to make the appropriate healthcare policy decisions when
the knowledge is not available. The model as designed could be used to diagnosis
remotely the COVID-19 patients.
4 Related Works
To deal with missing data especially in healthcare domain, many methods of treatment
are available. They are different but they can help to predict missing items in a specific
situation.
In this context, in [13] the authors tackled the issue of principle missing data
treatments. According to their study, they conclude that the dominated approach
used to address the missing data problem was deletion-based technique (36.7%).
The work [18] proposes an example of missing data in the work-family conflict
(WFC) context. The author purposes four approaches to deal with missing data (he
CWFC scores in this article), which are the following methods: Multiple Imputations
analysis, linear regression models, and logistic regression.
Moreover, the authors in [19] conducted a study on 20 pregnant responding to a
set of criteria, and they are in the end of their first trimester. The physical activity and
heart data of the samples were collected and transmitted through a gateway device
(Smartphone, PC) to make health decisions. Most of data analysis was performed
to extract useful information regarding to maternal heart rate. In this context, the
authors suggest an approach, based on imputation, to handle missing data issues that
occur when the sensor is incapable to provide data.
Another work [20] tackles two predictive approaches: The single imputation tech-
nique as a suitable choice when the missing value is not informative. The second one
is the multiple imputations that it is useful for a complete observation.
In [21], the authors highlight the benefit of the data and the electronic health
records available in healthcare centers which bring important opportunities for
advancing patient care and population health. The basic idea in [22] is to handle the
problem of missing data often occurred in the case of a sensor failure or the network
device problems, for example. To this end, the authors propose a new approach enti-
tled: a Dynamic Adaptive Network-Based Fuzzy Inference System (D-ANFIS). In
this study, the collected data was divided in two groups: complete data that be used
to train the proposed method and incomplete data to fill the missing values.
Furthermore, the authors in [23] describe the principal methods to handle missing
values in multivariate Data Analysis which can manage and treat missing data
according to the nature of information, especially categorical continuous, mixed or
structured variables. In this context, they introduce these methods: principal compo-
nent analysis (PCA), multiple correspondence analysis (MCA), factorial analysis for
mixed data (FAMD), and multiple factor analysis (MFA).
According to [24], the authors are mainly focused on the prediction of cere-
bral infarction risk since this is a fatal disease in the region of the study. To deal
with the issue, the authors propose a new convolutional neural network-based multi-
modal disease risk prediction (CNN-MDRP) for both structured and unstructured
data collected for their study. Another work [25] proposes a deep study to under-
stand the origin of a chronic and complex disease called the inflammatory bowel
disease (IBD). For this purpose, the authors introduce a new imputation method
centered on latent-based analysis combined with patients clustering in order to face
the issue of data missingness. The method as described allows to improve the design
of treatment strategies and also to develop the predictive models of prognosis.
On the other hand, the authors in [26] emphasize the importance of handling
missing data encountered in the field of clinical research. To deal appropriately with
the issue, they purpose to follow four mean steps particularly: (1) trying to reduce the
rate of missing data in the data collection stage; (2) performing a data diagnostic to
878 H. Bihri et al.
understand the mechanism of missing value; (3) handling missing data by application
of the appropriate method and finally (4) proceeding to the analyze of sensitivity when
required.
Moreover, the authors in [27] highlight the importance of data quality when estab-
lishing statistics. In this context, the authors suggest to deal with missing values using
cumulative linear regression as kind of imputation algorithm. The idea is to cumulate
the imputed variables in order to estimate the missing values in the next incomplete
variable. By applying the method on five datasets, the results obtained revealed that
the performances differ according to the size of data, missing proportion, and the
type of mechanisms used for this purpose.
In the next section, we describe our prototype to remotely monitor the patients’
diagnosis then we explain how to tackle missing values issue.
5 COVID-19 Diagnosis Monitoring
In what follows, we tackle the missing data issue in the e-health domain and
particularly in the COVID-19 context.
5.1 Architecture
Figure 1 describes the architecture of our diagnosis monitoring system.
Fig. 1 Diagnosis monitoring architecture

