Icicc 2017

Advances in Intelligent Systems and Computing 673
Subhash Bhalla
Vikrant Bhateja
Anjali A. Chandavale
Anil S. Hiwale
Suresh Chandra Satapathy Editors
Intelligent
Computing and
Information and
Communication
Proceedings of 2nd International
Conference, ICICC 2017
Advances in Intelligent Systems and Computing
Volume 673
Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: kacprzyk@ibspan.waw.pl
The series “Advances in Intelligent Systems and Computing” contains publications on
theory, applications, and design methods of Intelligent Systems and Intelligent Computing.
Virtually all disciplines such as engineering, natural sciences, computer and information
science, ICT, economics, business, e-commerce, environment, healthcare, life science are
covered. The list of topics spans all the areas of modern intelligent systems and computing.
The publications within “Advances in Intelligent Systems and Computing” are primarily
textbooks and proceedings of important conferences, symposia and congresses. They cover
significant recent developments in the field, both of a foundational and applicable character.
An important characteristic feature of the series is the short publication time and world-wide
distribution. This permits a rapid and broad dissemination of research results.
Advisory Board
Chairman
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
e-mail: nikhil@isical.ac.in
Members
Rafael Bello Perez, Universidad Central “Marta Abreu” de Las Villas, Santa Clara, Cuba
e-mail: rbellop@uclv.edu.cu
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
e-mail: escorchado@usal.es
Hani Hagras, University of Essex, Colchester, UK
e-mail: hani@essex.ac.uk
László T. Kóczy, Széchenyi István University, Győr, Hungary
e-mail: koczy@sze.hu
Vladik Kreinovich, University of Texas at El Paso, El Paso, USA
e-mail: vladik@utep.edu
Chin-Teng Lin, National Chiao Tung University, Hsinchu, Taiwan
e-mail: ctlin@mail.nctu.edu.tw
Jie Lu, University of Technology, Sydney, Australia
e-mail: Jie.Lu@uts.edu.au
Patricia Melin, Tijuana Institute of Technology, Tijuana, Mexico
e-mail: epmelin@hafsamx.org
Nadia Nedjah, State University of Rio de Janeiro, Rio de Janeiro, Brazil
e-mail: nadia@eng.uerj.br
Ngoc Thanh Nguyen, Wroclaw University of Technology, Wroclaw, Poland
e-mail: Ngoc-Thanh.Nguyen@pwr.edu.pl
Jun Wang, The Chinese University of Hong Kong, Shatin, Hong Kong
e-mail: jwang@mae.cuhk.edu.hk
More information about this series at http://www.springer.com/series/11156

Subhash Bhalla Vikrant Bhateja
•
Anjali A. Chandavale Anil S. Hiwale

•
Suresh Chandra Satapathy

Editors
Intelligent Computing
and Information
and Communication
Proceedings of 2nd International Conference,
ICICC 2017
123
Editors
Subhash Bhalla Anil S. Hiwale
Department of Computer Software Department of Information Technology
University of Aizu MIT College of Engineering
Aizuwakamatsu, Fukushima Pune, Maharashtra
Japan India
Vikrant Bhateja Suresh Chandra Satapathy

Department of Electronics and Department of Computer Science
Communication Engineering and Engineering
Shri Ramswaroop Memorial Group of Anil Neerukonda Institute
Professional Colleges of Technology and Sciences
Lucknow, Uttar Pradesh Visakhapatnam, Andhra Pradesh
India India
Anjali A. Chandavale
Department of Information Technology
MIT College of Engineering
Pune, Maharashtra
India
ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-981-10-7244-4 ISBN 978-981-10-7245-1 (eBook)
https://doi.org/10.1007/978-981-10-7245-1
Library of Congress Control Number: 2017959316
© Springer Nature Singapore Pte Ltd. 2018

This work is subject to copyright. All rights are reserved 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, express 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.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore
Preface
The second International Conference on Intelligent Computing and Communication

(ICICC 2017) was successfully organized by Dr. Vishwanath Karad, MIT World
Peace University, Pune, during August 2–4, 2017, at MIT College of Engineering,
Pune, India. The conference had technical collaboration with Computer Society of
India and College of Engineering Pune and was supported by All India Council for
Technical Education (AICTE) and Council of Scientific and Industrial Research
(CSIR). The objective of this international conference was to provide a platform for
academicians, researchers, scientists, professionals, and students to share their
knowledge and expertise in the field of intelligent computing, communication, and
convergence and address various issues to increase awareness of technological
innovations and to identify challenges and opportunities for the development of
smart cities using multidisciplinary research techniques. Research submissions in
various advanced technology areas were received, and after a rigorous peer-review
process with the help of program committee members and external reviewers, only
quality papers were accepted. The conference featured eight special sessions on
various cutting-edge technologies which were conducted by eminent professors and
many distinguished academicians like Dr. Vijay Bhatkar, Padma Bhushan,
Chancellor of Nalanda University, Bihar, India; Dr. Hemant Darbari, Executive
Director, CDAC, Pune, India; Dr. Bipin Indurkhya, Jagiellonian University,
Kraków, Poland; and Dr. Subhash Bhalla, University of Aizu, Japan.
Our sincere thanks to all special session chairs Prof. Dr. Priti Rege,
Prof. Dr. R. A. Patil, Dr. Anagha Kulkarni and distinguished reviewers for their
timely technical support. We would like to extend our special thanks here to our
v
vi Preface
publication chairs for doing a great job in making the conference widely visible.
Thanks to dynamic team members for organizing the event in a smooth manner.
Our sincere thanks to all sponsors, press, print, and electronic media for their
excellent coverage of this conference.
Aizuwakamatsu, Japan Subhash Bhalla

Lucknow, India Vikrant Bhateja
Pune, India Anjali A. Chandavale
Pune, India Anil S. Hiwale
Visakhapatnam, India Suresh Chandra Satapathy
Organizing Committee
Chief Patrons
Prof. Dr. Vishwanath Karad, India
Prof. Dr. Vijay Bhatkar, India
Patrons
Prof. Dr. Mangesh Karad, India
Prof. Dr. Rahul Karad, India
Organizing Chair
Dr. Ramchandra Pujeri, India
Organizing Co-chair
Dr. Anil Hiwale, India
Advisory Committee
Dr. Seeram Ramakrishna, Singapore
Dr. Subhash Bhalla, Japan
Dr. Shirish Sane, India
Dr. Suresh Borkar, USA
Dr. Shabnam Ivkovic, Canada
TPC Chair
Dr. Anjali A. Chandavale, India
Publication Chair
Dr. Suresh Chandra Satapathy, Vijayawada, India
vii
Contents
35 W GaN Solid-State Driver Power Amplifier for L-Band Radar

Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Vivek Ratnaparkhi and Anil Hiwale
Validation of Open Core Protocol by Exploiting Design Framework
Using System Verilog and UVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Gopika Rani Alekhya Pamarthy, M. Durga Prakash
and Avinash Yadlapati
Cellular Automata Logic Block Observer Based Testing for
Network-on-Chip Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Shaik Mohammed Waseem and Afroz Fatima
Structural Strength Recognizing System with Efficient Clustering
Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Sumedha Sirsikar and Manoj Chandak
Trajectory Outlier Detection for Traffic Events: A Survey . . . . . . . . . . . 37
Kiran Bhowmick and Meera Narvekar
A Privacy-Preserving Approach to Secure Location-Based Data . . . . . . 47
Jyoti Rao, Rasika Pattewar and Rajul Chhallani
Analysis of Blind Image Watermarking Algorithms . . . . . . . . . . . . . . . . 57
Chhaya S. Gosavi and Suresh N. Mali
ClustMap: A Topology-Aware MPI Process Placement Algorithm for
Multi-core Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
K. B. Manwade and D. B. Kulkarni
Mobile Agent-Based Frequent Pattern Mining for Distributed
Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Yashaswini Joshi, Shashikumar G. Totad, R. B. Geeta
and P. V. G. D. Prasad Reddy
ix
x Contents
A Hybrid Approach for Preprocessing of Imbalanced Data in Credit

Scoring Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Uma R. Salunkhe and Suresh N. Mali
Vision-Based Target Tracking Intelligent Robot Using NI myRIO with
LabVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Anita Gade and Yogesh Angal
Detection of Misbehaviors Nodes in Wireless Network with Help of
Pool Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Asha Chaudhary and Pournima More
Evaluation of Multi-label Classifiers in Various Domains Using
Decision Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
V. S. Tidake and S. S. Sane
An Effective Multilabel Classification Using Feature Selection . . . . . . . . 129
S. S. Sane, Prajakta Chaudhari and V. S. Tidake
High-Performance Pipelined FFT Processor Based on Radix-22 for
OFDM Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Manish Bansal and Sangeeta Nakhate
An Image Processing Approach to Blood Spatter Source
Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Abhijit Shinde, Ashish Shinde and Deepali Sale
Overlapping Character Recognition for Handwritten Text Using
Discriminant Hidden Semi-Markov Model . . . . . . . . . . . . . . . . . . . . . . . 163
Ashish Shinde and Abhijit Shinde
A Literature Survey on Authentication Using Behavioural Biometric
Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Rajvardhan Oak
Number System Oriented Text Steganography in English Language
for Short Messages: A Decimal Approach . . . . . . . . . . . . . . . . . . . . . . . 183
Kunal Kumar Mandal, Santanu Koley and Saptarshi Mondal
Novel Robust Design for Reversible Code Converters and Binary
Incrementer with Quantum-Dot Cellular Automata . . . . . . . . . . . . . . . . 195
Bandan Kumar Bhoi, Neeraj Kumar Misra and Manoranjan Pradhan
Routing with Secure Alternate Path Selection for Limiting the Sink
Relocation and Enhanced Network Lifetime . . . . . . . . . . . . . . . . . . . . . 207
Renuka Suryawanshi, Kajal Kapoor and Aboli Patil
Contents xi
Gene Presence and Absence in Genomic Big Data for Precision

Medicine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Mohamood Adhil, Mahima Agarwal, Krittika Ghosh, Manas Sule
and Asoke K. Talukder
A Survey on Service Discovery Mechanism . . . . . . . . . . . . . . . . . . . . . . 227
Gitanjali Shinde and Henning Olesen
Analysis of Multiple Features and Classifier Techniques Combination
for Image Pattern Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Smart and Precision Polyhouse Farming Using Visible Light
Communication and Internet of Things . . . . . . . . . . . . . . . . . . . . . . . . . 247
Krishna Kadam, G. T. Chavan, Umesh Chavan, Rohan Shah
and Pawan Kumar
Acceleration of CNN-Based Facial Emotion Detection Using NVIDIA
GPU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Bhakti Sonawane and Priyanka Sharma
Research Issues for Energy-Efficient Cloud Computing . . . . . . . . . . . . . 265
Nitin S. More and Rajesh B. Ingle
Implementation of REST API Automation for Interaction Center . . . . . 273
Prajakta S. Marale and Anjali A. Chandavale
Predictive Analysis of E-Commerce Products . . . . . . . . . . . . . . . . . . . . . 279
Jagatjyoti G. Tuladhar, Ashish Gupta, Sachit Shrestha, Ujjen Man Bania
and K. Bhargavi
User Privacy and Empowerment: Trends, Challenges, and
Opportunities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Prashant S. Dhotre, Henning Olesen and Samant Khajuria
Evolution Metrics for a BPEL Process . . . . . . . . . . . . . . . . . . . . . . . . . . 305
N. Parimala and Rachna Kohar
Development of Performance Testing Suite Using Apache JMeter . . . . . 317
Jidnyasa Agnihotri and Rashmi Phalnikar
Characterizing Network Flows for Detecting DNS, NTP, and SNMP
Anomalies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
Rohini Sharma, Ajay Guleria and R. K. Singla
Periocular Region Based Biometric Identification Using the Local
Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
K. Kishore Kumar and P. Trinatha Rao
xii Contents
Model-Based Design Approach for Software Verification Using

Hardware-in-Loop Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
Pranoti Joshi and N. B. Chopade
Silhouette-Based Human Action Recognition by Embedding HOG and
PCA Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
A. S. Jahagirdar and M. S. Nagmode
Unified Algorithm for Melodic Music Similarity and Retrieval in
Query by Humming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
Velankar Makarand and Kulkarni Parag
Predict Stock Market Behavior: Role of Machine Learning
Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
Uma Gurav and Nandini Sidnal
Stability of Local Information-Based Centrality Measurements Under
Degree Preserving Randomizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
Chandni Saxena, M. N. Doja and Tanvir Ahmad
Hybrid Solution for E-Toll Payment . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Ajinkya R. Algonda, Rewati R. Sonar and Saranga N. Bhutada
Enhancing Distributed Three Hop Routing Protocol in Hybrid
Wireless Network Through Data Weight-Based Scheme . . . . . . . . . . . . 415
Neha S. Rathod and Poonam Gupta
VLSI-Based Data Hiding with Transform Domain Module Using
FPGA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Latika R. Desai and Suresh N. Mali
A Novel Approach of Frequent Itemset Mining Using HDFS
Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
Prajakta G. Kulkarni and S. R. Khonde
Issues of Cryptographic Performance in Resource-Constrained
Devices: An Experimental Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Balaso Jagdale and Jagdish Bakal
Assessment of Object Detection Using Deep Convolutional Neural
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Ajeet Ram Pathak, Manjusha Pandey, Siddharth Rautaray
and Karishma Pawar
Contents xiii
Implementation of Credit Card Fraud Detection System with Concept

Drifts Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
Anita Jog and Anjali A. Chandavale
Intelligent Traffic Control by Multi-agent Cooperative Q Learning
(MCQL) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
Deepak A. Vidhate and Parag Kulkarni
Digital Tokens: A Scheme for Enabling Trust Between Customers and
Electronic Marketplaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
Balaji Rajendran, Mohammed Misbahuddin, S. Kaviraj
and B. S. Bindhumadhava
BMWA: A Novel Model for Behavior Mapping for Wormhole
Adversary Node in MANET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
S. B. Geetha and Venkanagouda C. Patil
AWGN Suppression Algorithm in EMG Signals Using Ensemble
Empirical Mode Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
Ashita Srivastava, Vikrant Bhateja, Deepak Kumar Tiwari
and Deeksha Anand
Visible-Infrared Image Fusion Method Using
Anisotropic Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
Ashutosh Singhal, Vikrant Bhateja, Anil Singh
and Suresh Chandra Satapathy
Fast Radial Harmonic Moments for Invariant Image
Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
Shabana Urooj, Satya P. Singh, Shevet Kamal Maurya
and Mayank Priyadarshi
A Wide-Area Network Protection Method Using PMUs . . . . . . . . . . . . . 539
Namita Chandra and Shabana Urooj
Analysis and Prediction of the Effect of Surya Namaskar on Pulse of
Different Prakruti Using Machine Learning . . . . . . . . . . . . . . . . . . . . . . 547
Jayshree Ghorpade-Aher, Abhishek Girish Patil, Eeshan Phatak,
Sumant Gaopande and Yudhishthir Deshpande
Cognitive Depression Detection Methodology Using EEG Signal
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
Sharwin P. Bobde, Shamla T. Mantri, Dipti D. Patil and Vijay Wadhai
xiv Contents
Biogas Monitoring System Using DS18B20 Temperature Sensor and

MQTT Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
Suruchi Dedgaonkar, Aakankssha Kaalay, Nitesh Biyani
and Madhuri Mohite
Time-Efficient and Attack-Resistant Authentication Schemes in
VANET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579
Sachin Godse and Parikshit Mahalle
Inferring User Emotions from Keyboard and Mouse . . . . . . . . . . . . . . . 591
Taranpreet Singh Saini and Mangesh Bedekar
An Overview of Automatic Speaker Verification System . . . . . . . . . . . . 603
Ravika Naika
Topic Modeling on Online News Extraction . . . . . . . . . . . . . . . . . . . . . . 611
Aashka Sahni and Sushila Palwe
Spam Mail Detection Using Classification Techniques and Global
Training Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
Vishal Kumar Singh and Shweta Bhardwaj
Smart Gesture Control for Home Automation Using Internet of
Things . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
Sunil Kumar Khatri, Govind Sharma, Prashant Johri and Sachit Mohan
A Proposed Maturity Model for Himachal Pradesh Government
e-Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
Alpana Kakkar, Seema Rawat, Piyush Gupta and Sunil Kumar Khatri
Malaria Detection Using Improved Fuzzy Algorithm . . . . . . . . . . . . . . . 653
Mukul Sharma, Rajat Mittal, Tanupriya Choudhury,
Suresh Chand Satapathy and Praveen Kumar
Feature Extraction Techniques Based on Human Auditory System . . . . 667
Sanya Jain and Divya Gupta
Movie Recommendation System: Hybrid Information Filtering
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
Kartik Narendra Jain, Vikrant Kumar, Praveen Kumar
and Tanupriya Choudhury
Eco-Friendly Green Computing Approaches for Next-Generation
Power Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687
Seema Rawat, Richa Mishra and Praveen Kumar
Implementing Test Automation Framework Using Model-Based
Testing Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695
Japneet Singh, Sanjib Kumar Sahu and Amit Prakash Singh
Contents xv
Improved Exemplar-Based Image Inpainting Approach . . . . . . . . . . . . 705

Hitesh Kumar, Shilpi Sharma and Tanupriya Choudhury
The Hidden Truth Anonymity in Cyberspace: Deep Web . . . . . . . . . . . 719
Saksham Gulati, Shilpi Sharma and Garima Agarwal
Test Case Optimization and Prioritization of Web Service Using
Bacteriologic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
Gaurav Raj, Dheerendra Singh and Ishita Tyagi
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745
About the Editors
Prof. Subhash Bhalla joined as a Faculty School of Computer and Systems

Sciences, Jawaharlal Nehru University (JNU), New Delhi, in 1986. He was a
Visiting Scientist at Sloan School of Management, Massachusetts Institute of
Technology (MIT), Cambridge, Massachusetts, USA (1987–1988). He is a Member
of the Computer Society of IEEE and SIGMOD of ACM. He is associated with the
Department of Computer Software at the University of Aizu, Japan. He has also
toured and lectured at many industries for conducting feasibility studies and for the
adoption of modern techniques. He has received several grants for research projects.
He currently participates in the Intelligent Dictionary System Project. He is
exploring database designs to support models for Information Interchange through
the World Wide Web. He is working with a study team on creating user interfaces
for Web users and transaction management system for mobile computing. He is
studying transaction management and algorithmic designs for distributed real-time
systems. He is also pursuing performance evaluation and modeling of distributed
algorithms. His research interests include managing components and application
services; distributed client/middleware/service-oriented computing; XML,
e-commerce; mobile database management systems; Web query and Web data
mining; synchronization and crash recovery; and integration of technologies.
Prof. Vikrant Bhateja is Associate Professor, Department of Electronics and
Communication Engineering, Shri Ramswaroop Memorial Group of Professional
Colleges (SRMGPC), Lucknow, and also the Head (Academics and Quality
Control) in the same college. His areas of research include digital image and video
processing, computer vision, medical imaging, machine learning, pattern analysis
and recognition, neural networks, soft computing, and bio-inspired computing
techniques. He has more than 90 quality publications in various international
journals and conference proceedings. He has been on TPC and chaired various
sessions from the above domain in the international conferences of IEEE and
Springer. He has been the track chair and served in the core technical/editorial
teams for the international conferences: FICTA 2014, CSI 2014 and INDIA 2015
under Springer-AISC Series, and INDIACom-2015 and ICACCI-2015 under IEEE.
xvii
xviii About the Editors
He is Associate Editor in International Journal of Convergence Computing

(IJConvC) and also serving in the editorial board of International Journal of Image
Mining (IJIM) under Inderscience Publishers. At present, he is guest editor for two
special issues floated in International Journal of Rough Sets and Data Analysis
(IJRSDA) and International Journal of System Dynamics Applications (IJSDA)
under IGI Global Publications.
Dr. Anjali A. Chandavale is currently working as Associate Professor
(Information Technology) in Maharashtra Academy of Engineering and Educational
Research’s Maharashtra Institute of Technology College of Engineering, Pune,
India. She has over 18 years of teaching and over 6 years of industrial experience in
the field of process instrumentation. Her research interests include Internet security,
image processing, and embedded systems. She has worked as Research and
Development Engineer in Lectrotek Systems (Pune) Pvt Ltd, Pune, for 4 years and
also as a Software Engineer (Design and Development) in Ajay Electronics Pvt Ltd
(1997–1998), Pune. She has several publications in journals as well as in conference
proceedings of international repute.
Prof. Anil S. Hiwale is currently working as Professor and Head of Department of
Information Technology, MIT College of Engineering, Pune. He completed his Ph.D.
from SGB Amravati University, Amravati. Prof. Hiwale has 28 years of teaching
experience and published over 50 research papers in National/International Journals
and conferences. His areas of research include signal processing, digital communi-
cations, multi-antenna systems and satellite communications. He is presently
supervising four Ph.D. scholars. He is a Fellow of Institution of Engineers (India) and
a member of IET, CSI, ISTE and Broadcast Engineering Society of India.
Prof. Suresh Chandra Satapathy is currently working as Professor and Head,
Department of Computer Science and Engineering, PVP Siddhartha Institute of
Technology, Andhra Pradesh, India. He obtained his Ph.D. in Computer Science
and Engineering from JNTU Hyderabad and M.Tech. in CSE from NIT Rourkela,
Odisha, India. He has 26 years of teaching experience. His research interests
include data mining, machine intelligence, and swarm intelligence. He has acted as
program chair of many international conferences and edited six volumes of pro-
ceedings from Springer LNCS and AISC series. He is currently guiding eight
scholars for Ph.D. He is also a Senior Member of IEEE.
35 W GaN Solid-State Driver Power
Amplifier for L-Band Radar Applications
Vivek Ratnaparkhi and Anil Hiwale
Abstract In this paper, 35 W driver power amplifier was designed and simulated
using GaN HEMT for L-band radar. GaN HEMT is used because it can provide
high output power and high gain as compared to other semiconductor technologies.
The 35 W output power is generated using CGHV40030 GaN HEMT which is
sufficient to drive further stages of power amplifier. The driver amplifier is designed
at 1.3 GHz of center frequency. This amplifier is designed in class AB and 60.5%
of PAE is achieved.
Keywords Class AB GaN HEMT L-band radar PAE
1 Introduction
Rapid development in the field of gallium nitride (GaN) semiconductor devices

since last two decades is changing the rules of power amplifier design for wireless
communications. To generate and amplify high microwave power, people were
using microwave tubes which are having certain limitations. It is all dependent on
particular application where microwave tube is used for signal generation and
amplification. For microwave applications, traveling wave tube tubes (TWT),
magnetrons, and klystrons were conventionally used. TWT amplifiers have certain
limitations such as high noise, shorter lifespan, and wide bandwidth which results in
interference with other wireless communication system operating in an adjacent
band. Consequently, there is a strong desire for solid-state power amplifiers (SSPA)
that are superior in long-term reliability and signal noise to replace conventional
TWT amplifiers at microwave frequencies [1–3]. GaN devices are becoming a
V. Ratnaparkhi (&)
Department of EXTC, SSGMCE, Shegaon, Maharashtra, India
e-mail: ratnaparkhi_vivek@yahoo.co.in
A. Hiwale
Department of IT, MITCOE, Pune, Maharashtra, India
e-mail: anil.hiwale@mitcoe.edu.in
© Springer Nature Singapore Pte Ltd. 2018 1

S. Bhalla et al. (eds.), Intelligent Computing and Information and Communication,
Advances in Intelligent Systems and Computing 673,
https://doi.org/10.1007/978-981-10-7245-1_1
2 V. Ratnaparkhi and A. Hiwale
Table 1 Comparative analysis of different device types

Device type Pmax Efficiency Gain Bias voltage Operation
(kW) (%) (dB) (kV) time
GaN devices 0.8 50–80 10–20 0.025–0.1 High
Power grid tubes 0.5–10 50–60 10–15 0.5–10 Less
Electron beam 0.1– 25–60 20–45 25–100 Moderate
devices 2000
promising choice for SSPA design for microwave applications. Most of the tran-
sistor manufacturers are now developing GaN high-power transistors with attractive
performance characteristics. Table 1 shows the comparison of power devices family
on the basis of device characteristics. It is apparent that new GaN semiconductor
devices are capable of producing almost 1 KW of output power with the highest
efficiency of 65%. If several stages of solid-state power amplifiers are combined
properly, very high output power in the range of few kilowatts can be generated. It
is also important to note that lifetime of power amplifiers using SSPDs is very high
when compared with the power grid tubes (PGTs) and electron beam devices
(EBDs) such as traveling wave tubes (TWTs) and klystrons. From the reliability
point of view, SSPDs are more reliable than PGTs and EBDs.
GaN solid-state transistors have the potential to disrupt very large vacuum tube
market and can replace some conventional vacuum tubes. There is high demand for
GaN SSPAs for microwave applications.
2 GaN Device Technology
For radar applications, traveling wave tubes (TWT) such as magnetrons and kly-
strons were conventionally used because of the required power level as high as
1 kw [4]. GaN transistors were first demonstrated in the 1990s and have started to
become commercially available in last decade [5]. GaN transistors have many
advantages when it is compared with other semiconductor materials like Si, GaAs,
and SiC. Table 2 compares material characteristics of Si, GaAs, SiC, and GaN. It is
apparent that GaN has higher breakdown voltage which allows GaN HEMTs to
Table 2 Material parameters comparison

Property Si GaAs SiC GaN
Band gap energy (eV) 1.11 1.43 3.2 3.4
Critical breakdown field (MV/cm) 0.3 0.4 3.0 3.0
Thermal conductance (W/cm K) 1.5 0.5 4.9 1.5
Mobility(cm2/V s) 1300 6000 600 1500
Saturated velocity (107 cm/s) 1.0 1.3 2.0 2.7
JFOM versus Si 1.0 1.7 20 27
35 W GaN Solid-State Driver Power Amplifier for L-Band … 3
operate at biasing voltages about 50VDC. Large drain voltages lead to high output
impedance per watt of RF power and result in easier matching circuit design.
Since GaN HEMT devices have higher impedance than other semiconductor
devices, this can help to enhance the PAE and bandwidth of the required SSPA.
High-saturated drift velocity results in higher saturation current densities and watts
per unit periphery. This ensures GaN HEMTs suitability for switched-mode power
amplifiers [6].
It is evident that high-power amplifiers made by Gallium arsenide produce high
efficiency than silicon high-power amplifiers. GaN devices even provide higher
efficiencies than GaAs power amplifiers. Due to such high-efficiency capabilities,
GaN high-power SSPAs are being used extensively in microwave applications.
GaN devices can operate at higher voltage and have higher saturated velocity which
allows GaN transistors to generate higher power in smaller space which results in
high power density. Consequently, smaller size high-power amplifiers can be
fabricated with GaN transistors.
There are many researchers and manufacturers working in the field of
GaN SSPA design for microwave applications. Recently, solid-state power
amplifiers designed using GaN HEMT devices have replaced many conventional
vacuum tubes used for radar and space applications. It is expected that GaN
technology will certainly grow and will be promising choice for high output power
applications at microwave frequencies with high reliability and lesser noise when
compared to conventional vacuum tubes.
3 GaN SSPA Design
Over the last decade, GaN-based high electron mobility transistors (HEMTs) have
emerged as excellent devices for a number of applications [7]. GaN HMTs tech-
nology has been shown to provide high power and high efficiency making it the
perfect candidate for this next-generation radar system [8]. George Solomon [9] and
his group from Communications and Power Industries, LLC Beverly Microwave
Division have reported the VSX3622, a 1.5 kW X-B, and GaN power amplifier for
radar applications. In this section, a step-by-step design of GaN HEMT driver
power amplifier for L-band radar applications is explained.
Figure 1 shows topology of proposed solid-state power amplifier for L-band
radar applications. Using proposed topology, at least 500 W output power can be
achieved. In this paper, design and simulation of GaN solid-state driver power
amplifier are discussed.
Fig. 1 Topology of proposed GaN SSPA
4 Results
In this section, simulation results of GaN driver power amplifier are presented.
Keysigth’s ADS software is used for design and simulation of GaN driver power
amplifier. Since our requirement is to generate at least 30 W output power from
driver stage, CGHV40030 GaN HEMT is selected.
4.1 DC Analysis
Cree’s CGHV40030 GaN HEMT is used for designing driver power amplifier [10].
First stage of PA design is to carry out DC analysis and find out the operating point
of proposed amplifier. Since maximum output power is expected, class AB is the
best choice for proposed driver power amplifier. Upon performing DC analysis in
ADS, VDS = 48 V and VGS = −3 V selected as operating points, which ensures
IDS = 20 mA. The device power consumption at this bias is 88 mW.
4.2 Stability Analysis
It is very important to perform and ensure stability of the device before proceeding
for further design. Any instabilities of the device may generate unwanted oscilla-
tions and amplifier performance will affect significantly. Device may have
unconditionally stable or conditionally stable. There are two stability checks
available to check stability of the device: stability factor and stability measure.
m2
2.4
2.2
m2
freq=1.300GHz
2.0 StabFact1=1.902
StabMeas1
StabFact1
StabMeas1=1.407
1.8
1.6
1.4
1.2
1.00 1.05 1.10 1.15 1.20 1.25 1.30 1.35 1.40 1.45 1.50
freq, GHz
Fig. 2 Stability factor and stability measure
Figure 2 shows stability factor greater than one and stability measure greater than
zero which ensures that device is unconditionally stable. This stability analysis is
performed after proper DC biasing to the device.
4.3 Load Pull Analysis
Load pull analysis is the important step in PA design. Having a good nonlinear
model is essential for starting PA design. We have obtained nonlinear model of
CGHV40030 from device manufacturer. Load pull analysis is carried out using
nonlinear models of the device to find out optimum impedance to be presented to
device for specific output power, PAE, IMD, etc. Once optimum load and source
impedances are found using load pull, impedance matching network is designed.
For CGHV40030 GaN HEMT, load pull analysis is performed and results are
shown in Fig. 3.
4.4 Impedance Matching Network Design and Complete

Amplifier Performance
Once load pull analysis is completed and obtained values for source and load
impedances, input and output impedance matching networks can be designed.
Results of complete amplifier with input and output matching networks after some
optimization are shown in this section. Figure 4a shows output power of designed
GaN SSPA, which is almost 35 W. Figure 4b shows PAE of designed GaN SSPA
and it is acceptable for our proposed driver amplifier.
Fig. 3 Results of load pull simulation at maximum power and gain
(a) 50 m1
40
m1
freq=1.300GHz
dBm(Vout)=45.328
30
dBm(vout)
20
10
-10
-20
0 1 2 3 4 5 6 7
freq, GHz
(b) freq PAE1

1.300 GHz 60.494
Fig. 4 a Output power, b power added efficiency

5 Conclusion
Solid-state driver power amplifier using CGHV40030 GaN HEMT is designed at

1.3 GHz for L-band radar applications. Output power of 45 dBm and PAE of
60.494% are achieved. Load pull analysis is carried out to determine device output
and input impedances. The nonlinear model of the device is obtained from device
manufacturer. Complete GaN SSPA with four parallel cascaded stages with power
divider and combiner can be designed for generating 500 W of output power.
References
1. Allen Katz and Marc Franco “GaN comes of Age” in IEEE microwave magazine, December
2005. pp. 524–534.
2. David Schnaufer “GaN: The Technology of the Future http://www.rfglobalnet.com/doc/gan-
the-technology-of-the-future-0001.
3. A. Katz et al. “A linear GaN UHF SSPA with record high efficiency,” in IEEE MTT-S Int.
Microwave Symp. Dig., Boston, MA, June 7–12, 2009, pp. 769–772.
4. Ken Kikkuchi et al., “An X-Band 300-Watt Class High Power GaN HEMT Amplifier for
Radar Applications”, SEI Technical Review, Number 81, October 2015, pp. 40–44.
5. Francesco Fornetti “Characterisation and performance optimization of GaN HEMTs and
amplifiers for Radar applications” December 2010.
6. Andrew Moore and Jose Jimenez “GaN RF Technology For Dummies”, TriQuint Special
Edition Published by John Wiley & Sons, Inc. 111 River St. Hoboken, NJ 07030-5774,2014.
7. Guest Editorial Special Issue on GaN Electronic Devices “IEEE Transactions on Electron
Devices,” vol. 60, no. 10, October 2013, pp. 2975–2981.
8. T. Thrivikraman and J. Hoffman, “Design of an ultrahigh efficiency GaN high-power
amplifier for SAR remote sensing,” in IEEE Aerospace Conference, pp. 1–6, 2013.
9. George Solomon et al., “The VSX3622, a 1.5 kW X-band GaN Power Amplifier for Radar
Application”, Communications & Power Industries LLC, Beverly Microwave Division.
10. Cree’s Datasheet Online available at: http://www.wolfspeed.com/cghv40030.
Validation of Open Core Protocol
by Exploiting Design Framework Using
System Verilog and UVM
Gopika Rani Alekhya Pamarthy, M. Durga Prakash

and Avinash Yadlapati
Abstract Today’s scenario of semiconductor technology is a tremendous inno-

vation; it includes a large number of intellectual property (IP) cores, interconnects,
or buses in system on chip (SOC) design and based upon the necessity its com-
plexity keeps on increasing. Hence, for the communication between these IP cores,
a standard protocol is developed. The necessity of IP reuse, abridging the design
time and the complexity makes large-scale SOC more challenging in order to
endorse IP core reusability for SOC designs. An efficient non-proprietary protocol
for communication between IP cores is open core protocol (OCP). OCP comes
under socket-based interface and openly licensed core concentric protocol. This
paper addresses on the verification of implemented design of OCP. The proposed
paper is to verify the implemented design by using System Verilog and Universal
Verification Methodology (UVM) in SimVision tool.

Keywords Open core protocol (OCP) Intellectual property (IP)

System on chip (SOC) Socket-based interface Core concentric
1 Introduction
Open core protocol (OCP) is a competent protocol for communication on SoC.