In order to achieve the physical distancing recommended by the World Health

Organization (WHO) in the context of COVID-19, we propose a diagnosis monitoring
architecture that can be used to diagnose patients remotely without having to move
to the medical unit and have direct contact with the doctor or other person of the
medical staff.
In fact, the system proposes to diagnose cases probably infected with the virus
through the use of several tools able to transmit information about the patients’
condition to the medical unit.
Therefore, data is exchanged between the patient and intensive care unit using
some remote monitoring tools. However, the data provided could become unavailable
or incomplete due to failure in the data collection step or during the data transmission.
This can occur when the connection to the medical unit server is broken down or
interrupted. Otherwise, if the device doesn’t work correctly or is disconnected from
the network.
A scenario that we propose for the remote diagnosis of COVID-19 is the one used
in Morocco. In fact, the patient who doubts or has symptoms such as fever, respira-
tory symptoms, or digestive symptom, such as lack of appetite, diarrhea, vomiting,
and abdominal pain, can use this system to contact the intensive care unit. After-
ward, a variety of accurate questions based on a checklist of common symptoms are
communicated to the patient through our system to determine and identify whether
the patient needs to seek immediate medical attention or not. Indeed, the application
of such measure can contribute effectively to deal with the widespread of the disease
by decreasing the high transmission rate, and then limit the spread of the virus. We
suggest in the next subsection our prototype to face such situation and to handle
missing values using the mean imputation approach.
5.2 Missing Data Handling: Prototype
Figure 2 describes the processes of making a decision using our diagnosis monitoring
system. In fact, the proposed system aims to automate collection of data related to
people likely be infected by COVID-19. Afterward, the system ensures the transfer
of information to the intensive care unit. Thus, it will guarantee also the individuals
distancing and protecting the medical staff from any probable contamination.
Subsequently, data is treated and stored in the appropriate server in order to be
analyzed and to extract useful information necessary for making suitable decision.
Therefore, the system is subdivided into four main steps:
• Data collection phase;
• Data prediction phase;
• Data processing phase;
• Decision-making phase.
880 H. Bihri et al.
Fig. 2 Missing data handling prototype
5.2.1 Data Collection
This step aims to collect data using various tools for information exchanging such
as smartphones, webcam-enabled computers, smart thermometers, etc. These data
sources must provide a real situation of the patient and collect information that will
be transmitted to the medical unit for treatment. We distinguish two types of data:
• Complete information refers to data measured and collected correctly. The
observed entries are then automatically redirected for storage in a local server.
• Incomplete data refers to data collected through the monitoring tools. In such
situations, data is usually biased or loosed due to the devices, or if the network
connection is broken down for example.
5.2.2 Data Prediction
As described above, the input data is sent to the intensive unit care server to be treated.
It refers to the information collected using monitoring tools (medicine sensors and
mobile device). The data provided is then sorted, and two groups are distinguished:
missing data and complete data.
For complete data, it will be redirected automatically to the next phase: Data
processing. However, missing data need to be treated before being processed. In a
single imputation method, missing data is filled in by some means and the resulting
completed data set is used for inference [28]. In our case, we suggest replacing
missing values using the mean of the available cases using the mean imputation
approach as a kind of single imputation method. Even if the variability in the data
is reduced, which can lead to underestimate standard deviations as well as variance
estimates, the method is still easy to use and gives sufficiently interesting results in
our case study.
5.2.3 Data Processing
The estimated data as well as complete data are both redirected to the storage handling
phase. The system proceeds at this stage to the following operations:
• Data storage: In this case, we need to use the appropriate hardware/software tools
to ensure data storage for both complete and predicted data. In fact, we need to
take into account the types of data collected (recordings, video file…) as well
as the way such data increase to better size the required capacity for the storage
needs.
• Data analysis: It aims to analyze the data collected using the appropriate data
analysis tools. This will help to improve the decision making in the next steps.
5.2.4 Making Decision
The objective is to give a screening of the patient state to the health staff and clinician
using reporting and dashboard. The goal is to help them taking the right decision and
to improve their prevention’s strategy.
To meet the specific needs of physicians, the key performance indicators (KPIs)
and useful information to be displayed must be carefully defined in collaboration with
the medical professionals. Therefore, the implementation of this real-time monitoring
system reduces the time and effort required to search and interpret information which
will improve significantly the medical decision making.
5.3 Discussion
Missing data can present a significant risk of drawing erroneous conclusions from
clinical studies. It is common practice to impute missing values, but this can only
provide an approximate result to the actual expected outcome.
Many recent research works are devoted to handle missing data in the healthcare
domain and to avoid the problems described previously. The review of some articles
tackling different approaches used in the field reveals a variety of techniques to deal
with this issue such as deletion-based techniques and recovering-based techniques.
882 H. Bihri et al.
Therefore, the use of an inappropriate method according to the missing item can
bias results of the study. Hence, the identification of the suitable method depends
mainly on whether the data is missing completely at random (MCAR), missing at
random (MAR), or not missing at random (NMAR) as explained previously.
In the prototype proposed in this paper and according to the reasons leading to
incomplete data, we consider MCAR as the appropriate type of missing variable. In
addition, we opt for the use of mean imputation in order to predict missing values
from the observed one. Whereas we exclude the use of techniques based on deletion
due to the loss of data that occurs when such type of techniques is applied.
Moreover, even if mean, median or mode imputations are simples and easy to
implement, we are aware that such techniques could underestimate variance and
ignore the relationship with other variables. Therefore, further investigations need to
be done and deeply analyzed to understand such relations and to enhance our model
in order to obtain reliable results.
Furthermore, even if data security remains outside the scope of this paper, it is
still highly recommended to take into consideration such aspects during the design
of the application. In fact, the importance of data security in healthcare is becoming
increasingly critical and nowadays, it is imperative for healthcare organizations to
understand the risks that they could encounter in order to ensure protection against
online threats.
On the other hand, the first experiments show that our proposition lack perfor-
mance and could be limited if the data collected is not sufficient to execute the logic
of the imputation function. In this context, many studies confirm, using the simu-
lation of missing values, that a significant drop in performance is observed even in
cases where only a third of records were missing or incomplete. Thus, it is strongly
recommended to use large datasets to deal with this problem and promote prediction
of missing items.
Finally, the completeness of the data should be assessed using a monitoring system
that provides reports to the entire study team on a regular basis. These reports can
be used then to improve the conduct of the study.
6 Conclusion
According to the current situation in the world related to the COVID-19 pandemic,
we aim in this study to support the physical distancing by minimizing the contact of
the health staff. To this end, we describe our monitoring prototype to deal with such
situation and to perform patients’ diagnosis remotely in order to reduce contamination
risks.
However, healthcare analytics depends mainly on the availability of data and
presence of missing values during data collection stage which can lead to bias or loss
of precision. In this context, we suggest in this paper to use a prediction technique to
avoid loss of data and by the way to predict missing information in order to take the
right and valid decision. The method proposed for prediction is the mean imputation
technique used in the collection stage to fill our dataset by estimated values.
As prospects, we plan to conduct more experimental studies regarding the perfor-
mance of our prototype in other medical cases such as mammographic mass and
hepatitis datasets. We will also enhance our model in the future to take other
imputation methods, especially the multiple imputation method.
References
1. Balakrishnan, S.M., Sangaiah, A.K.: Aspect oriented modeling of missing data imputation for
internet of things (IoT) based healthcare infrastructure. Elsevier (2018)
2. Wells, B.J., et al.: Strategies for handling missing data in electronic health record derived
data. eGEMs (Generating Evidence & Methods to Improve Patient Outcomes), 1(3), 7 (2013).
https://doi.org/10.13063/2327-9214.1035
3. Donders, A.R.T., et al.: Review: a gentle introduction to imputation of missing values. J. Clin.
Epidemiol. 59(10), 1087–1091 (2006). https://doi.org/10.1016/j.jclinepi.2006.01.014
4. Ebada, A., Shehab, A., El-henawy, I.: Healthcare analysis in smart big data analytics: reviews.
Challenges Recommendations (2019). https://doi.org/10.1007/978-3-030-01560-2_2
5. Yang, G.Z., et al.: Combating COVID-19-the role of robotics in managing public health and
infectious diseases. Sci. Robot. 5(40), 1–3 (2020). https://doi.org/10.1126/scirobotics.abb5589
6. Engla, N.E.W., Journal, N.D.: New England J. 1489–1491 (2010)
7. Hsaini, S., Bihri, H., Azzouzi, S., El Hassan Charaf, M.: Contact-tracing approaches to
fight COVID-19 pandemic: limits and ethical challenges. In: 2020 IEEE 2nd International
Conference on Electronics, Control, Optimization and Computer Science. ICECOCS 2020,
(2020)
8. Wang, Y., et al.: United States Patent 2(12) (2016). Interfacial nanofibril composite for selective
alkane vapor detection. Patent No.: US 10,151,720 B2. Date of Patent: 11 Dec 2018
9. Park, M.: United States Patent 1(12) (2010). Method and apparatus for adjusting color of image.
Patent No.: US 7,852,533 B2. Date of Patent: 14 Dec 2010
10. Salgado, C.M., Azevedo, C., Proença, H., Vieira, S.M.: Missing data. In: Secondary Analysis
of Electronic Health Records. Springer, Cham (2016)
11. Haldorai, A., Ramu, A., Mohanram, S., Onn, C.C.: EAI International Conference on Big Data
Innovation for Sustainable Cognitive Computing (2018)
12. Sterne, J.A.C., et al.: Multiple imputation for missing data in epidemiological and clinical
research: potential and pitfalls. BMC 339(July), 157–160 (2009). https://doi.org/10.1136/bmj.
b2393
13. Lang, K.M., Little, T.D.: Principled missing data treatments. Prev. Sci. 19, 284–294 (2018).
https://doi.org/10.1007/s11121-016-0644-5
14. Heymann, D., Shindo, N.: COVID-19: what is next for public health? Lancet 395 (2020).
https://doi.org/10.1016/S0140-6736(20)30374-3
15. Huang, C., Wang, Y., Li, X., et al.: Clinical features of patients infected with 2019 novel
coronavirus in Wuhan, China. Lancet 395(10223), 497–506 (2020)
16. Huang, C., Wang, Y., Li, X., et al.: Clinical features of patients infected with 2019 novel
coronavirus in Wuhan, China. Lancet 395, (2020). https://doi.org/10.1016/S0140-6736(20)301
83-5
17. Kiesha, P., Yang, L., Timothy, R., et al.: The effect of control strategies to reduce social mixing
on outcomes of the COVID-19 epidemic in Wuhan, China: a modelling study. Lancet Public
Health 5, (2020). https://doi.org/10.1016/S2468-2667(20)30073-6
18. Nguyen, C.D., Strazdins, L., Nicholson, J.M., Cooklin, A.R.: Impact of missing data strategies
in studies of parental employment and health: missing items, missing waves, and missing
mothers. Soc. Sci. Med. 209, 160–168 (2018)
884 H. Bihri et al.
19. Azimi, I., Pahikkala, T., Rahmani, A.M., et al.: Missing data resilient decision-making for
healthcare IoT through personalization: a case study on maternal health. Future Gener. Comput.
Syst. 96, 297–308
20. Josse, J., Prost, N., Scornet, E., Varoquaux, G.: On the consistency of supervised learning with
missing values. arXiv preprint arXiv:1902.06931 (2019)
21. Stiglic, G., Kocbek, P., Fijacko, N., Sheikh, A., Pajnkihar, M.: Challenges associated with
missing data in electronic health records: a case study of a risk prediction model for diabetes
using data from Slovenian primary care. Health Inform. J. 25(3), 951–959 (2019)
22. Turabieh, H., Mafarja, M., Mirjalili, S.: Dynamic adaptive network-based fuzzy inference
system (D-ANFIS) for the imputation of missing data for internet of medical things applications.
IEEE Internet Things J. 6(6), 9316–9325 (2019)
23. Josse, J., Husson, F.: missMDA: a package for handling missing values in multivariate data
analysis. J. Stat. Softw. 70(i01), (2016)
24. Chen, M., Hao, Y., Hwang, K., Wang, L., Wang, L.: Disease prediction by machine learning
over big data from healthcare communities. IEEE Access 5, 8869–8879 (2017). https://doi.org/
10.1109/ACCESS.2017.2694446
25. Abedi, V., et al.: Latent-based imputation of laboratory measures from electronic health records:
case for complex diseases. bioRxiv, pp. 1–13 (2018). https://doi.org/10.1101/275743
26. Papageorgiou, G., et al.: Statistical primer: how to deal with missing data in scientific
research? Interact. Cardiovasc. Thorac. Surg. 27(2), 153–158 (2018). https://doi.org/10.1093/
icvts/ivy102
27. Mostafa, S.M.: Imputing missing values using cumulative linear regression. CAAI Trans. Intell.
Technol. 4(3), 182–200 (2019). https://doi.org/10.1049/trit.2019.0032
28. Jamshidian, M., Mata, M.: Advances in analysis of mean and covariance structure when data
are incomplete. In: Handbook of Latent Variable and Elated Models, pp. 21–44 (2007). https://
doi.org/10.1016/B978-044452044-9/50005-7
An Analysis of the Content in Social
Networks During COVID-19 Pandemic
Mironela Pirnau
Abstract During the COVID-19 pandemic, Internet and SN technologies are an

effective resource for disease surveillance and a good way to communicate to prevent
disease outbreaks. In December 2019, the frequency of the words COVID-19, SARS-
CoV-2, and pandemic was very low in online environment, being only few posts
informing that, “the mysterious coronavirus in China could spread.” After March 1,
2020, there have been numerous research projects that analyze the flows of messages
in social networks in order to perform real-time analyses, to follow the trends of
the pandemic evolution, to identify new disease outbreaks, and to elaborate better
predictions. In this context, this study analyzes the posts collected during [August–
September 2020], on the Twitter network, that contain the word “COVID-19,” written
both in Romanian and English. For the Romanian language posts, we obtained a
dictionary of the words used, for which it was calculated their occurrence frequency
in the multitude of tweets collected and pre-processed. The frequency of words
for non-noisy messages was identified from the multitude of words in the obtained
dictionary. For the equivalent of these words in English, we obtained the probability
density of words in the extracted and pre-processed posts written in English on
Twitter. This study also identifies the percentage of similarity between tweets that
contain words with a high frequency of apparition. The similarity for the collected and
pre-processed tweets that have “ro.” in the filed called Language has been computed
making use of Levenshtein algorithm. These calculations are intended to quickly help
find the relevant posts related to the situation generated by the COVID-19 pandemic.
It is well known that the costs of analyzing data from social networks are very low
compared to the costs involved in analyzing data from the centers of government
agencies; therefore, the proposed method may be useful.
M. Pirnau (B)
Faculty of Informatics, Titu Maiorescu University, Bucharest, Romania
e-mail: mironela.pirnau@prof.utm.ro
https://doi.org/10.1007/978-981-16-3637-0_62
886 M. Pirnau
1 Introduction
Social networks are widely used by people not only to share news, states of minds,
thoughts, photos, etc. but also to manage disasters by informing, helping, saving,
and monitoring health. The results obtained by processing the huge quantities of
data that circulate in social networks allow identifying and solving several major
problems that people have. In the present, Big data systems have come to enable
multiple processing of data [1]. Thus, by means of the obtained information, Big
data systems contribute to the management of risk in climate changes [2], to the
analysis of the traffic in big locations [3], to the personalization of health care [4], to
early cancer detection [5], to solving some social problems by building houses using
3D printer technology [6], to social media analysis [7], and to developing methods
that contribute to identify the feelings and emotions of people [8]. Social networks
have contributed to the creation and continuous development of massive amounts of
data at a relatively low cost. There is a lot of research that analyzes large amounts
of data and highlights both the role of computing tools and the manner of storing
them [9, 10]. Big data in media communications and Internet technology have led to
the development of alternative solutions to help people in case of natural disasters
[11]. There are research studies that monitor different types of disasters based on the
data extracted from social platforms that have highlighted real solutions [12–17]. By
extension, these studies can also be used for the situation generated by the pandemic,
which the whole of humanity is facing. The relevance of the words [17–19] can
greatly help the analysis of data processed in the online environment, in a disaster
situation. This study analyzes data collected in August–September 2020 from the
Twitter network [20], to find the most common words relevant to the situation of
COVID-19. In this respect, the study consists of (1) Introduction, (2) Related Works,
(3) presentation of the data set used, (4) the results obtained, and (5) Discussions
and Conclusions. Because, throughout this period, it has been a proven fact that the
Public Health District Authority cannot handle the communication with the people
who need its help, an automatic system for processing and correct interpretation
of email messages received during the COVID-19 crisis would contribute to the
decrease in the response and intervention time. In this sense, this study tries to prove
that the identification of relevant words in online communication can contribute to
the efficient development of a demand–response system.
2 Related Works
People have become more and more concerned about the exponential evolution of
illness cases, about the deaths caused by them, as well as about the severe repercus-
sions of the COVID-19 pandemic’s evolution on the daily life. There are numerous
studies that analyze the magnitude of disasters and their consequences [12–14, 21–
24]. At the same time, there are studies that show an exponential pattern of increase
An Analysis of the Content in Social Networks … 887
in the case of the number of messages during the first interval after a sudden catas-
trophe [24], but there are also studies that analyze the behavior of news consumers
on social networks [25–27]. Some research [28], based on analyzing the data taken
from social networks, characterizes users in terms of the use of controversial terms
during COVID-19 crisis. Knowing the most frequently used words in posts referring
to a certain category of events enables the determination of the logical conditions
for searching the events corresponding to emergencies [29–31]. Analyzing the data
collected from social networks in order to identify emergencies, it is essential to
establish the vocabulary used for a regional search, as shown in the studies [13, 29,
31, 32]. According to the studies [30, 33], the analysis of information with high
frequency of occurrence from the social networks contributes to the rapid decrease
in the effects of disasters. Online platforms allow the understanding of social discus-
sions, of the manner in which people cope with this unprecedented global crisis.
Informative studies on posts from social networks should be used by state institutions
for developing intelligent online communication systems.
3 Data Set
Taking into account the rapid evolution of the COVID-19 pandemic, as well as
the problems that have arisen regarding the people’s concern about the significant
increase in cases of illness and death caused by the infection of the population with
Sarv-Cov_2, we have extracted tweets from the Twitter network during [August–
September 2020] using the topic COVID-19 as a selective filter. The data have been
extracted using the software developed in the research [20] and then cleaned to
be consistent in the processing process. An important task of fundamental natural
language processing (NLP) is lemmatization. Natural language processing (NLP)
has the role of contributing to the recognition of speech and natural language. This is
the most common text pre-processing technique used in natural language processing
(NLP) and machine learning. Because lemmatization involves deriving the meaning
of a word from a dictionary, it is time-consuming, but it is the simplest method. There
are lemmatizers that are based on the use of a vocabulary and a morphological analysis
of words. These work well for simple flexible forms, but for large compound words,
it is necessary to use a rule-based system for machine learning from an annotated
corpus [34–38]. Statistical processing of natural language is largely based on machine
learning. From the collected messages, only the ones that have the language field
completed with “en” and “ro” were used. The messages were cleaned and prepared
for processing [39–41]. Many messages are created and posted by robots [42–44],
automated accounts that enhance certain discussion topics, in contrast to the posts
that focus on the public health problems, so that the human operators should find it
difficult to manage such a situation. In the previous context, the noisy tweets were
removed [45], but the hardest task was to rewrite the tweets that had been written
using diacritics (s, , t, , ă, î, â). After all this processes, 811 unique tweets written in
Romanian and 43,727 unique tweets written in English were obtained for processing.
888 M. Pirnau
The pre-processing procedure and the actual processing procedure of these tweets
were performed using the PHP 7 programming environment. MariaDB was used for
the data management.
4 Results
Using regular expression, a database including all the words used for writing the
811 posts in Romanian was generated. Using my own application written in PHP
for content parsing, only the words in Romanian were extracted. Thus, only 1103
words were saved. In the collected tweets, it was noticed the fact that, even if the
language used for writing the tweets was “ro,” the users also operate with words
originated from other languages when writing messages and words. The following
were not taken into account: names of places, people’s names and surnames, names
of institutions, prepositions, and articles (definite and indefinite). In Table 1, the
average number of words used in one post was determined. The average value of
14 words, significant for the posts related to COVID-19 pandemic, indicates the fact
that the average number of words used is enough to convey a real state or situation.
In the dictionary of words, only the ones with the occurrence frequency of at least
5 times in the analyzed posts were selected. Vector V ro contains the words (written
in Romanian language) that occur at least 5 times in posts, and vector F ro contains
the number of occurrences corresponding to vector V ro .
V ro {} = {activ; afectat; ajutor; analize; antibiotic; anticorp; anti-covid; aparat;
apel; azil; bilant; boala; bolnav; cadre; cauza; cazuri; central; centru; confirmat;
contact; contra; control; convalescent; coronavirus; decedat; deces; depistat; diag-
nostic; disparut; donatori; doneaza; echipamente; epidemia; fals; pozitive; focar;
forma; grav; gripa; imbolnavit; imun; infectat; informeaza; ingrijiri; inregistrat;
intelege; laborator; localitate; lume; masca; medic; medicament; merg; moarte;
mondial; mor; negativ; oameni; pacient; pandemia; plasma; pneumonia; negativ;
post-covid; post-pandemic; pre-covid; pulmonary; raportat; rapus; reconfirmat;
restrictii; rezultat; risc; scolile; sever; sicriu; situatia; spital; test; tragedie; transfer;
tratament; tratarea; tratez; urgenta; vaccin; virus;. The English translation for the
terms written above is as follows: active; affected; help; analysis; antibiotic; anti-
body; anti-covid; camera; call; asylum; balance sheet; disease; sick; frames; cause;
Table 1 Determination of an average number of words from the analyzed posts written in Romanian
Analyzed elements Found values Number of used characters Average number of
characters
Tweets 811 78,261 96.49
Words 1103 7378 6.68
Average number of tweets 14.42
words =
cases; central; center; confirmed; contact; against; control; convalescent; coron-