OCP [1] compliance IP cores can be reversed by the designer, depending on system
integration and verification approach in multiple designs without reinstallation,
G. R. A. Pamarthy (&) M. Durga Prakash

Department of Electronics and Communication Engineering, K L University, Guntur,
Andhra Pradesh, India
e-mail: alekhya.rani.gopika@gmail.com
M. Durga Prakash
e-mail: mdprakash82@gmail.com
A. Yadlapati
CYIENT Ltd, Hyderabad, India
e-mail: avinash.amd@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_2
10 G. R. A. Pamarthy et al.
reducing the development time, cost, and design risk. OCP is an interface for
communication between IP cores on an SOC. OCP defines a bus-independent
configurable interface.
OCP renovates IP cores making them independent of the architecture and design
of the systems in which they are used and shortens system verification and testing
by providing a secure boundary around each IP core. OCP is simple, synchronous,
point-to-point, highly scalable, and configurable to match the communication
requirements associated with different IP cores. Even complex high-performance
cores can be accommodated capably with OCP extensions. Cores with OCP
interfaces enable true plug-and-play [2] approach and automated design processes,
thus allowing the system integrator to choose the best cores and best interconnect
system.
Unlike bus approach, with reference to the standard communication approach,
there are mainly two protocols: VCI (Virtual Component Interface) and OCP. OCP
which is scalable and bus independent is a superset of VCI which reports only data
flow aspects; moreover, OCP supports sideband control signaling and tests harness
signals which are configurable. OCP is the only protocol which unifies all the
inter-core communication. OCP establishes a point-to-point interface between two
IP cores [3]. One of them acts as the master who is the controller and generates the
commands and other as the slave responding to commands generated by the master,
either by accepting or giving data to the master.
The OCP defines a point-to-point interface between two communicating entities
such as IP cores and bus interface modules (bus wrappers). One entity acts as the
master of the OCP instance and the other as the slave. Only the master can present
commands and is the controlling entity. The slave responds to commands presented
to it, either by accepting data from the master or presenting data to the master. For
two entities to communicate in a peer-to-peer fashion, there need to be two
instances of the OCP connecting them—one where the first entity is a master, and
one where the first entity is a slave.
2 Project Scope
Implementation is carried out using behavioral Verilog HDL [4] simulation envi-
ronment. The design implements a simple memory read and write operations and
Burst transactions between two IP cores. The design complies with subset of
OCPIP handshake signals. The implemented design, i.e., DUT, is verified using
system Verilog and UVM. As the system Verilog is superset of Verilog and is based
upon OOPS [5] (Object Oriented Programming) concepts, the environment can be
extended without modifying the intention of the original existing code by adding all
the required new features.
Validation of Open Core Protocol by Exploiting Design Framework … 11
3 OCP Protocol
OCP is a configurable protocol that defines one of the communicating entities as

master and other as slave. Master initiates the operation by generating a request
signal to the slave, and in turn slave responds by sending the acknowledgement to
the master. Once master receives the acknowledgement [5] from slave, it transmits
the data to the slave; this phenomenon illustrates the typical handshaking process of
communication as shown in Fig. 1.
3.1 OCP Signal Description
The dataflow signals consist of a small set of required signals and a number of
uncompelled signals that can be configured to support additional requirements [1].
The dataflow signals are grouped into basic signals, burst extensions (support for
bursting). The naming conventions for dataflow signals use the prefix M for signals
driven by the OCP master and S for signals driven by the OCP slave [3].
• Clk: Input clock signal for the OCP clock.
• Maddr: Input address to the master in which data has to be written into the
corresponding memory location and is to be accessible later during read oper-
ation. Maddr width is configurable.
• MCmd: Mode assignment command of width 3. This signal specifies the mode
of OCP which is requested by the master. Usually, depending upon the com-
mand generated by the master, slave responds and performs that operation.
If OCP is in idle state, it does not perform any mode of transfer, whereas during
non-idle state depending on the direction of data flow, OCP [2] performs either
read/write operation.
• Mdata: The data sent by the master that has to be written into the prescribed
memory location of slave which is a configurable one.
• MDatavalid [6]: This is the indication to the slave that the data sent by the
master is valid only. When it is set, it specifies that the Mdata field is valid.
• MRespAccept: Acknowledgment from master that the response from slave was
accepted by it when MRespAccept is set.
Fig. 1 Signal transmission between master and slave

• SCmdAccept: Acknowledgment from slave that the response from master was
accepted by it when SCmdAccept is set.
• Sresp: Response from the slave to a transfer request from the master. It is of
width 2.
• SData: It transfers the requested data by master as read data from the slave to
the master.
• MBurstLength: This illustrates configurable number of transfers in a burst.
MBurstLength of value 0 is illegal.
• MBurstPrecise: This specifies whether the precise burst length is known at the
start of the burst or not.
• MBurstSeq: This 3-bit-width field indicates the sequence of addresses for
requests in a burst.
• SRespLast: Last response from the slave in a burst.
• MReqlast: Last request from the master in a burst.
4 Functional Description
The data transfer between master and slave is communicated non-serially.

Commands are generated by the master to the slave. Based on the command given
by the master, slave responds and decides the mode of operation need to be per-
formed that is requested by the master. Data is either written or read based upon the
command from master into the particular memory location which is specified by the
master [6].
4.1 Master
Master is the one who starts the transactions by providing data and address to the
slave. It is the commander and controller of the entire design. It makes the slave to
function on what it needs. The basic block diagram of master with all the input and
output signal specifications is shown in the figure.
Master activates the OCP by sending command, address, and data to the slave
when clock is active. Slave responds to the master’s command and sends an
acknowledgment to the master indicating the acceptance of command from master
as SCmdAccept [7]. Write/read mode of transmission is performed based on the
master’s request. In order to indicate that the data sent by master is valid, it sends a
valid bit on MDatavalid to the slave; on receiving it, slave starts writing the data in
the corresponding memory location. Similarly, master reads the data from slave by
activating data to be read on SData by specifying the exact memory location.
Master block with all its inputs and outputs driving to slave are shown in Fig. 2.
Fig. 2 OCP master block
4.2 Slave Block
Slave simply performs the operation depending on the command sent by the master.
It activates itself to respond on to the master request. Data handshaking process of
communication is adopted to have a proper and efficient communication [8]. Slave
acknowledges for each and every signal from master as a correspondence or
acceptance of request from the master. The detailed slave block is shown in Fig. 3.
Slave responds to the request sent by the master [9]. It performs read/write mode
of transfer by responding to the signals from master. Slave acknowledges the master
for each and every transfer at enabled clock. During read mode of operation, it
acknowledges the master that the data sent by it is a valid one by sending its
Fig. 3 OCP slave block

response by means of Sresp signal. Both master and slave are controlled by the
same clock [1].
5 Timing Analysis
To connect two IP cores together by allowing them to communicate over an OCP

interface, the protocols, signals, and pin-level timing must be compatible. OCP
establishes the communication based on the applied clock to the master [3]. Timing
plays a vital role while sending the data to the slave or from the slave.
5.1 Basic Mode of OCP
The timing analysis for simple write and simple read of OCP is illustrated in Fig. 4.
The diagram shows a write with no response enabled on the write [5].
Fig. 4 Timing analysis of basic mode OCP

5.2 Burst Mode of Operation
OCP burst mode for a burst of four 32-bit words, incrementing precise burst write,
with optional burst framing information (MReqlast) is illustrated [4]. As the burst is
precise (with no response on write), the MBurstLength signal is constant during the
whole burst. MReqlast flags the last request of the burst, and SRespLast flags the
last response of the burst. The slave monitors MReqlast for the end of burst. The
timing diagram is shown in Fig. 5.
6 Simulation Results
The implemented design of OCP is verified by creating the UVM verification

environment by using SimVision, cadence tool. SimVision Debug can be used to
debug digital, analog, or mixed-signal designs. These can be written in Verilog,
SystemVerilog, e, VHDL, and SystemC® languages. It supports all IEEE standard
designs of signal-level and transaction-based flows, testbench, and assertion lan-
guages. SimVision Debug provides a unified simulation and debug environment. It
allows Incisive Enterprise Simulator to manage multiple simulation runs easily [8].
To analyze both design and testbench at any point in the verification process is quiet
easy.
Fig. 5 Timing analysis of burst mode OCP

6.1 Simple Mode of Operation (Both Write and Read)
Basic mode of OCP with simple write and read is implemented and verified [2].
Master initiates the transfer by sending the address and data to the slave on Maddr
and MCmd, respectively. Slave acknowledges the master by asserting the
SCmdAccept to the master. It performs either read or write based upon the com-
mand generated by the master. The simulation results obtained for the applied
transaction between master and slave when RTL is verified by using UVM
Methodology on SimVision tool are shown in Fig. 6.
6.2 Burst Mode of Operation (Both Write and Read)
Precise burst mode was implemented and verified. MBurstLength,

MBurstSequence, and MBurstPrecise are considered constant throughout the
transaction. Master initiates precise burst write operation by sending the request to
slave (MCmd), address (Maddr), and data (Mdata). Mreqlast = 1 indicates the end
of request from master. Slave acknowledges the master by enabling SCmdAccept.
A high on SRespLast represents the end of precise burst read [6]. The simulation
result obtained for the applied transaction between master and slave when RTL is
verified by using UVM methodology on SimVision tool is shown in Fig. 7.
Fig. 6 Simulation results of basic mode OCP

Fig. 7 Simulation results of burst mode OCP
7 Conclusion
This paper is mainly about the implementation and verification of basic modes of
OCP. This work delivers that OCP is an efficient protocol for secured data core
communication. OCP is capable enough in reducing the design time and risk by
simultaneously designing the cores and working in the system. Based upon the
real-time applications, IP cores are supposed to be designed such that they can be
redesigned by cores which are not having inbuilt system logic that can be reused
with no additional time for cores to be re-created.
References
1. Technical Information on Open Core Protocol, [online] Available: http://www.ocpip.org/,

Open Core Protocol - International Partnership (OCP-IP).
2. Shihua Zhang, Asif Iqbal Ahmed and Otmane Ait Mohamed, “A Re-Usable Verification
Framework of Open Core Protocol (OCP)”. OCP-IP, 2008.
3. Elina Rajan Varughese and Rony Antony, “Implementation of extended open core protocol
interface memory system using Verilog HDL,” P. 978-1-4673- 6126-2/13/$31.00 c 2013 IEEE.
4. Open Core protocol (OCP) Specifications 2.2, Revision 1.0, Accellera 2004, [online]
Available: http://www.ocpip.org/home.
5. OCP-IP, The Importance of Sockets in SoC Design, [online] Available: http://www.ocpip.org/
white papers.php.
6. Chin-Yao Chang, Yi-Jiun Chang, Kuen-Jong Lee, Jen-Chieh Yeh, Shih-Yin Lin and
Jui-liangma “Design of On-Chip Bus with OCP Interface,”.
7. R. Usselmann, “Implementation of Re-configurable Open Core Protocol Compliant Memory
System using VHDL 2010 5th International Conference on Industrial and Information
Systems,” ICIIS 2010, Jul 29–Aug 01, 2010, India.
8. Chris Spear, “System Verilog for verification: A Guide to Learning the test bench for Language
Features”: Springer, second edition.
9. S. Palnitkar, “Verilog HDL: A Guide to Digital Design and Synthesis”, Upper Saddle River,
New Jersey: Prentice Hall, Jan. 1996.
Cellular Automata Logic Block Observer
Based Testing for Network-on-Chip
Architecture
Shaik Mohammed Waseem and Afroz Fatima
Abstract Necessity to test the logic circuits has increased rapidly due to the
increase in number of applications being hosted on a single chip. This in turn has
demanded the design of testing architectures which are capable of providing high
fault coverage with less resource utilization and minimal power usage. Cellular
Automata Logic Block Observer (CALBO), a technique homologous to Built-In
Logic Block Observer (BILBO), has been considered in this paper to test the routers
which are considered as important components of a Network-on-Chip
(NoC) architecture. The resource utilization and power report of the design have
been successfully generated to list out the advantages of the CALBO in comparison
to BILBO for the architecture considered.
Keywords CALBO BILBO Network-on-chip 7-port router

Cellular automata Testing
1 Introduction
Increase in demand for miniaturization of the electronic devices across the globe
has triggered a challenge for design engineers to look at various alternatives and in
such an attempt to design the low-power devices with reduced chip area, the
probability for occurrence of faults has proportionately increased [1]. Different
techniques and architectures have been under practice for providing testability to
the SoCs and one among such is the CALBO which is a technique similar to
BILBO and comes with the advantages of cellular automata-based rules being used
for test pattern generation and response compaction [2]. CALBO has its presence in
S. M. Waseem (&)
IF&S, Kadapa, Andhra Pradesh, India
e-mail: waseem.vlsi@gmail.com
A. Fatima
Mosaic, London, ON, Canada
e-mail: afatima.es@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_3
20 S. M. Waseem and A. Fatima
testing for quite a long time and in [3], small ternary circuits has been tested using
CALBO to demonstrate its ability as a good technique for providing testability. The
ability of CALBO to act as a simple register, test pattern generator, and response
compactor depending upon the different modes of operation forced on it through
control signals has made it a peculiar technique when compared to others [4].
2 Background
G. Jervan et al. in [5] analyzed the impact of NoC parameters on the test scheduling
through quantization methods. In [6], C. L. Hsu et al. proposed a low-power
BILBO structure based on the LPTM approach, which can functionally reduce the
switching activity during test operation. K. Namba et al. in [7] proposed a con-
struction of FF that can effectively correct soft errors during system operations and
can work as a BILBO FF on manufacturing testing. In [8], E. Sadredini et al.
proposed IP-BILBO that generates test data locally and makes use of two tech-
niques, namely direct and indirect reseedings of internal registers for generating
pseudorandom test patterns. L. Gao et al. in [9] proposed a linear hybrid cellular
automata using Cellular Automata (CA) Rule 90 and 150 and used it successfully as
test pattern generator and signature analyzer for implementing built-in self-test
capability. In [10], S. M. Waseem et al. observed the behavior of a 7-port router
architecture for three-dimensional NoC, by implementing the same for a mesh
topology and compared the results of its behavior for algorithms like Hamiltonian
routing and conventional routing.
3 Cellular Automata Logic Block Observer
CALBO is homologous to BILBO register [11] in its use but slightly differs in its
mechanism as it hosts cellular automata-based test pattern generator and
multiple-input signature register. Among the 256 rules of CA and their usage
proposed by Stephen Wolfram in [2], Rule 45 of CA is considered in this paper due
to its peculiar cyclic nature and good account of randomness in patterns generated
that could be applied to Circuit Under Test (CUT), here being the router archi-
tecture for NoC. The uniqueness of CALBO is due to its ability to act as a Cellular
Automata Test Pattern Generator (CATPG) in one test cycle and Cellular Automata
Multiple-Input Signature Register (CAMISR) in another. In normal operation, it can
also be used as a simple register to store data, if required. The different modes of
operation of CALBO are shown in Table 1.
Cellular Automata Logic Block Observer Based Testing … 21
Table 1 Modes of operation for CALBO

Control signal-1 (B1) Control signal-2 (B2) Mode of operation
0 1 Cellular automata test pattern generator
0 0 Serial scan chain
1 1 Cellular automata MISR
1 0 Normal D flip-flop
3.1 CALBO as Test Pattern Generator
The ease to detect the presence of faults often depends on the extent of randomness
in the generated patterns to be applied to CUT in this criterion, CA-based test
pattern generator has overcome the conventional LFSR-based TPGs. Rule 45 of CA
has been considered in this paper, for test pattern generation of CALBO due to its
higher degree of randomness [12, 13]. Rule 45 of CA is represented through a
logical correlation of its neighbors as shown in Eq. (1):
p
C n ¼ Ci1 XOR Cip OR NOT Cipþ 1 ; ð1Þ
where
Cn Succeeding stage output state of the cell, which is under consideration
p
Ci1 Existing state of the adjacent (left) cell
Cip Existing state of the cell which is under consideration
Cipþ 1 Existing state of the adjacent (right) cell
The pictorial representation of Eq. (1) could be best described with the help of
Fig. 1, where the black cell can be considered as “logical 1” and white cell as
“logical 0”.
Numerical representation of the same is shown as below:
{{1, 1, 1} ! 0, {1, 1, 0} ! 0, {1, 0, 1} ! 1, {1, 0, 0} ! 0, {0, 1, 1} ! 1,
{0, 1, 0} ! 1, {0, 0, 1} ! 0, {0, 0, 0} ! 1}
p
Numerical correspondence can, respectively, be evaluated as {Ci1 , Cip , and
p n
Ci þ 1 }! C
Fig. 1 Pictorial cell representation for rule 45 of cellular automata

3.2 CALBO as Multiple-Input Signature Register
In this paper, to implement CALBO, in order to act as CA-based MISR (CAMISR),

CA-based MISR from [14] has been considered in which the signature is formed by
first doing XOR of each cell with the corresponding circuit output and then
incrementing the cellular automaton. This method can be mathematically repre-
sented as in (2):
C ðt þ 1Þ ¼ ðC ðtÞXOR OðtÞÞ0 ð2Þ
where
C(t) Contents of cellular automaton at time “t”
O(t) Output of the circuit at time “t”
C(t + 1) Contents of cellular automaton at time “t + 1” and symbol (0) indicates
incremented value at respective time
4 NoC Architecture Testing with CALBO
4.1 NoC Architecture
Router is often considered as the main component responsible to route data on a

chip and hence is of greater importance to successfully implement the
network-on-chip architecture. The router designed in this paper consists of seven
ports and hence could possibly route data in seven different directions with a scope
for three-dimensional NoC implementation. The arbitration mechanism is of round
robin in nature and the crossbar switch being the multiplexer based, one which
drives the data to related FIFO buffer. The data is routed to the concerned port with
the information from LUT (Look Up Table) which is responsible for hosting
information related to routing logic. The router architecture can be best described as
shown in Fig. 2.
4.2 CALBO-Based Testing for NoC
In an attempt to test the NoC architecture with CALBO, in this paper, a

four-router-based NoC architecture is considered and testing has been accomplished
in a scheduled three test sessions as detailed in Table 2. In the first test session,
Router 1 and Router 4 have been tested simultaneously with CALBO 1 and
CALBO 3 made to act as CAMISR and CATPG, respectively, by forcing appro-
priate control signals as stated in Table 1. Test session-2 was aimed at successfully
In E
In W
In N
ROUND
In S
ROBIN
In U
ARBITER
In_D
In L
Out_E
FIFO 1
Out_W
FIFO 2
ROUTING
LOGIC
Out N
FIFO 3
Out_S
FIFO 4
Out_U
CROSSBAR
FIFO 5
SWITCH
Out_D
FIFO 6
Out_L
FIFO 7
Fig. 2 7-port router architecture for NoC implementation
Table 2 Test session for Test session-1(Router-1 and Router-4)

routers in NoC architecture
CALBO1-CAMISR
for CALBO-based testing
CALBO3-CATPG
Test Session-2(Router-2)
CALBO1-CATPG
CALBO2-CAMISR
Test Session-3(Router-3)
CALBO2-CATPG
CALBO3-CAMISR
testing Router 2 by forcing CALBO 1 and CALBO 2 to work under modes of

CATPG and CAMISR respectively. Router 3 in third test session has been tested by
making CALBO 2 to operate in CATPG mode and CALBO 3 in CAMISR mode as
shown in Table 2 and Fig. 3.
Fig. 3 Testing of four routers

in an NoC architecture with CATPG
CALBO …
T1
ROUTER 1
…
CALBO 1
…
T2
ROUTER 2
…
CALBO 2
…
T3
ROUTER 3
…
CALBO 3
…
T1
ROUTER 4
…
CAMISR
5 Results and Discussion
Four-router-based NoC architecture has been considered for CALBO-based testing

and the design has been implemented on Kintex Ultrascale device of Xilinx Inc.
with the help of Xilinx Vivado 2016.2 V suite. The generated power and resource
utilization reports for the design have been tabulated in Tables 3 and 4, respec-
tively, along with that of the BILBO-based testing statistics. From the graphical
representation of the statistics in Figs. 4 and 5, it can be noticed that the
CALBO-based testing is beneficial in both the important criteria, i.e., power and
resource utilization which are rigorously being pursued by researchers working on
System-on-Chip (SoC) architectures. Even though the difference in the numerical
for some of the factors of CALBO in power and resource utilization report is almost
same or is less statistically comparable to that of BILBO, still the difference in total
is considerably acceptable and it is sought to be increased with increase in the size
of the target application (i.e., the number of routers being taken into consideration
due to increase in the size of the NoC architecture).
Table 3 Power report for CALBO- and BILBO-based testing

Type of testing Dynamic power (W)
Clocks (W) Signals (W) Logic (W) I/O (W) Total (W)
CALBO-based testing 0.017 0.014 0.004 0.322 0.356
BILBO-based testing 0.018 0.016 0.004 0.349 0.387
Table 4 Resource utilization Type of testing Resource utilization

report for CALBO- and
LUT FF I/O BUFG
BILBO-based testing
CALBO-based testing 1253 2021 482 1
BILBO-based testing 1460 2656 483 1
Fig. 4 Graphical 0.4

representation of power report
Watt
for CALBO and BILBO 0.2 CALBO based Testing

testings BILBO based Testing
0
Clocks
Signals
Logic
I/O
Total
Power Components
Fig. 5 Graphical 3000

Number of
representation of utilization
report for CALBO and 2000 CALBO based Testing
BILBO testings
1000 BILBO based Testing
Resource Utilization Components
6 Conclusion
The CALBO-based testing though demands higher level of complexity in the

design, the degree of randomness that it provides in order to detect the faults, is
often beneficial in comparison to BILBO. Further, the statistical and graphical
reports of power and resource utilization tend to prove the advantages of
CALBO-based testing over BILBO, for the considered application.
References
1. G. Strawn, C. Strawn, “Moores law at fifty,” IT Professional, vol. 17, no. 6, pp. 69–72, 2015.
2. Stephen Wolfram, “A New Kind of Science”, Wolfram Media Inc., 2001.
3. C. Rozon and H. T. Mouftah, “Pseudo-random testing of CMOS ternary logic circuits,”
[1988] Proceedings. The Eighteenth International Symposium on Multiple-Valued Logic,
Palma de Mallorca, Spain, 1988, pp. 316–320.
4. P. D. Hortensius, R. D. McLeod and B. W. Podaima, “Cellular automata circuits for built-in
self-test,” in IBM Journal of Research and Development, vol. 34, no. 2.3, pp. 389–405, March
1990.
5. G. Jervan, T. Shchenova and R. Ubar, “Hybrid BIST Scheduling for NoC-Based SoCs,” 2006
NORCHIP, Linkoping, 2006, pp. 141–144.
6. Chun-Lung Hsu and Chang-Hsin Cheng, “Low-Power built-in logic block observer
realization for BIST applications,” 2007 7th International Conference on ASIC, Guilin,
2007, pp. 994–997.
7. K. Namba and H. Ito, “Soft Error Tolerant BILBO FF,” 2010 IEEE 25th International
Symposium on Defect and Fault Tolerance in VLSI Systems, Kyoto, 2010, pp. 73–81.
8. E. Sadredini, M. Najafi, M. Fathy and Z. Navabi, “BILBO-friendly hybrid BIST architecture
with asymmetric polynomial reseeding,” The 16th CSI International Symposium on Computer
Architecture and Digital Systems (CADS 2012), Shiraz, Fars, 2012, pp. 145–149.
9. L. Gao, Y. Zhang and J. Zhao, “BIST using Cellular Automata as test pattern generator and
response compaction”, Consumer Electronics, Communications and Networks (CECNet),
2012 2nd International Conference on, Yichang, 2012, pp. 200–203.
10. S. M. Waseem and A. Fatima, “Pursuance measures of CRA & HRA for 3D networks on a
7-port router schema,” 2015 International Conference on Communications and Signal
Processing (ICCSP), Melmaruvathur, 2015, pp. 0055–0061.
11. M. L. Bushnell and V. D. Agrawal, Essentials of Electronic Testing for Digital, Memory, and
Mixed-Signal VLSI Circuits. Springer, 2000.
12. Stephen Wolfram, “Random sequence generation by cellular automata”, Advances in Applied
Mathematics, vol. 7, Issue 2, 1986, pp. 123–169.
13. A. Fatima and S. M. Waseem, “Cellular Automata based Built-In-Self Test implementation
for Star Topology NoC,” 2017 11th International Conference on Intelligent Systems and
Control (ISCO), Coimbatore, 2017, pp. 45–48.
14. P. D. Hortensius, R. D. McLeod and H. C. Card, “Cellular automata-based signature analysis
for built-in self-test,” in IEEE Transactions on Computers, vol. 39, no. 10, pp. 1273–1283,
Oct 1990.
Structural Strength Recognizing System
with Efficient Clustering Technique
Sumedha Sirsikar and Manoj Chandak
Abstract Internet of Things (IoT) visualizes future, in which the objects of

everyday life are equipped with sensor technology for digital communication. IoT
supports the concept of smart city, which aims to provide different services for the
administration of the city and for the citizens. The important application of IoT is
Structural Strength Recognition (SSR). This approach is becoming popular to
increase the safety of buildings and human life. Proper maintenance of historical
buildings requires continuous monitoring and current conditions of it. Sensor nodes
are used to collect data of these historical buildings or large structures. Structural
strength recognition covers huge geographical area and it requires continuous
monitoring of it. It involves more energy consumption during these activities.
Hence, there is need for efficient energy management technique. Clustering is one
of the important techniques for energy management in Wireless Sensor Networks
(WSN). It helps in reducing the energy consumed in wireless data transmission. In
this paper, SSR system is designed with efficient clustering algorithm for wide
network and also finds out optimum number of clusters.

Keywords Structural strength recognition Energy management

Optimal clustering Energy efficient WSN
1 Introduction
The idea of Internet of Things was developed with combination of Internet and
wireless sensor networks. The term Internet of things was devised by Kevin Ashton
in 1999 [1]. It refers to identify objects and its virtual representations in
S. Sirsikar (&)
Sant Gadge Baba Amaravati University, Amravati, Maharashtra, India
e-mail: sumedha.sirsikar@mitpune.edu.in
M. Chandak
Shri Ramdeobaba College of Engineering and Management, Nagpur, Maharashtra, India
e-mail: chandakmb@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_4
28 S. Sirsikar and M. Chandak
“Internet-like” structure uniquely. These objects can be anything like huge build-
ings, parts of a large system, industrial plants, planes, cars, machines, any kind of
goods and animals, etc. Wireless communication technologies play a major role in
IoT. Major advancements in WSNs will be carried out by integrating sensors with
objects. IoT is developed based on this principle.
One of the aims of IoT is to design and develop cities smartly. It will offer better
services for the administration of the city and citizens. It is used in variety of
applications such as structural strength recognition, agriculture, military, biodi-
versity mapping, disaster relief operations, and medical health care.
1.1 Structural Strength Recognition
Structural strength recognition is a system used to evaluate real-time condition of

buildings. This application has enormous importance in safety of human life. The
data collected from sensor devices is used to recover the operation. This mainte-
nance operation of the system is to repair and replace the structures. Detection of
current damages can be used to distinguish eccentricities from the design perfor-
mance. Monitored data needs to be integrated in structural management systems. It
will help to take superior decisions by providing reliable and accurate information.
The defective houses and multistorey buildings have serious implications.
Sometimes, buildings get collapse just because of irregular monitoring and no
maintenance which is very dangerous for human life.
The structural strength recognizing system design requires focus on the partic-
ular requirements. The first thing is the probable reasons for degradation and the
related risks, second the expected requirements for this degradation, and finally, a
suitable structural strength recognizing system to detect such conditions. In SSR,
the appropriate sensors are required for appropriate monitoring. If the process of
data collection, updation, and data analysis is correct, then recognizing strength of
structures becomes really easy.
If the selection of the sensors for specific risks associated with a given structure
is done properly, then it is necessary to combine different parameters to analyze the
building. It requires maintenance system to insure proper data collection process,
integrated data acquisition, and management system.
Multistorey buildings have complex constructions with a combination of dif-
ferent elements. Generally, building structure differs in size, operational system,
geometry, construction material, and foundation characteristics. Due to the stress of
natural actions, these elements create impact on a building performance.
Few years before, implementation of sensor nodes in large civil structures was not
suitable, in terms of cost and power requirements [1]. Today, due to the tremendous
progress of low-power semiconductor devices and high functionality microcon-
troller units results in the great development of sensor nodes which are used in SSR
[2]. A large amount of energy is consumed for data transmission over the sensor
network. Sensor nodes are powered through small batteries that possess short life.
Structural Strength Recognizing System with Efficient … 29
The network lifetime can be improved using efficient energy management

technique. This can be done using efficient clustering techniques [2] that serves an
important role in large SSR systems. The proposed model would be useful to carry
out survey of large geographical area. But large area consumes more energy [3].
Hence, efficient clustering technique with optimum number of clusters [4, 5] is
required for proper utilization of energy because it is very difficult to recharge or
replace nodes in sensor network as it is complex in nature [6–8].
1.2 Structural Damage Detection
Building damages structural damage detection can be analyzed from different

parameters of building such as columns, beams, and flooring. Nonstructural damage
detection can be analyzed by observing damage or partition in ceilings, walls, and
glasses.
2 Related Work
2.1 SSR Techniques
Civil structural strength recognizing system consists of various popular techniques

such as acoustic emission (AE) analysis, vibration study, and optical sensing. The
most popular technology is optical sensing, accurate in detecting small distortions
in structures. Variety of fiber optic strain sensors are available. Due to their char-
acteristic as small physical size as compared to the civil structures and multiplexing
capability, the fiber optic sensors have become very widespread [9]. Moreover, it is
a perfect choice for civil structures of any shape to embed such cost- and
size-effective sensors. Passive Radio Frequency Identifier (RFID) system was
developed by Ikemoto et al. which does not require to care about battery lifetime
problems of sensor nodes [9]. Each sensor node holds a passive RFID tag that is
embedded into the structure in duration of construction or can be attached later on
surface. The main aim of the development of this device is zero battery power of
sensor node.
2.2 Strength Recognizing Components
There are different components to recognize structural strength. These are classified
into several categories such as Scale (Local scale, Global scale, and Network scale),
Periodicity (Periodic, Semi-continuous, and Continuous), Parameter (Mechanical,

Chemical, Physical, and Environmental actions), and Data collection (Manual,
None, Online, Off-line, and Real time).
2.3 Advantages of SSR
The benefits of a structural strength recognizing systems are increased in security,

concealed structural assets determination, increase in safety, and long-term quality
and structural management preservation.
2.4 System Integration
The SSR system is designed as an integrated system. In the system integration, all
data flow into a single database from buildings and represent through a user
interface. Various sensors can be attached to the same data logger single data
management system. As shown in Fig. 1, different types of sensors are positioned
into buildings that collect actual data of buildings. This data will be stored in the
database for further analysis. After analysis, results are displayed either on site or at
a remote location.
Fig. 1 SSR system integration

3 Problem Statement
Technical monitoring of buildings and its structures are the basic needs nowadays.
The bulk of buildings are growing every year. SSR covers large geographical
area. The data collection from sensor nodes to database requires large amount of
energy. The main objective is to determine optimum number of clusters for a large
sensor network so that energy of sensor network will be saved and thus it increases the
network lifetime with the help of efficient clustering techniques.
4 Proposed Solution
Energy management is an important factor in WSN. Transmission of data from one

location to other requires energy consumption. There are different techniques to
manage the network energy efficiently. Clustering is one of the techniques to
decrease the energy depletion and increase network lifetime. SSR requires large
area for monitoring and analysis. This large area needs huge amount of energy to
collect information and to transmit it to the base station. Clustering technique is
used to solve issues related to scalability, energy, and lifetime of sensor networks.
At the same time, clustering process has problem of determining the optimum
number of clusters that will minimize energy consumption. Number of clusters are
directly proportional to the amount of energy consumed by the network. The
solution is proposed to design SSR system and find out the optimum number of
clusters for wide network as shown in Fig. 2.
5 Proposed Model
This section describes mathematical model for finding Optimum Number of

Clusters (ONC) and average energy consumption.
Fig. 2 Proposed system architecture

ONC = f N, C, NC, Rg
N = Set of all nodes present in the network = {n1, n2, n3, …}

C = Cluster which contains group of sensor nodes
NC = Number of cluster = {nc1, nc2, nc3, …}
R = Range
Optimum number of clusters is obtained from the intersection of R and NC as
shown in Fig. 3.
Ei = Initial energy of each node
Energy consumed by node i initially requires some amount of energy for certain
events, for example, send, receive, and drop.
Ici = Energy consumed by ith node = Initial Energy − Residual Energy
Total Consumed Energy (TCE)
X
N
TCE ¼ ðInitial Energy Residual EnergyÞ of node
i¼1
Average Energy Consumption (AvgEC) = Total energy consumed by nodes in the

network
X
Nt
AvgEC ¼ Ici/N
i¼1
Fig. 3 Intersection of
number of clusters and range
6 Proposed Methodology
One of the important constraints of WSN is lifetime of network and energy. In this
research work, this problem is kept in mind and a solution for finding optimum
number of clusters and efficient clustering is described. The proposed solution
contains the following steps:
I. Calculation of optimum number of clusters
II. Formation of clusters and selection of cluster head
III. Reformation of cluster
Step I: Calculation Optimum Number of Clusters

Optimum number of clusters for a sensor network is calculated using proposed
method by plotting graph, where x-axis represents the number of clusters and y-axis
represents the range. Two lines are drawn that covers the maximum points on the
curve. Each line must contain at least two points and must start at either end of the
data. The intersection point of these two lines is used as optimum number of
clusters as shown in Fig. 3.
Step II: Formation of Clusters and Selection of Cluster Head
After getting exact number for clusters, divide the circular network or make par-
titions of circular network according to number of clusters.
Degree of partition = 360/Optimum Number of Clusters
Degree of partition = 360/Optimum Number of Clusters

Cluster Head (CH) is selected after formation of clusters. The cluster head is
selected by Base Station (BS). Here, it is assumed that the BS knows all the
information of sensor nodes. BS will select CH based on this information. Node
which has high energy level and suitable geographical location is selected as CH.
BS then broadcasts this information to all nodes [10].
Step III: Reformation of Cluster
During the continuous operation of sensor network, the battery level of sensor
nodes as well as cluster heads decreases. Due to this, sensor nodes and cluster heads
become inactive. If the node energy is below the threshold value, then the sensor
network should be re-organized into new clusters. In this process, the node with
highest residual energy and degree will become new cluster head [10].
6.1 SSR Application Execution Based on Proposed

Methodology
As per the phases, working of application of SSR is described in Fig. 4. Here,

circular geographical area is considered, which contains certain number of houses
and buildings. In the houses, single node is deployed but for buildings single sensor
Fig. 4 Deployment of nodes in houses
Fig. 5 Formation of clusters
node is not sufficient. Building contains multiple sensor nodes as shown in Fig. 4.
Now the number of required clusters for that circular area is calculated as shown in
Fig. 5. Here, yellow circles represent single sensor node for house and blue star
represents group of sensor nodes of a particular building.
Initially, calculate optimum number of clusters for network and partition the
circular network into that much number of clusters. Example is shown in Fig. 5.
After obtaining intersection point as per Fig. 3, divide network into four clusters for
example.
Now the cluster head is selected for data collection process. Finally, the CH will
gather data from sensor nodes and send it to base station. Later, the data acquisition
system collects data from BS. Data processing system will monitor that data for
further analysis. Alarm system generates notifications and alert alarms on computer
system and mobile devices as shown in Fig. 6.
Fig. 6 Data aggregation and sensing system
7 Conclusion
The lifetime of a sensor network is one of the most important parameters in WSNs.
It depends on energy level of battery-powered sensor nodes. The network is left
unattended to perform monitoring and reporting functions. Energy of sensors in
WSN is depleted due to direct transmission of data between sensors and a base
station, which can be avoided by clustering the nodes in WSN. It further enhances
scalability of WSN in real-world applications while conserving limited energy.
Thus, proposed algorithm aims to minimize the number of clusters and in turn
energy consumption per cluster. Hence, lifetime of WSN can be increased with
efficient clustering technique. Our proposed algorithm gives optimum number of
clusters that certainly help to improve the performance of network. Next step is to
implement the proposed solution in NS2 and to observe the performance in terms of
energy and network lifetime.
References
1. Zanella, A., Bui, N., Castellani, A., Vangelista, L., & Zorzi, M. (2014), Internet of things for
smart cities. IEEE Internet of Things Journal, 1(1), pp. 22–32.
2. Sasikumar, P., & Khara, S. (2012), K-means clustering in wireless sensor networks. Fourth
international conference on Computational Intelligence and Communication Networks
(CICN), IEEE, pp. 140–144.
3. Navjot Kaur Jassi, Sandeep Singh Wraich, (2014), A Review: An Improved K-means
Clustering Technique in WSN, Proceedings of the International Conference on Advances in
Engineering and Technology (ICAET).
4. Bholowalia, P., & Kumar, A. (2014). EBK-means: A clustering technique based on elbow
method and k-means in Wireless Sensor Network, International Journal of Computer
Applications, 105(9).
5. Alrabea, A., Senthilkumar, A. V., Al-Shalabi, H., & Bader, A. (2013), Enhancing k-means
algorithm with initial cluster centers derived from data partitioning along the data axis with
PCA. Journal of Advances in Computer Networks, 1(2), pp. 137–142.
6. Bhawna, Pathak, T., & Ranga, V., (2014), A Comprehensive Survey of Clustering
Approaches in Wireless Sensor Networks, Elsevier Publications.
7. Solaiman, B. F., & Sheta, A. F. Energy optimization in wireless sensor networks using a
hybrid K-means PSO clustering algorithm.
8. Tong, W., Jiyi, W., He, X., Jinghua, Z., & Munyabugingo, C. (2013), A cross unequal
clustering routing algorithm for sensor network. Measurement Science Review, 13(4),
pp. 200–205.
9. Alahakoon, S., Preethichandra, D. M., & Ekanayake, E. M. (2009), Sensor network
applications in structures–a survey. EJSE Special Issue: Sensor Network on Building
Monitoring: From Theory to Real Application, pp. 1–10.
10. Sirsikar, S., Chunawale, A., & Chandak, M. (2014), Self-organization Architecture and Model
for Wireless Sensor Networks. International Conference on Electronic Systems, Signal
Processing and Computing Technologies (ICESC), IEEE, pp. 204–208.
Trajectory Outlier Detection for Traffic
Events: A Survey
Kiran Bhowmick and Meera Narvekar
Abstract With the advent of Global Positioning System (GPS) and extensive use
of smartphones, trajectory data for moving objects is available easily and at cheaper
price. Moreover, the use of GPS devices in vehicles is now possible to keep a track
of moving vehicles on the road. It is also possible to identify anomalous behavior of
vehicle with this trajectory data. In the field of trajectory mining, outlier detection of
trajectories has become one of the important topics that can be used to detect
anomalies in the trajectories. In this paper, certain existing issues and challenges of
trajectory data are identified and a future research direction is discussed. This paper
proposes a potential use of outlier detection to identify irregular events that cause
traffic congestion.