avirus; deceased; death; diagnosed; diagnosis; missing; donors; donate; equipment;
epidemic; fake; false positives; outbreak; form; serious; flu; ill; immune; infected;
inform; care; registered; understand; laboratory; locality; world; mask; doctor; drug;
go; death; world; die; negative; people; patient; pandemic; plasma; pneumonia; nega-
tive; post-covid; post-pandemic; positive; pre-covid; pulmonary; reported; killed;
reconfirmed; restrictions; result; risk; schools; severe; coffin; situation; hospital; test;
tragedy; transfer; treatment; treating; treats; emergency; vaccine; virus}.
F ro {} = {17; 6; 7; 8; 5; 5; 16; 9; 6; 5; 14; 18; 10; 8; 27; 87; 14; 5; 25; 5; 16; 6; 9;
55; 8; 32; 7; 7; 5; 5; 5; 5; 10; 7; 5; 6; 20; 8; 8; 13; 5; 54; 8; 5; 12; 5; 7; 45; 5; 15; 14;
7; 6; 5; 8; 34; 8; 10; 8; 15; 9; 6; 25; 7; 6; 20; 6; 5; 14; 6; 7; 5; 5; 7; 13; 6; 6; 23; 21;
62; 6; 5; 8; 5; 6; 5; 47; 69}.
For the number of words in vector V ro , the number of occurrences is represented
by the sum of elements of vector F ro , meaning the 88 words that are used 1220
times in the 811 unique selected and cleaned tweets. The occurrence probability of
the words in vector V ro within the 811 posts may be calculated according to the
following Eq. (1).
No posts
P = n (1)
1 Fro
where No posts is 811, n represents the number of words in vector V ro with the
value of 88, and F contains the number of occurrences of these words. Thus, P =
66.48%. This value indicates the fact that vector V ro of the obtained words provides
an occurrence possibility of more than 20%, which demonstrates that these words
are relevant for the tweets analyzed in the context of COVID-19 pandemic.
If we represent vectors V ro and F ro graphically, Fig. 1 is obtained, namely
the Pareto graph. It indicates that the words with an occurrence frequency more
than 50% are Vrelevant {cazuri; virus; test; coronavirus; infectat; vaccin; local-
itate; mor; deces; cauza; confirmat; positiv/pozitiv; situatia; spital; forma; negatic;
boala;}/corresponding to the English words {cases; virus; test; coronavirus; infected;
vaccine; locality; die; death; cause; confirmed; positive; situation; hospital; form;
negative; disease;}.
For these words, Table 2 indicates the distribution of occurrence frequencies. The
Pareto graph can be seen in Fig. 2.
Figure 2 highlights that both the groups of words {cazuri; virus; test; coronavirus;
infectat;} and the group of words {vaccin; localitate; mor; deces; cauza; confirmat;
pozitiv; situatia; spital; forma; negativ; boala} have the occurrence frequency of 50%
in the analyzed posts.
Similarly, for the unique collected tweets written in English, vector V en was deter-
mined. It contains the same number of words as vector V ro , but in English, and their
occurrence frequency was determined.
In Table 3, only the words written in English with the occurrence frequency of
over 2% in tweets were kept.
890 M. Pirnau
Fig. 1 Distribution of V ro words in the analyzed posts
Table 2 Occurrence frequency in tweets for V relevant