Keywords Trajectory data Map matching Trajectory outlier detection

GPS data Similarity measures
1 Introduction
Trajectory data is the data about moving objects like vehicles, animals, and people.
With the advent of technologies of mobile computing and location-aware services,
there is a massive generation of this trajectory data. Trajectory data analysis helps in
identifying several real-world phenomena like, for example, identifying the buying
patterns of humans in shopping malls, understanding the migratory patterns of
animals and birds, and detecting the travel path of hurricanes and tornadoes.
Vehicle trajectory analysis is used in various traffic dynamics like traffic jam pre-
diction, roadmap construction, and real-time traffic flow; in route navigation like
suggest optimum route, suggest places to visit, and identify frequently visited
K. Bhowmick (&) M. Narvekar

Dwarkadas J. Sanghvi College of Engineering, Vile Parle, Mumbai, Maharashtra, India
e-mail: kiran.bhowmick@djsce.ac.in
M. Narvekar
e-mail: meera.narvekar@djsce.ac.in

https://doi.org/10.1007/978-981-10-7245-1_5
38 K. Bhowmick and M. Narvekar
locations; in itinerary navigation like discover popular routes, predict next location;
and in other applications like predicting human flow, suggesting alternate transport
means, etc. [1–4].
In recent years, there has been an increase in the research of outlier analysis of
trajectory data. This outlier analysis is used to identify anomaly in the regular data.
Outlier detection can also be used to identify events that occur in regular traffic.
This paper discusses different outlier detection techniques, the research literature
available, and suggests the use of trajectory outlier detection for detecting irregular
events such as traffic congestion due to road accidents or roadblocks or any other
reason.
2 Related Concepts
2.1 Basic Concepts
This section describes the concepts related to trajectory mining and outlier detection
basics.
2.1.1 Trajectory data
Definition 1: A trajectory T is an ordered list of spatiotemporal samples p1, p2, p3,

…, pn. Each pi = (xi, yi, ti), where xi, yi are the spatial coordinates of the sampled
point and ti is the timestamp at which the position of the point is sampled, with
t1 < t2 < … < tn [5].
In general, a trajectory depicts the time and location of a moving object where it
can be tracked. A single trajectory can include many trajectory points and can have
any length for a single moving object. Basically, the length depicts the places that
the object has traveled from point A to point B before it comes to a halt.
A sub-trajectory depicts a partition or a segment of a full trajectory.
2.1.2 Trajectory outlier
Definition 6: An outlier in a trajectory data is an item that is significantly different

from other items in terms of some form of similarity [1].
Outliers are objects that are inconsistent with remaining data items. Outliers are
also sometimes considered to be noise. But the noise for someone can be someone
else’s data [6]. For example, in the following, figures t4 and t5 are outliers in the
path of the regions R1 and R2.
However, t4 can be considered as an alternate path, while t5 can be a long detour
(Fig. 1).
Trajectory Outlier Detection for Traffic Events: A Survey 39
Fig. 1 Outliers between

regions
2.2 Similarity Measures
The distance between two trajectories is usually measured by some kind of

aggregation of distances between trajectory points. The various similarity measures
are defined in [1, 7]. These include closest pair distance (CPD), sum of pairs
distance (SPD), Euclidean distance (ED), dynamic time warping (DTW and
PDTW), edit distance measures (ERP and EDR), and LCSS. The CPD, SPD, and
ED require equal length of the trajectories and calculate the minimum distance
between two trajectories. Edit distances, DTW and LCSS, assume variable length
trajectories.
2.3 Map Matching
Map matching is a preprocessing technique in the trajectory data mining used to

convert the sequence of GPS points to a sequence of road segments. There are
various categories of map matching algorithms depending upon the type of infor-
mation considered and the range of sampling [8]. The category with the type of
information includes geometrical, topological, weight-based, probabilistic, and
advanced, whereas the range of sampling has two types local/incremental and
global methods.
2.4 Outlier Detection Techniques
This section introduces different outlier techniques.

Distance-based outlier detection techniques These techniques consider the dis-
tance between trajectory and its neighbors. If the trajectory does not have neighbors
above a particular threshold, it is an outlying trajectory [9, 10].
Density-based outlier detection techniques This means a trajectory is an outlier if
its density is relatively much lower than that of its neighbors [9, 11].
The distance-based outlier techniques detect global outliers due to the way the
distance is measured between the points. It also requires uniform distribution of
data. With different density distributions of data, the distance-based techniques
cannot be used. The density-based outlier techniques detect local outliers and can
do analysis in different density distributions but require a large number of k-nearest
neighborhood queries and are computationally very intensive.
Historical similarity-based outlier detection techniques This technique suggests
the use of historical similarity to identify trends between data points. At each time
step, a road segment is compared with its historical recorded temporal neighbor-
hood values to identify similarity values. A drastic change in the values results in
the outlier being detected.
Motifs-based outlier detection techniques A motion classifier to detect trajectory
outliers using an objects motion feature is called as motifs. These motifs are a
sequence of motion features with values related to time and location. The classifier
works on this high-dimensional feature space to distinguish an anomalous trajectory
with a normal one.
3 Related Work
Although it is an upcoming field, a lot of work have already been done in the outlier
detection and overall trajectory mining. With respect to the different similarity
measures, H. Wang et al. in their paper [7] have done a comparative study about
various similarity measures under different circumstances.
Map matching helps overcome the problem of uncertainty, but the existing map
matching algorithms also require high sampling rate of the data which is not always
the case. Yin Lou et al. in [12] have used spatiotemporal matching and analysis to
find the candidate graph which helps to map low sampling digital data to a
roadmap. K. Zheng et al. [13] proposed the use of similarity between query data
points and historical data to infer path for low sampling data. The inference is done
using traversal graph and nearest-neighbor-based heuristics approach. In [14], the
authors propose a technique to consider the complex nature of the urban roads and
address the problem of identifying a data point on elevated roads. The problem of
identifying the candidate segment selection for the first GPS point is addressed in
[15], where the authors employ a heuristic A* algorithm to find the shortest path
between a previous point and the candidate segment. Paolo Cintiaa and Mirco
Nannia in [16, 17] argue that the shortest path should not be the measure to find the
best matching point on the road network when the data samples are low and that a
time-aware Dijkstra’s shortest path is suggested to match the points. G. Hu et al. in
[18] suggested the use of an information fusion of data to accurately perform map
matching as well as consider the complexity of road networks. And lastly, the
authors in [19] suggest using LCSS and clustering algorithm totally avoiding the
use of any map matching algorithm. They conclude with their experimental results
that none of the measures are better than others. In all, each similarity measure is
affected by the decrease in the sampling rate. None of the related techniques
consider time-aware heuristics methods together with the complexity of road net-
works like multilayer and parallel roads, complex interchanges, and elevated roads.
Outlier detection has also been researched in different scenarios. Next, paragraph
review of literature on the research is been discussed.
Lee et al. first proposed a trajectory outlier detection algorithm TRAOD [19] to
detect outlying sub-trajectories using a partition and detect framework. It uses
density as well as distance-based measures and does not suffer from local density
problem and has minimal overhead for density measures. However, it has a com-
plexity of O(n2) and considers only spatial data. Piciarelli et al. in [20] used a
single-class SVM classifier on fixed-dimensional vectors of spatial data to identify
outliers. But the algorithm uses a training data with outliers. Outliers not seen in
training data will not be detected. In [21], Li et al. proposed a temporal outlier
detection TOD algorithm that considers time dimension and historical data to detect
outliers. The authors suggest the use of exponential function to detect outliers. This
exponential function, however, cannot be used for spatial dimension. Yong Ge et al.
in [22] proposed TOP-EYE to detect an evolving trajectory using a decay function.
If a trajectory is going to be an outlier, then it will be detected at an early stage by
this method. Wei Liu et al. in [23] suggested the use of frequent pattern tree to
identify the causal interactions between the detected outliers. They have considered
temporal information about current and historical time frames trajectory to compare
using Euclidean distance for detecting outliers. Daqing Zhang et al. proposed their
algorithm iBAT [24] for detecting outliers based on the isolation techniques. An
iTree based on the data-induced random partition is generated for all the trajecto-
ries. If a trajectory is an outlier, then it will have comparatively shorter paths than
those of a normal trajectory. Alvares et al. in [25] detected an outlier between
surveillance camera and inbound trajectories by analyzing the behavior of trajec-
tories. If a trajectory seems to avoid the object in certain patterns, then it is detected
as outlier. But it fails to identify if the avoidance was forced or intentional. Chawla
et al. in [26] used PCA analysis to identify anomalies and then L1 optimization to
infer possible routes whose flow traffic caused the outlier. Zhu et al. in [27, 28] used
time-dependent transfer graphs to identify outliers online.
From the above literature review of outlier, it is observed that, while considering
trajectory data, its inherent characteristics affect the outlier detection method.
Preprocessing of the data can reduce most of the challenges. But the fact that the
trajectory data is not equally distributed on all roads, it is important to consider
historical data. The choice of hybrid outlier detection technique and suitable sim-
ilarity measure can improve the accuracy majorly.
4 Issues and Challenges Identified
Trajectory data can be efficiently used in traffic management. Considering the

characteristics of trajectory data and the literature review, the following issues and
challenges related to trajectory data, map matching, and outlier detection are
identified.
4.1 Trajectory Length and Sampling
Every trajectory have different lengths, i.e., the number of data points is different for
each, i.e., T1 = (p1, p2, …, pn) and T2 = (q1, q2, …, qm) and n and m can be
different. It is important to identify the region of interest to consider the start and
end of each trajectory.
4.2 Low and Uneven Sampling Rates
The GPS data of the moving vehicles usually have a low sampling rate to avoid the
overhead of communication cost, data storage, and battery life of devices. So the
time interval between two consecutive GPS points can be very large (average
sampling rate 2 min). This leads to an uncertainty of the path taken by a trajectory
between these time gaps. Also, the data sampling rates are different across all
trajectories, i.e., T1 may be sampled for every 1 m and T2 is sampled every 2 m.
4.3 Trajectory Directions, Regions, and Road Networks
Trajectories moving in different directions and different regions should be con-

sidered different. Trajectories moving in different directions but close proximity and
those moving in the same direction but different regions should be considered as
different. Also for vehicles, the trajectories are bound by the underlying road net-
works, so the data points should match the road networks map.
4.4 Similarity Measures
There are different types of similarity measures as mentioned in Sect. 2.2. However,
with different lengths, low samplings, uncertainties, and noises, it is difficult to
identify a suitable similarity measure to compare trajectories together. All the
measures described in above section are sensitive to decrease in sampling rate. It is

a challenge to process data with low sampling rate.
4.5 Map Matching Challenges
Most of the existing map matching algorithms assume high sampling of data as they
usually perform local or incremental approach. Also, they consider only spatial data
for mapping to road networks. However, there is a lot of information available that
can improve the map matching considerably like time, speed, and direction. The
complexity of urban road networks is another issue. The topology of road networks
in urban areas has many challenges like parallel roads, multilayer roads, complex
interchanges, and elevated roads. Considering all these to perform map matching
increases the complexity, but will be realistic and accurate.
4.6 Outlier Detection Challenges
Different types of outlier detection techniques exist, viz., distance-based and

density-based. The existing trajectory outlier detection algorithms have different
techniques and each tries to solve the outlier detection with a different objective.
The issue is to devise an algorithm that can overcome the problem of uncertainty,
low sampling, and uneven length data with mixed distribution of trajectories and
considering spatial as well as temporal similarity constraints efficiently identifies
outliers even with large amount of data. A data structure that can help to deal with
the variety of information and the length of the data is needed. The algorithm
should be able to identify sparse as well as dense trajectories for online data. Also,
there needs to be a way to propagate the effects of the outlier to linked traffic nodes.
5 Research Problem
The outlier detection is still an upcoming research due to its application in traffic
management. A huge number of trajectory data are generated for a single moving
object and there are thousands of moving objects in the form of GPS enabled
vehicles and human beings carrying smartphones. The large number of sampling
points and the uneven length requires a suitable data structure like a modified R-tree
Fig. 2 a Preprocessing b Outliers between regions
for fast retrieval and indexing of the trajectory data. The R-tree should be able to
store temporal and related information (Fig. 2).
The trajectory data has uneven rate of sampling and an inherent uncertainty
about the locations between two sampled points. This uncertainty between sam-
pling points can be reduced by using map matching algorithms. But the existing
map matching algorithms consider only spatial data and work well with high
sampling. Also, it is important to consider the urban road complexity like multi-
layer, complex interchanges, and elevated roads. With low sampling, this makes it
more challenging. The use of shortest path to find the actual location of the tra-
jectory is not helpful. There has to be a consideration of hybrid technique like
combining the global method with weight-based technique or global method with
advanced or probabilistic method. A time-aware heuristic method can improve the
accuracy of map matching algorithms in low sampling data. Since the distribution
of trajectory data is skewed, using only density-based or distance-based method of
outlier detection is not the solution. A hybrid approach is used by many existing
algorithms but temporal information is not considered. Using time-dependent
popular routes graph based on historical data to deal with the changing nature of
outliers is beneficial. The detected outliers can be used to identify events by per-
forming pattern matching and comparing it with similar patterns that were seen
earlier when an irregular event occurred.
6 Conclusion
This paper is a part of an on-going research on identification of traffic events using

vehicle trajectory. This paper mainly focusses on the identifying problems in tra-
jectory data and outlier detection. A detailed literature survey on existing map
matching techniques and outlier detection technique is discussed. The paper also
discusses the proposed research problem for identifying traffic events using outlier
detection techniques.
7 Future Scope
In the future, we plan first to enhance our research by designing a novel outlier
detection technique focussing on the issue of low sampling and uncertainty using
taxi trajectory data and second to design pattern matching algorithms to identify
traffic events.
References
1. Zheng Y. 2015. “Trajectory Data Mining: An Overview”, ACM Transactions on Intelligent

Systems and Technology (TIST), Volume 6 Issue 3, May 2015 Article No. 29
2. Zaiben Chen et. al., “Discovering Popular Routes from Trajectories”, ICDE Conference 2011,
IEEE 2011, DOI 978-1-4244-8960-2/11
3. Lin, Miao and Wen-Jing Hsu. “Mining GPS Data for Mobility Patterns: A Survey”. Pervasive
and Mobile Computing 12, Elsevier, pp. 1–16, July 2013
4. Li Z, et. al., MoveMine: Mining Moving Object Databases. SIGMOD ’10 ACM, pp. 1203–
1206, June 2011
5. Wagner et. al., “Mob-Warehouse: A Semantic Approach for Mobility Analysis with a
Trajectory Data Warehouse”, Springer, pp. 127–136, Nov 2013
6. J. Han and M. Kamber, Data Mining: Concepts and Techniques, 2nd ed. Morgan Kaufmann,
2006
7. H. Wang et.al., “An Effectiveness Study on Trajectory Similarity Measures”, Proceedings of
the Twenty-Fourth Australasian Database Conference (ADC 2013), Adelaide, Australia, vol.
137, 2013
8. Mahdi Hashemi and Hassan A. Karimi, “A critical review of real time map matching
algorithms: Current issues and future directions”, Computers, Environment and Urban
Systems, Elseveir, vol 48, pp. 153–165, Aug 2014
9. Gupta, Manish et al. “Outlier Detection for Temporal Data: A Survey”. IEEE Transactions on
Knowledge and Data Engineering 26.9, vol. 25, no. 1, 2014: 2250–2267
10. Knorr, Edwin M., Raymond T. Ng, and Vladimir Tucakov. “Distance-Based Outliers:
Algorithms And Applications”. The VLDB Journal, 8.3–4 (2000): 237–253
11. Liu, Zhipeng, Dechang Pi, and Jinfeng Jiang. “Density-Based Trajectory Outlier Detection
Algorithm”. Journal of Systems Engineering and Electronics 24.2 (2013): 335–340
12. Y. Lou, C. Zhang, Y. Zheng, X. Xie, W. Wang and Y. Huang, “Map-matching for
low-sampling-rate GPS trajectories”, Proceedings of the 17th ACM SIGSPATIAL—GIS ‘09
13. K. Zheng, Y. Zheng, X. Xie, and X. Zhou, “Reducing uncertainty of low-sampling-rate
trajectories”, 28th IEEE International Conference on Data Engineering. IEEE, 2012, pp:1144–
1155
14. Z. He, S. Xi-wei, P. Nie and L. Zhuang, “On-line map-matching framework for floating car
data with low sampling rate in urban road networks”, IET Intelligent Transport Systems, vol.
7, no. 4, pp. 404–414, 2013
15. M. Quddus and S. Washington, “Shortest path and vehicle trajectory aided map-matching for
low frequency GPS data”, Transportation Research Part C: Emerging Technologies, Elsevier,
vol. 55, pp. 328–339, March 2015
16. Paolo Cintiaa, and Mirco Nannia “An effective Time-Aware Map Matching process for low
sampling GPS data”, Elsevier, March 2016
17. G. Hu, J. Shao, F. Liu, Y. Wang and H. Shen, “IF-Matching: Towards Accurate
Map-Matching with Information Fusion”, IEEE Transactions on Knowledge and Data
Engineering, vol. 29, no. 1, pp. 114–127, Oct 2016
18. J. Kim and H. Mahmassani, “Spatial and Temporal Characterization of Travel Patterns in a
Traffic Network Using Vehicle Trajectories”, Transportation Research Procedia, vol. 9,
pp. 164–184, July 2015
19. J. G. Lee, J. W. Han and X. L. Li. “Trajectory outlier detection: a partition and detect
framework”. 24th International Conference on Data Engineering ICDE, IEEE, pages 140–
149, 2008
20. Piciarelli, C., C. Micheloni, and G.L. Foresti. “Trajectory-Based Anomalous Event
Detection”. IEEE Trans. Circuits Syst. Video Technol. 18.11 (2008): 1544–1554
21. X. L. Li, Z. H. Li, J. W. Han and J. G. Lee. “Temporal outlier detection in vehicle traffic
data”. 25th International Conference on Data Engineering, pages 1319–1322, 2009
22. Yong Ge, et. al., “TOP-EYE: Top-k Evolving Trajectory Outlier Detection”, CIKM’10 ACM,
2010
23. Wei Liu et. al., “Discovering Spatio-Temporal Causal Interactions in Traffic Data Streams”,
KDD’11, ACM, August 2011, California, USA
24. Daqing Zhang et. al., “iBAT: Detecting anomalous taxi trajectories from GPS traces”,
UbiComp’11, ACM 978-1-4503-0603-0/11/09, September 2011, Beijing, China
25. L. Alvares, A. Loy, C. Renso and V. Bogorny, “An algorithm to identify avoidance behavior
in moving object trajectories”, Journal of the Brazilian Computer Society vol. 17, no. 3,
pp. 193–203, 2011
26. S. Chawla, Y. Zheng, and J. Hu, Inferring the root cause in road traffic anomalies, 12th IEEE
International Conference on Data Mining. IEEE, 141–150, 2012
27. Zhu, Jie et al. “Time-Dependent Popular Routes Based Trajectory Outlier Detection”.
Springer International Publishing (2015): pp. 16–30, Switzerland, https://doi.org/10.007/978-
3-319-26190-4_2
28. Zhu, Jie et al. “Effective And Efficient Trajectory Outlier Detection Based On
Time-Dependent Popular Route”. Springer Science (2016), NY, https://doi.org/10.007/
s11280-016-000-6
A Privacy-Preserving Approach to Secure
Location-Based Data
Jyoti Rao, Rasika Pattewar and Rajul Chhallani
Abstract There are number of sites that provide location-based services. Those
sites use current location of user through the web applications or from the Wi-Fi
devices. Sometimes, these sites will get permission to the user private information
and resource on the web. These sites access user data without providing clear detail
policies and disclosure of strategies. This will be used by the malicious sites or
server or adversaries breaches the sensitive data and confidentiality of the user. User
shares original context of the location. An adversary learns through the user’s
original context. Due to the lack of secure privacy-preserving policies, it has shifted
them to specific goals for various hazards. In order to secure or preserve privacy of
user, new privacy-preserving technique called FakeIt is proposed. In FakeIt, system
works around privacy, security to satisfy privacy requirements and the user decides
context before sharing. If the current location context is sensitive to the user, then
user decides to share the fake location context to location-based services instead of
original. System restricts the adversaries to learn from the shared sensitive location
context of the user.
Keywords Privacy Protection Location-based service Security

Sensitive context
J. Rao (&) R. Pattewar R. Chhallani

Department of Computer Engineering, Dr. D.Y. Patil Institute of Technology,
Pimpri, Pune, India
e-mail: jyoti.aswale@gmail.com
R. Pattewar
e-mail: pattewar.rb@gmail.com
R. Chhallani
e-mail: rajuldchhallani@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_6
48 J. Rao et al.
1 Introduction
The increasing popularity of the location-based service provides new opportunities

to users. They provide personalized services to the user based on their location
details provided to the LBSN sites. For example, some sites provide remainder if
user is at a particular location. Due to large number of location-sharing applications,
the large amount of privacy issues are raised. User trusts on an application
location-sharing web application. But sometimes these types of service may use
users personal data for other purposes like to sell third parties or may use adver-
saries for malicious purpose, etc., privacy leakage without consent of user or
without user’s permissions. Sometimes, this concerns to threat user privacy and
location sensitivity. To prevent such type of privacy leakage, user has
privacy-preserving approach to restrict the adversaries. To restrict the adversaries
from getting the personal information of user, sensitive details are disclosed to
adversaries harmful to the user. The need of policies in user decides when or where
to share location or share other random location to the location-sharing sites [1].
Current system aims to limit the location-sharing site from accessing sensitive
data, which would prevent adversaries from inferring sensitive details of user.
Location context control policy gets privileges depending on the user. To restrict
adversaries on policy used called as deception policy, this policy is used in location
sharing to decide whether to release fake or not; if location is sensitive, then shared
location will be fake location. Adversaries may get know from previous user
context; for these temporal correlations, we use MaskIt policy. MaskIt is privacy
checking policy that decides releasing context of user, to share or to conceal current
location. FakeMask is a new technique to control the location sharing in web
application. Decide while sharing location with the application or fake location
context shared to web applications [1].
In existing privacy-preserving techniques, a simple technique is used. In this
technique, all the private information is secreted and other is transmitted. Though
the private information not transmitted, an attacker can infer some secreted private
information and resources from the transmitted information. For applying those
preventions related to information that distinguishes among nearly situated subar-
eas, the main reason first is that the public and private locations may unexpectedly
release their secrete context and hazarded their individual security for locations,
concerning on preventing adversary from accessing sensitive information and
resources. But many times it shortfall approaches security [1].
2 Review of Literature
Cai et al. [1] presented FakeMask technique. FakeMask chooses whether to spread
the present information or release fake information. FakeMask is useful for pro-
tecting secrecy of locality. The information distributed under FakeMask creates
A Privacy-Preserving Approach to Secure Location-Based Data 49
trouble to attacker to collect original information. Deception technique is acquired

inside FakeMask because of non-secret information releasing like secret informa-
tion. Semi-Morkov model and secrecy examining methods are used for modeling.
Privacy of user is protected from dangerous software which behaves like mediator.
Present information should not be revealed to the adversary. Handling with the
FakeMask and understanding information spread from user are not simple for the
attacker. Attacker is not capable of understanding this kind of information.
Shen et al. [2] studied structural design, communication progression, protection,
and privacy of network. Authors studied three groups of smartphone applications
accompanied by a center of attention on two independent smartphone approaches.
Business card plus assistance assessment are approaches that have learned. They
search practical method to relate with the connected security and confidentiality
challenge. In previous approach, there are deficiencies of the technique, so author
presents various capable research guidelines.
Jedari et al. [3] presented mechanism which mainly focuses to give logical group
on safety challenge and an enormous study on some present way in mobile sensor
networks. This exertion decreases safety troubles and resolution techniques
throughout opportunistic networks and stoppage tolerable networks to mobile
sensor networks containing anticipation of wrapping all effort started about pro-
tection, privacy, and confidence in mobile networks. Correlation among social
belief and reputation system is unstated. One problem is energy custom of points
and humanity outcome on points. Process of applying matchmaking procedures by
innovative people to realize additional concerning usage of mobile sensor network
approaches is not determined.
Kelley et al. [4] observed users’ risk and advantage perception associated with
practice of these apparatuses and privacy restrict of current location-sharing
approaches. They studied an online survey and observed that although the great
figure of their responders had heard of location-sharing approaches, they do not yet
understand the probable value of mentioned approaches; also they are worry
regarding sharing current location context online. Next to evaluating present
commercial locality-sharing approaches’ secrecy controls, author analyzed that
location-sharing approaches never offer their users different group of rules to
manage disclosure of user’s place; they offer only modicum of privacy.
Gao et al. [5] presented a technique connected toward locality susceptible
hashing toward partition consumer area in groups each consisting at minimum
number of K users, also designed an efficient algorithm to reply kNN queries for
any point in the spatial cloaks of arbitrary polygonal shape. Great replication study
shows that two algorithms provide superior effect having normal measuring com-
plexity. Two drawbacks associated with mentioned technique are present. First, it
fragments dataset; it is not locality preserving. Next, time complexity is more.
Ying et al. [6] proposed a motivational technique having privacy conservation
within smartphone crowdsourcing systems. Combining profits of offline incentive
techniques and online incentive techniques, a motivation apparatus presents that
choose the worker public continually, and then energetically prefer defenders after
bidding. The proposed motivation technique includes two algorithms. Algorithms
50 J. Rao et al.
are improved two-stage auction algorithm also straightforward online status

upgrading algorithm. By simulation, author validates capacity and benefits of the
represented incentive technique. Represented motivational technique can determine
freeriding problem and also enhance the capability and usefulness of mobile
crowdsourcing system sufficiently.
Gehrke et al. [7] presented MaskIt approach. This approach keeps secrecy which
is helpful to choose a user information series. The chosen text information could be
passed to approaches. It may be useful to feedback requirements from these kinds
of approaches. User explained privateness which is associated with a group of
secrete context mentioned. Although adversaries are powerful and understand
context about choosing system and time-related cooperations in the context
sequence, information realized by adversaries based on filtered sequence about to
user being in sensitive data is limited by MaskIt. Inside MaskIt, privacy inspect
determines whether to deliver or suppress the present user’s context. This technique
gives two novel privacy inspects and explains how to choose the check with more
utility for a user.
Zhang et al. [8] at the beginning acknowledged that leak privacy issue has a
belief of data dynamics and hazardous third parties who have potentials of tuning
their rushing policies, and then represent the correlative contest within users,
adversaries as a zero-sum stochastic game. As well, the author presents
well-planned least learning algorithm to obtain best protection strategy.
Weiss et al. [9] defined BlurSense, a tool which allow secure, customized per-
mission to all sensors in smartphones, tablets, and same type of gadgets. The
present permission restrict to mobile tools, like sensor information, is consistent and
coarse-grained. BlurSense is not static, fine-grained, feasible permission restrict
approach, behaving as a line of protection that permits users to describe and join
secrecy filters. As an output, the user may reveal filtered sensor information to
untrusted applications, and scientists can gather data in a manner that guards users’
privacy.
Oluwatimi et al. [10] presented an access control approach. In mentioned
approach, benefits would be dynamically allowed or disallowed to mechanisms
related to particular context of user. This application of context differentiates
between closely located subareas in a similar location. After a change in operating
system, context-based access control restrictions can be defined and enforced.
Ghafoor et al. [11, 12] presented a scheme. It is accuracy-forced privacy-
preserving access control scheme. The scheme is a composition of access reticence
and privacy protection methods. The access reticence method grants only allowed
query predicates on secrete information. The privacy-preserving unit anonymizes
information to reach privacy needs and imprecision restrictions on premise identified
by the permission restrict method. The permission restrict policies describe that
chosen premises were able to obtain the roles, whereas the privacy requirement is to
fulfill k-anonymity or l-diversity. An additional restriction which wants it should be
fulfilled by Privacy Protection Mechanism (PPM) is the imprecision bound for each
chosen premise.
3 System Overview
Figure 1 shows that users share location to Location-Based Service Network

(LBSN) through application. After obtaining the current location of user, FakeIt
runs and location given by FakeIt is shared. Adversary can take information when
interaction between user and LBSN is performing.
Whereas the third party server is not trustworthy, this type of server directly
attacks the user’s privacy or sells user private information to individual. An
adversary refers to malicious parties, which aim at user location details or hack data
and loose privacy. System is to expand the complication of extrapolate private
information; very difficult policies are acquired in secrecy protecting. To provide
high level of user security and manageability, it also reduces the transmission risk
and enhances security and privacy preserving.
4 System Architecture
In this system, a technique is proposed to preserve privacy of user-sensitive loca-

tion. Figure 2 shows proposed system architecture.
Using Google, API’s user can share his current location to the application
through the GPS or Wi-Fi. User can decide whether to release his location or not.
FakeIt decides before location shared with application. User can share current
location but if user’s current location is sensitive, then user decided that not to
release current real location. On such situations, system releases fake location. With
the FakeIt approach, we use to release fake location where KNN algorithm and
ranking algorithm are used. The use of KNN algorithm adds more privacy to the
Fig. 1 FakeIt approach