No current Words in “ro” Number of Frequency of Frequency of Frequency of
occurrences occurrence in occurrence in occurrence in
the dictionary tweets (%) the key word
(%) set (%)
1 Cazuri 87 7.89 10.73 98.86
2 Virus 69 6.26 8.51 78.41
3 Test 62 5.62 7.64 70.45
4 Coronavirus 55 4.99 6.78 62.50
5 Infectat 54 4.90 6.66 61.36
6 Vaccin 47 4.26 5.80 53.41
7 Localitate 45 4.08 5.55 51.14
8 Mor 34 3.08 4.19 38.64
9 Deces 32 2.90 3.95 36.36
10 Cauza 27 2.45 3.33 30.68
11 Confirmat 25 2.27 3.08 28.41
12 Pozitiv 25 2.27 3.08 28.41
13 Situatia 23 2.09 2.84 26.14
14 Spital 21 1.90 2.59 23.86
15 Forma 20 1.81 2.47 22.73
16 Negativ 20 1.81 2.47 22.73
17 Boala 18 1.63 2.22 20.45
Fig. 2 Distribution of V relevant words in the analyzed posts
Table 3 Words in V EN with the frequency of occurrence in tweets > 2%

Words in English Number of occurrences The frequency of occurrence from 43,727 tweets
(%)
Test 4171 9.54
Cases 3122 7.14
People 2364 5.41
Death 2010 4.60
Virus 1778 4.07
Help 1627 3.72
Form 1597 3.65
Positive 1533 3.51
Report 1387 3.17
Pandemic 1274 2.91
Situation 1273 2.91
Disease 1175 2.69
Vaccine 1168 2.67
World 1168 2.67
School 1122 2.57
Coronavirus 1011 2.31
Died 944 2.16
Risk 941 2.15
Mask 911 2.08
892 M. Pirnau
Table 4 Statistics indicators

Mean 39.06
Standard error 4.95
Median 32.00
Mode 25.00
Standard deviation 20.41
Sample variance 416.68
Kurtosis 0.13
Skewness 0.97
Range 69.00
Minimum 18.00
Maximum 87.00
Sum 664.00
Count 17.00
Confidence level (95.0%) 10.50
For the words in Table 2, the main statistical indicators were determined in Table
4. The value of the Kurtosis peakedness parameter indicates the fact that the curve
is flatter than the normal one. The value of skewness parameter shows that the right
side of the average is asymmetric.
One can notice that mode, the statistic indicator, is 25, which corresponds to the
words of the group formed by “confirmat; pozitiv.”
The distribution function of continuous probability was calculated based on the
statistical indicators (the average and the standard deviation), and the group of words
“vaccine, locality, die, death” has the greatest occurrence density, as seen in Fig. 3.
Table 5 is created by intersecting the set of values in Tables 2 and 3. It indicates
the fact that there is a number of words with high occurrence frequency for the posts
written both in English and in Romanian.
Because Pearson’s r correlation coefficient is a dimensionless index between −
1.0 and 1.0 that indicates the extent of the linear relationship between two data sets,
0.02
0.015
0.01
0.005
Fig. 3 Probability density for V relevant

Table 5 Words with high occurrence frequency

English word Romanian word Frequency of English word Frequency of Romanian word
(%) (%)
Disease Boala 2.69 2.22
Form Forma 3.65 2.47
Hospital Spital 2.08 2.59
Situation Situatia 2.91 2.84
Positive Pozitiv 3.51 3.08
Death Deces 4.60 3.95
Die Mor 2.16 4.19
Vaccine Vaccin 2.67 5.80
Coronavirus Coronavirus 2.31 6.78
Test Test 9.54 7.64
Virus Virus 4.07 8.51
Cases Cazuri 7.14 10.73
we have noticed that the correlation between the Frequency of English word and the
Frequency of Romanian word for data in Table 5 is 0.609, which means that there is
a moderate to good correlation among the words found. The determined coefficient
was calculated by Eq. (2) for Pearson’s r correlation coefficient:

(x − x̄)(y − ȳ)
r= (2)
2 2
(x − x) (y − y)
where x and y are Frequency of English word and Frequency of Romanian word,
respectively (in Table 5). For the unique tweets in Romanian, the similarity of
contents was calculated for the posts, using the Levenshtein algorithm [46–48].
Levenshtein distance is a string metric that measures the difference between two
sequences (Levenshtein, 1966) and represents the minimum number of operations,
so that string X can be converted into string Y. Consider strings Xa = x 1 x 2 ..x i and Yb
= y1 y2 ..yj , where X and Y are sets of tweets referring to the same subject. If we define
D [a, b] the minimum number of operations by which Xa can be converted into Yb,
then D [i, j] is the Levenshtein editing distance looked for. Dynamic programming
is the method by which this algorithm is implemented.
Thus, it was noticed that 190 posts have a similarity ranging between 50 and 90%,
which represents a percent of 23,42% from the analyzed tweets, as seen in Fig. 4. The
tweets with similarity over 90% were not considered because their content varied
depending on the punctuation characters. Thus, they were considered informational
noise. In 2017, Twitter’s decision to double the number of characters from 140 to
280 allows users enough space to express their thoughts in their posts. Basically,
identifying the similarity of tweets becomes relevant, because the user no longer has
to delete words.
894 M. Pirnau
Similarity tweets
1.0000
0.9000
0.8000
0.7000
Similarity
0.6000
0.5000
0.4000
0.3000
0.2000
0.1000
0.0000
1
8
15
22
29
36
43
50
57
64
71
78
85
92
99
106
113
120
127
134
141
148
155
162
169
176
183
190
Number tweets
Fig. 4 Variation in content similarity
5 Discussions and Conclusions
Social networks play a vital role in the real-world events, including those incidents
which happen in critical periods such as earthquakes, hurricanes, epidemics, and
pandemics. It is a well-known fact that the social network messages have both posi-
tive and negative effects regarding the media coverage or the excessive publicity of
disasters. During disasters, the messages from social networks could be used success-
fully by authorities for a more efficient management of the actual calamity. The same
messages also represent a tool for malicious persons to spread false news. The study
of social networks provides informative data that helps identifying the manner used
by people to cope with a disaster situation. If, for the COVID-19 period, these data
were replaced with real messages received by empowered bodies—governments—
an automatic system could be built. This system could significantly contribute to
diminishing the waiting time when receiving a reply from these types of institutions.
Moreover, the empirical data from Table 3 indicate that a dictionary with common
terms, regardless of the language it uses, could be used to implement an efficient call–
response system, which would replace the human factor when communicating with
the authorities that are overwhelmed by the situation created by the 2020 pandemic.
The ICT systems, which use such dictionaries for their “communication” with people,
must be highly complex to function efficiently in unexpected disaster conditions.
Acknowledgements I thank Prof. H.N. Teodorescu for the suggestions on this research and for
correcting several preliminary versions of this chapter.
References
1. Avci, C., Tekinerdogan, B., Athanasiadis, I.N.: Software architectures for big data: a systematic
literature review. Big Data Anal. 5(1), 1–53 (2020). https://doi.org/10.1186/s41044-020-000
45-1
2. Guo, H.D., Zhang, L., Zhu, L.W.: Earth observation big data for climate change research. Adv.
Clim. Chang. Res. 6(2), 108–117 (2015)
3. Zhao, P., Hu, H.: Geographical patterns of traffic congestion in growing megacities: big data
analytics from Beijing. Cities 92, 164–174 (2019)
4. Tan, C., Sun, L., Liu, K.: Big data architecture for pervasive healthcare: a literature review. In:
Proceedings of the Twenty-Third European Conference on Information Systems, pp. 26–29.
Münster, Germany (2015)
5. Fitzgerald, R.C.: Big data is crucial to the early detection of cancer. Nat. Med. 26(1), 19–20
(2020)
6. Moustafa, K.: Make good use of big data: a home for everyone, Elsevier public health emergency
collection. Cities 107, (2020)
7. Kramer, A., Guillory, J., Hancock, J.: Experimental evidence of massive scale emotional
contagion through social networks. PNAS 111(24), 8788–8790 (2014)
8. Banerjee, S., Jenamani, M., Pratihar, D.K.: A survey on influence maximization in a social
network. Knowl. Inf. Syst. 62, 3417–3455 (2020)
9. Yue, Y.: Scale adaptation of text sentiment analysis algorithm in big data environment: Twitter
as data source. In: Atiquzzaman, M., Yen, N., Xu, Z. (eds.) Big Data Analytics for Cyber-
Physical System in Smart City. BDCPS 2019. Advances in Intelligent Systems and Computing,
vol. 1117, pp. 629–634. Springer, Singapore (2019)
10. Badaoui, F., Amar, A., Ait Hassou, L., et al.: Dimensionality reduction and class prediction
algorithm with application to microarray big data. J. Big Data 4, 32 (2017)
11. Teodorescu, H.N.L., Pirnau, M.: In: Muhammad Nazrul Islam (ed.) Cap 6: ICT for Early
Assessing the Disaster Amplitude, for Relief Planning, and for Resilience Improvement (2020).
e-ISBN: 9781785619977
12. Shan, S., Zhao, F.R., Wei, Y., Liu, M.: Disaster management 2.0: a real-time disaster damage
assessment model based on mobile social media data—A case study of Weibo (Chinese Twitter).
Saf. Sci. 115, 393–413 (2019)
13. Teodorescu, H.N.L.: Using analytics and social media for monitoring and mitigation of social
disasters. Procedia Eng. 107C, 325–334 (2015)
14. Pirnau, M.: Tool for monitoring web sites for emergency-related posts and post analysis. In:
Proceedings of the 8th Speech Technology and Human-Computer Dialogue (SpeD), pp. 1–6.
Bucharest, Romania, 14–17 Oct (2015).
15. Wang, B., Zhuang, J.: Crisis information distribution on Twitter: a content analysis of tweets
during hurricane sandy. Nat. Hazards 89(1), 161–181 (2017)
16. Eriksson, M., Olsson, E.K.: Facebook and Twitter in crisis communication: a comparative
study of crisis communication professionals and citizens. J. Contingencies Crisis Manage.
24(4), 198–208 (2016)
17. Laylavi, F., Rajabifard, A., Kalantari, M.: Event relatedness assessment of Twitter messages
for emergency response. Inf. Process. Manage. 53(1), 266–280 (2017)
18. Banujan, K., Banage Kumara, T.G.S., Paik, I.: Twitter and online news analytics for enhancing
post-natural disaster management activities. In: Proceedings of the 9th International Conference
on Awareness Science and Technology (iCAST), pp. 302–307. Fukuoka (2018)
19. Takahashi, B., Tandoc, E.C., Carmichael, C.: Communicating on Twitter during a disaster:
an analysis of tweets during typhoon Haiyan in the Philippines. Comput. Hum. Behav. 50,
392–398 (2015)
20. Teodorescu, H.N.L., Pirnau, M.: Analysis of requirements for SN monitoring applications in
disasters—a case study. In: Proceedings of the 8th International Conference on Electronics,
Computers and Artificial Intelligence (ECAI), pp. 1–6. Ploiesti, Romania (2016)
896 M. Pirnau
21. Ahmed, W., Bath, P.A., Sbaffi, L., Demartini, G.: Novel insights into views towards H1N1
during the 2009 pandemic: a thematic analysis of Twitter data. Health Inf. Libr. J. 36, 60–72
(2019)
22. Asadzadeh, A., Kötter, T., Salehi, P., Birkmann, J.: Operationalizing a concept: the systematic
review of composite indicator building for measuring community disaster resilience. Int. J.
Disaster Risk Reduction 25, 147–162 (2017)
23. Teodorescu, H.N.L., Saharia, N.: A semantic analyzer for detecting attitudes on SNs. In:
Proceedings of the International Conference on Communications (COMM), pp. 47–50.
Bucharest, Romania (2016)
24. Teodorescu, H.N.L.: On the responses of social networks’ to external events. In: Proceedings of
the 7th International Conference on Electronics, Computers and Artificial Intelligence, pp. 13–
18. Bucharest, Romania (2015)
25. Gottfried, J., Shearer, E.: News use across social media platforms 2016. White Paper, 26. Pew
Research Center (2016)
26. Gupta, A., Lamba, H., Kumaraguru, P., Joshi, A.: Faking sandy: characterizing and identi-
fying fake images on twitter during hurricane sandy. In WWW’13 Proceedings of the 22nd
International Conference on World Wide Web, pp. 729–736 (2013)
27. Allcott, H., Gentzkow, M.: Social media and fake news in the 2016 election. J. Econ. Perspect.
31(2), 211–236 (2017)
28. Lyu, H., Chen, L., Wang, Y., Luo, J.: Sense and sensibility: characterizing social media users
regarding the use of controversial terms for COVID-19. IEEE Trans. Big Data (2020)
29. Teodorescu, H.N.L., Bolea, S.C.: On the algorithmic role of synonyms and keywords in
analytics for catastrophic events. In: Proceedings of the 8th International Conference on
Electronics, Computers and Artificial Intelligence, ECAI, pp. 1–6. Ploiesti, Romania (2016)
30. Teodorescu, H.N.L.: Emergency-related, social network time series: description and anal-
ysis. In: Rojas, I., Pomares, H. (eds.) Time Series Analysis and Forecasting. Contributions
to Statistics, pp. 205–215. Springer, Cham (2016)
31. Bolea, S.C.: Vocabulary, synonyms and sentiments of hazard-related posts on social networks.
In: Proceedings of the 8th Conference Speech Technology and Human-Computer Dialogue
(SpeD), pp. 1–6. Bucharest, Romania (2015)
32. Bolea, S.C.: Language processes and related statistics in the posts associated to disasters on
social networks. Int. J. Comput. Commun. Control 11(5), 602–612 (2016)
33. Teodorescu, H.N.L.: Survey of IC&T in disaster mitigation and disaster situation manage-
ment, Chapter 1. In: Teodorescu, H.-N., Kirschenbaum, A., Cojocaru, S., Bruderlein, C. (eds.),
Improving Disaster Resilience and Mitigation—IT Means and Tools. NATO Science for Peace
and Security Series—C, pp. 3–22. Springer, Dordrecht (2014)
34. Kanis, J., Skorkovská, L.: Comparison of different lemmatization approaches through the
means of information retrieval performance. In: Proceedings of the 13th International
Conference on Text, Speech and Dialogue TSD’10, pp. 93–100 (2010)
35. Ferrucci, D., Lally, A.: UIMA: an architectural approach to unstructured information processing
in the corporate research environment. Nat. Lang. Eng. 10(3–4), 327–348 (2004)
36. Jacobs, P.S.: Joining statistics with NLP for text categorization. In: Proceedings of the Third
Conference on Applied Natural Language Processing, pp. 178–185 (1992)
37. Jivani, A.G.: A comparative study of stemming algorithms. Int. J Comp Tech. Appl 2, 1930–
1938 (2011)
38. Ingason, A.K., Helgadóttir, S., Loftsson, H., Rögnvaldsson, E.: A mixed method lemmatization
algorithm using a hierarchy of linguistic identities (HOLI). In: Raante, A., Nordström, B. (eds.),
Advances in Natural Language Processing. Lecture Notes in Computer Science, vol. 5221,
pp. 205–216. Springer, Berlin (2008)
39. Krouska, A., Troussas, C., Virvou, M.: The effect of preprocessing techniques on Twitter senti-
ment analysis. In: Proceedings of the International Conference on Information, Intelligence,
Systems & Applications, pp. 13–15. Chalkidiki, Greece (2016)
40. Babanejad, N., Agrawal, A., An, A., Papagelis, M.: A comprehensive analysis of preprocessing
for word representation learning in affective tasks. In: Proceedings of the 58th Annual Meeting
of the Association for Computational Linguistics, pp. 5799–5810 (2020)
41. Camacho-Collados, J., Pilehvar, M.T.: On the role of text preprocessing in neural network
architectures: an evaluation study on text categorization and sentiment analysis. In: Proceedings
of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks
for NLP, pp. 40–46. Association for Computational Linguistics (2018)
42. Davis, C.A., Varol, O., Ferrara, E., Flammini, A., Menczer, F.: BotOrNot: a system to evaluate
social bots, a system to evaluate social bots. In: Proceedings of the 25th International Conference
Companion on World Wide Web, pp. 273–274 (2016)
43. Ferrara, E.: COVID-19 on Twitter: Bots, Conspiracies, and Social Media Activism. arXiv
preprint arXiv:2004.09531 (2020)
44. Metaxas, P., Finn, S.T.: The infamous#Pizzagate conspiracy theory: Insight from a Twitter
Trails investigation. Wellesley College Faculty Research and Scholarship (2017)
45. Teodorescu, H.N.L.: Social signals and the ENR index—noise of searches on SN with keyword-
based logic conditions. In: Proceedings of the International Symposium on Signals, Circuits
and Systems. Iasi, Romania (2015)
46. Aouragh, S.I.: Adaptating Levenshtein distance to contextual spelling correction. Int. J.
Comput. Sci. Appl. 12(1), 127–133 (2015)
47. Kobzdej, P.: Parallel application of Levenshtein’s distance to establish similarity between
strings. Front. Artif. Intell. Appl. 12(4) (2003)
48. Rani, S.; Singh, J.: Enhancing Levenshtein’s edit distance algorithm for evaluating document
similarity. In: Communications in Computer and Information Science, pp. 72–80. Springer,
Singapore (2018)
Author Index
A Benghabrit, Asmaa, 705

Abdoun, Otman, 231 Beni-Hssane, Abderrahim, 311, 845
Abid, Meriem, 423 Benmammar, Badr, 117
Abid, Mohamed, 327, 829 Benouini, Rachid, 737
Abouchabaka, Jâafar, 231 Bensaali, Faycal, 603
Abou El Hassan, Adil, 177 Bensalah, Nouhaila, 87
Abourahim, Ikram, 215 Bentaleb, Youssef, 343
Adib, Abdellah, 87 Berehil, Mohammed, 561
Ahmed, Srhir, 549 Bessaoud, Karim, 423
Ali Pacha, Adda, 635 Bihri, Hayat, 873
Alsalemi, Abdullah, 603 Birjali, Marouane, 311, 845
Amadid, Jamal, 147, 161 Bohorma, Mohamed, 795
Amghar, Mustapha, 215 Bordjiba, Yamina, 75
Amira, Abbes, 603 Bouattane, Omar, 59
Ammar, Abderazzak, 59 Boubaker, Mechab, 33
Asaidi, Hakima, 691 Boudhir, Anouar Abdelhakim, 577
Asri, Bouchra El, 45 Bouhaddou, Imane, 705
Assayad, Ismail, 815 Bouhdidi El, Jaber, 197
Ayad, Habib, 87 Bouhorma, Mohammed, 19, 453
Azzouzi, Salma, 509, 873 Boulmakoul, Azedine, 243, 439, 749
Boulouird, Mohamed, 147, 161
Bounabat, Bouchaib, 619
B
Bouzebiba, Hadjer, 133
Baba-Ahmed, Mohammed Zakarya, 117
Bamiro, Bolaji, 815
Bannari, Rachid, 257
Barhoun, Rabie, 465 C
Belkadi, Fatima Zahra, 667 Cavalli-Sforza, Violetta, 763
Belkasmi, Mohammed Ghaouth, 409 Chadli, Sara, 409
Bella, Kaoutar, 243 Chaibi, Hasna, 275
Bellouki, Mohamed, 691 Charaf, My El Hassan, 509, 873
Benabdellah, Abla Chaouni, 705 Chehri, Abdellah, 275
Ben Abdel Ouahab, Ikram, 453 Cherradi, Ghyzlane, 439
Ben Ahmed, Mohamed, 577 Cherradi, Mohamed, 679
Bencheriet, Chemesse Ennehar, 75 Chillali, Abdelhakim, 351
© The Editor(s) (if applicable) and The Author(s), under exclusive license 899
to Springer Nature Singapore Pte Ltd. 2022
https://doi.org/10.1007/978-981-16-3637-0
900 Author Index
D Hsaini, Sara, 873