52 J. Rao et al.
Fig. 2 Proposed system architecture
over user-sensitive location. KNN algorithm used to find nearest locations from the
current location of the user. By using KNN algorithm, system finds n number of
nearest location. These n locations then reversed by the ranking algorithm. From
reversed ranked n locations, select top location and share it to the location-based
service. This has great difficulty to adversary to find users sensitive context from
release user context.
5 Mathematical Model
S ¼ ffu1 ; u2 ; . . .un gCL; f ðFLÞ; lat; lng; FIfFL; Topk; RevTopk; PCgg
CL ¼ lat þ ln
ð1Þ
FL 2 RevTopk
Zk
RevTopk ¼ ReverseOrderðCL 2 Topk Þ ð2Þ
0
FL ¼ TopðRevTopkÞ ð3Þ
X
k
f ðFLÞ ¼ TopReversek ð4Þ
i¼0
FK ¼ PCfFLg; ð5Þ
where
S = Start of the program, f ðFLÞ ¼ Fake Location, CL ¼ Current Location,
lat ¼ Lattitude, lng ¼ Longitude, Topk ¼ Top k Results,
RevTopk ¼ Reverse order of Top k Results,
FK = Fake Location, FI ¼ FakeIt, and PC ¼ Privacy check.
Equation (1) gives current location of user where user presents at the stage.
Location is the combination of ðlatÞ latitude and ðlngÞ longitude on the geospatial
data. Here, consider latitude as the X-axis and longitude as the Y-axis. By using a
KNN algorithm, find the nearest neighbor to current location. Consider Topk results
are found from the current location. Equation (2) gives reverse order of result and
from these results, Eq. (3) gives fake location where Eq. (4) gives function for
obtaining finally Eq. (5) that checks and gives the fake location.
6 Expected Results
There are privacy utility approaches for the efficiency of the different privacy
checks. Shown in table the average times for making the decision to privacy checks
and filter the current instances of traces. The existing approaches do not prevent
adversaries from inferring sensitive context.
The current context data may be breached to this exceed the belief than certain
level. This has a normal privacy guarantee.
Table 1 shows average processing time for existing approaches. Mask sensitive
shows lowest initialization time but it breaches privacy. The FakeIt provides
maximum utility than mask sensitive in which approach conceals data without
breaching privacy for more sensitive context.
Table 1 Avarage processing time

Algorithm Initialization time on PC (min) Release time on web (ms)
Mask sensitive 0 <1
MaskIt 18 <128
Naive fake >60 <1
Efficient fake 3 <1
FakeIt <3 <1
54 J. Rao et al.
7 Conclusion
In this, the problem of user’s privacy in location-based web application is dis-

cussed. Proposed work is related to the privacy preserving of user’s location.
Proposed work conceals the user-sensitive context being leaked from the
location-based services. The fake location context will be released for sensitive
context. An attacker has great trouble to learn original information from these
shared contexts.
Acknowledgements Success is never achieved single-handed. Apart from our humble efforts,
this paper is outcome of the help, co-operation, and guidance from various corners. I would like to
add a few heartfelt words for the people who were part of this in numerous ways and the people
who gave unending support right from the stage of ideas.
References
1. Lichen Zhang, Zhipeng Cai, Xiaoming Wang, “FakeMask: A Novel Privacy Preserving
Approach for Smartphones” IEEE Transactions on Network and Service Management, 2016.
J. Clerk Maxwell, A Treatise on Electricity and Magnetism, 3rd ed., vol. 2. Oxford:
Clarendon, 1892, pp. 68–73.
2. X. Liang, K. Zhang, X. Shen, and X. Lin, “Security and privacy in mobile social networks:
challenges and solutions,” IEEE Wireless Communications, vol. 21, no. 1, pp. 33–41, 2014.
3. Y. Najaflou, B. Jedari, F. Xia, L. T. Yang, and M. S. Obaidat, “Safety challenges and
solutions in mobile social networks,” IEEE Systems Journal, vol. 9, no. 3, pp. 834–854, 2015.
4. J. Tsai, P. G. Kelley, L. F. Cranor, and N. Sadeh, “Location sharing technologies: Privacy
risks and controls,” I/S: A Journal of Law and Policy for the Information Societ, vol. 6, no. 2,
pp. 119–317, 2010.
5. K. Vu, R. Zheng, and J. Gao, “Efficient algorithms for k-anonymous location privacy in
participatory sensing,” in Proceedings of the 31st Annual IEEE International Conference on
Computer Communications (INFOCOM’12), Orlando, FL, USA, March 25–30 2012,
pp. 2399–2407.
6. Y. Wang, Z. Cai, G. Ying, Y. Gao, X. Tong, and G. Wu, “An incentive mechanism with
privacy protection in mobile crowdsourcing systems,” ComputerNetwork., p. In Press, 2016.
7. M. Gotz, S. Nath, and J. Gehrke, “Maskit: Privately releasing user context streams for
personalized mobile applications,” in Proceedings of the 2012 ACM SIGMOD International
Conference on Management of Data (SIGMOD’12), Scottsdale, Arizona, USA, May 20–24
2012, pp. 289–300.
8. W. Wang and Q. Zhang, “A stochastic game for privacy preserving context sensing on mobile
phone,” in Proceedings of the 33rd Annual IEEE International Conference on Computer
Communications (INFOCOM’14), Toronto, Canada, April 27–May 2 2014, pp. 2328–2336.
9. J. Cappos, L. Wang, R. Weiss, Y. Yang, and Y. Zhuang, “Blursense: Dynamic fine-grained
access control for smartphone privacy,” in Proceedings of the IEEE Sensors Applications
Symposium (SAS’14), Queenstown, New Zealand, Febrary 18–20 2014, pp. 329–332.
10. B. Shebaro, O. Oluwatimi, and E. Bertino, “Context-based access control systems for mobile
devices,” IEEE Transactions on Dependable and Secure Computing, vol. 12, no. 2, pp. 150–
163, 2015.
11. Z. Pervaiz, W. G. Aref, A. Ghafoor, and N. Prabhu, “Accuracy constrained privacy preserving
access control mechanism for relational data,” IEEE Transactions on Knowledge and Data
Engineering, vol. 26, no. 4, pp. 795–807, 2014.
12. Rasika Pattewar and Jyoti Rao, “A Survey on Privacy Preserving Approaches for Location
Based Data”, International Journal of Advanced Research in Computer and Communication
Engineering, 240–243, Vol. 5, Issue 12, December 2016.
Analysis of Blind Image Watermarking
Algorithms
Chhaya S. Gosavi and Suresh N. Mali
Abstract This paper presents an overview of blind image watermarking algorithms.

In this paper, we analyzed these algorithms for different criteria like robustness,
security, and imperceptibility. We also compared pros and cons of using blind
method for embedding and extraction of watermark. Most of these algorithms are
implemented using MATLAB 2011 and tested on the standard image dataset. We
used true color images of size 256 256 and binary watermarks of size 32 32 for
testing. This paper will help watermarking researcher to choose the particular
algorithms depending on their need for the application they are working on.
Keywords Blind Watermark DCT DWT SVD
1 Introduction
A watermark is irremovable, imperceptibility, and robustly embedded in the mul-

timedia data. The process of embedding watermark is called watermarking.
A watermark is embedded for protection of digital data from illegally used. The
main applications of watermarking are copyright protected, authentication, piracy
detection, etc. Many researchers have been working in this area since two decades.
Image watermarking is explored by many researchers as it is easy compared with
audio and video data and also having a lot of redundant information which can be
used for embedding.
Watermarking process is mainly divided into three parts. First, watermark is
added into an image. In the next part, it is extracted from the watermarked image
and lastly authentication is done by comparing original and extracted watermarks.
Watermark insertion can be done in two ways, first is spatial domain and the second
C. S. Gosavi (&) S. N. Mali

Department of Computer Engineering, DYPIET, Pimpri, Pune, Maharashtra, India
e-mail: chhaya.gosavi@gmail.com
S. N. Mali
e-mail: snmali@rediffmail.com

https://doi.org/10.1007/978-981-10-7245-1_7
58 C. S. Gosavi and S. N. Mali
is transform domain. Watermarked image can be tampered with using different

attacks like common image processing attacks, geometric attacks, or compression
attacks. The watermark should be robust to such attacks. Spatial domain embedding
is fragile and cannot withstand even simple manipulations. Transform domain
technique is generally used for robust embedding.
The data is needed to remove the watermark partition; the watermarking algo-
rithm is divided into two types, blind algorithms and non-blind algorithms. At the
time of watermark removal if we need original image, then the algorithms are called
non-blind algorithms otherwise blind algorithm. In non-blind algorithms, it is easy
to compare original images and watermarked image and generated watermark data.
The sender has to transfer original image along with watermarked image to the
receiver. But practically, it is not practicable and this does not assist the purpose of
watermarking. Non-blind algorithms can be more robust compared with blind
algorithms but not good.
Hence, in this paper, we compare different methods for blind insertion and
detection of watermarks.
2 Literature Review
Let us consider I1 is the image, W1 is the watermark, I2 is watermarked image, and

W2 is extracted watermark. Here, we discuss algorithms to embed W1 in I1 and
extract W2 from I2.
Algorithm 1 In [1–3], authors have inserted watermark in third level Discrete
Wavelet Transform (DWT) of an image, that is, in LH3 sub-band. Maximum value
of pixel’s eight neighbors is “t”. Every coefficient is replaced by the following
equation:
New LH3 ¼ LH3 þ W1 LH3 ;

W1 ¼ 1 if W ¼ 1 and LH3 [ t
or W ¼ 1 and LH3\t ð1Þ
W1 ¼ 1 if W ¼ 1 and LH3 \t
or W ¼ 1 and LH3 [ t
for extraction, authors have used W as follows:
W2 ¼ 1 if W ¼ 1 and LH3 [ t
or W ¼ 1 and LH3 \t ð2Þ
¼ 1 otherwise:
The algorithm that uses original watermark for extraction is known as non-blind
algorithm.
Analysis of Blind Image Watermarking Algorithms 59
Algorithm 2 In [4], DWT and SVD techniques are used for embedding. First,
DWT of image is obtained by the SVD of the low-frequency part that is calculated
and then a watermark is embedded into highest singular value. These steps are
explained by following equations:
½LL; LH; HL; HH ¼ DWTðI1Þ

½U S V ¼ svdðLLÞ
if ðW1 ¼¼ 1Þ
ð3Þ
if ððS%QÞ\¼ Q=4Þ
S ¼ S ðS%QÞ Q=4;
else
S ¼ S ðS%QÞ þ 3 Q=4;
end
ð4Þ
else
if ððS%QÞ [ ¼ 3 Q=4Þ
S ¼ S ðS%QÞ þ 5 Q=4;
ð5Þ
else
S ¼ S ðS%QÞ þ Q=4;
end ð6Þ
end
Here, Q is the embedding strength.
Extraction Procedure:
NewLL ¼ DWTðI2Þ
½U1 S1 V1 ¼ svdðNewLLÞ
ð7Þ
W2 ði,jÞ ¼ 1 if ððS1%QÞ [ Q=2Þ
¼ 0 otherwise
Q is needed for extraction. The algorithm is semi-blind.

Algorithm 3 In [5], watermark is embedded by the following method:
C ¼ DWTðI1Þ;
if W1½j ¼ 1;
ð8Þ
interchange C½i with max of ðC½i; C½i þ 1. . .:C½i þ 5Þ
Else
Interchange C½i with min of ðC½i; C½i þ 1; . . .. . .C½i þ 5Þ

For the extraction of watermark; they have used ð9Þ
W2 ¼ DWTðI1Þ
if W2 [ medianðWC½i; WC½i þ 1; . . .. . .WC½i þ 5;

W2 ¼ 1
ð10Þ
Else
W2 ¼ 0
This algorithm is a pure blind algorithm.

Algorithm 4 In [6], watermark is embedded using
if W1ðkÞ ¼¼ 1
I2 ð4; 4; 1Þ ¼ I1ð4; 4; 1Þ þ k; ð11Þ
end
if W1ðkÞ ¼¼ 0
I2 ð4; 4; 1Þ ¼ I1 ð4; 4; 1Þ k; ð12Þ
End
for extraction
If ððI2ð5; 5; 1ÞÞÞ [ ¼ I2ð4; 4; 1Þ

W2ðiÞ ¼ 0;
else ð13Þ
W2ðiÞ ¼ 1;
end
This algorithm is also pure blind algorithm.

Algorithm 5 In [7], embedding in HH band of DWT
If ðW1ðiÞ ¼¼ 1Þ then
blocHH ¼ blocHH þ K 1
Else
blocHH ¼ blocHH þ K 1
End If
Here, K is the robustness factor = 20

for extraction,
find gradient of each HH block, i.e., GHH
If ðSum Of GHH Coefficients [ 0Þ then

W2ðiÞ ¼ 1
Else
If ðSum Of GHH Coefficients\ 0Þ
W2ðiÞ ¼ 0
End If
End If
Algorithm 6 In [8, 9], singular value decomposition is used for embedding the
watermark:
½s1; s2; s3. . . ¼ svdðI1Þ;

s1; s2; s3. . . are singular values
Watermark is embedded by the following method:
Si ¼ 0:5½ðSi 1 þ Si þ 1Þ þ W1ðiÞ ðSi 1Si þ 1Þ ð14Þ
Extraction process is as follows:
W2ðiÞ ¼ 0 if Si [ 0:5ðSi 1 þ Si þ 1Þ
¼ 1 otherwise
In [10], they are using the same key for embedding and extraction of watermark
it makes the algorithm semi-blind.
In [11], embedding procedure is not explained here. They have applied genetic
algorithms and extracted watermark by doing fitness test.
In [12], image I1 is decomposed into three-level DWT. SVD is applied on HH
band to get singular values. These are replaced by singular values of watermark.
Reverse method is applied for extraction.
Algorithm 7 Rajab [9] uses LSB method to embed watermark as follows:
HL ¼ DWTðI1Þ
½U SHL V ¼ svdðHLÞ
LSB ðSHL ði; iÞÞ ¼ W1ðiÞ
Extraction method
HL1 ¼ DWTðI2Þ
½U SHL1 V ¼ svdðHL1Þ
W2ðiÞ ¼ LSB ðSHL1 ði; iÞÞ
Algorithm 8 Karmakar [13, 14] uses the following algorithm for embedding:
if ðW1ðkÞ ¼¼ 1Þ
if ððI1Þ%dÞ\a)
ð16Þ
I2 ¼ I1ðI1%dÞ a
else
I2 ¼ I1ðI1%dÞ þ c
endif
ð17Þ
elseif ðW1ðkÞ ¼¼ 0Þ
if ððI1%dÞ [ cÞ
I2 = I1 ðI1%dÞ þ e
ð18Þ
else
I2 ¼ I1 ðI1%dÞ þ a
ð19Þ
endif
Extraction process is
if ððI2%dÞ [ bÞ
W2ðbÞ ¼ 1
else ð20Þ
W2ðbÞ ¼ 0
endif
This algorithm is also semi-blind as it uses “d” for embedding as well as
extraction.
As seen from the literature survey, three things are important while embedding a
watermark into the image.
1. Watermark embedding position should be known at extraction process.
2. Embedded bit should survive after different manipulations and attacks.
3. Embedding watermark should not affect quality of the cover.
3 Proposed Algorithm
The proposed algorithm for embedding and extraction is explained below:

Embedding algorithm:
Input—Image I1 (256 256), watermark W1 (32 32);
Output—Embedded Image I2 (256 256).
1. Divide I1 into 8 8 blocks Iblk;
2. Calculate [U T] = SCHUR(Iblk);
3. Max = Maximum(T);
4. Min = Minimum (T);
5. if(W1(i) == 1)
T(0,0) = Max
else
T(0,0) = Min;
6. Continue till all watermark bits get embedded in the image; and
7. Save I2 as watermarked image.
Extraction algorithm:
Input—Watermarked Image I2 (256 256);
Output—Extracted watermark W2 (32 32).
1. Divide I2 into 8 8 blocks Iblk;
2. Calculate [U T] = SCHUR(Iblk);
3. Max = Maximum(T);
4. Min = Minimum (T);
5. if T(1,1) = Max
W2(i) = 1
else
W2(i) = 0;
6. Continue till all watermark bits are extracted from the image; and
7. Save W2 as extracted watermark.
Table 1 summarizes the results.
PSNR is calculated to check imperceptibility of watermarked image by the
following equation:
1 XM X N
MSE ¼ ðxði; jÞ yði; jÞÞ2 ð21Þ
MN i¼1 j¼1
2
ð2n 1Þ
PSNR ¼ 10 log10 pffiffiffiffiffiffiffiffiffiffi ð22Þ
MSE
Normalized cross-correlation (NK) is calculated to test robustness of extracted

watermark by the following equation:
Table 1 Comparison of blind watermarking algorithms

Algorithm Algo1 Algo2 Algo3 Algo4 Algo5 Algo6 Algo7 Algo8 Proposed
Blindness Non Semi Pure Pure Pure Pure Pure Semi Pure
PSNR (dB) 45.44 43.21 47.32 41.53 54.43 41.03 49.7 40.6 60.20
NK—No 0.9593 0.9231 0.9421 0.9133 0.9741 0.9265 0.9434 0.9121 0.9982
attack
Median 0.9382 0.9072 0.9112 0.8913 0.9512 0.9051 0.9211 0.8943 0.9895
filtering
(3 3)
Smoothing 0.9322 0.9245 0.9176 0.9063 0.9476 0.9143 0.9212 0.8987 0.9863
Cropping 0.9002 0.8910 0.8941 0.8921 0.9384 0.8965 0.9138 0.8912 0.9754
(10%)
Compression 0.8893 0.8928 0.8942 0.8933 0.8974 0.8926 0.8944 0.8912 0.9482
P
M P
N
ðxði; jÞ yði; jÞ
i¼1 j¼1
NK ¼ ð23Þ
P
M P
N
2
ðxði; jÞÞ
i¼1 j¼1
4 Conclusion
The algorithm presented here is a pure blind algorithm. It is a block-based algorithm

so one bit of watermark is embedded in 8 8 block of cover image which helped
to maintain quality of watermarked image. This method is robust for all signal
processing attacks like median filtering, smoothing, cropping, and compression.
References
1. A. Essaouabi and F. Regragui: A Blind Wavelet-Based Digital Watermarking for Video, In:
International Journal of Video & Image Processing and Network Security, Vol 09, No 09,
(2009), pp. 37–41
2. Majid Masoumi and Shervin Amiri: A Blind Video Watermarking Scheme Based on 3D
Discrete Wavelet Transform In: International Journal of Innovation, Management and
Technology, Vol. 3, No. 4, (2012) pp. 487–490
3. DING Hai-yang, ZHOU Ya-jian, YANG Yi-xian, ZHANG Ru: Robust Blind Video
Watermark Algorithm in Transform Domain Combining with 3D Video Correlation In:
Journal of Multimedia, Vol. 08, NO. 2 (2013) pp. 161–166
4. Chetan K. R Raghavendra K.: A Blind Video Watermarking Algorithm Resisting to Rotation
Attack. In: IEEE International Conference on Computer and Communications Security,
(2009) pp. 111–114
5. Chetan K. R Raghavendra K: DWT Based Blind Digital Video Watermarking Scheme for
Video Authentication, In: International Journal of Computer Applications (0975—8887),
Volume 4– No. 10, (2010), pp. 19–26
6. D. Niranjan babu, D. Jagadeesh: A Blind and Robust Video Water Marking Technique in
DCT Domain, In: International Journal of Engineering and Innovative Technology (IJEIT),
Volume 2, Issue 2, (2012) pp. 128–132
7. Henri Bruno Razafindradina and Attoumani Mohamed Karim: Blind and Robust Image
Watermarking based on Wavelet and Edge insertion In: International Journal on
Cryptography and Information Security (IJCIS), Vol. 3, No. 3, (2013) pp. 23–30
8. S. Naveen kumar, U. Nageswar Rao: Robust & Blind Video Watermarking Algorithm
Using SVD Transform, In: International Journal of Engineering Research & Technology
(IJERT), ISSN: 2278-0181, Vol. 3 Issue 5, (2014) pp. 1359–1364
9. Lama Rajab, Tahani Al-Khatib, Ali Al-Haj: A Blind DWT-SCHUR Based Digital Video
Watermarking Technique, In: Journal of Software Engineering and Applications, 8, (2015),
pp. 224–233
10. Md. Asikuzzaman, Md. Jahangir Alam, Andrew J. Lambert, Mark Richard Pickering,:
Imperceptible and Robust Blind Video Watermarking Using Chrominance Embedding: A Set
of Approaches in the DT CWT Domain, In: IEEE Transactions of Information Forensics and
Security, Vol. 09, No. 09, (2014), pp. 1502–1517
11. Nitin A. Shelke, Dr. P. N. Chatur: Blind Robust Digital Video Watermarking Scheme using
Hybrid Based Approach, In: International Journal of Computer Science and Information
Technologies, Vol. 5 (3), (2014) pp. 3619–3625
12. Sakshi Batra, Rajneesh Talwar: Blind Video Watermarking based on SVD and
Multilevel DWT, In: European Journal of Advances in Engineering and Technology, 2(1),
(2015), pp. 80–85
13. Amlan Karmakar, Amit Phadikar, Baisakhi Sur Phadikar, Goutam Kr. Maity: A blind video
watermarking scheme resistant to rotation and collusion attacks, In: Elsevier, Journal of King
Saud University—Computer and Information Sciences, 28, (2016) pp. 199–210
14. X. Y. Wang, Q. L. Shi, S. M. Wang, H. Y. Yang: A blind robust digital watermarking using
invariant exponent moments, In: Elsevier, Int. J. Electron. Commun. (AEÜ), 70, (2016)
pp. 416–426
ClustMap: A Topology-Aware MPI
Process Placement Algorithm
for Multi-core Clusters
K. B. Manwade and D. B. Kulkarni
Abstract Many high-performance computing applications are using MPI

(Message Passing Interface) for communication. The performance of MPI library
affects the performance of MPI applications. Various techniques like communica-
tion latency reduction, increasing bandwidth, and increasing scalability are avail-
able for improving the performance of message passing. In multi-core cluster
environment, the communication latency can be further reduced by topology-aware
process placement. This technique involves three steps: finding communication
pattern of MPI application (application topology), finding architecture details of
underlying multi-core cluster (system topology), and mapping processes to cores. In
this paper, we have proposed novel “ClustMap” algorithm for the third step. By
using this algorithm, both system and application topologies are mapped. The
experimental results show that the proposed algorithm outperforms over existing
process placement techniques.

Keywords High-performance computing Topology-aware process placement

System topology MPI application topology
K. B. Manwade (&)
Department of Computer Science and Engineering, Ph.D. Research Center, Walchand
College of Engineering, Sangli, Maharashtra, India
e-mail: mkarveer@gmail.com
K. B. Manwade
Shivaji University, Kolhapur, Maharashtra, India
D. B. Kulkarni
Department of Information Technology, Walchand College of Engineering, Sangli,
Maharashtra, India
e-mail: d_b_kulkarni@yahoo.com

https://doi.org/10.1007/978-981-10-7245-1_8
68 K. B. Manwade and D. B. Kulkarni
1 Introduction
MPI is a dominant message passing library used in high-performance clusters.

Therefore, improving its performance is very much essential. Computer architec-
tures are evolved from uni-core to multi-core as illustrated in [1]. Also, modern
interconnect technologies provide bandwidth in gigabytes [2]. The development in
these two areas gives an opportunity to take advantages of shared memory, pro-
cessor locality, and high communication bandwidth which are available in
multi-core clusters. Current MPI implementations map processes on processors or
cores without considering topology of application or topology of the underlying
system. While placing processes in multi-core clusters, the affinity between cores
and processes should be known. This leads to the development of new
topology-aware process placement technique as suggested in [3]. In this paper, we
suggest new topology-aware process placement algorithm.
The rest of this paper is organized as follows. In Sect. 2, concepts of
topology-aware process placement, the data structure for system, and application
topology are illustrated. The topology-aware process placement algorithm,
its working, and time complexity are given in Sect. 3. In Sect. 4, works related
to topology-aware process placement are listed. The results of the “ClustMap”
algorithm for benchmark applications are given in Sect. 5. The summary of our
work, conclusion, and future work are given in Sect. 6.
2 Topology-Aware Process Placement
2.1 Process Placement Techniques
As shown in Fig. 1, there are two approaches for process placement: static and
dynamic. Static approach places the process before the start of the execution,
whereas dynamic approach reschedules the process at run time. There are two
sub-techniques of static approach. The first technique, core binding [4], determines
physical core on which a specific MPI process should be placed so that the com-
munication latency will be reduced. In this technique, modification of MPI appli-
cation is not required. So the performance of legacy as well as new MPI application
can be improved by using this technique. But before process placement, both
system and application topologies should be known. Second technique, rank
reordering [5] maps processes with default policy and then by using communication
pattern of processes the ranks of processes are reordered under the communicators.
The constraint for rank reordering is that it should be performed before application
data is loaded and for that purpose few new lines need to be added to MPI code.
In the second approach, processes are remapped to another core or processor
during runtime. Two techniques are available in this approach; process migration
and rank reordering. Process migration is more complicated and time-consuming
ClustMap: A Topology-Aware MPI Process Placement Algorithm … 69
Fig. 1 Taxonomy of topology-aware MPI process placement techniques
technique. Implementation of rank reordering is easy but in time slice-based rank,

reordering handling of intermediate result is difficult. We have used core binding
technique for process placement.
2.2 Data Structure for System and Application Topology
Topology-aware process placement is an NP-complete problem as mentioned in [6].

The mapping time depends on data structure used for the representation of the
system topology and application topology. We have used a data structure called as
“dendrogram” [7] to represent both system and application topologies. The key
benefits of this data structure are as follows:
1. It explores the hierarchical nature of the system.
2. It forms clusters of cores based on the various nodes as well as network
parameters and hence finds the nearest cores.
3. Its tree-like structure reduces the mapping time.
4. It is compatible with existing work as it works on data stored in matrix format.
The sample network-based system is shown in Fig. 2a and its topology in
“dendrogram” format is shown in Fig. 2b.
Figure 2a shows that four nodes are connected in a network. Each node has eight
cores that are situated on two sockets. In this case, cores can be grouped hierar-
chically according to their location. This grouping is socket wise (groups G1–G8),
node wise (groups G9–G12), and cluster-wise (group G13) as shown in Fig. 2b.
Fig. 2 a Sample network-based system b Sample network-based system using “dendrogram”

3 Topology-Aware Process Placement Algorithm
3.1 The Process Mapping Problem
We have treated process to core mapping as cluster mapping problem. Let P and
C are the set of clusters of processes and cores, respectively.
P¼f p1 ; p2 ; p3 ; . . .; pn g ð1Þ
C¼fc1 ; c2 ; c3 ; . . .; cm g ð2Þ
The mapping function “F” is defined as
F:P !C ð3Þ
The mapping function maps processes from set P to cores in set C, if the
following conditions are satisfied:
jPi j jCk j ð4Þ
where 1 < i < n, 1 < k < m, and number of processes in cluster Pi and number of
cores in cluster Ck are same. The first condition from Eq. (4) states that if the size of
sub-cluster Ck is greater than or equal to sub-cluster Pi, then one-to-one mapping of
the process to the core is possible. Otherwise, more processes need to be scheduled
on a core instead of one process per core:
Pn
j¼1 E Pj
E ðPi Þ ; ð5Þ
ð n 1Þ
where i 6¼ j, E is the affinity function, and process from Pi have more communi-
cation affinity than process from Pj. The second condition in Eq. (5) indicates that
processes from Pi are chosen for mapping if they have more affinity between them
than processes from Pj.
Pn
k¼1 E ðCl Þ
E ðCk Þ ; ð6Þ
ð m 1Þ
where k 6¼ l, E is the affinity function, and cores from Ck have more affinity than
cores from Cl. The third condition from Eq. (6) indicates that cores in Ck have more
affinity than cores from Cl
Pi and Ck are unmapped clusters of processes and cores respectively ð7Þ
According to the fourth condition from Eq. (7), only unmapped clusters, Pi and
Ck, are chosen for mapping.
3.2 The “ClustMap” Algorithm
---------------------------------------------------------------------------
Algorithm1: ClustMap
---------------------------------------------------------------------------
1. Input:
2. S //Hierarchical cluster of cores
3. A // Hierarchical cluster of processes
4. Si and Ai //Sub-clusters of S & A respectively
5. Output:
6. M //Process to core mapping
7. Procedure: ClustMap(S,A)
8. If(Size of Cluster(S) = Size of Cluster(A))
9. MapClusters(S,A); //Perform mapping
10. ElsIf(Size of Cluster(S) < Size of Cluster(A))
11. Remap(S, A); //Perform mapping with more processes on each node
12. ElsIf(Size of Cluster(S) > Size of Cluster(A))
13. For (Each sub cluster Si of S)
14. ClustMap(Si, A); //Recursively check for the mapping
15. EndFor
16. EndIf
17. EndProcedure
---------------------------------------------------------------------------
18. Procedure: ReMap(S,A)
19. For (Each process Pi from A and core Cj from S)
20. M[0][i]=Pi and M[1][i]=Cj //Map Pi to Cj as 1:1
21. EndFor
22. For (Each remaining process Pj from A)
23. If (Affinity(M[0][i] , Pj) is more)
24. M[1][i]= Pj //Map remaining process
25. EndIf
26. EndFor
27. EndProcedure
---------------------------------------------------------------------------
28. Procedure: MapClusters(S,A)
29. For (Each leaf node of A)
30. M[0][j]=Sj
31. M[1][j]=Aj
32. EndFor
33. EndProcedure
---------------------------------------------------------------------------
The algorithm 1 takes “dendrogram” of system and application topology as

inputs and produces a process to core mapping as output. If the size of both system
and application “dendrograms” is equal, then process to core mapping is performed
(Lines 8–9). If the size of system “dendrogram” is less than application “dendro-
gram”, then map processes to cores with more processes on each node based on
process affinity (Lines 10–11). If the size of system “dendrogram” is greater than
application “dendrogram”, then it recursively goes to next level of “dendrogram”
until the application “dendrogram” maps to system “dendrogram” (Lines 12–14).
The “ClustMap” algorithm maps processes to cores hierarchically, and therefore
the complexity varies linearly with a number of cores in the system and number of
processes in the application. If M be a number of cores in the system and N be a
number of processes in the application, then the complexity of this algorithm is

O(M + N).
4 Related Work
In literature, various topology-aware process placement algorithms are given. The

process placement is treated as graph mapping or tree mapping problem. MPI
standard [8] itself provides set of routines to create and manipulate topologies, but
all MPI implementations do not support these features. Also, MPI standard does not
support mapping of application topology on system topology.
A recursive bi-partitioning algorithm [9] is used to map application topology on
system topology. In [10], graph-based mapping algorithm is used. For process
mapping, lists of subgraphs of both application and system topologies are created
and by using permutation, topologies are mapped. In [11], author treated
topology-aware process placement as graph embedding problem. The virtual
topology of processes is mapped on physical topology of cores. To represent the
physical topology of system, the tree data structure is used.
In [12], authors treated the mapping as graph isomorphism problem. They have
developed a “mapper” algorithm for processor array and mapped processes in
pair-wise manner. In [13], authors treated process mapping as graph mapping
problem. They have considered weighted graph to indicate the volume of data
exchanged between processes and distance between processors. Initially, processes
are mapped using greedy approach and then pair-wise exchange is used. Similarly,
in [14], authors treated mapping as graph mapping problem. They developed
GRASP heuristic algorithm for mapping. To evaluate the quality of mapping,
dilation metric is used.
In [15, 16], authors proposed different process mapping strategies for Blue Gene
system. This implementation is specific to Blue Gene system.
In [17], authors constructed application topology using MPI Graph create
function and to construct system topology, they have used hwloc tool and network
discovery tool. For process mapping, rank reordering technique is used.
In [18], authors have used TreeMatch algorithm for process mapping. To opti-
mize mapping function, two parameters, network contention and load balance,
among processors are used. To collect hardware topology, hwloc tool is used and
for application topology, a profiling tool is used.
5 Results
5.1 Experimental Setup
For performance evaluation, we have used “SimGrid” [19] simulator. The MPI
programs are compiled using “smpicc” compiler and are executed using “smpirun”
command. A platform file of four nodes with 64 cores each is defined. The simu-
lation system contains totally 1024 cores. The “ClustMap” algorithm gives
topology-aware mapping of MPI processes in host file and it is used as one of the
parameters for “smpirun” command. For testing purpose, programs from OSU
micro-benchmarks version 5.3.2 [20] are used. This benchmark contains programs
for point-to-point and collective communications.
The performance of topology-aware process placement technique is measured as
the extent of latency reduced through effective process placement. The performance
of the “ClustMap” algorithm is measured as the time required for mapping pro-
cesses to cores.
5.2 Latency Reduction for MPI Communication
To compare the performance of topology-aware process placement with default

process placement, the benchmark programs are used. The communication latency
of these programs is measured by varying data size from “0” to “1024” bytes. As
shown in Fig. 3a, b topology-aware placement significantly reduces latency of MPI
programs having point-to-point communication.
In collective communication, all processes are involved; therefore,
topology-aware process placement does not reduce communication latency.
Figure 4a shows that latency of the benchmark program containing only collective
communication is not reduced. Thus, topology-aware process placement alone is
not sufficient to improve performance collective communication.
5.3 Performance Evaluation of “ClustMap” Algorithm
The time complexity of “ClustMap” algorithm is O(M + N). Its performance is

measured by varying number of processes. The process of core mapping time is
measured.
As shown in Fig. 4b, the “ClustMap” algorithm is scalable as it shows linear
performance for increasing number of processes in the application.
Fig. 3 a Latency reduction for point-to-point communication b Latency reduction for

point-to-point and collective communication
6 Conclusion
To reduce the latency of MPI program and improve its performance by using
topology-aware process placement, three steps are required: finding application
topology, finding system topology, and topology-aware process placement. This
paper contributes to the third step. We have proposed “ClustMap” algorithm which
places processes efficiently.
It is observed that topology-aware process placement approach gives better
performance than default process placement. The extent of improvement depends
Fig. 4 a Latency reduction for collective communication b Scalability of “ClustMap” algorithm
on the affinity between the processes to be mapped. Thus, the performance

improvement of topology-aware process placement is proportional to the affinity
between the processes. The complexity of “ClustMap” varies linearly with number
of processes. By using “ClustMap” algorithm, latency of only point-to-point
communication can be reduced. In collective communication, all processes are
involved therefore for collective communication latency reduction though
topology-aware placement is not possible.
We have planned to improve the performance of collective communication by
dividing it into inter-node and intra-node communications. For inter-node com-
munication, point-to-point communication functions and for intra-node communi-
cation, shared memory technique will be used.
References
1. Balaji Venu, “Multi core processors: An overview”, Department of Electrical Engineering and
Electronics, University of Liverpool, Liverpool, UK, (2010).
2. Dennis Abts and John Kim, “High Performance Datacenter Networks: Architectures,
Algorithms, and Opportunities”, Synthesis Lectures on Computer Architecture, Book series,
Pages 1–115, Vol. 6, No. 1, Morgan Claypool Publishers, (2011).
3. Mohammad Javad Rashti, Jonathan Green, Pavan Balaji, Ahmad Afsahi and William Gropp,
“Multi-core and Network Aware MPI Topology Functions, EuroMPI’11, ISBN:
978-3-642-24448-3 (2011), pp. 50–60.
4. Joshua Hursey, Jerey M. Squyres and Terry Dontje, “Locality-Aware Parallel Process
Mapping for Multi-Core HPC Systems”, IEEE International Conference on Cluster
Computing, (2011).
5. B. Brandfass, T. Alrutz and T. Gerhold, “Rank reordering for MPI communication
optimization”, 23rd Int. Conf. on Parallel Fluid Dynamics, (2013), pp. 372–380.
6. Torsten Hoefler, Emmanuel Jeannot, Guillaume Mercier, “An Overview of Process Mapping
Techniques and Algorithms in High-Performance Computing”, High Performance Computing on
Complex Environments, Wiley, pp. 75–94, 2014, 978-1-118-71205-4 <hal-00921626>, (2014)
7. Alberto Fernndez and Sergio Gmez, “Solving Non-Uniqueness in Agglomerative Hierarchical
Clustering Using Multi-dendrograms”, Journal of Classification 25:43-65, ISSN: 0176–4268,
(2008).
8. MPI, “MPI 2.2: A Message-Passing Interface Standard”, MPI Forum, www.mpi-forum.org/
docs/mpi-2.2/mpi22-report.pdf (2009).
9. B. W. Kernighan and S. Lin, “An efficient heuristic procedure for partitioning graphs”, Bell
System Technical Journal, 59 (1970), pp. 291–307.
10. T. Hoeer and M. Snir, “Generic topology mapping strategies for large-scale parallel
architectures”, Int. Conf. on Supercomputing, (2011), pp. 75–84.
11. T. Hatazaki, “Rank reordering strategy for MPI topology creation functions”, Recent
Advances in Parallel Virtual Machine and Message Passing Interface, EuroPVM/MPI 1998,
Lecture Notes in Computer Science, vol 1497, pp. 188–195, (1998).
12. S. H. Bokhari, “On the mapping problem”, IEEE Transactions on Computers, Volume 30
Issue 3, Pages 207–214, ISSN: 0018-9340, (1981).
13. S. Y. Lee and J. K. Aggarwal, “A mapping strategy for parallel processing”, IEEE
Transactions on Computers, Volume 36 Issue 4, Pages 433-442, ISSN: 0018–9340, (1987).
14. C. Sudheer and A. Srinivasan, “Optimization of the hop-byte metric for effective
topology-aware mapping”, 19th Int. Conf. on HPC, (2012), pp. 19.
15. P. Balaji, R. Gupta, A. Vishnu, and P. H. Beckman, “Mapping communication layouts to network
hardware characteristics on massive-scale Blue Gene systems”, Journal of Computer Science -
Research and Development, Volume 26 Issue 3-4, Pages 247–256, ISSN:1865-2034 (2011).
16. B. E. Smith and B. Bode, “Performance effects of node mappings on the IBM Blue-Gene/L
machine”, Euro-Par 2005. Lecture Notes in Computer Science, vol 3648, Pages 1005–1013,
ISBN: 978-3-540-28700-1, (2005).
17. J. L. Traff, “Implementing the MPI process topology mechanism”, Proceedings of the 2002
ACM/IEEE conf. on Supercomputing, IEEE Computer Society Press, (2002), pp. 114.
18. Guillaume Mercier, Jerome Clet-Ortega, “Towards an efficient process placement policy for
MPI applications in multi-core environments”, Recent Advances in Parallel Virtual Machine
and Message Passing Interface. EuroPVM/MPI 2009, pp. 104–115, ISBN:
978-3-642-03769-6 (2009).
19. Henri Casanova, Arnaud Giersch, Arnaud Legrand, Martin Quinson, Frederic Suter,
“Versatile, Scalable, and Accurate Simulation of Distributed Applications and Platforms”,
Journal of Parallel and Distributed Computing, Volume 74, Issue 10, pp. 2899–2917, (2014).
20. http://mvapich.cse.ohio-state.edu/benchmarks/.
Mobile Agent-Based Frequent Pattern
Mining for Distributed Databases
Yashaswini Joshi, Shashikumar G. Totad, R. B. Geeta and

P. V. G. D. Prasad Reddy
Abstract In today’s world of globalization, business organizations produce

information from many branch offices of their business while operating across the
globe and hence lead to large chunk of distributed databases. There is an innate
need to look at this distributed information that leverages the past, monitors the
present, and predicts the future with accuracy. Mining large distributed databases
using client–server model is time-consuming and sometimes impractical because it
requires huge databases to be transferred over very long distances. Mobile agent
technology is a promising alternative that addresses the issues of client–server
computing model. In this paper, we have proposed an algorithm called MADFPM
for frequent pattern mining of distributed databases that use mobile agents. We have
shown that the performance of MADFPM is better compared to the conventional
client–server approach.