Dadi, Sihem, 327
Dhaiouir, Ilham, 521
I
Ibnyaich, Saida, 287
E
Eddabbah, Mohamed, 301
Ed-daibouni, Maryam, 465 J
Eddoujaji, Mohamed, 795 Jadouli, Ayoub, 3
Elaachak, Lotfi, 453
El Allali, Naoufal, 691
El Amrani, Chaker, 3 K
El-Ansari, Anas, 311 Karim, Lamia, 439
Elboukhari, Mohamed, 381 Kerrakchou, Imane, 409
Eleuldj, Mohsine, 215 Khabba, Asma, 287
El Fadili, Hakim, 737 Khaldi, Mohamed, 521
El Gourari, Abdelali, 535 Khankhour, Hala, 231
EL Haddadi, Anass, 679 Korachi, Zineb, 619
Elkafazi, Ismail, 257 Kossingou, Ghislain Mervyl, 653
Elkaissi, Souhail, 749
El Kamel, Nadiya, 301
El Kettani, Mohamed El Youssfi, 103 L
EL Makhtoum, Hind, 343 Lafifi, Yassine, 479
El Mehdi, Abdelmalek, 177 Lakhouaja, Abdelhak, 763
El Ouariachi, Ilham, 737 Lamia, Mahnane, 479
El Ouesdadi, Nadia, 495 Lamlili El Mazoui Nadori, Yasser, 593
Errahili, Sanaa, 287 Lmoumen, Youssef, 301
Esbai, Redouane, 593, 667 Loukil, Abdelhamid, 635
Ezziyyani, Mostafa, 521
M
F Maamri, Ramdane, 775
Fariss, Mourad, 691 Mabrek, Zahia, 75
Farouk, Abdelhamid Ibn El, 87 Mandar, Meriem, 439
Ftaimi, Asmaa, 393 Mauricio, David, 365
Mazri, Tomader, 393, 549
M’dioud, Meriem, 257
G Mehalli, Zoulikha, 635
Ghalbzouri El, Hind, 197 Mikram, Mounia, 45
Gouasmi, Noureddine, 479 Mouanis, Hakima, 351
Grini, Abdelâli, 351
Grissette, Hanane, 859
N
Nait Bahloul, Sarah, 423
H Nassiri, Naoual, 763
Habri, Mohamed Achraf, 593 Ndassimba, Edgard, 653
Hadj Abdelkader, Oussama, 133 Ndassimba, Nadege Gladys, 653
Hammou, Djalal Rafik, 33 Nejjari, Rachid, 873
Hankar, Mustapha, 311, 845 Nfaoui, El Habib, 859
Hannad, Yaâcoub, 103
Harous, Saad, 775
Hassani, Moha Mâe™Rabet, 147, 161 O
Himeur, Yassine, 603 Olufemi, Adeosun Nehemiah, 723
Houari, Nadhir, 117 Ouafiq, El Mehdi, 275
Author Index 901
Ounasser, Nabila, 45 Slalmi, Ahmed, 275

Ouya, Samuel, 653 Smaili, El Miloud, 509
Soussi Niaimi, Badr-Eddine, 19
Sraidi, Soukaina, 509
P
Pirnau, Mironela, 885
T
Torres-Calderon, Hector, 365
R Touahni, Raja, 301
Raoufi, Mustapha, 535
Rhanoui, Maryem, 45
Riadi, Abdelhamid, 147, 161
V
Rochdi, Sara, 495
Velasquez, Marco, 365
Routaib, Hayat, 679
S Y
Saadane, Rachid, 275 Youssfi, Mohamed, 59
Saadna, Youness, 577
Saber, Mohammed, 177, 409
Sah, Melike, 723 Z
Samadi, Hassan, 795 Zarghili, Arsalane, 737
Sassi, Mounira, 829 Zekhnini, Kamar, 705
Sayed, Aya, 603 Zenkouar, Khalid, 737
Sbai, Oussama, 381 Zeroual, Abdelouhab, 287
Seghiri, Naouel, 117 Zigh, Ehlem, 635
Semma, Abdelillah, 103 Zili, Hassan, 19
Skouri, Mohammed, 535 Zitouni, Farouq, 775

Networking, Intelligent Systems and Security (Proceedings of NISS 2021)

Uploaded by

Copyright:

Available Formats

Networking, Intelligent Systems and Security (Proceedings of NISS 2021)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Networking, Intelligent Systems and Security (Proceedings of NISS 2021)

Uploaded by

Copyright:

Available Formats

Smart Innovation, Systems and Technologies 237

Mohamed Ben Ahmed

More information about this series at http://www.springer.com/series/8767

Anouar Abdelhakim Boudhir

ISSN 2190-3018 ISSN 2190-3026 (electronic)

Tomader Mazri, ENSA, Ibn Tofail University, Kenitra, Morocco

Conference General Chairs

Mohamed Ben Ahmed, FST, Tangier UAE University, Morocco

Conference Technical Programme Committee Chair

Wassila Mtalaa, Luxembourg Institute of Science and Technology, Luxembourg

Keynote and Panels Chair

Domingos Santos, Polytechnic Institute Castelo Branco, Portugal

İsmail Rakıp Karas, o, Karabuk University

Special Issues Chair

Senthil Kumar, Hindustan College of Arts and Science, India

Local Organizing Committee

Tomader Mazri, ENSA, UIT, Morocco

Technical Programme Committee

Ismail Rakip Karas, Karabuk University, Türkiye

Ben Yahya Sadok, Faculty of Sciences of Tunis, Tunisia

Loncaric Sven, Zagreb University, Croatia

In an age of explosive worldwide growth of electronic data storage and communica-

Tangier, Morocco Mohamed Ben Ahmed

Artificial Intelligence for Sustainability

Network and Cloud Technologies

Smart Digital Learning

Smart Information Systems

Smart Agriculture Solution Based on IoT and TVWS for Arid

Security and Privacy Protection in the e-Health System: Remote

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 899

Mohamed Ben Ahmed is an associate professor of computer sciences at Abdel-

Prof. Horia-Nicolai L. Teodorescu teaches intelligent systems at Gheorghe

of Applied Sciences of Kenitra, a permanent member of Electrical and Telecom-

Parthasarathy Subashini has also received Ph.D. in Computer Science in 2009

Anouar Abdelhakim Boudhir is currently an associate professor at the Faculty

Ayoub Jadouli and Chaker El Amrani

1.1 Wildfire and Machine Learning

A. Jadouli (B) · C. El Amrani

1.2 Deep Learning and Remote Sensing

3 Proposed Solution and Methods

3.1 Architecture and Implementation of the Design

3.2 Conventional Neural Network Model

3.3 Transfer Learning

Transfer learning is a technique where information learned in a primary model of

Fig. 3 Approach 1 and

Table 1 Approach 1 model’s

Table 2 Approach 2 model’s

Table 3 Approach 3 model’s

Fig. 4 Approach 3 model’s

Fig. 5 Sample images from UC merced land dataset

Fig. 6 Sample images from the approach 1 dataset

See Table 4; Fig. 9.

See Table 5; Fig. 10.

Fig. 7 Sample images from

Fig. 8 Sample images from the approach 3 dataset

Fig. 9 Approach 1 validation accuracy by number of epochs

Fig. 10 Approach 2 validation accuracy by number of epochs

Badr-Eddine Soussi Niaimi, Mohammed Bouhorma, and Hassan Zili

B.-E. Soussi Niaimi (B) · M. Bouhorma · H. Zili

The data-driven approach is a relatively new approach, because it is a data-based