Keywords Business intelligence Distributed database Data mining

Mobile agents Frequent patterns FP-tree Parallel computing

Association rules Decision-making
Y. Joshi (&) S. G. Totad

Department of Computer Science and Engineering, BVBCET, Hubballi, India
e-mail: yashaswinijoshi@gmail.com
S. G. Totad
e-mail: totad@bvb.edu
R. B. Geeta
Department of Information Science, KLEIT, Hubballi, India
e-mail: geetatotad@yahoo.co.in
P. V. G. D. Prasad Reddy
Department of CS & SE, Andhra University, Visakhapatnam, India
e-mail: prasadreddy.vizag@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_9
78 Y. Joshi et al.
1 Introduction
Business intelligence is all about making right decision at right time. A lot of
business intelligence is hidden in the enterprise database that helps in making wise
decisions. Association rules ([1, 2]) are very useful for business promotions.
Furthermore, in the today’s world of globalization and internet, related data is
more often distributed and shared among multiple parties across remote locations.
In case of large distributed databases, the conventional centralized client–server
model of computation is time-consuming and sometimes impractical, because it
requires huge databases to be transmitted over very long distances. Chattratichat
et al. [3], in their work, explained how the distributed systems support parallel
computation while describing the demands of the distribute databases for data
mining. Mobile agents are a promising alternative technology attracting interest of
researchers from the fields of distributed systems that address the issues of cen-
tralized client-–server computing model. In this paper, we propose a frequent
pattern mining approach for distributed databases that uses mobile agents. The
proposed approach not only minimizes the data transmission but also reduces time
required to compute frequent patterns by the way of inherent parallel execution.
2 Related Works
The notations given by Ruan et al. [4] are as follows: LFI denotes the set of local
frequent itemsets, GFI is the set of global frequent itemsets, and CGFI denotes the
set of candidate GFI. The works proposed by Kulkarni et al. [5] and Ruan et al. [4]
obtain frequent patterns by computing LFIs, GFIs, and CGFI and performing
parallel computations. The parallel algorithm for Mining Frequent Itemsets (PMFI)
proposed by You-Lin Ruan et al. constructs FP-tree at each of the distributed sites
DBi. It computes LFI and sends it to the central site. The central site receives LFIs
from all distributed sites and computes GFI and CGFI. It sends CGFI back to each
of the sites. An itemset X CGFI may be frequent or infrequent at a distributed
site. Each of the distributed sites computes support count of all the itemsets from
CGFI that are infrequent at that site and sends it back to the central site for
consideration to include into the GFI.
Realizing the negative effect of heavy network traffic on the performance of
distributed/parallel data mining applications, in their work proposed a programmable
data parallel to primitive D-DOALL, a generic mechanism to schedule a set of
independent tasks on a network of workstations. A framework called MAD-ARM is
proposed by Saleem Raja and George Dharma Prakash Raj [6], which attempts to
reduce the communication overhead. Research work published by Liu [7] demon-
strates how set of mobile agents having specialized expertise can be used to gather
relevant information and how they cooperate with each other to arrive at timely
Mobile Agent-Based Frequent Pattern Mining … 79
decisions, in dealing with various supply chain finance–business analysis scenarios.

The technique proposed by Kulkarni et al. [5] also computes LFIs concurrently like
PMFI algorithm, but using mobile agents. Also, Folorunso et al. [8] proposed an
improved cost model for mining association rule for distributed databases.
3 Proposed MADFPM Algorithm
The two approaches are proposed by Kulkarni et al. [5] and Ruan et al. [4] and
discussed above in Sect. 2, indulge thrice in transfers of itemsets between central
site and distributed sites or between distributed sites themselves. This transfer of
itemsets, particularly the transfer of LFIs between the distributed sites, results in
heavy network traffic leading to network congestion. This in turn reduces the
overall computation speed of the algorithm, as the distributed sites will have to wait
long for the receipt of the LFIs from all other neighbors. The technique that we
present in this section transfers itemsets only once and avoids the multiple transfers
of itemsets.
The technique proposed has a main component called Parent that performs
operations of central site. It is responsible for creating mobile agents, dispatching
them to remote distributed sites, receiving results from retracted mobile agents, and
processing the received results. The Parent creates and dispatches mobile agents
twice in its course of action. First time, it dispatches mobile agents called
SupportCount_MobileAgent to distributed sites to get Local Support Count (LSC)
of items. It computes Global Support Count (GSC) by adding the LSCs of corre-
sponding items received from distributed sites. Then, it computes L-order from the
computed GSCs of the items. L-order is an ordering of items with support count
minSup that is arranged in the decreasing order of item support count. Second
time, the Parent dispatches mobile agents called FP-tree_MobileAgents to each of
the distributed sites for getting frequent itemsets. FP-tree_MobileAgents carry the
L-order computed by the Parent while migrating. Each mobile agent migrates to its
respective distributed site and constructs FP-tree by accessing the distributed
database located there. Mobile agents explore the constructed FP-tree [9–11] and
retract back to the Parent with LocalDisjoint Matrix (LDM) containing “Composite
Itemsets”. A Composite Itemset contains multiple itemsets hidden within it. It
represents a path from a child of the root node to a leaf node of the FP-tree.
Below-mentioned property and the theorem, defined by Tanbeer et al. [12], are
equally applicable and hold good here as well.
Property The total frequency count of any node in the FP-tree is greater than or
equal to the sum of the total frequency counts of its children.
Theorem If C = {C1 U C2,…, Cx} is a pattern where x is a positive integer and Cj,
j 2 ½1; x is a sub-pattern of C, then
80 Y. Joshi et al.
P
supðCÞ i minfsupi ðC1 Þ; supi ðC2 Þ; . . .; supi ðCx Þg;
where i 2 ½1; n and n = number of sites.
Transmission of the LDM containing Composite Itemsets reduces the network
traffic (in terms of number of itemsets), as the composite itemset contains multiple
itemsets hidden within it. Finally, it computes Global Disjoint Matrix (GDM) by
merging all the LDMs received from FP-tree_MobileAgents. This GDM that con-
tains composite itemsets can be further used for obtaining frequent itemsets.
Further, here in our approach, as we avoid computation of CGFI, we also avoid
transfer of CGFI back to distributed sites. Instead, we sent mobile agents to dis-
tributed sites to get back LSC of the items. Here again, much of the network traffic
is reduced, as support count of the items only is transferred instead of CGFI which
is a huge set of itemsets. Algorithm given below explains the sequence of opera-
tions performed by the Parent.
The FP-tree_MobileAgent is a mobile agent that migrates to the specified remote
distributed site carrying with it the information required and the code to be executed
at the remote location. After migrating, FP-tree_MobileAgent connects to the
remote database, accesses the transactions and constructs disjoint matrix repre-
sentation of FP-tree of the remote database.
FP-tree_MobileAgent may store its LDM into a file and further transfer it to the
parent by the way of FTP, if LDM is too large. Algorithm given below explains the
sequence of operations performed by an FP-tree_MobileAgent.
4 Performance Analysis
The proposed MADFPM algorithm mines distributed database by performing

independent and parallel mining of the partitions DB1, DB2,…, DBn at distributed
sites S1, S2,…, Sn, respectively, and thus adopts divide-and-conquer technique. It
also avoids computation and hence transfer of CGFI, thereby reducing the cost of
computation and communication significantly. Tests have been carried out by
running the algorithms on the IBM Almaden Quest research group’s dataset.
4.1 Comparison with Client–Server Approach
Client–server approach for distributed frequent pattern mining involves transfer of

all the distributed database partitions DB1, DB2,.., DBn located at sites S1, S2,…, Sn
to the central site and then building FP-tree/disjoint matrix at the central site for
obtaining GFI. Therefore, time, TCS, required to construct FP-tree/disjoint matrix in
client–server approach is
X
n
TCS ¼ Time to transfer each database DBi þ TB-GDM
i¼1
where TB-GDM—Time to build global disjoint matrix.

All transactions take approximately the same time to transfer:
n X
X p
TCS ¼ Time to transfer each transaction þ TB-GDM
i¼1 j¼1
Let, TT-Tranx be the time to transfer each transaction. Then,
TCS ¼ n p TT-Tranx þ TB-GDM ð1Þ
(Note: n p = D, Total distributed database)

Let TMADFPM be the time required to construct disjoint matrix for the given
distributed database DB, using MADFPM approach:
) TMADFPM ¼ n TS-MA-LSC þ TC-LSC þ n TR-LSC þ n TS-MA-LDM

ð2Þ
þ TC-LDM þ n TR-LDM þ TM-GDM ;
where TS-MA-LSC—Time to send mobile agent to a distributed site to compute LSC,

TC-LSC—Time to compute LSC for all items at a distributed site, TR-LSC—Time to
receive LSC from a distributed site, by the central site, TS-MA-LDM—Time to send
mobile agent to a distributed site to compute LDM, TC-LDM—Time to compute
LDM at a distributed site, TR-LDM—Time to receive LDM from a distributed site, by
the central site, and TM-GDM—Time to merge n LDMs at central site.
Computations of LSC and LDM are done carried out simultaneously at all sites.
Rewriting the equation by moving the related terms in Eq. (2) toward one end, we
get
TMADFPM ¼ n TS-MA-LSC þ n TR-LSC þ n TS-MA-LDM þ n TR-LDM

þ TC-LSC þ TC-LDM þ TM-GDM
But TR-LSC = m Time to receive support count of one item.

Assuming that it takes 1 unit of time to receive support count of one item, then,
TR-LSC = m.
But TC-LSC + TC-LDM + TM-GDM = Time to build GDM = TB-GDM
) TMADFPM ¼ n þ ðn mÞ þ n þ ðn TR-LDM Þ þ TB-GDM

ð3Þ
i:e:; TMADFPM ¼ 2n þ ðn mÞ þ n TR-LDM þ TB-GDM
82 Y. Joshi et al.
But
TR-LDM ¼ q Time to transfer a Composite Itemset

¼ q TT-CI;
where q is the number of composite itemsets.

As Composite Itemsets are formed by merging transactions, we may suitably
assume TT-CI ≅ TT-Tranx:
) TR-LDM ¼ q TT-Tranx
Substituting this in Eq. (2), we get
TMADFPM ¼ ð2 nÞ þ ðn mÞ þ ðn q TT-Tranx Þ þ TB-GDM ð4Þ
where m and n are the constants for a given distributed database system.
) TMADFPM ffi n q TT-Tranx þ TB-GDM ð5Þ
Comparing Eqs. (1) and (5), it is evident that the run times of client–server and
distributed systems for frequent patterns mining are governed by the parameters
p and q, respectively. p is the number of transactions (i.e., database size) and q is the
number of Composite Itemsets. It is to be noted that q is always less than p. Hence,
as the new transactions are added, most of the time only frequencies of existing
itemsets increase and not the number of itemsets. Figure 1 below represents a graph
of number of transactions of databases versus number of itemsets of the FP-tree of
transaction databases. It may be observed from the graph that, as the database size
increases, the number of itemsets also increases. But increase in number of itemsets,
which is little less compared to the increase in database size initially, drops further
as the database size increases which eventually reduces to half of the size of the
database. Hence, as Composite Itemset contains multiple itemsets hidden within it,
as the number of transactions increases, number of Composite Itemsets also.
Therefore, MADFPM always takes much less time than its client–server version.
Fig. 1 Number of itemsets No of Transactions

No of Itemsets / Transactions
versus number of transactions 100000 No of Itemsets

80000
60000
40000
20000
0
10 20 30 40 60 70 80 90 100
Dataset Size(in million transactions)
4.2 Comparison with PMFI-A Approach
As discussed in Sect. 2, PMFI-A, while distributed sites send their LFIs to the
neighbors, they also receive LFIs from the neighbors at the same time, that is,
distributed sites perform overlapped sending and receiving operations. Further, each
distributed site starts computation of PGFI only after it completes sending and
receiving of LFIs, whereas the central site only receives LFIs (and does not send).
Hence, probability of central site completing its computations of GFI and CGFI
before distributed sites complete their computation of PGFI is very high. This fact
is ignored while measuring performance. Therefore, sending of PGFI is sequential
operation done after the distributed sites complete computation of their PGFI and
not a parallel operation mkl;’;l.,;’. Hence, we have considered the time required to
receive PGFIs by the central site (from all distributed sites) for calculating overall
time required to compute GFI. Let TPMFI-A be the time required to construct GFI for
the given distributed database DB, using PMFI-A approach.
)TPMFI-A ¼ n TS-MA-LFI þ TC-LFI þ n TR-LFI þ TC-C=GFI þ n TR-PGFI þ

ð6Þ
n TPGFI- [ GFI
where TS-MA-LFI—Time to send mobile agent to a distributed site to compute LFI,

TC-LFI—Time to compute LFI at a distributed site, TR-LFI—Time to receive LFI
from a distributed site, by the central site, TC-C/GFI—Time to compute GFI and
CGFI at central site, TR-PGFI—Time to receive PGFI from a distributed site, by the
central site, and TPGFI->GFI—Time to compute final GFI using PGFI count.
Comparing Eqs. (6) and (2), we have
TS-MA-LFI ffi TS-MA-LSC ; TC-LFI ffi TC-LSC þ TC-LDM ;

TR-LFI ffi TR-LDM ; TPGFI- [ GFI ffi TM-GDM :
Furthermore, TC-C/GFI + n TR-PGFI n TR-LSC + n TS-MA-LDM, because

LSC is an array of items’ support count that depends on the number of items “m”,
which is very small compared to size of PGFI. PGFI is a set of itemsets where each
itemset contains many items. Hence, transfer of PGFI (i.e., TR-PGFI) takes much more
time compared to that of LSC. It needs no explanation to say that the computation of
CGFI and GFI, using all of the LFIs, received from distributed sites, at the central site
taking more time compared to the time required to dispatch “n” mobile agents [13].
Further, though we have assumed that TR-LFI ≅ TR-LDM, considering the fact that the
Composite Itemset contains multiple itemsets hidden within it, size of the LDM is
much less compared to that of LFI.
84 Y. Joshi et al.
) TMADFPM
TPMFI-A :
In PMFI-A, every distributed site transfers set of itemsets twice to the central site
in the course of execution, namely LFI first time and then PGFI later. This transfer
of itemsets twice contributes to network traffic apart from contributing to the
computational time. Whereas, in case of MADFPM, each distributed site transfers
LDM, which is a set of Composite Itemsets, only once in the course of execution.
Hence, the performance of MADFPM is superior to PMFI-A.
5 Conclusion
Our proposed algorithm MADFPM for frequent pattern mining of distributed

databases uses mobile agents and found to be better in terms of performance.
Performance of MADFPM is better than conventional client–server approach as
preprocessed compact data in the form of disjoint matrix is transferred to the central
site rather than moving the entire data to the central site and then processing it there.
Performance of MADFPM is also better compared to the other mobile agent-based
approaches, PMFI and PMFI-A, as it reduces the computational cost and network
traffic. Computational cost is reduced because the distributed sites compute local
itemsets in the form of LDM only once, unlike PMFI-A which computes local
itemsets (at the distributed sites) twice, in the form of LFI and PGFI. Network
traffic is reduced because only the LDMs are transferred from distributed sites to the
central site once. Moreover, in case of PMFI-A, each distributed site transfers LFI
to all neighbors along with the central site contributing to the network traffic further.
References
1. Agrawal R., Imielinski, T., and Swami, A. (1993), “Mining association rules between sets of
items in large databases”. In Proc. of ACM-SIGMOD, (SIGMOD’93), pp. 207–216.
2. Paul S. Bradley, J. E. Gehrke, Raghu Ramakrishnan and Ramakrishnan Srikant (2002),
‘Philosophies and Advances in Scaling Mining Algorithms to Large Databases”.
Communications of the ACM.
3. Chattratichat, J., Darlington, J, et al. (1999), “An Architecture for Distributed Enterprise Data
Mining”, 7th Intl. Conf. on High Performance Computing and Networking.
4. You-Lin Ruan, Gan Liu, Quin-Hua Li (2005), “Parallel Algorithm for Mining Frequent
Items”, Proceedings of the Fourth International Conference on Machine Learning and
Cybernetics, Guangzhou, pp-18–21.
5. U.P. Kulkarni, P.D. Desai, Tanveer Ahmed, J.V. Vadavi, A.R. Yardi (2007), “Mobile Agent
Based Distributed Data Mining”. International Conference on Computational Intelligence and
Multimedia Applications, pp. 18–24.
6. Saleem Raja, George Dharma Prakash Raj, (2013), “Mobile Agent based Distributed
Association Rule Mining”, International Conference on Computer Communication and
Informatics (ICCCI), 2013.
7. LIU Xiang (2008), “An Agent-based Architecture for Supply Chain Finance Cooperative
Context-aware Distributed Data Mining Systems”. 3rd International Conference on Internet
and Web Applications and Services.
8. Ogunda A.O., Folorunso O., Ogunleye G.O., (2011), “Improved cost models for agent-based
association rule mining in distributed databases, Anale. SeriaInformatică. Vol. IX fasc. 1 –
2011.
9. J. Han, J. Pei, Y. Yin and R. Mao (2004), “Mining Frequent Patterns without Candidate
Generation: A Frequent-Pattern Tree Approach”. Data Mining and Knowledge Discovery, 8
(1), pp. 53–87.
10. Keshavamurthy B.N., Mitesh Sharma and DurgaToshniwal (2010), “Efficient Support
Coupled Frequent Pattern Mining Over Progressive Databases”, International Journal of
Database Systems, Vol.-2, No-2, pp-73–82.
11. Mengling Feng, Jinyan Li, Guozhu Dong, Limsoon Wong (2009), “Maintenance of Frequent
Patterns: A Survey”, published in IGI Global, XIV Chapter, pp-275–295.
12. Syed K. Tanbeer, C. F. Ahmed, B-S Jeong (2009), “Parallel and Distributed Algorithms for
Frequent PatternMining in Large Databases”. IETE Technical Review, Vol. 26, Issue 1,
pp-55–66.
13. Raquel Trillo, Sergio Ilarri, Eduardo Mena (2007), “Comparison and Performance Evaluation
of Mobile Agent Platforms”, Third International Conference on Autonomic and Autonomous
Systems (ICAS’07), pp. 41.
A Hybrid Approach for Preprocessing
of Imbalanced Data in Credit Scoring
Systems
Uma R. Salunkhe and Suresh N. Mali
Abstract During the last few years, classification task in machine learning is
commonly used by various real-life applications. One of the common applications
is credit scoring systems where the ability to accurately predict creditworthy or
non-creditworthy applicants is critically important because incorrect predictions can
cause major financial loss. In this paper, we aim to focus on skewed data distri-
bution issue faced by credit scoring system. To reduce the imbalance between the
classes, we apply preprocessing on the dataset which makes combined use of
random re-sampling and dimensionality reduction. Experimental results on
Australian and German credit datasets with the presented preprocessing technique
has shown significant performance improvement in terms of AUC and F-measure.

Keywords Credit scoring Imbalanced data Preprocessing Classifier ensemble
1 Introduction
Classification techniques in data mining are applied to diverse data in order to

predict the category to which the given instance belongs. With the rapid growth of
data, usage of classification techniques by various applications has also increased.
One of the most common applications of classification is credit scoring systems
where applicants are classified as creditworthy or non-creditworthy based on the
analysis of their characteristics. Credit risk management is the major concern for the
financial organizations so that the probable risk associated with repayments can be
decreased. Credit scoring systems are designed to handle this issue by categorizing
U. R. Salunkhe (&)
Department of Computer Engineering, Smt. Kashibai Navale College
of Engineering, Pune, Maharashtra, India
e-mail: umasalunkhe@yahoo.com
S. N. Mali
Sinhgad Institute of Technology and Science, Pune, Maharashtra, India
e-mail: snmali@rediffmail.com

https://doi.org/10.1007/978-981-10-7245-1_10
88 U. R. Salunkhe and S. N. Mali
the applicants into good and bad applicants [1]. Traditionally, human experts used
to do this categorization by applying their experience and following the guidelines
given by the organization. Basel Committee on Banking Supervision has also
specified some guidelines for this purpose [2]. Five characteristics namely char-
acter, capacity, capital, collateral, and conditions were considered for making the
decision of whether applicant is good or bad [3]. But this could lead to incorrect
decisions and decision was expert dependent. Hence, those traditional models are
replaced by credit scoring models that have the following advantages:
1. Credit scoring models are cheaper than traditional methods.
2. The decision-making is faster.
3. Those models are consistent in decision-making.
4. Those models are adaptive to policy changes.
2 Related Work
Marques et al. [1] presented two-level classifier ensembles for credit scoring sys-
tems in order to improve performance in terms of AUC compared to single-level
ensemble. The proposed method introduces diversity using different training
datasets as well as different feature sets. Authors conclude that two-level ensemble
that combines Bagging and Rotation Forest gives the best performance. Wu et al.
[4] proposed a prediction model based on hybrid re-sampling approach that com-
bines undersampling with SMOTE and AdaBoost. Presented method shows per-
formance improvement in terms of evaluation measures like AUC, accuracy, and
G_mean.
Xiao et al. [5] proposed a classifier ensemble technique to improve the perfor-
mance of credit scoring system by diversifying the base classifiers. K-means
clustering divides instances of each class into different clusters. Clusters from two
classes are combined to form various diverse training subsets which are then used to
train base classifiers. Abellan et al. [6] presented an experimental study wherein
different classifier ensembles are tested with different base classifiers in order to
select the best base classifier. Authors suggest that credal decision tree
(CDT) which is very unstable classifier is a good choice for classifier ensembles.
Pozzolo et al. [7] made an experimental comparison of many existing classifi-
cation techniques on fraud detection datasets. Their focus is on three important
issues: unbalancedness, nonstationary, and assessment. Study recommends average
precision, AUC, and precision rank as preferable performance measures. Oreski
et al. [8] introduced hybrid genetic algorithm with neural network that proves to
increase the accuracy and scalability of credit scoring system. Presented method
first removes redundant attributes by dimensionality reduction which is then further
analyzed for refinement and applied with incremental learning.
Han et al. [9] addressed the issue of curse of dimension faced by logistic
regression and support vector machine. The orthogonal transform process is
A Hybrid Approach for Preprocessing … 89
introduced for the feature extraction. Grid search method is used and then
dimensionality reduction is done. Experimental results show better results in terms
of sensitivity, specificity, and prediction accuracy. Kim et al. [10] resolved multi-co
linearity problem using genetic algorithm-based technique.
Xiao et al. [11] proposed a dynamic classifier ensemble method for imbalanced
data (DCEID) that makes combined use of classifier ensemble and cost-sensitive
learning approaches. Dynamic classifier ensembles use either dynamic classifier
selection (DCS) or dynamic ensemble selection (DES). The DCEID adaptively
chooses the best of these two for each test customer. Also, a new evaluation
criterion called cost-sensitive symmetric regularity criterion (CS-SRC) is
constructed.
3 Proposed Model
Credit scoring systems suffer from the issue of skewed distribution of dataset, i.e.,
the available training dataset has imbalanced distribution of data. This is because
customers that are likely to be defaulter for making repayments are very less
compared to the customers that are creditworthy. The classification model that is
trained with such imbalanced data is likely to make biased decision toward the
creditworthy class. This is due to unavailability of sufficient data that belongs to
“bad” (non-creditworthy) applicant’s class. Thus, the applicant who is likely to be
defaulter in repayments is classified as “good” (creditworthy) and will be granted
the credit. This can lead to major financial loss to the organization.
In order to handle above issue, classification model should be designed such that
it tries to focus on the accuracy of the class with less number of samples, i.e.,
minority class. Traditional classifiers are not effective when applied to imbalanced
dataset. Recent studies present different ways to handle imbalanced dataset which
can be categorized into classes, namely data-level approaches, algorithm-level
approaches, cost-sensitive algorithms, and classifier ensemble approaches.
In this paper, we present an approach that can handle the imbalance between two
classes of credit scoring system by making use of data-level approach. To achieve
this, initially noisy instances and instances with missing values are identified. Then,
re-sampling of training data is done and imbalance between them is reduced to
some extent. The term “Imbalance Ratio” is used to indicate the imbalance between
classes and is defined as ratio of number of samples of majority class to that of
minority class.
Figure 1 represents the experimental setup in which initially original credit
scoring dataset is provided as input to the preprocessing stage. It involves handling
instances with missing values and noisy instances. The instances which are having
more than half nearest neighbors of different classes are known as noisy instances
and are removed from the dataset. Then, the dimension reduction is done in order to
remove the redundant attributes. The aim is to select the appropriate subset of
Fig. 1 Proposed method
attributes that is sufficient to train the model [8]. This is done to overcome the issues
raised by the existence of redundant attributes as listed below [9].
1. If model is trained with redundant attributes, it may predict incorrectly as pre-
dictions are skewed by those attributes.
2. Redundant attributes may lead to over-fitting issue.
3. Removal of such attributes reduces the size of the training dataset which in turn
reduce the time required to train model.
The output of previous step is applied with random sampling with replacement.
This generates the re-sampled credit scoring dataset which is then provided as input
to train the classifier model.
4 Experimental Design
This section discusses the experimental design of the work that is presented in this
paper.
The experiments were carried out using Weka environment with its default
parameters. In this work, we implemented a preprocessing technique that selects
suitable subset of attributes of the training dataset and then applies re-sampling to it.
Preprocessed data is provided as input to the classification algorithm. Different
individual classifiers such as J48, Naïve Bayes, Multilayer Perceptron (MLP), and
SVM are used for the experimentation. Also, ensemble techniques namely
AdaBoostM1, Bagging, random subspace, and rotation forest are tested. All
experiments were carried out using tenfold cross-validation. Results of the different
classifiers when trained with preprocessed data are compared with the results when
classifier is trained on original imbalanced training dataset.
4.2 Datasets
For experimentation, we have chosen two publicly available credit scoring datasets,
namely Australian and German datasets from UCI repository. Details of those
datasets are given in Table 1.
4.3 Evaluation Parameters
Classification algorithms can be evaluated by using different evaluation parameters

such as accuracy, G-mean, F-measure, AUC, Type-I error, etc. [12]. To derive these
parameters, we need to use confusion matrix for binary classification problem
which is shown in Table 2.
Previous studies [13] suggest that AUC is more appropriate measure than
accuracy for the datasets that have imbalanced distribution of data. In this paper, we
have used AUC and F-measure as our evaluation measures.
AUC (Area under ROC curve)
AUC can be defined as arithmetic average of the mean predictions for each class
[1]. AUC is represented in Eq. (1):
sensitivity þ specificity
AUC ¼ ; ð1Þ
2
where sensitivity and specificity are computed using Eqs. (2) and (3),
TP
Sensitivity ¼ ð2Þ
TP þ FN
TN
Specificity ¼ ð3Þ
FP þ TN
Table 1 Characteristics of Dataset # Examples # Minority # Majority # Attr.

the datasets used in the
experiments Credit-a 690 307 383 14
Credit-g 1000 300 700 21
Table 2 Confusion matrix Predicted as Predicted as

for a two-class problem positive negative
Positive class True positive (TP) False negative (FN)
Negative False positive (FP) True negative (TN)
class
F-measure F-measure represents accuracy by using combination of two measures,

precision and recall. It is computed by using [14] Eq. (4),
2 precision recall
F - measure ¼ ; ð4Þ
precision þ recall
where precision and recall are computed using Eqs. (5) and (6),
TP
Precision ¼ ð5Þ
TP þ FP
TP
Recall ¼ ð6Þ
TP þ FN
The experimental results in this paper are presented in two subsections. The first
subsection describes the performance in terms of AUC with different classifier
models. Results of the classifiers are compared between the one that is trained with
re-sampled credit scoring dataset and another trained with dataset without
re-sampling. The next section shows performance in terms of F-measure. Table 3
summarizes the AUC performance of different classification models on two credit
scoring datasets.
Figure 2 plots the performance of the different classifier models in terms of AUC
when they are trained with original credit scoring dataset and then with prepro-
cessed dataset.
Table 4 summarizes the performance of different classification models using
F-measure as evaluation measure on two credit scoring datasets.
Figure 3 plots the performance of the different classifier models in terms of
F-measure when they are trained with original imbalanced credit scoring dataset
and then with preprocessed dataset.
Analysis of the graphs shown in Figs. 2 and 3 clearly shows that preprocessing
has enhanced the performance of most of the tested classifiers. Although the
Table 3 Performance evaluation using AUC on the credit scoring datasets

Classification AUC
algorithm Credit-a Credit-g
Original Re-sampled Original Re-sampled
dataset dataset dataset dataset
J48 0.867 0.879 0.639 0.716
Naïve Bayes 0.892 0.904 0.787 0.737
Multilayer 0.89 0.936 0.733 0.741
perceptron
Logistic 0.906 0.935 0.715 0.743
SVM 0.853 0.853 0.671 0.681
Rotation forest 0.906 0.954 0.777 0.786
Bagging 0.919 0.936 0.752 0.761
AdaBoostM1 0.902 0.963 0.724 0.742
Random subspace 0.912 0.952 0.754 0.744
Fig. 2 Performance evaluation using AUC on the credit scoring datasets
improvement in AUC or F-measure values seems smaller, it is beneficial for credit

scoring systems where correct identification of the customer who is likely to be
defaulter is critically important and may save huge financial loss. While experi-
menting, we also recorded the training time for those classifiers in different sce-
narios. It helped us to conclude that classifier ensemble techniques require
significantly higher training time compared to individual classifiers. When we
applied proposed preprocessing to the dataset, we could achieve considerable
performance improvements in individual classifiers. Hence, it could be beneficial to
save the time and achieve good performance with the help of combined use of the
proposed preprocessing and individual classifier method. Multilayer preceptron
classifier model is exception to this as it requires very high time for training than
that required for the classifier ensemble technique.
Table 4 Performance evaluation using F-measure on the credit scoring datasets

Classification F-measure
algorithm Credit-a Credit-g
Original Re-sampled Original Re-sampled
dataset dataset dataset dataset
J48 0.861 0.883 0.692 0.73
Naïve Bayes 0.767 0.77 0.746 0.753
Multilayer 0.831 0.874 0.715 0.75
perceptron
Logistic 0.854 0.858 0.744 0.751
SVM 0.847 0.847 0.741 0.738
Rotation forest 0.862 0.917 0.737 0.746
Bagging 0.864 0.89 0.731 0.741
AdaBoostM1 0.843 0.91 0.694 0.748
Random subspace 0.863 0.901 0.696 0.673
Fig. 3 Performance evaluation using F-measure on the credit scoring datasets
6 Conclusion
We have presented an approach to enhance performance of credit scoring systems

which face challenge of skewed dataset. In order to resolve this issue, we initially
applied preprocessing to the imbalanced credit scoring dataset and tried to convert it
to balanced form to some extent. While doing this, redundant attributes were
removed from the feature space which helped to improve performance. Then, it was
given as input to the classification algorithm and has shown significant improve-
ments in classifying the customers correctly. Another concern of this work was to
test this model on different classifiers and compare their behavior. Comparison of
performance improved and amount of time to train the model indicates that it is
better to achieve considerable improvement with less training time for individual
classifiers.
References
1. Marqués, A. I., Vicente García, and Javier Salvador Sánchez. “Two-level classifier ensembles
for credit risk assessment.” Expert Systems with Applications 39.12 (2012): 10916–10922.
2. BIS. Basel III: a global regulatory framework for more resilient banks and banking systems.
(2011). Basel Committee on Banking Supervision, Bank for International Settlements, Basel.
ISBN print: 92-9131-859-0. <http://www.bis.org/publ/bcbs189.pdf>.
3. Marqués, A. I., Vicente García, and Javier Salvador Sánchez. “Exploring the behaviour of
base classifiers in credit scoring ensembles.” Expert Systems with Applications 39.11 (2012):
10244–10250.
4. Wu, Xiaojun, and SufangMeng. “E-commerce customer churn prediction based on improved
SMOTE and AdaBoost.” Service Systems and Service Management (ICSSSM), 2016 13th
International Conference on. IEEE, 2016.
5. Xiao, Hongshan, Zhi Xiao, and Yu Wang. “Ensemble classification based on supervised
clustering for credit scoring.” Applied Soft Computing 43 (2016): 73–86.
6. Abellán, Joaquín, and Javier G. Castellano. “A comparative study on base classifiers in
ensemble methods for credit scoring.” Expert Systems with Applications 73 (2017): 1–10.
7. Dal Pozzolo, Andrea, et al. “Learned lessons in credit card fraud detection from a practitioner
perspective.” Expert systems with applications 41.10 (2014): 4915–4928.
8. Oreski, Stjepan, and Goran Oreski. “Genetic algorithm-based heuristic for feature selection in
credit risk assessment.” Expert systems with applications 41.4 (2014): 2052–2064.
9. Han, Lu, Liyan Han, and Hongwei Zhao. “Orthogonal support vector machine for credit
scoring.” Engineering Applications of Artificial Intelligence 26.2 (2013): 848–862.
10. Kim, Myoung-Jong, and Dae-Ki Kang. “Classifiers selection in ensembles using genetic
algorithms for bankruptcy prediction.” Expert Systems with applications 39.10 (2012): 9308–
9314.
11. Xiao, Jin, et al. “Dynamic classifier ensemble model for customer classification with
imbalanced class distribution.” Expert Systems with Applications 39.3 (2012): 3668–3675.
12. Salunkhe, Uma R., and Suresh N. Mali. “Classifier Ensemble Design for Imbalanced Data
Classification: A Hybrid Approach.” Procedia Computer Science 85 (2016): 725–732.
13. Liu, Xu-Ying, Jianxin Wu, and Zhi-Hua Zhou. “Exploratory under-sampling for
class-imbalance learning.” Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE
Transactions on 39.2 (2009): 539–550.
14. Kamalloo, Ehsan, and Mohammad SanieeAbadeh. “An artificial immune system for
extracting fuzzy rules in credit scoring.” Evolutionary Computation (CEC), 2010 IEEE
Congress on. IEEE, 2010.
Vision-Based Target Tracking Intelligent
Robot Using NI myRIO with LabVIEW
Anita Gade and Yogesh Angal
Abstract Robots are worked to do tasks that are risky to people, for example,
defusing bombs and discovering survivors in unsteady environments and investi-
gation. The rising exploration field on scaled-down programmed target following
robot is given significance in hazardous, military and industrial areas for navigation,
observation, safety, and goal acknowledgment by image processing. The real-time
vision-based technique is used for target tracking. The objective of this system is to
develop the proto robot capable of following a target using image processing.
Normally, the way of a mobile robot is controlled by a pre-identified data about
objects in nature. In this developed system, the robot tracks the desired destination
goal in a systemized way.
Keywords Camera LabVIEW myRIO Target tracking Vision
1 Introduction
The world we live in is a tridimensional one. That is the reason it is critical for
artificial vision framework to see the world in three dimensions. Ordinary sensors,
utilized as a part of video cams, are just bidimensional. In spite of the fact that, by
consolidating at least two cameras in a stereoscopic design, the tridimensional parts
of the world might be evaluated, the development speaks to on its hand a critical
wellspring of data over a succession of images. Consolidating the information
separated from tridimensional space utilizing the stereovision with those in regards
to development in bidimensional arrangement permits acquiring some performing
arrangement of tracking objects [1]. The video stream is handled by
A. Gade (&) Y. Angal

Department of E&TC Engineering, JSPM’s Bhivarabai Sawant Institute of Technology &
Research, SPPU, Pune, India
e-mail: anita.gade@yahoo.co.in
Y. Angal
e-mail: yogeshangal@yahoo.co.in

https://doi.org/10.1007/978-981-10-7245-1_11
98 A. Gade and Y. Angal
LabVIEW-VDM to decide the way that the robot must take after to achieve the
objective. Depending upon the adjustment in position of target, LabVIEW recal-
culates the way and commands are sent to the robot to move toward the new area of
target. The elements of the moving target like shape, size, and color can be
observed for tracking it. In this venture, color and shape of the target were utilized
to track the moving target. In this paper, we will present an incorporated
vision-based strategy for adaptable robots that are able to track the target using
monocular camera images.
2 Problem Definition
In a substitute method, a robot is furnished with sequence of images within space.

By differentiating these pre-recorded pictures and the camera pictures, the robot is
able to decide its area [2]. Previous methodologies have identified ultrasonic sen-
sors or IR sensors for target tracking but by using these sensors, we cannot estimate
color, size, and shape of object. Considering this problem, we have implemented
this system using stereovision camera for target detection and tracking.
3 System Overview
Target recognition and tracking are carried out in LabVIEW using vision assistant
module. Two cameras are used to acquire real target image. NI myRIO device is
control and processing element of this system [3]. Four-wheel drive robot is used to
track the object. Opt-interrupter encoder is used to count the pulses. Camera is
calibrated to acquire image of the entire arena (Fig. 1).
Robot with
Cameras wheel drive
NI myRIO
PC with LabVIEW Pulse counting
circuit
Fig. 1 System overview

Vision-Based Target Tracking Intelligent Robot … 99
4 Algorithm
Two cameras (Left and Right) have been configured for continuous acquisition of
image. The hypothesis made for ideal stereovision may not be done for the
application of stereovision in the real world. Even the best cameras and lenses will
present certain distortion level of acquired image, so in order to compensate this
distortion stereovision system needs to be calibrated [4]. Calibration processes the
use of matrix of alignment with the assistance of sets of images which are acquired
from various angles to figure the distortion of the image. The calibration grill
contains in vision development module. Disparity information provides relative
depth information. The disparity information is made by using semi-block matching
algorithm from NI vision because it provides more details and works in lesser
region in case of proposed system. We have used two servomotors for pan/tilt
camera support: The development of the cameras in horizontal arrangement and the
development of the cameras in vertical arrangement.
• System Execution Flow:
1. Adjustment of camera is done in camera calibration setup, i.e., Focus.
2. Image is acquired using left and right cameras. Two images of the same
scene are captured at the same time by two different cameras.
3. Quality of the image is improved using image preprocessing.
Image acquisition using left Image acquisition using

camera right camera
Camera calibration Object detection

Image preprocessing
Pan and tilt system

control
NI myRIO
Disparity path Stereo vision Object tracking
Fig. 2 Algorithm
4. Particle analysis of image is done.

5. Analyzed output of image is given to myRIO controller.
6. Control signals are generated for hardware to track the object (Fig. 2).
• Binocular Stereovision Principle
We have used binocular stereovision algorithm to calculate the position of
objects within the frame of the scene and in particular, the depth of the elements
contained in the two images [5]. The principle of binocular stereo vision is illus-
trated in Fig. 3, where f is the central length of left and right cameras, b is the
baseline, P (x, y, z) is a real-world point, uQ is the projection of this present reality
point P in an image got by the left camera, and uR is the projection of this present
reality point P in an image got by the right camera. Since the two cameras are
separated by “d”, both cameras see a same real-world point P. The x-directions of
point’s uQ and uR are given by relations (1) and (2):
x
uQ ¼ f ð1Þ
z
xd
uR ¼ f ð2Þ
z
Distance between the two projected points is known as “disparity” and it is

computed using relation (3):
d
Disparity ¼ uQ uR ¼ f ð3Þ
z
The depth information and the distance between real-world point “P” and the
stereovision system, is given by the relation (4):
d
Depth ¼ f ð4Þ
disparity
Fig. 3 Stereovision system Left Camera Right Camera

d
principle
f x y
y x
z f
z
Image uQ uR Image
P (x, y, z)
Fig. 4 Stereovision VI
The coordinates of focus of the object (x1, y1) and (x2, y2) obtained from left and
right cameras, respectively, are changed relatively to the center of the images of
sizes 640 480.
5 Experimental Results
1. Calibrate stereovision system (Fig. 4)

Disparity image indicates the disparity between left and right rectified images.
Depth image provides comprehensive real-world depth information about the scene.
Error map gives the inherent error in measurement for each pixel in the depth
image.
2. Stereovision front panel (Fig. 5)
A calibration grid of dots was introduced to both cameras and images are
acquired simultaneously. Brighter pixels indicate points that are closer to the
camera and darker pixels indicate points that are farther away from the camera.
3. Target tracking VI
4. Robot and target tracking by image processing
Figure 6 shows the block diagram VI in LabVIEW for stereovision. Figure 7
shows images captured from left and right camera which are taken precisely at same
time for stereovision. At the point when target is detected by image processing
Fig. 5 Stereovision front panel
Fig. 6 Target tracking VI
techniques, we have determined coordinates of mass focal point of it in the two

images [6].
6 Conclusion
We have implemented target tracking robot using image processing for developed
proto system which detects the target by vision-based method. This system is
implemented with wireless communication using LabVIEW-VDM image pro-
cessing techniques to reach target destination. We have developed the application
with remote computer in which the user gives a command to capture an image,
process, and control remotely.
Left Camera Images Right Camera Images
Fig. 7 Target tracking robot
Acknowledgements Authors put on record and warmly recognize to Savitribai Phule Pune
University for partially funded by BCUD Research Grant vide, BCUD Sanction Letter, OSD/
BCUD/113/48. I take it a deemed privilege to express thanks to JSPM’S Bhivarabai Sawant
Institute of Technology and Research, Wagholi, to complete this research work.
References
1. Francois Guerin, Simon G. Fabri, Marvin K. Bugeja “Double Exponential Smoothing for
Predictive Vision Based Target Tracking of a Wheeled Mobile Robot”, IEEE 2013.
2. Chung-Hao Chen, David Page, Chang Cheng, Andreas Koschan, and Mongi Abidi “A Moving
Target Tracked by A Mobile Robot with Real-Time Obstacles Avoidance Capacity” IEEE,
0-7695-2521-0/06, 2006.
3. You-Wei Lin and Rong-Jong Wai, “Design of Adaptive Moving-Target Tracking Control for
Vision-Based Mobile Robot, IEEE 2013.
4. Anita Gade, Yogesh Angal, “Automation in Library Management Using LabVIEW”,
International Conference on Computing Communication Control and automation, IEEE
2016. https://doi.org/10.1109/ICCUBEA.2016.7860133.
5. Robert Bichsel and Paulo V K Borges, “low–Obstacle Detection Using Stereo Vision”,
International Conference on Intelligent Robots and Systems, IEEE 2016.
6. Prajin Palungsuntikul, Wichian Premchaiswadi, “Object Detection and Keep on a Mobile
Robot by using a Low Cost Embedded Color Vision System”, IEEE 2010.
Detection of Misbehaviors Nodes
in Wireless Network with Help of Pool
Manager
Asha Chaudhary and Pournima More
Abstract As in recent time, wireless networks were developed in various extreme

ends but then also facing from the problem of malicious nodes in the network. In
this case, networks drop the packets at the receiver side and create various routing
attacks. By the help of Dijkstra’s algorithm, network can escape from the problem
of dropping packets. In that, nodes can calculate shortest path on the basis of Fuzzy
logic. Biggest downsides of existing propose are that it is unable to handle
occurring of blackhole and grayhole in the network and nodes can easily befool
with each other. To avoid occurring of attacks, network can use a scheme in which
it creates pool manager at the starting of communication by which load of the nodes
get reduces and routing attacks can easily detect. And at the time of further com-
munication, that node is banned and does not engage for delivery packets from
source to destination. By ejection of malicious nodes from the network with the
help of pool manager that can iterative more established data by comparing hash
function.
Keywords Blackhole Grayhole Pool manager MANETs Titlemapping

DOS
1 Introduction
Mobile adhoc networks (MANET) is a scenario in which a group of mobile nodes

comes together to give communication between them through wireless connection.
In this case, there is no need of any centralized administration; it can be easily
resemble in the situations where it is not possible to set up any infrastructure.
A. Chaudhary (&) P. More

Computer Network Department, G.H. Raisoni College of Engineering and Management,
Wagholi, Pune, India
e-mail: Asha92chaudhary@gmail.com
P. More
e-mail: Pournima.more1@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_12
106 A. Chaudhary and P. More
Advantages of MANET are dynamic in nature, by which different nodes of the

network can participate or exit at any time.
In MANETs, each mobile node operates not only as a host but also acts like
router which forward packets from one node to other mobile nodes that require
cooperation of nodes in the network. As described earlier, it is dynamic in nature
and there is a high risk of security. By this way, DSR (Dynamic Source Routing)
protocol is used instead of using routing table at very hop; its divides the operation
into route discovery and route maintenance. One of the biggest drawbacks faced by
the network is when nodes of the network start misbehaving to each other due to
that a concept of malicious nodes arises. By this effectiveness, approaches become
weak and nodes collude with each other that results in more devastating damages of
the network [1]. A weakness of infrastructure leads to highly susceptibility to
routing attacks (blackhole and grayhole) in the network. Blackhole attack is one
kind of Denial of Service (DOS) attacks. In this type of attacks, shortest path and
sequence number of the nodes, the message is intercepting to its destination, but
when the malicious (fake) node made a use of the vulnerabilities of the route
discovery packets of routing protocols at that time, there is a arise of blackhole
attack in the network. A node which acts as malicious node has a tendency to attract
packet through false route reply packet (RREP) to falsely claim fake shortest route
from source to destination; instead of forwarding to destination, it discards the
packets in between the routes [1] (Fig. 1).
But detection of grayhole attack was very difficult to identify because in this
there is a continuous dropping of the packet data. In this case, packet may drop data
for specific IP address; meanwhile, other nodes of the network receive correct
information. Due to selectively drops of the data it is difficult to identify. To avoid
this, nodes have to take more responsibility and detect grayhole attack by their own
due to that proper data transmission can be held. All this attack mainly affects the
routing speed and performance of the network itself (Fig. 2).
To avoid creation of malicious nodes in the network, concept of pool manager
was that condition network creates pools that consist of number of mobile nodes. In
this type of network, pool manager is all time in active state whenever transmission
Fig. 1 Blackhole attack

Detection of Misbehaviors Nodes in Wireless Network … 107
Fig. 2 Grayhole attack
process is going on first; all the information pass through pools of the network, due
to that network can easily identify all nodes that are misbehaving to each other due
to the creation of attacks. DTN (Delay Tolerant Network) also plays an important
role in the detection of blackhole and grayhole in the network. It provides a solution
of resistant network that faces the problem of intermittent connectivity, long delay,
or frequency disrupt. It also makes use of hop-to-hop routing and store-and-forward
model to overcome the problem of end-to-end paths [2]. Now, the question arises
how DTN helps in network to escape that nodes which cause attacks; in that case,
there is a concept of intermediate nodes. These nodes receive, store, and forward
packets to other destination which sometimes works as intermediate nodes. Suppose
intermediate nodes are not in the path, then those node store entire packets. By the
help of probabilistic behavior of nodes, it can predict and measure all information
which is stored by the node in contact of its potential forwarder in terms of deliver
message to the destination node. Most of the methodologies in wireless paradigm
are to identify the vulnerability issues like grayhole and blackholes. Most of those
are assigned this task to the routing nodes which can create the more havoc at the
node end to decrease the routing performance of the nodes. By keeping this issue as
of priority, proposed system assigns this task to a node’s head of the pool and called
it as “pool manager”.
This paper is segmented into various sections in which Sect. 2 tell about related
work of how another scheme helps to find malicious nodes. Section 3 details about
our proposed methodology of how efficiently network can remove all the malicious
nodes from the network by the creation of pool manager with help of title mapping.
The consequence of our methodology is performed in Sect. 4. Section 5 concludes
the paper with some scope of future.
2 Related Work
Various numbers of methods were discovered to detect black- and grayhole nodes
in the network, in which some of them work during route discovery process while
others at the time of data transmission. Those who work during route discovery
have more advantage, because in that case they do not lose the data.
2.1 Detection of Blackhole Attacks
In this, author proposed a method in which it can design a network which can use
destination unique ID to detect malicious node of blackhole attack and then it keeps
apart the node between source and destination [3]. In this route, node has to check
and analyze all neighbors node by comparing response time of sender by the help of
AODV route request message. Author describes a detection method that is known
as cooperative bait detection Scheme (CBDS). In this scheme, source node has the
tendency to select an adjacent node to whom it has to cooperate, and then after this,
address is used for bait destination address to bait malicious nodes to send a reply
RREP message. With the help of the reverse tracing technique, network can detect
all malicious nodes and then prevent them from participating in the routing oper-
ation of the network for the transmission of data packets [1]. The main advantage of
this scheme is that it can work for both proactive detection and reactive detection.
2.2 Detection of Grayhole Attacks
In this paper, author uses a concept of sequence number. For detection of grayhole
node, use sequence number for normal nodes by which it eliminates itself to enter in
the blacklist. Algorithm works to check the peak value of the sequence number
whether it is greater or less than reply packet [4]. Author describes an algorithm
named channel aware detection (CAD) which consists of two techniques: one is
called hop-by-hop observation and other is called traffic overhearing due to that
network can easily detect malicious nodes which belong to that particular path,
where transfer of packet is going on. By this, mesh node of the network can be
easily detected by appropriate subject of attack [5]. Describe a scheme which is
called as course-based detection method. Whatever nodes involved in the routing
route keep total focus on the neighboring node. In this, all neighbors do not include
monitoring [6]. Here, only neighboring node of the route gets the packet which was
sent by source node, which consists of itself as a copy of packets in the form of
buffer named Fwd packet buffer. By this source, node can analyze overhear rate and
get percentage of data packets transferred from source and received by destination.
2.3 Use of DTN for Detection of Routing Attacks
Author ensures a mechanism for the contrast analysis of blackhole attack in DTN
that is known as Combined Faith Value (CFV) [7]. By checking the node history
separately, CFV value can be calculated for each node of the network. Node history
can be stored in the form of node buffers. By this history of nodes information gets
regarding to the data transmission, number of packets which was created, how
much packets received, dropped and sent numbers of packets. In this blackhole,
nodes have maximum packets to drop and deliver wrong message based on CFV
mechanism. Author proffers how DTN performs store–carry–forward method. DTN
only works when both sender and receiver nodes come in contact with each other.
Till they come in contact, nodes store the packet in buffer and then carries to next
node when it gets the range for transmission. Distributed scheme is introduced for
detection of packet dropping in that node select contains in the form of signed
contact record of its previous contact. At the time of collecting record of nodes,
contact can easily detect which node dropped the packet [8].
2.4 Detection of Blackhole and Grayhole Attacks on Same

Network
Author recommends a cryptographic system which consists of secure data transfer

mechanism that protects the data from encryption and hashing during the time of
attackers attacking the system. In this system, if the receiver is unable to decrypt the
cipher data correctly, then taint propagation detects it, and immediately sends to
sender that particular node start acts like malicious node [9]. Author represents a
probabilistic misbehavior detection scheme (iTrust) that can reduce the detection
overhead more effectively. For this, iTrust creates Trusted Authority (TA) to
determine the node’s behavior based on collecting routing evidences and proba-
bilistically checking. It works on the basis of two phases: one is called evidence
generation phase and other is called routing evidence phase [10]. Author suggests a
scheme that is known as statistical-based detection of blackhole and grayhole
attackers (SDBG) by which network can address both individual and collusion
attacks. In this scheme, nodes are required to record their nodes histories by which
it helps to evalute forwarding behaviors. To aviod continous dropping of messages,
scheme promotes metrices at the same time by which attackers have to create fake
encounter records frequently [11]. Robust algorithm helps to detect colluding
attacks for that it exploits abnormal pattern for appearance frequency and number of
message sent in fake encounter [12].
In this section, it is discussed how pool title system works regarding providence of
perfect network in which there is no chance of occurrence blackhole and grayhole
in the network by which network detects all nodes that behave like malicious nodes
and for further communication it escapes from the network (Fig. 3).
Following are the steps which were used in the pool title method for the wireless
communication by that our system is more effective for its working.
In the below diagram, it is easily understood that it shows the relation between
user and admin. For appropriate system, there should be a healthy relationship
between admin and its users. In this system, all the parameter of the system should
be considered as perfect organizer, due to which information of system should reach
its proper destination.
3.1 Create Pool Manager
In the initial step of the proposed system, first all the nodes of the network should
register to wireless router and make scenario in which we have to make our own
user ID and passwords by which selective nodes of the system can only take part in
the communication and no further nodes take part. Create pool manager and all the
nodes which were taking part for communication have to register to it. For that,

nodes have to send their node names and IP address. To check the IP address, first
go to command prompt and put IP configure; then the system generates its IP
address and Mac Address. After this, whatever information used for communication
is stored in pool manager for further transactions.
3.2 Select Destination Node
For this step, source node of the network can view the entire registered node which
was required for communication in our system. After source node is selected,
destination node information is sent. After the destination node is selected by the
source node, then a routing path is requested to the pool manager to update its
profile. In this, systems have option to select all the nearby nodes due to which
communication can be easy and short. What nodes select by source node to send
data those nodes only receive the data rest all are also active but at that time not
used for communication.
3.3 Shortest Path Estimation
After receiving the routing path request from the source node, pool manager assigns
the random integer value for each node of the network, which involves the com-
munication of data. Network uses Dijkstra’s algorithm for calculating the shortest
path for routing process, where each node assigned a random weight and then based
on the random weights, node distances are been evaluated for instance nodes. As
Dijkstra’s algorithm not only work for shortest path between nodes but it can put
common variant fixes a single nodes as the source node and finds the shortest paths
from all source to all the other nodes in the graph and produce shortest path tree.
3.4 Use of Hash Key
As early as shortest path is calculated by the help of Dijkstra’s algorithm with help
of prior step in which data is start routing based on the route map provided by the
pool manager. And for every node, a signature of the data is sent in the form of hash
code that will be generated by using MD5 algorithm. Hash function is used to
convert the data of arbitrary length to a fixed length and it provides much faster than
a symmetric encryption.
3.5 Identification of Malicious Nodes
Here, in the last step, a strict analysis of the hash signature is carried out for the
pattern of the blackhole and grayhole attacks though avalanche effect for a fix time
slot is called as “tile”. By the help of avalanche effect, system can describe cryp-
tography in which there is a slight change in the input value (message) that causes a
significant change in the output (hash value). If a node is behaving as the grayhole,
then it partially drops the data or it malfunctions the data, whereas for the blackhole
it completely drops the data so that only one hash signature will be received by the
pool manager.
The whole process of blackhole and grayhole detection process can be shown in
Algorithm 1
Algorithm 1: Gray- and Blackhole Identification Using Pool Tile Method
//Input: Sender Data bytes B
//Destination Node Nst
//Output: Successful identification of Gray- and Blackhole node
Step 0: Start
Step 1: Activate pool managers Tm
Step 2: Register Node Ni with pool manager Tm
Step 3: Set Tile T for Pool managers Tm(Where tile is Time in Seconds)
Step 4: Choose Data bytes B by source node Src
Step 5: Set Destination node Nst
Step 6: Identify the shortest path Pth
Step 7: WHILE B 62Bn
Step 8: for each tile T
Step 9: Bd ! Snt2B (Pd = previous data in Hash and Snt = Run time Source node)
Step 10: Cd ! Tnt2B(Cd = Current data in hash and Tnt = Run time Destination
node)
Step 11: check IFCd 6¼ Pd(Avalanche Effect)
Step 12: Label Current instance nodeCn as Grayhole
Step 13: END IF
Step 14: IF Cd = NULL
Step 15: Label Current instance node Cn as Blackhole
Step 16: END IF
Step 17: END WHILE
Step 18: Stop
Table 1 Comparison between pool title method and cluster analysis method
Number of Number of packets Average transmission Average transmission
hops sent time(ms) time(ms)
Tile bitmap process DSDV protocol
1 5334 675.64 1000
2 5334 728.98 1000
3 5334 1268.38 1000
4 5334 1369.06 1000
5 5334 1102.36 1000
6 5334 693.42 1000
7 5334 728.98 1000
8 5334 746.76 1000
9 5334 656.86 1000
10 5334 497.84 1000
4 Results and Discussions
An explicit experiment is conducted for the performance evaluation of the system

based on the time taken to identify the blackhole and grayhole in the wireless
network. System is set to detect the blackhole and grayhole for much number of
runs to record the timings of performance, which are tabled in Table 1.
When it was compared with the pool tile and cluster analysis method, following
difference occurs due to which it can easily analyze pool tile mapping that is far
better than any other method. Cluster analysis method works for the destination
sequence distance vector routing protocol in which time requires more to transmit
and analysis blackhole and grayhole of the network (Fig. 4).
5 Conclusion and Future scope
Due to increased use of wireless networks, there is problem which faces a huge
entry point to the attacker in the network, in which blackhole and grayhole attacks
are most common and highly dangerous forms of threats to the wireless network. In
this paper, it introduced pool tile method to identify the attacker node which effi-
ciently works than that of most of the other methods. This is because proposed
system eradicates the burden of detection of malicious nodes from the routing nodes
which come under shortest path sequence, thereby assigning the entire detection job
to high configured pool manager. In the future, this method can work for more
efficiently incorporated by assigning the detection of blackhole and grayhole jobs to
the multiple pool manager on increasing of pool size.
Fig. 4 Performance comparison of different methods
References
1. Jian-Ming Chang, Po-Chun Tsou, Isaac Woungang, Han-Chieh and Chin-Feng Lai,
“Defending Against Collaborative Attacks by Malicious Nodes in MANETs: A
Cooperative Bait Detection Approach” IEEE System Journal, Vol. 9, No. 1, March 2015
2. Pham Thi Ngoc Diep, Chai kiat Yeo, “Detecting Colluding Blackhole and Greyhole Attacks
in Delay Tolerant Network”, IEEE Transactions on Mobile Computing, https://doi.org/10.
1109/TMC.2015.2456895
3. S. Sankara Narayanan and Dr. S. Radhakrishnan, “Secure AODV to Combat BlackHole
Attack in MANETs” 2013 International Conference on Recent Trends in Information
Technology (ICRTIT) ISBN:978-1-4799-1024-3/13/$31.00 ©2013 IEEE
4. A. M. Kanthe, D. Simunic, R. Prasad, “A Mechanism for Detection of Gray Hole Attack in
Mobile Ad Hoc Networks” International Journal of Computer Applications, Volume 53,
Sep. 2012
5. Devu Manikantan Shila, Yu Cheng and Tricha Anjali, “Channel Aware Detection of Grayhole
Attacks in wirless Mesh Networks”, 978-1-4244-4148-8/09 IEEE “Globecom” 2009
6. Disha G. Kariya, Atul B. Kathole, Sapna R. Heda, “Detecting Blackhole and Gray hole
Attacks in Mobile Adhoc Network using an Adaptive method”, International Journal of
Emerging Technology and Advanced Engineering, Volume 2, Issue 1, January 2012
7. Shuchita Upadhyaya and Karishma, “A Co-operative Mechanism to contrast Blackhole
Attacks in Delay Tolerant Networks”, IOSR Journal of Computer Engineering, E-ISSN:
2278-0661, P-ISSN: 2278-8727
8. Sangeetha. R and Krishnammal. N, “Detection of Routing Attacks in Disruption Tolerant
Networks”, The International Journal of Engineering and Science, Volume 2, 2013 ISSN:
2319-1813, ISBN: 2319-1805
9. Susan Basil, Panchami V, “Attack Identification in a Cryptographic system using Dynamic
Taint Propagation”, International Journal of Science, Engineering and Technology Research,
Volume 4, Issue 4, April 2015
10. Haojin Zhu, Suguo Du, Zhaoyu Gao, Mianxiong Dong and Zhenfu Cao, “A Probabilistic
Misbehavior Detection Scheme toward Efficient Trust Establishment in Delay-Tolerant
Networks” IEEE Transcations on Parallel and Distributed System, Volume 25, Number 1,
January 2014
11. Yinghui Guo, Sebastian Schildt and Lars Wolf, “Detecting Blackhole and greyhole Attacks in
Vehicular Delay Tolerant Networks”, IEEE, 978-1-4673-5494-3/2013
12. Surpriya Pustake, Dr. S. J. Wagh, D. C. Mehetre, “Grayhole Detection and Removal in
MANETs by Pool Tile Method”, International Journal of Science and Research, Volume 5,
Issue 7, July 2016
Evaluation of Multi-label Classifiers
in Various Domains Using Decision Tree
V. S. Tidake and S. S. Sane
Abstract One of the commonly used tasks in mining is classification, which can be
performed using supervised learning approach. Because of digitization, lot of
documents are available which need proper organization, termed as text catego-
rization. But sometimes documents may reflect multiple semantic meanings, which
represents multi-label learning. It is the method of associating a set of predefined
classes to an unseen object depending on its properties. Different methods to do
multi-label classification are divided into two groups, namely data transformation
and algorithm adaptation. This paper focuses on the evaluation of eight algorithms
of multi-label learning based on nine performance metrics using eight multi-label
datasets, and evaluation is performed based on the results of experimentation. For
all the multi-label classifiers used for experimentation, decision tree is used as a
base classifier whenever required. Performance of different classifiers varies
according to the size, label cardinality, and domain of the dataset.

Keywords Machine learning Multi-label classification Data transformation

Algorithm adaptation Decision tree Label cardinality
1 Introduction
One commonly used task in mining is classification. If a set of known instances,

called train set, is used to train the model, then it is referred as supervised learning.
Once the training and testing of the model are complete, it is useful for classification
V. S. Tidake (&)
Department of Computer Engineering, MCERC, Savitribai Phule Pune
University, Nashik, India
e-mail: vaishalitidake@yahoo.co.in
S. S. Sane
Department of Computer Engineering, KKWIEER, Savitribai Phule Pune
University, Nashik, India
e-mail: sssane@kkwagh.edu.in

https://doi.org/10.1007/978-981-10-7245-1_13
118 V. S. Tidake and S. S. Sane
of unseen instances. Several distinct domains [1–8] like TC use supervised learning.
Sometimes, a document may reflect multiple semantic meanings. Hence, unlike
traditional classification, it may be associated with one or more than one class
labels, which represents multi-label learning. It is the method of associating a set of
predefined classes to an unseen document depending on its contents. Association of
each input example with single-class label is termed as SL (single-label) classifi-
cation or just classification. Depending on the total count of class labels involved,
SL classification is either referred as a binary single-label (BSL) classification when
the label space has only two class labels or multi-class single-label
(MSL) classification if the label space includes more than two class labels. For
example, a news document represented as a square in Fig. 1 may be related to either
education (+) or health (−) category representing BSL (Fig. 1a) or one of education,
health, and economy (^) categories representing MSL (Fig. 1b) [2, 6]. A news
saying that “Yoga and meditation are crucial for the stress management of students”
is related to education as well as health categories (+ −) representing MLC
(multi-label) classification (Fig. 1c). Already many tools and algorithms are
available to handle SL classification problems. Use of MLC in the recent past has
been done for TC, prediction of gene function, tag recommendation, discovery of
drug, [2–6], etc. So in the area of machine learning, it has gained the position of an
upcoming research field.
This paper deals with a comparative study of MLC. Sections 2, 3, and 4 describe
the metrics used for evaluation, two approaches used for MLC, about the experi-
ments and results, respectively. Section 5 gives the concluding remarks.
2 Multi-label Classification (MLC)
A. Definition
Like SL, MLC uses supervised approach for learning. It is the task which relates
an unseen instance considering its features to a set of predefined labels. Let
C represents a set of disjoint labels. Let an instance be described by a vector fj of
(a) BSL (b) MSL classification, (c) ML classification

classification,
Fig. 1 Classification
Evaluation of Multi-label Classifiers in Various Domains … 119
features and belongs to a set yj of labels. Let Q denotes a training set (xj, yj), and
then obtain a function f(x) for mapping each fj vector to a set yj of labels, where
yj C and j = 1, 2… |Q|.
B. Metrics used to measure performance
Let PLi and ALi denote set of predicted labels by a classifier and a set of actual
labels for training instance xi. Let T and C denote a test set and a set of disjoint
labels, respectively. Let f denotes a classifier. ML learning uses metrics following
metrics.
Hamming loss Most commonly used which is used to measure the number of
times an instance and its associated label is not correctly classified. Expected value
of hamming loss metric is small [2].
1 X T
jBðPLi H ALiÞj
HLð f Þ ¼ ; ð1Þ
jT j i¼1 jC j
where B(.) = 0 if ALi and PLi are same for all labels of instance i, else B(.) = 1.
Here, H is used for symmetric difference.
Ranking loss This metric measures performance of ranking task which generates
all labels in the order of relevance. It is used to measure the number of times an
irrelevant label has been ranked above the relevant labels. Expected value of
ranking loss metric is small [6].
1 X
jT j
1
RLð f Þ ¼ jfðy1; y2Þjlðy1; xiÞ lðy2; xiÞgj; ð2Þ
jT j i¼1 jALijALi
where y1 2 ALi and y2 2 ALi. Assume lðq; r Þ denotes relevance of label q for an
instance r and smaller value denotes more relevance.
One-error It counts the number of times a label generated by the classifier at the
top rank does not appear in the correct labels associated with an input instance. The
smaller the one-error, the better it is [6, 7]. Here, B(.) = 1 if (.) is true, else B(.) = 0.
1 X T
OE ð f Þ ¼ B arg miny2C lðy; xiÞÞ 62 ALi ð3Þ
jT j i¼1
Coverage It measures how much down the list of labels generated by the classifier
should be traversed to include all the labels relevant to an example assuming
top-most labels appear at the start of the list. The less the value, the better is the
result [6, 7].
1 X
jT j
CGð f Þ ¼ max lðy; xiÞ 1 ð4Þ
jT j i¼1 y2ALi
Average precision It computes an average proportion of relevant labels which are

ranked above a particular relevant label. The bigger the value, The better is the
result [6, 7].
1 X T
1 X jfz 2 ALijlðz; xiÞ lðy; xiÞgj
APð f Þ ¼ ð5Þ
jT j i¼1 jALij y2ALi lðy; xiÞ
Subset Accuracy It is an average over all the instances which checks whether
predicted label set of an instance is same as its actual label set [3, 5, 9].
1 X T
SAð f Þ ¼ BðPLi ¼ ALiÞ; ð6Þ
jT j i¼1
where B(.) = 1 if ALi and PLi are same for all labels of instance i, else B(.) = 0.
Example-Based Recall, Precision, and F-Measure [2, 6, 7]:
1 X jPLi \ ALij
jT j
ExRcð f Þ ¼ ;
jT j i¼1 jALij
1 X T
jPLi \ ALij
ExPr ð f Þ ¼ ; ð7Þ
jT j i¼1 jPLij
1 X 2jPLi \ ALij
jT j
ExF1ð f Þ ¼
jT j i¼1 jALij þ jPLij
3 Various Methods
In the literature, various methods to perform multi-label learning have been

developed and reported. Two broad categories used to perform MLC are the data
transformation and the algorithm adaptation [4]. The data transformation approach
involves transformation of an input instance into data which suits for many
single-label traditional classifiers, whereas the algorithm adaptation approach
involves transformation of SL classifier algorithm which suits multi-label
data [2, 3].
So far various SL algorithms are developed by researchers. The data transfor-
mation approach utilizes these existing SL algorithms. The approach transforms
data representation from multi-label to single-label which is acceptable by existing
SL classification algorithms. In different words, the data transformation operates on

the fundamental concept of “fit data to an algorithm” [2]. Transformation does not
change the algorithm, hence is said to be “independent of an algorithm” [3].
Algorithms like BR, LP, CC, ECC, RAkEL, and HOMER use this approach. The
adaptation approach modifies the existing SL algorithms for managing multi-label
data appropriately. In different words, the algorithm adaptation operates on the
fundamental concept of “fit an algorithm to data” [2]. Since an algorithm, not data,
is updated, this approach is said to be “dependent on an algorithm” [3]. Algorithms
like BRkNN, MLkNN, ML-C4.5, and BP-MLL use this approach. Let C represents
a set of labels.
Binary Relevance (BR)
It is the most widely used method for data transformation in which a multi-label
problem is converted into |C| binary SL classification problems. Each of the binary
classifiers contributes its vote separately to do classification [4, 5]. BR has one
disadvantage of not considering the association between labels (if any) as it treats
every label individually [2, 6].
Label Powerset (LP)
Overcoming the drawback of BR for treating every label individually is removed in
LP. It considers each different group of labels as a separate class and treats the
entire problem as a multi-class single-label (MSL) problem [7].
Multi-label data is treated as multi-class data. For example, multi-label data
having |C| labels forms 2|C| classes with different label combinations. Thus, LP
considers multiple labels simultaneously and overcomes the drawback of BR [8].
However, the number of groups of classes formed increases with |C|. It results in
distribution of the original data into different groups of classes. This distribution
may result in scenario similar to class imbalance where few classes may belong to
more number of instances, whereas some classes may belong to less number of
instances. The situation may affect classifier accuracy. Also, higher value of
C causes time complexity of LP to become worst.
Random k-Label sets (RAkEL) It is actually an ensemble of multiple LP clas-
sifiers having different combinations of all labels referred as label sets [1]. These
k-size label sets help to remove the class imbalance drawback in LP. For multi-label
data with |C| labels, N label sets each of size k are formed randomly and separate LP
classifier models are designed for them. Average of votes obtained from N models
for each label is used for classification of an unseen instance. If it is more than a
threshold, then prediction of that label is P; otherwise, it is A to represent the
presence or absence of that label. However, classifier accuracy depends on the label
sets which are selected randomly. Also, choosing N and k values may also affect the
classifier performance.
Classifier Chain (CC) A weakness of BR not considering association between
labels is removed in CC [7, 8]. Like BR, it transforms ML problem into |C| SL
problems and for each label Cj, a separate binary classifier Bj is designed. But the
input for each classifier Bj is different. Each classifier Bj takes as input all feature
vectors f1…D of all instances and predictions of all earlier classifiers also. In general,
output Oij of each classifier represents prediction of classifier Bi for Classi for
instance j. Oij takes values either P or A for class Ci of instance j. Accordingly,
output of all classifiers is obtained. Thus, label information is passed from classifier
Bi to Bj, and so on. Such organization takes into account associations among labels
and thus overcomes the weakness of BR described earlier. But an important con-
cern in CC is that sequence of considering labels may result in different classifier
accuracies [8] affecting its performance to a great extent and guessing the best
possible order is difficult.
Ensemble of Classifier Chains (ECC) Instead of depending on single chain of
labels, ECC takes benefit of using multiple different order chains as well as
ensemble. It obtains votes from a group of classifiers each using different chains and
different set of instances, which improves the accuracy of prediction [8].
Calibrated Label Ranking (CLR) It is a modification of Ranking by Pairwise
Comparison (RPC) [2, 6]. It augments the label set with a virtual label Lv. Then, it
constructs C(C−1)/2 binary classifiers as in RPC. Each classifier Bij outputs P for an
instance if it contains label Ci and A if it contains label Cj, and does not consider
instances having both or none of these labels in the pair (Ci, Cj) [5]. CLR also
constructs C binary classifiers to represent relationship between each label Ci and a
virtual label Lv. While classifying an unseen instance, votes are obtained from all
these constructed classifiers to generate ranking of all labels having relevant and
irrelevant labels separated by a virtual label.
Multi-Label k-Nearest Neighbors (MLkNN) It updates traditional kNN algo-
rithm to process multi-label data. For classification of an unseen instance, it finds k-
nearest neighbors. Then, statistical data like count of nearest neighbors for a
training instance x associated with particular label and not associated with particular
label is obtained for each label using computed k-nearest neighbors. Next, a rule
based on Bayes theorem is applied to labels of an unseen instance [2, 10]. Further, it
computes label information from obtained nearest neighbors with the help of
posterior and prior probabilities. The MLkNN exhibits a limitation of not consid-
ering label relationship by processing each label separately.
Hierarchy Of Multi-labEl leaRners (HOMER) Each individual classifier works
on smaller size distinct label set as compared to the original one, where each label
set contains related labels together. Hierarchical distribution of these labels is an
important feature of this classifier [2].
4 Experimentation and Results
A. Multi-label Datasets
MEKA is a WEKA-based project. An open-source library Mulan uses Java to
perform multi-label data mining. Data sets from various domains are made available
in MEKA, MULAN, and LibSVM [11–14]. Table 1 briefs some multi-label
datasets used for experimentation along with their information [11]. Table 1 shows
label cardinality of all datasets which denotes the average number of labels per
example [2].
B. Parameter initialization
For BR, LP, CLR, CC, and ECC, C4.5 decision tree algorithm [15, 12] is used as
base SL classifier. For HOMER and RAkEL, LP with C4.5 is used as a base
classifier. HOMER is run with three clusters and random method. RAkEL [1] runs
with ix models, 3 as size of subset and 0.5 as threshold. MLkNN is executed with
10 neighbors and 1 as smoothing factor. Cross-validation is used for evaluation
with tenfolds [9, 11].
C. Results and Discussion
Experiments are carried out using Intel(R) Core(TM) i5-6200U CPU
@2.30 GHz having 8 GB RAM and Windows10 and Java for programming with
libraries from Mulan 1.5 [11] and WEKA 3.8.1 [12]. Figures 2, 3, 4, 5, 6, 7, 8, 9,
and 10 show results obtained from execution of eight classifiers on eight datasets to
measure nine performance metrics. Information in brackets shows criteria expected
for that metric value. Legend for all charts is same as that shown in Fig. 10.
LP classifier produced memory error for mediamill dataset when running on the
above-mentioned hardware. An attempt is done to compare results obtained with
other work reported in the literature [7, 11, 16]. The variation in the results may be
due to different base classifiers used or different parameter settings or different
default parameters in different tools used for experimentation. For multimedia
Table 1 Multi-label datasets

Domain Dataset #attributes #labels #instances Label Label
cardinality density
Biology Yeast 103 14 2417 4.237 0.30
Biology Genbase 1186 27 662 1.252 0.05
Text Medical 1449 45 978 1.245 0.03
Text Enron 1001 53 1702 3.378 0.06
Multimedia Scene 294 6 2407 1.073 0.18
Multimedia Corel5 k 499 374 5000 3.522 0.01
Multimedia Emotions 72 6 593 1.868 0.31
Multimedia Mediamill 120 101 43,907 4.375 0.04
Fig. 2 Hamming loss (small) 0.300

0.200
0.100
0.000
BR
LP
MLkNN
CLR
RAkEL
HOMER
CC
ECC
Fig. 3 Ranking loss (small) 0.800
0.600
0.400
0.200
0.000
BR
LP
CLR
RAkEL
MLkNN
HOMER
CC
ECC
Fig. 4 Subset accuracy 1.500
(large)
1.000
0.500
0.000
BR
LP
CLR
RAkEL
MLkNN
HOMER
CC
Fig. 5 One-error (small) 1.500 ECC
1.000
0.500
0.000
BR
LP
CLR
RAkEL
MLkNN
HOMER
CC
ECC
Fig. 6 Coverage (small) 400.000

300.000
200.000
100.000
0.000
BR
LP
CLR
RAkEL
MLkNN
HOMER
CC
ECC
1.500
1.000
0.500
0.000
BR
LP
CLR
RAkEL
MLkNN
HOMER
CC
ECC
Fig. 7 Average precision (large)
1.500
1.000
0.500
0.000
BR
LP
CLR
RAkEL
MLkNN
HOMER
CC
ECC
Fig. 8 Example-based precision (large)
1.500
1.000
0.500
0.000
BR
LP
CLR
RAkEL
MLkNN
HOMER
CC
ECC
Fig. 9 Example-based recall (large)
1.500 emoƟons
1.000 scene
yeast
0.500
genbase
0.000 medical
enron
mediamill
corel5k
Fig. 10 Comparison of various classifiers for example-based F-measure (large)
domain, MLkNN achieved best results in terms of hamming loss followed by subset
accuracy. Especially, mediamill and corel5 k datasets of multimedia domain having
more label cardinality achieved better results on hamming loss and subset accuracy
among all. Emotions dataset in multimedia domain showed good performance for
all MLC methods using C4.5 [15] except LP and HOMER for all example-based
metrics, namely hamming loss, example-based recall, precision, and F-measure as
compared to ranking-based metrics, namely coverage, one-error, and average pre-
cision [7]. LP, RAkEL, and HOMER performed poor in all domains for all metrics.
The reason can be that the decision tree classifier may not be able to utilize the label
relationship to the required extent when used with LP. Drawback of LP showing
poor performance if different classes are associated with different numbers of
examples should not be ignored. Also, subset size and number of models used with
RAkEL [1] may be too small for not considering label correlation properly.
Already, LP with decision tree has shown poor results and same is taken as base
classifiers for HOMER and RAkEL, which may be the reason behind their poor
performance. ECC performed much better than CC, LP, CLR, RAkEL, and
HOMER for all domains next to MLkNN for ranking loss, hamming loss, coverage,
and one-error, but poor for the remaining metrics. It can be due to the power of
ensemble which has been already proved to be better than single-label classifier in
the literature [8]. For biology domain, BR, CC, and ECC showed best performance
on all example-based measures followed by CLR for genbase dataset only. Yeast
dataset has shown poor performance than genbase dataset. We can hypothesize that
it may be due to higher label cardinality for yeast dataset as compared to genbase
dataset. Similar may be the case for the text domain. Medical dataset in text domain
has achieved better metric values especially in BR, CLR, CC, and ECC for the same
reason as compared to enron dataset. MLkNN and CLR showed less misclassifi-
cation of instance label pairs by giving smaller hamming loss than [7]. The cause
may be different in number of neighbors and the base classifier selected in both
classifiers, respectively.
5 Conclusion
There are two ways to design algorithms for ML classification. Data transformation
methods transform data having multiple labels into single-label aiding traditional
single-label methods. The adaptation approach updates the existing algorithms of
learning for processing multi-label data. For all the multi-label classifiers used in
this work, decision tree is used as a base SL classifier wherever necessary.
Mediamill and corel5 k of multimedia domain having more label cardinality
achieved good results on hamming loss and subset accuracy. For biology domain,
BR, CC, and ECC showed best performance on all example-based measures on
both datasets followed by CLR for genbase only. Medical in text domain has
achieved better metric values especially in BR, CLR, CC, and ECC as compared to
enron. ECC also showed effectiveness next to MLkNN but better than other
methods, thus giving their votes to ensemble and adaptation, respectively. It will be
interesting to see the effect of other base classifiers on different ML classifiers.
References
1. G. Tsoumakas, I. Katakis, and I. Vlahavas, “Random k-labelsets for multilabel classification,”

IEEE Trans. Knowl. Data Eng., vol. 23, no. 7, pp. 1079–1089, Jul. 2011.
2. M. L. Zhang and Z. H. Zhou, “A review on multi-label learning algorithms”, IEEE
Transactions On Knowledge And Data Engineering, Vol. 26, No. 8, August 2014.
3. A. de Carvalho and A. A. Freitas, “A tutorial on multi-label classification techniques,” in
Studies in Computational Intelligence 205, A. Abraham, A. E. Hassanien, and V. Snásel, Eds.
Berlin, Germany: Springer, 2009, pp. 177–195.
4. G. Tsoumakas and I. Katakis, “Multi-label classification: An overview,” Int. J. Data
Warehousing Mining, vol. 3, no. 3, pp. 1–13, 2007.
5. G. Tsoumakas, I. Katakis, and I. Vlahavas, “Mining multilabel data”, Data Mining and
Knowledge Discovery Handbook, O. Maimon and L. Rokach, Eds. Berlin, Germany:
Springer, 2010, pp. 667–686.
6. G. Tsoumakas, M. -L. Zhang, and Z. -H. Zhou, “Tutorial on learning from multi-label data,”
in ECML PKDD,Bled,Slovenia,2009[Online].Available: http://www.ecmlpkdd2009.net/
wpcontent/uploads/2009/08/learning-from-multi-label-data.pdf.
7. G. Madjarov, D. Kocev, D. Gjorgjevikj, and S. Džeroski, “An extensive experimental
comparison of methods for multi-label learning,” Pattern Recognit., vol. 45, no. 9, pp. 3084–
3104, 2012.
8. J. Read, B. Pfahringer, G. Holmes, E. Frank, “Classifier chains for multi-label classification”,
in: Proceedings of the 20th European Conference on Machine Learning, 2009, pp. 254–269.
9. Nasierding, Gulisong, and Abbas Z. Kouzani. “Comparative evaluation of multi-label
classification methods.” In Fuzzy Systems and Knowledge Discovery (FSKD), 2012 9th
International Conference on, pp. 679–683. IEEE, 2012.
10. M. -L. Zhang and Z. -H. Zhou, “ML-KNN: A lazy learning approach to multi-label learning,”
Pattern Recognit., vol. 40, no. 7, pp. 2038–2048, 2007.
11. G. Tsoumakas, E. Spyromitros-Xioufis, J. Vilcek, and I. Vlahavas, “MULAN: A Java library
for multi-label learning,” J. Mach. Learn. Res., vol. 12, pp. 2411–2414, Jul. 2011.
12. M. Hall et al., “The WEKA data mining software: An update”, SIGKDD Explor., vol. 11, no.
1, pp. 10–18, 2009.
13. C. -C. Chang and C. -J. Lin, “LIBSVM: A library for support vector machines,” ACM Trans.
Intell. Syst. Technol., vol. 2, no. 3, Article 27, 2011 [Online]. Available: http://www.csie.ntu.
edu.tw/*cjlin/libsvm.
14. Tidake, Vaishali S., and Shirish S. Sane. “Multi-label Learning with MEKA”, CSI
Communications (2016).
15. Ross Quinlan (1993). C4.5: Programs for Machine Learning. Morgan Kaufmann Publishers,
San Mateo, CA.
16. Cerri, Ricardo, Renato RO da Silva, and André CPLF de Carvalho. “Comparing methods for
multilabel classification of proteins using machine learning techniques.” In Brazilian
Symposium on Bioinformatics, pp. 109–120. Springer Berlin Heidelberg, 2009.
An Effective Multilabel Classification
Using Feature Selection
S. S. Sane, Prajakta Chaudhari and V. S. Tidake
Abstract Recently, multilabel classification has received significant attention

during the past years. A multilabel classification approach called coupled k-nearest
neighbors algorithm for multilabel classification (called here as CK-STC) reported
in the literature exploits coupled label similarities between the labels and provides
improved performance [Liu and Cao in A Coupled k-Nearest Neighbor Algorithm
for Multi-label Classification, pp. 176–187, 2015]. A multilabel feature selection is
presented in Li et al. [Multi-label Feature Selection via Information Gain, pp. 346–
355, 2014] and called as FSVIG here. FSVIG uses information gain that shows
better performance when used with ML-NB, ML-kNN, and RandSvm when
compared with existing multilabel feature selection algorithms.This paper investi-
gates the performance of FSVIG when used with CK-STC and compares its per-
formance with other multilabel feature selection algorithms available in MULAN
using standard multilabel datasets. Experimental results show that FSVIG when
used with CK-STC provides better performance in terms of average precision and
one-error.

Keywords Algorithm adaptation Coupled label similarity Feature selection

Information gain Multilabel classification
S. S. Sane P. Chaudhari (&)

Department of Computer Engineering, K. K. Wagh Institute of Engineering
Education and Research, SPPU, Nashik, India
e-mail: p.prajakta42@gmail.com
S. S. Sane
e-mail: sssane@kkwagh.edu.in
V. S. Tidake
Computer Engineering, Matoshri College of Engineering and Research Center,
Savitribai Phule Pune University, Nashik, India
e-mail: vaishalitiake@yahoo.co.in

https://doi.org/10.1007/978-981-10-7245-1_14
130 S. S. Sane et al.
1 Introduction
Previously in supervised learning, each example is associated with only one

instance and single label is assigned to this instance. As single label is assigned to
the instance, this classification is known to be single-label classification. In some
situations, where real-world object cannot be handled by single label, this problem
can be solved by using multilabel classification. The purpose of multilabel classi-
fication is to discover the set of labels for unseen instances. In multilabel classifi-
cation, input data instances are related to multiple class labels. For example,
particular news article can be associated with economical article, political article,
and business article.
Multilabel classification is used in several areas such as bioinformatics, text
categorization, tag recommendation, image classification, direct marketing, medical
diagnosis, query categorization, and protein function prediction [3, 4].
2 Related Work
Multilabel classification techniques are grouped into two categories that are prob-
lem transformation and algorithm adaptation method [5].
Problem transformation approach transforms the multilabel problem into a set of
single-label problem. Problem transformation approach is algorithm independent.
Several problem transformation methods such as binary relevance (BR) [6], label
powerset (LP) [5], and classifier chain (CC) [7] are used in multilabel classification.
The algorithm adaptation method extends the available machine learning algo-
rithms so as to handle multilabel data. The several algorithm adaptation approaches
have been proposed that are multilabel k-nearest neighbor (ML-kNN) [8], binary
relevance k-nearest neighbor (BR-kNN) [9], and IBLR [10].
ML-kNN and BR-kNN do not consider label correlation whereas IBLR con-
siders label correlation but not good in terms of hamming loss.
Liu and Cao [1] proposed coupled k-nearest neighbor algorithm for multilabel
classification (CK-STC). CK-STC is based on coupled label similarity [11] which
updates ML-kNN algorithm which can handle label correlation. An advantage of
CK-STC is that it considers label correlations and provides better performance than
that of ML-kNN, IBLR, BSVM, and BR-kNN but it is more complex.
Same as traditional classification, multilabel classification also suffers from curse
of dimensionality which can be handled by feature selection. Feature selection
selects the relevant or more efficient features from original set of features.
Traditional feature selection methods handle multilabel data by transforming
multilabel problem into single-label problem and then apply feature selection.
Lee et al. [12] proposed multivariate mutual information-based feature selection.
Due to transformation of multilabel problem into single-label problem, there is
damage to original label structures and which reduces the performance of classifier.
An Effective Multilabel Classification Using Feature Selection 131
So new method called information gain feature selection for multilabel data [2] can
perform feature selection on multilabel data directly.
Multilabel classification tools MEKA [13] and MULAN [14] are available for
handling multilabel data. MEKA is a multilabel extension to WEKA data mining
tool. MULAN provides framework for implementations of many multilabel
learning algorithms. In MULAN feature selection methods, multilabel data is
transformed into single-label data and then traditional single-label dimensionality
reduction technique is applied on these single-label data.
3 Implementation Details
3.1 FSVIG Multilabel Feature Selection Algorithm Details
FSVIG multilabel feature selection algorithm is implemented using information

gain measure. Feature selection method removes the redundant and irrelevant
features from database. Feature selection technique used in this work is reported in
[2]. In this feature, selection filter approach is used. Information gain is used as
information metric for measuring correlation degree between features and labels
that are present in database. Information gain (IG) between features and entire label
set is used to quantify the importance of features. IG is calculated with the help of
label entropy H(L) and feature entropy H(fi). A bigger value of IG represents
greater correlation between feature and labels. But in some situation, information
gain of each label set may not be in same range of measurement. So for comparison,
the normalized processing was performed on information gain (SU). Figure 1
shows pseudo-algorithm for feature selection FSVIG which is as follows.
3.2 CK-STC Multilabel Classification Algorithm Details
CK-STC algorithm for multilabel classification is reported in [1, 15]. In CK-STC

algorithm, coupled label similarity is estimated using intra-coupling similarity
between label value wxi and wyj and inter-coupling label similarity between label
value wxi and wyj w:r:t feature value wzp . Inter-coupling label similarity is calculated
with the help of co-occurrence frequency (CF) and intra-coupling similarity is
calculated with the help of occurrence frequency (F). For every instance in training
data, k-nearest neighbor is calculated. Coupled label similarity is considered while
calculating prior probabilities and frequency arrays. Estimation of k-nearest
neighbor for unseen instance is done. For each label, statistical information is
calculated with the help of k-nearest neighbors of unseen instances. Unseen
instance labels are evaluated via MAP (Maximum a posterior) rule. MAP rule is
based on Bayes theorem.
Input:
F { f1,f2,f3,………fn}
L {l1,l2,l3………….lq}
Output:
Relevant set of features
1.begin
2.Initialize IGS = null
3.for(i=1 to m )
4.IG(fi,L)= H(fi)+H(L)-H(fi,L)
5.SU( fi,L) =
6.IGSi = SU( f i,L)
7.IGS = IGS IGSi
8.End for
9.threshold =
10.Relevant feature = F
11.for(i=1 to m)Do
12.if IGSi < threshold then
13.Relevant feature = Relevant feature - {fi}
14.end for
15.end
Fig. 1 Pseudo-algorithm for feature selection FSVIG
Figure 2 shows pseudo-algorithm for multilabel classification CK-STC which is

stated as follows:
4 Experimental Setup
Table 1 shows benchmark multilabel datasets information including number of

features, total number of labels, and number of instances. Experiments were per-
formed on Genbase, Medical, and Enron dataset.
Experiments were performed on i5 processor and on Windows 7 operating
system. Implementation of multilabel classification algorithm CK-STC and multi-
label feature selection algorithm FSVIG was done in MS VISUAL STUDIO.NET
and development tool is visual studio 10. In experiments, a proposed (CK-STC with
feature selection) method was compared with existing CML-kNN method.
Euclidean distance was used to calculate k-nearest neighbors. The experiments
were performed on k = 5, 7, 9 and then the average is calculated. Tenfold
cross-validation is performed on the above dataset.
Input:
Unlabeled instance (ut ), multilabel data.
Output:
Label set of unlabeled instance.
1. begin
2. Calculate intra coupling similarity
3. Calculate inter coupling similarity

4. Calculate coupled label similarity (CLS)
5. for i=1 to m do
6. Estimate k nearest neighbor k(ui) for each ui
7. end for
8. for j=1 to q do
9. Estimate prior probability P(Hj) and P(-Hj)
10. Maintain frequency array
11. End for
12. Estimate k nearest neighbor k(ut ) for unseen instance ut
13. for j = 1 to q
14 .Estimate unseen instance label statics
15. End for
16. Return label set of unlabeled instance ut according to MAP rule
17. end
Fig. 2 Pseudo-algorithm for multilabel classification CK-STC
Table 1 Dataset Description

Datasets Number of Number of Number of Label Label Feature
instances features total labels cardinality density type
Genbase 662 1185 27 3.378 0.064 Categorical
Medical 978 1449 45 1.252 0.046 Categorical
Enron 1702 1001 53 1.245 0.028 Categorical
Performance of multilabel classification is measured in terms of hamming loss,

one-error, and average precision. For each evaluation metric, “#” indicates “smaller
value has better results” and “"” indicates “bigger value has better results”. Bold
value indicates winner of the classifier. Experiments were carried out in four sets as
follows:
(1) Multilabel classification using CK-STC without feature selection
Table 2 shows results of CK-STC algorithm.
Table 2 Results of CK-STC

Dataset One-error # Hamming loss # Average precision "
Enron 0.303 0.087 0.597
Genbase 0.008 0.003 0.991
Medical 0.158 0.013 0.874
Average 0.156 0.034 0.821
(2) Multilabel feature selection using FSVIG algorithm

Table 3 indicates results of feature selection FSVIG.
(3) Multilabel classification using MULAN feature selection technique binary
relevance attribute evaluator and its comparison with CK-STC with fea-
ture selection FSVIG
The results of multilabel feature selection algorithm FSVIG and MULAN
attribute selection algorithms such as binary relevance attribute evaluator algorithm
represented in Tables 4, 5 and 6.
Table 4 indicates that CC algorithm with binary relevance attribute evaluator
performs better than other algorithms w.r.t. hamming loss. Table 5 indicates that
CK-STC algorithm with FSVIG feature selection technique performs better than
other algorithms w.r.t. one-error. Table 6 shows that CK-STC algorithm with
FSVIG feature selection technique performs better than other algorithms w.r.t
average precision.
(4) Multilabel classification using MULAN feature selection technique multi-
class attribute evaluator and its comparison with CK-STC with feature
selection FSVIG
The results of multilabel feature selection algorithm FSVIG and MULAN
attribute selection algorithms such as multiclass attribute evaluator algorithm rep-
resented in Tables 7, 8, and 9.
Table 7 indicates that CK-STC algorithm with FSVIG feature selection tech-
nique performs better than other algorithms w.r.t. hamming loss. Table 8 indicates
that CK-STC algorithm with FSVIG feature selection technique performs better
than other algorithms w.r.t. one-error. Table 9 indicates that CK-STC algorithm
with FSVIG feature selection technique performs better than other algorithms w.r.t
average precision.
Table 3 Results of feature selection as per FSVIG

Dataset Number of features before feature Number of features after feature
selection selection
Enron 1001 596
Genbase 1185 316
Medical 1449 400
Table 4 Comparisons of CK-STC, feature selection FSVIG with MULAN binary relevance attribute evaluator w.r.t. hamming loss #
Dataset Number of attribute CK-STC FSVIG MULAN (binary relevance attribute evaluator)
used CK-STC CK-STC BR CC RAKEL ML-kNN LP BR-kNN IBLR
Enron 596 0.087 0.054 0.052 0.049 0.052 0.047 0.052 0.069 0.056 0.055
Genbase 316 0.003 0.002 0.004 0.001 0.001 0.001 0.005 0.001 0.003 0.002
Medical 400 0.013 0.012 0.012 0.011 0.001 0.011 0.012 0.013 0.014 0.013
Average – 0.034 0.022 0.023 0.020 0.018 0.020 0.023 0.027 0.025 0.024
An Effective Multilabel Classification Using Feature Selection
135
136
Table 5 Comparisons of CK-STC, feature selection FSVIG with MULAN binary relevance attribute evaluator w.r.t. one-error #
Enron 596 0.303 0.225 0.285 0.369 0.397 0.268 0.282 0.832 0.441 0.332
Genbase 316 0.008 0.007 0.013 0.003 0.003 0.010 0.006 0.006 0.016 0.013
Medical 400 0.158 0.146 0.192 0.190 0.197 0.238 0.200 0.262 0.283 0.231
Average – 0.156 0.126 0.163 0.187 0.200 0.172 0.163 0.366 0.247 0.192
S. S. Sane et al.
Table 6 Comparisons of CK-STC, feature selection FSVIG with MULAN binary relevance attribute evaluator w.r.t. average precision "
Enron 596 0.597 0.617 0.643 0.613 0.595 0.622 0.643 0.211 0.578 0.626
Genbase 316 0.991 0.996 0.986 0.992 0.992 0.989 0.989 0.990 0.982 0.986
Medical 400 0.874 0.913 0.855 0.839 0.820 0.772 0.850 0.727 0.803 0.829
Average – 0.821 0.842 0.828 0.815 0.802 0.794 0.827 0.642 0.788 0.813
137
138
Table 7 Comparisons of CK-STC, feature selection FSVIG with MULAN binary relevance attribute evaluator w.r.t. hamming loss #
Dataset Hamming loss #
Number of attribute CK-STC FSVIG MULAN (multiclass attribute evaluator)
Enron 596 0.087 0.054 0.056 0.056 0.060 0.055 0.056 0.070 0.061 0.060
Genbase 316 0.003 0.002 0.004 0.001 0.001 0.004 0.004 0.001 0.038 0.002
Medical 400 0.013 0.012 0.012 0.010 0.010 0.011 0.011 0.011 0.013 0.013
Average – .034 0.022 0.024 0.022 0.023 0.024 0.024 0.027 0.037 0.025
S. S. Sane et al.
Table 8 Comparisons of CK-STC, feature selection FSVIG with MULAN multiclass attribute evaluator w.r.t one-error #
Dataset One-error #
Number of attribute CK-STC FSVIG MULAN (multiclass attribute evaluator)
Enron 596 0.303 0.225 0.366 0.442 0.473 0.383 0.367 0.954 0.565 0.431
Genbase 316 0.008 0.007 0.013 0.003 0.003 0.010 0.006 0.004 0.016 0.013
Medical 400 0.158 0.146 0.117 0.162 0.167 0.184 0.185 0.215 0.233 0.217
Average – 0.156 0.126 0.185 0.202 0.214 0.192 0.186 0.391 0.271 0.220
139
140
Table 9 Comparisons of CK-STC, feature selection FSVIG with MULAN multiclass attribute evaluator w.r.t average precision "
Dataset Number of attribute CK-STC FSVIG MULAN (multiclass attribute evaluator)
Enron 596 0.597 0.617 0.587 0.568 0.549 0.515 0.590 0.108 0.555 0.569
Genbase 316 0.991 0.996 0.986 0.992 0.992 0.989 0.989 0.991 0.982 0.986
Medical 400 0.874 0.913 0.862 0.858 0.838 0.818 0.855 0.768 0.825 0.829
Average – 0.821 0.842 0.812 0.806 0.793 0.774 0.812 0.622 0.787 0.795
S. S. Sane et al.
Tables 6, 7, 8, and 9 indicate that feature selection technique FSVIG with

CK-STC gives better performance in terms of one-error and average precision than
feature selection algorithms reported in MULAN. Multilabel classification with
FSVIG feature selection does not provide best performance in terms of hamming
loss because of weak connection between labels in the datasets. Feature selection
algorithms in MULAN library transform multilabel problem into single-label
problem and on single-label dataset, feature selection methods are applied; so due to
transformation there is damage to original label structures which reduces the per-
formance of classifier.
For evaluating the statistical significance of the results, the Friedman test for
paired data is used. The level of significance of the Friedman test was determined at
a = 0.05.
6 Conclusions
A lazy learning approach CK-STC has been reported in the literature for multilabel
classification using coupled similarity between labels that improves the prediction
performance, but irrelevant features present in database affect the prediction
accuracy of classifier.
A multilabel feature selection FSVIG is incorporated in CK-STC algorithm.
FSVIG uses information gain that shows better performance than existing multil-
abel feature selection algorithms when used with ML-NB, ML-kNN, and RandSvm
classifier. This paper investigates a performance of FSVIG when used with
CK-STC and compares its performance with other multilabel feature selection
algorithms available in MULAN using standard multilabel datasets.
Experimental results show that FSVIG when used with CK-STC provides better
performance in terms of average precision and one-error.
References
1. Chunming Liu and Longbing Cao, “A Coupled k-Nearest Neighbor Algorithm for Multi-label
Classification”, Springer PAKDD, Part I, LNAI 9077, pp. 176–187, (2015).
2. Ling Li, Huawen Liu, Zongjie Ma, Yuchang Mo, Zhengjie Duan, Jiaqing Zhou, Jianmin
Zhao, “Multi-label Feature Selection 6a Information Gain”, 345–355, (2014).
3. Tsoumakas, G., Katakis, I. “Multi-label classification: An over6ew” International Journal of
Data Warehousing and Mining (IJDWM) 3(3), 1–13 (2007).
4. Gjorgji Madjarov, DragiKocev, Dejan Gjorgje6kj, Saso Dzeroski, “An extensive experimen-
tal comparison of methods for multi-label learning”, Pattern Recognition 45 (2012) 3084–
3104.
5. Grigorios Tsoumakas, Ioannis Katakis, and Ioannis Vlahavas, “Mining Multi-label Data”,
Data Mining and Knowledge Discovery Handbook, Springer, 2010, pp. 667–685.
6. Boutell, M.R., Luo, J., Shen, X., Brown, C.M. “Learning multi-label scene classification”,
Pattern recognition 37(9), 1757–1771 (2004).
7. Read, J., Pfahringer, B., Holmes, G., Frank, E.: Classifier chains for multi-label classification.
Machine Learning 85(3), 333–359 (2011).
8. Zhang, M.L., Zhou, Z.H., “ML-kNN: A lazy learning approach to multi-label learning”,
Pattern recognition 40(7), 2038–2048 pp. 401–406. Springer (2007).
9. Spyromitros, E., Tsoumakas, G., Vlahavas, I., “An empirical study of lazy multilabel
classification algorithms”, Artificial Intelligence: Theories, Models and Applications,
pp. 401–406. Springer (2008).
10. Cheng, W., HÂ´lullermeier, E., “Combining instance-based learning and logistic regression
for multi-label classification”, Machine Learning 76(2–3), 211–225 (2009).
11. Cao, L., “Coupling learning of complex interactions”, Information Processing and
Management (2014).
12. Lee, J., Kim, D.W. “Feature selection for multi-label classification using multivariate mutual
information”, Pattern Recognition Letters 34(3), 349 – 357 (2013).
13. Tidake Vaishali S. And Shirish S. Sane, “Multilabel Learning with MEKA”, CSI
communications (2016).
14. G. Tsoumakas, E. Spyromitros-Xioufis, J. 6lcek, and I. Vlahavas, “MULAN: A Java library
for multi-label learning” Machine Learning. Res., vol. 12, pp. 2411–2414, Jul.(2011).
15. Prajakta Chaudhari and S. S. Sane, “Multilabel Classification Exploiting Coupled Label
Similarity with Feature Selection”, IJCA ETC (2016).
High-Performance Pipelined FFT
Processor Based on Radix-22
for OFDM Applications
Manish Bansal and Sangeeta Nakhate
Abstract This paper introduces high-performance pipelined 256-point FFT pro-

cessor based on Radix-22 for OFDM communication systems. This method uses
Radix-2 butterfly structure and Radix-22 CFA algorithm. Radix-2 butterfly’s
complexity is very low and Radix-22 CFA algorithm reduces number of twiddle
factors compared to Radix-4 and Radix-2. The proposed design is implemented in
VHDL language, synthesized using XST of Xilinx ISE 14.1, and simulated using
ModelSim PE Student Edition 10.4a successfully. Also, MATLAB code has been
written and simulated with MATLAB R2012a tool. The computation speed of
proposed design is observed to be 129.214 MHz after the synthesis process and
SQNR is 50.95 dB.
Keywords FFT CFA Radix-22 Complex multiplier SDF

OFDM
1 Introduction
The FFT is the extensively used algorithms in digital signal communication sys-
tems. Nowadays, FFT is the prime building block for the communication systems
especially for the orthogonal frequency division multiplexing receiver systems.
OFDM is a multi-carrier technology for high-speed and high rate reliable data
transmission in the wireless and wire communication systems such as LAN and
wireless local area network (Fig. 1). WLANs are one of the most modern based on
IEEE 802.11 standards [1]. Fast computational schemes for the implementation of
FFT architectures have fascinated many researchers. Some of the approach to
M. Bansal (&) S. Nakhate

ECE Department, Maulana Azad National Institute of Technology,
Bhopal, Madhya Pradesh, India
e-mail: manishbansal2008@gmail.com
S. Nakhate
e-mail: sanmanit@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_15
144 M. Bansal and S. Nakhate
Time domain sum

Of Sinusoids
Transmitter Receiver
IFFT FFT
Fig. 1 OFDM system block diagram
design FFT is memory-based, pipeline-based, and general-purpose DSP. Memory

based is most area efficient but it needs many computation cycles. Pipeline-based
architecture possesses small chip area and high throughput rate with a lower
frequency.
All fast Fourier transform designs can be isolated basically into three types of
pipelined architectures—single-path delay commutator (SDC), multiple-path delay
commutator (MDC), and single-path delay feedback (SDF). Among these archi-
tectures, the SDF architecture is more convenient as—(1) The SDF structure is
more relevant for the fixed length FFT implementation and (2) SDF structure
requires less number of delay elements as compared to SDC and MDC
architectures [2].
This paper presents the design of pipelined single-path delay feedback FFT
processor based on Radix-22 CFA for 256 point which can be used for OFDM
applications. The Radix influences FFT architecture. A small Radix means simple
butterfly structure and increases number of twiddle factors. A higher Radix means
complex butterfly structure and reduces count of twiddle factors. In the proposed
FFT Radix-22 architecture, Radix-2 butterfly is used which is simplest butterfly
structure and Radix-22 uses less number of twiddle factors as Radix-4.
The controller of proposed design is simpler than other designs as overall
architecture and all components are controlled by count signal only.
This paper is structured as follows—Sect. 2 represents the proposed FFT pro-
cessor design based on Radix–22 Common Factor Algorithm (CFA). Section 3
represents data flow of proposed 256 point FFT processor. Section 4 represents the
comparison. Finally, Sect. 5 represents the conclusion.
2 Proposed FFT Processor Design Based on Radix-22 CFA
Proposed processor is designed using Radix-22 common factor algorithm and

Radix-2 butterfly structure which can be used for performing for 256-point FFT
computation. Proposed processor is designed in SDF (Single delay feedback)
pipeline architecture which is one the most efficient architectures and reduces
complexity of FFT processor. Each stage of proposed design consists of Radix-2
butterfly structure (BU) and shift register (SRE). In this section, Radix-22 common
High-Performance Pipelined FFT Processor Based … 145
factor algorithm, proposed FFT processor design, and required components are
described for 256-point.
The DFT is defined for sequence of N as [2]
X
N 1
j2pnk
X ðk Þ ¼ x ð nÞ e N ; ð1Þ
n¼0
j2pnk
where WNnk ¼ e N twiddle factor which represents the N-th root. n and k are time
and frequency indexes, respectively. The Radix-22 CFA for 256 point is formulated
by using three-dimensional linear index mapping and CFA and is expressed as
follows.
Divide and conquer 3-D linear index mapping is represented as
n ¼ 128n1 þ 64n2 þ n3 ð2Þ
k ¼ k1 þ 2k2 þ 4k3 ; ð3Þ
where
n1 ; n2 ¼ 0; 1 and n3 ¼ 0; 1; . . .; 63
k1 ; k2 ¼ 0; 1 and k3 ¼ 0; 1; . . .; 63
The common factor algorithm (CFA) forms [2]
63 X
X 1 X
1
X ðk1 þ 2k2 þ 4k3 Þ ¼ xð128n1 þ 64n2 þ n3 Þ WNnk
n3 ¼0 n2 ¼0 n1 ¼0
63 X
X 1 X
1
¼ xð128n1 þ 64n2 þ n3 Þ
n3 ¼0 n2 ¼0 n1 ¼0
ð128n1 þ 64n2 þ n3 Þðk1 þ 2k2 þ 4k3 Þ

WN ð4Þ
The twiddle factor can be expressed as
ð128n1 þ 64n2 þ n3 Þðk1 þ 2k2 þ 4k3 Þ

WNnk ¼WN
ð1Þn1 k1 ðjÞn2 k1 ð1Þn1 k2 W256
n ðk þ 2k Þ
3 1 2
W n 3 k3 ð5Þ
|fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl} |fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl} 64
BU PE BU PE
where n1 ; n2 ; n3 are the sampled terms of the input n and k1 ; k2 ; k3 are the sampled
terms of the output k. ð1Þn1 k1 and ð1Þn1 k2 are butterfly elements. ðjÞn2 k1 and
n3 ðk1 þ 2k2 Þ
W256 are processing elements.
The detailed structure of proposed Radix-22 FFT processor for 256 point is
shown in Fig. 2 which retains eight stages (S1, S2, S3, S4, S4, S5, S6, S7, and S8).
SRE indicates shift register which stores values. In case of 256 point, first shift
register stores 128 values, second shift register stores 64 values, third shift register
stores 32 values, fourth shift register stores 16 values, fifth shift register stores 8
values, sixth shift register stores 4 values, seventh shift register stores 2 values, and
last eighth shift register stores 1 value. Twiddle ROM has 189 number of twiddle
factors values which are calculated based on W256 . Swap ðjÞn2 k1 performs real sign
inversion and swaps signed inverted real and imaginary values. Butterfly
(BF) performs addition, subtraction, and by-pass operations. Complex multiplier
multiplies input complex values with twiddle factor values.
Figure 3 shows the butterfly (BU) structure which performs operation between
n-th value and ðn þ N=2Þth value. On first 128 cycles, the signal T of Mux1, 2, 3
and 4 switch to position “0” in butterfly unit. The input data is moved to the shift
registers until shift registers are filled. On next N/2 cycles, the signal T of Mux1, 2,
3 and 4 switch to position “1”, and then butterfly structure starts addition/
subtraction operation to compute 2-point DFT between the data stored in the SREs
and incoming data [3]. The subtraction results are stored in same register; first half
number of addition results are stored in next stage register and second half number
of addition results start addition/subtraction with stored first half number of addition
Controller (Clock) Twiddle ROM
Swap
128 64 32 16 8 4 2 1
SRE SRE SRE SRE SRE SRE SRE SRE
BU BU BU BU BU BU BU BU
I/P O/P
X X X
Complex Complex Complex
Multiplier Multiplier Multiplier
S1 S2 S3 S4 S5 S6 S7 S8
Fig. 2 Proposed FFT processor based on Radix-22 for 256-point
Fig. 3 Butterfly Mux1

(BU) structure Mux2
Mux3
Mux4
T
Fig. 4 Trivial (-j) multiplier Mux1

(swap) I/P Real
O/P Real
Mux2
I/P Imaginary
O/P Imaginary
2nd complement
S
in the same stage. Thus, this process runs from current stage to next stage till last
stage and butterfly element is connected in pipelined structure in order to compute a
result in every clock cycle.
In Fig. 4, trivial (−j) multiplier performs swapping of imaginary real value and
sign inversion of real value. This swapping between imaginary and real values is
controlled by multiplexer control signal “S” and sign inversion of real value is done
by second compliment. Whenever −j multiplication is required as per signal flow
graph, Mux1 and Mux2 multiplexors switch to position “1” and then first, sign of
real value is inverted and then inverted real value is swapped to imaginary and
imaginary value is swapped to real. In this way, sign inversion and swapping are
done instead of multiplication in case of −j multiplication. This swap approach
increases the performance and reduces the hardware logic.
3 Data Flow of Proposed 256-Point FFT Processor
Proposed FFT processor is designed for 256-point FFT. As shown in Fig. 2, all
components—BFs, stages, SREs, Twiddle ROM, and swap—are configured by
count control signal based on value of n1 ; n2 ; n3 ; k1 ; k2 ; k3 and all operations are
controlled by count signal. Here, proposed FFT is designed which takes 33% less
number of twiddle factors than conventional FFT processor [4].
Proposed FFT design uses eight stages, eight shift registers, and one twiddle
ROM. First shift register stores 128 values, second shift register stores 64 values,
third shift register stores for 32 values, fourth shift register stores 16 values, fifth
shift register stores 8 values, sixth shift register stores 4 values, seventh shift
register stores 2 values, and eighth shift register stores 1 value. Twiddle ROM has
189 number of twiddle factors value which are calculated based on W256 and
n3 ðk1 þ 2k2 Þ n ðk þ 2k Þ n ðk þ 2k Þ
generate W256 , W643 1 2 , and W163 1 2 twiddle factors value. Swap
n ðk þ 2k Þ
ðjÞn2 k1 in first stage and twiddle factors W256
3 1 2
operations in second stage are
performed as per 256 point samples. Similarly, Swap ðjÞn2 k1 in third stage and
n ðk þ 2k Þ
twiddle factors W643 1 2 operations in fourth stage are performed as per 64 points
n ðk1 þ 2k2 Þ
samples. Similarly, Swap ðjÞn2 k1 in fifth stage and twiddle factors W163
Table 1 Processing elements value after stages at each position for FFT processor based on
Radix-22 CFA
PE S1 S2 stage S3 S4 stage S5 S6 stage S7 stage
N stage n ðk1 þ 2k2 Þ
W N3 stage n ðk1 þ 2k2 Þ
W N3 stage n ðk1 þ 2k2 Þ
W N3 ðjÞn2 k1
ðjÞn2 k1 ðjÞn2 k1 ðjÞn2 k1
256 n1 ; n2 ; k2 ; k3 ¼ 0; 1 n1 ; n2 ; k2 ; k3 ¼ 0; 1 n1 ; n2 ; k2 ; k3 ¼ 0:1 n1 ; n2 ; k2 ; k3 ¼ 0; 1
n3 ; k3 ¼ 0; 1; n3 ; k3 ¼ 0; n3 ; k3 ¼ n3 ; k3 ¼ 0&N ¼ 4
2; . . .63&N ¼ 256 1; 2; . . .15&N ¼ 64 0; 1; 2; 3&N ¼ 16 and and repeating for
and repeating for repeating for every every next 4 points
every next 64 next 16 points
points
operations in sixth stage are performed as per 16 points samples. Similarly, swap
ðjÞn2 k1 in seventh stage is performed as per four points samples as mentioned in
Table 1.
In the 256-point FFT design, first 128 point is stored in stage 1 register using
butterfly structure in each clock and takes 128 clock. At 129th clock as x(128) point
inputted into stage 1, butterfly start addition and subtraction between x(0) and
x(128) point and this addition is stored in stage 2 register and subtraction is stored
in stage 1 register. At 130th clock as x(129) point inputted into stage 1, butterfly
again starts addition and subtraction between x(1) and x(129) point and this
addition is stored in stage 2 register and subtraction is stored in stage 1 register.
This process runs up to 192th clock. At 193th clock as x(192) point inputted into
stage 1, stage 1’s butterfly again starts addition and subtraction between x(64) and x
(192) point and this addition is stored in stage 2 register and subtraction is stored in
stage 1. At this clock, stage 2’s butterfly also starts addition and subtraction and
stores addition in stage 3 register and subtraction in stage 2 register. This process
continuously runs up to 256th clock. At 256th clock, as x(255) point inputted stage
1’s butterfly again starts addition and subtraction between x(127) and x(255) point
and this addition is stored in stage 2 register and subtraction is stored in stage 1 and
at this clock, buttery structure of stages 2, 3, 4, 5, 6, 7, and 8 also starts addition and
subtraction and stores addition in stages 3, 4, 5, 6, 7, and 8 registers and subtraction
in stages 2, 3, 4, 5, 6, and 7 registers. Stage 8 output is the final FFT design values.
The twiddle ROM in the proposed processor has 189 number of twiddle factors
values which are used at even stages based on values of n1 ; n2 ; n3 ; k1 ; k2 ; k3 at each
point and multiplied with incoming data signal as mentioned in Table 1. This
approach is reduced extra hardware as same twiddle factors values of twiddle ROM
k þ N=2
are used at even stages by using symmetric WN ¼ WNk and periodicity
kþN
WN ¼ WN property. In the proposed processor, twiddle factor WN0 value is
k
bypassed instead of multiplying wherever is required to multiply as WN0 value is 1.

These all approaches are not only reducing hardware but also increasing the speed
of the proposed processor (Fig. 5).
Fig. 5 Twiddle factors at even stages
4 Comparison
The architecture of proposed design is implemented in VHDL language, synthe-

sized with xc7vh290t-2hcg1155 device and simulated in ModelSim and MATLAB
to verify its results.
Table 2 shows proposed FFT design with conventional FFT design comparison.
The results show that proposed architecture design used 513 twiddle factors and 18
real multipliers, while conventional design use 768 twiddle factors and 16 complex
multiplier. 16 complex multiplier means 64(16*4) real multiplier. Thus, proposed
design uses less twiddle factors compared to conventional design which causes less
number of multiplier and less area.
Table 3 shows synthesized results of proposed design with conventional FFT
design [4] between max clock frequency and other parameters. Max clock fre-
quency of proposed design is 129.21 MHz which is higher speed than conventional
design.
Figure 6a–c represents comparison of max frequency, number of twiddle factors,
and number of multiplier, respectively, between proposed and conventional design.
It represents 70% high frequency, 33% less number of twiddles factor, and 71% less
number of multiplier of proposed design compared to conventional.
Figure 7 shows the register-transfer level (RTL) schematic view of proposed
FFT processor which is generated while synthesizing VHDL code using
xc7vh290t-2hcg1155 FPGA device by Xilinx 14.1 synthesis XST tool.
Table 2 FFT architectures parameters

Architecture Number Method Radix Number Number Number Memory
of point of stages of twiddle of real
factors multiplier
FFT [4] 256 Pipeline Radix-24 8 768 16*4 255
Proposed 256 Pipeline Radix-22 8 513 18 255
FFT
Table 3 Implementation parameters comparison

Parameters FFT [4] Proposed FFT
Number of point (N) 256 256
FPGA Kit Xilinx Virtex2 1500 xc7vh290t-2hcg1155
Number of slices used – 725
Number of slices Flip Flops used – 538
Number of four input LUTs used – 1345
Number of GCLK used – 6
Max frequency (in MHz) 35.76 129.21
SQNR – 50.95
(a) (b) (c)

150 1000 768 80 64
Twiddle Factors
Max Frequency
800 513 60
Number of
Number of
100
Multiplier
129.21
in MHz
600 40
35.76 18
50 400
20
200
0 0
0
Fig. 6 Proposed and conventional FFT design comparison
Fig. 7 RTL schematic view of proposed FFT processor
5 Conclusions
The purpose of this research is to implement FFT processor for 256 point based on
Radix-22 CFA for OFDM applications and use less number of twiddle factors than
conventional FFT which reduces the memory and area and reduces number of
multiplications. The less number of multiplications reduces processing time and
increases maximum frequency. This maximum frequency of proposed design is
higher which shows higher speed than conventional design.
References
1. Ahmed Saeed, M. Elbably, G. Abdelfadeel and M. I. Eladawy, “FPGA implementation of

Radix-22 Pipelined FFT Processor,” Proceedings WAV’09 Proceedings of the 3rd WSEAS
international symposium on Wavelets theory and applications in applied mathematics, signal
processing & modern science, 2009, pp. 109–114.
2. Jiang Wang and Leif Arne Ronningen, “An Implementation of Pipelined Radix-4 FFT
Architecture on FPGAs,” Journal of Clean Energy Technologies, Vol. 2, No. 1, January 2014,
pp. 101–103.
3. Wei-Hsin Chang and Truong Nguyen, “An OFDM-specified lossless FFT architecture,” IEEE
Transactions on Circuits and Systems I, 2006, pp. 1235–1243.
4. Chih-Peng Fan1, Mau-Shih Lee and Guo-An Su, “Efficient low multiplier cost 256-point FFT
design with Radix-24 SDF architecture,” Journal of Engineering, National Chung Hsing
university, Vol. 19, No.2, 2008, pp. 61–74.
An Image Processing Approach to Blood
Spatter Source Reconstruction
Abhijit Shinde, Ashish Shinde and Deepali Sale
Abstract Blood spatter analysis is a part of Criminal Justice system and subpart of
the forensic science. Traditional blood spatter analysis has a problem with the crime
scene contamination. The crime scene contamination can led to the unacceptance of
the evidences. The blood spatter analysis is the process which heavily relies on the
expertise of the forensic scientist. The human intervention also creates the problem
of errors and misjudgments. Use of image processing to the whole will deal with
automation of removing human factor. The proposed method takes the image from
blood spatter using image processing and reconstructs the source of the blood. The
proposed methodology uses the Otsu’s method for thresholding and Hough trans-
form for edge detection.

Keywords Blood spatter trajectory Image processing Trajectory analysis

Blood spatter analysis Forensic image processing
1 Introduction
Forensic science is an integral part of the justice system because it helps in col-
lecting the evidences on a scientific basis. The blood pattern analysis or Blood
spatter analysis is one of the tools used for reconstruction of the crime scene.
A. Shinde (&)
Bhima Institute of Management and Technology, Kagal, Kolhapur 416202, India
e-mail: abhijitvshinde@gmail.com
A. Shinde
Sinhgad College of Engineering, Pune-41, Maharashtra, India
e-mail: shindeashishv@gmail.com
D. Sale
Dr. D. Y. Patil Institute of Engineering Technology, Pimrpi, Pune- 18, India
e-mail: deepalisale@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_16
154 A. Shinde et al.
The BPA traditionally has predominant skill of the Forensic Scientist. The National
Academy of Sciences, USA has emphasized the need of more standardization in the
field and the better training for the analysts, so that the reliability of BPA cannot be
questionable in the courts [1, 2].
The blood spatter originates in the violent crime scenes. The source of blood is
an exit wound of blood. The blood which impacts on the wall leaves the stain
pattern which is blood spatter. The blood droplets will be circular if the impact
angle is 90°, otherwise they will be elliptical. The shape of the blood droplets
depends upon the gravity, the speed, blood viscosity, air drag, and surface tension
of the blood drop [1, 2].
The blood spatter analysis has been neglected by the researchers of the image
processing and only few research is available which was highlighted by Abhijit and
Sale [3].
Abhijit and Sale [4] has explained all steps needed to be carried out in blood
spatter analysis but did not used any dataset as a standard dataset that is absent. So
the dataset was generated to test the output of the proposed methodology. Giovanni
Acampora et al. stressed the need for the automation of blood spatter analysis
process. Authors had developed the cognitive robot for the purpose of the blood
spatter analysis but the research work fails to propose the methodology blood
spatter analysis [5].
The basis blood spatter scenario is explained in Fig. 1; projections l1 and l2 are
extensions of major axes of blood droplet, respectively. Projections l1 and l2 meet at
point P1 which is on y-z plane. Point P1 is the projection of the blood source on yz
plane. Angles h1 and h2 are the angles between sources l1 and l2, respectively. The
traditional blood spatter analysis calculates these angles by taking tan inverse of
major and minor axes of blood drop [1].
Fig. 1 3D projection
representation of blood spatter
analysis
An Image Processing Approach to Blood Spatter … 155
2 Methodology
2.1 Orthorectification and Interpolation
The plane of the camera and the plane of the surface of the blood spatter may not be
parallel to each other which results in the error in determining the dimensions of the
ellipses. To bring the images in the same plane, the orthorectification is performed,
using reference point [6]. The blue dots are placed in four corners; these corners
form the square. As the planes of the camera and surface are not parallel, the dots on
corner form the trapezoid which is rectified, as shown in Fig. 2.
2.2 Blood Droplet Segmentation
The next step in the process will be blood droplet segmentation. This step finds out
the belonging of the pixel to its corresponding blood droplet. This step performs not
only the segmentation but also the background removal.
The method which will be suitable for the above-mentioned condition is adap-
tive Otsu’s Method. This method will not only segment the blood droplets but also
make the background black. The one more advantage of using Otsu’s method is that
it eliminates the small droplet and the large drops in the process of segmentation.
The large drops do not take the shape of the ellipses but takes the shape of the
circle, whereas smaller droplets cannot be used for the mathematical processing of
finding major and minor axes. So if we considered that the spatter creates the
Gaussian distribution, then after applying area criteria we use blood droplets which
are segmented and their distribution will be in very narrow range centered around
the mean of all segmented areas [7]. This has been illustrated in Fig. 3.
Fig. 2 The blood spatter

image before and after the
orthorectification
Fig. 3 Blood droplet

segmentation
2.3 Ellipse Detection
The Hough transform is used to detect the ellipse. The ellipse detection and fitting
help us to find only the point which satisfies the equation of the ellipse, whereas we
need to develop the algorithm to find out the needed parameters, which are lengths
of major and minor axes, and the orientation of the ellipse [8].
We need two pints to find out the centroid of the ellipse and the length of the
major axis. It can be given by Eq. 1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
a ¼ ð x2 x1 Þ 2 þ ð y2 y1 Þ 2 ð1Þ

y2 y1
a ¼ tan ð2Þ
x2 x1
Here, a is the length of the major axis and the a is the orientation of the ellipse
(McLaughlin [8]). We need the second point to find out the length of the minor axis.
It can be given by equations
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
a2 d 2 sin2 s
b¼2 ð3Þ
a2 d 2 cos2 s
a2 þ d 2 f 2
coss ¼ ð4Þ
2ad
Here, b is the length of the minor axis (Yuen et al. [9]). Other parameters can be
observed in Fig. 4.
Fig. 4 Ellipse geometry, f1

and f2 are the ellipses foci [10]
2.4 Major and Minor Axes Analysis of Individual Ellipses
The major axis and minor axis will help us to draw the projection in 3D. Here, we
have to find out the projection for the y-z plane and x-z plane separately. The y-
z plane is the plane of the surface which has blood spatter; the x dimension is the
dimension coming out of the y-z plane. For the y-z plane, we need the orientation of
each of the major axes, whereas for the x dimension projection, we need the lengths
of major and minor axes. Figure 5 shows the major and minor axes detection along
with ellipses.
2.4.1 The y-z Plane Projection
As we discussed in the literature survey, the major axis is useful for the determining
the direction of the travel of the blood droplet. If we extend the major axis of each of
the blood droplets, then they will meet each other. Theoretically, projections are
supposed to meet at single point but practically they do not. So we need to make
some approximation. So we put forward the concept of the profile line. We draw the
mean line with mean orientations of all projections, and draw a perpendicular line to
it, known as profile line. We will move the profile line in x-y direction. The profile
line will measure the cumulative distances of each line with other. Wherever the
cumulative distance is minimum, we will take that point as the common intersection
point. It can be expressed in Eq. 5 and can be represented as shown in Fig. 6a, b [2].
" #!
1 X
N 1
min ðy0 ynÞ 0 ð5Þ
N 1 n¼1
The projection then made with a single point as the intersecting point. This is
shown in Fig. 6c. This is the projection of the source on the x-z plane, i.e., on the
surface of the blood spatter.
Fig. 5 Major and minor axes tracking

Fig. 6 a Plotting of profile line. b Schematic showing profile plane. c The common point of
conversions
2.4.2 Projection in the x Dimension
Once we found out the 2D projection, we need to convert it to 3D. To get the
position in 3D, we need to draw projection from the common intersecting point and
a projection making an angle with major axis. The shape of the blood droplet
changes with impact angle as shown in Fig. 7.
We can see at 90° blood drop is perfectly 90°, whereas with an angle other than
90° it forms ellipse and the major axis length increases with increase in angle.
The traditional method calculates the angle by Eq. 6:
b
h ¼ sin1 ð6Þ
a
We have considered that the angle will be proportional to the ratio of the major
and minor axes:
b
h/ ð7Þ
a
We found out the ratio of a=b for the angles from 10° to 70° with interval of 5°
and by training the program we found out the ratios for intermediate angles.
2.4.3 Plotting of the Projections
Figure 8 shows the plotting of the paths in y-z plane and x-z plane. The coordinates
got from this are used to generate the 3D projection as shown in figure.
Fig. 7 Angle of impact affection of the shape of the blood drops

Fig. 8 3D projection
2.5 Gravity Correction
The gravity affecting the blood droplet makes its path parabolic. The above
assumed steps have considered the flight of the blood as the straight line which is
needed to be corrected. This flight under the influence of gravity make blood drop
to strike the wall at a position lower than it would had been under no gravity
condition. This height adjustment has to be done. This height is added in z direction
as the z direction represents the height. The processes have been illustrated using
Fig. 9, and the output after correction is shown in Fig. 10.
The calculation for the gravity correction can be given by Eq. 7 [4]:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
u( 2 )
u 1 Y
ZA ¼ ZD þ t z 9:8 Y2 ð7Þ
2 3:2
3 The Dataset
The dataset was generated using the method described in research apparatus
(Acampora et al. [5]). The dataset consists of 35 samples of blood spatter describing
various scenarios reconstructed. Figure 11a–c represents groups which contain
perfectly shaped ellipses; Fig. 11d, e shows very few uneven and non-elliptical
shapes, while 11f very high degree of inconsistency in the formation of elliptical
spatter.
Fig. 9 The schematic of gravity correction
Fig. 10 3D projection with gravity correction
Fig. 11 a, b and c Blood spatter with perfectly shaped ellipses. d and e Blood spatter with one
blood drop having uneven shape. f Messy blood spatter resembling actual crime scene blood
spatter
Table 1 Error estimation

Sr Image Actual Estimated Error in Actual Estimated Error
no. distance distance distance angle angle (%)
(mm) (mm) estimation (%)
1 Figure 11a 1300 1343.7814 3.367 30° 36.2349° 20.783
2 Figure 11b 1300 1235.7948 −4.9389 40° 35.473° −11.31
3 Figure 11c 1300 1317.6864 1.36049 30° 29.5632° −1.45
4 Figure 11d 1300 1374.6556 5.74274 35° 40.6501° 16.14
5 Figure 11e 1300 1241.3125 −4.5144 35° 34.6944° −0.87
6 Figure 11f 1300 1436.5138 10.5011 50° 55.0622° 10.12
Mean error in distance 5.2 Mean error in angle 10.2
estimation (%) estimation (%)
4 Result
Table 1 shows estimations of errors for the mean distance from the source and the
blood droplet and angles of projection. We have considered the constant distance of
1300 mm, varying the values of angle for the estimation of error.
Average estimation error for the angle of projection for various samples
approximates at 10.2%. While in a case it has reached to 20%, the error is a certain
degree of prediction which is used to compensate mathematical errors. The average
estimation error in distance calculation is 5.02% only and has highest error per-
centage of 10.5%.
5 Conclusion and Future Scope
Results of proposed methodology have replicated the manual process of blood

spatter analysis. The result also reflects the less error margin, whereas the biggest
success of the proposed methodology is that it has eliminated the human factor from
the process which will help in avoiding crime scene contamination and human
error. This will help in making strong evident in criminal justice. The 3D modeling
proposed will also help in increase in reliability.
The absence of standard database has limited result analysis. In a further step, the
development of the methodology will be in the generation of extensive dataset,
checking results, an extension of a model for spatter stains on multiple planes, and
generation of parabolic flight trajectory.
Disclaimers The fake blood used for experimentation was made up of non-animal food products
such as corn syrup, sugar, coco-powder, and artificial food coloring. The resulting product is
usually used for forensic experiments. No human being or animal was harmed for the experiment.
The procedure for generating blood spatter is similar to one followed by many other authors [1, 2].
References
1. Anita Wonder BS, “Blood Dynamics”, Anita Y. Wonder, Academic Press, 2001 ISBN-10:
0127624570, 1st edition August 2001, pp. 36–51.
2. P. Joris, W. Develter, E. Jenar, D. Vandermeulen, W. Coudyzer, J. Wuestenbergs, B. De
Dobbelaer, W. Van de Voorde, E. Geusens, P. Claes, “A novel approach to automated
bloodstain pattern analysis using an active bloodstain shape model”, Journal of Forensic
radiology and image, Volume 2, Issue 2, April 2014 pp. 95.
3. Abhijit Shinde, Dr. Deepali Sale, “Blood Spatter Trajectory Analysis for Spatter Source
Recognition Using Image Processing”, 2016 IEEE Conference on Advances in Signal
Processing (CASP), Cummins College of Engineering for Women, Pune, pg. 375–380, Jun 9–
11, 2016, DOI: 10.1109/CASP.2016.774619934-39.
4. Abhijit Shinde, Dr. Deepali Sale, “Blood Spatter Source Reconstruction Using Image
Processing”, IETE National Journal of Innovation and Research (NJIR), ISSN 2320–8961,
Volume III, Issue II, December 2015.
5. Giovanni Acampora, Autilia Vitiello, Ciro Di Nunzio, Maurizio Saliva, Luciano Garofano,
“Towards Automatic Bloodstain Pattern Analysis through Cognitive Robots”, conference—
IEEE International Conference on Systems, Man, and Cybernetics, 2015.
6. Wen-Yuan Chen, Yan-Chen Kuo, Chen-Chung Liu and Yi-Tsai Chou, “Region-Based
Segmentation Algorithm for Color Image Using Generated Four-Connectivity and Neighbor
Pixels Comparing Criterion”, Proceedings of the 2005 Workshop on Consumer Electronics
and Signal Processing, 2005.
7. G. Gordon, T. Darrell, M. Harville, J. Woodfill, “Background estimation and removal based
on range and color”, Proceedings of the IEEE Computer Society Conference on Computer
Vision and Pattern Recognition, (Fort Collins, CO), June 1999.
8. Robert A. McLaughlin, Technical Report, “Randomized Hough Transform: Improved Ellipse
Detection with Comparison”, University of Western Australia, 1997.
9. H. K. Yuen, J. Illingworth, J. Kittler, “Detecting partially occluded ellipses using the Hough
transform”, Image and Vision Computing—4th Alvey Vision Meeting archive, Volume 7
Issue 1, February 1989, pp. 31–37.
10. Yonghong Xie; Dept. of Comput. Sci., Nevada Univ., Reno, NV, USA; Qiang Ji, “A new
efficient ellipse detection method”, Pattern Recognition, 2002. Proceedings 16th International
Conference on (Volume: 2), 2002, pp. 957–960.
Overlapping Character Recognition
for Handwritten Text Using Discriminant
Hidden Semi-Markov Model
Abstract The field of handwritten character recognition has always attracted a

large number of researchers. The proposed methodology uses Discriminant Hidden
Semi-Markov Model for tackling the problem of recognition of handwritten char-
acters. Preprocessing on the input image such as denoising and adaptive thresh-
olding is done for input conditioning, followed by segmentation for finding the area
which contains text. The text image is then passed through the second stage of
segmentation, which separates overlapping characters. Then, these segmented
characters are digitized using feature extraction. For feature extraction,
Discriminant Hidden Semi-Markov Model is used. For feature matching and
character extraction, the proposed methodology uses KNN Classifier. The training
feature library, consisting 180 samples of each character in the capital and small,
processed using training algorithm of Discriminant HsMM. Paragraphs of 80–120
characters are processed in recognition module. The 86% average accuracy rate is
achieved for a large set of characters.
Keywords Overlapping character segmentation Word recognition

Preprocessing techniques Handwritten text
1 Introduction
The pathway to higher end digital handheld devices like mobile, PDAs, etc. Hassle
free is one of the most researched areas of interest in digital devices nowadays. One
of such areas is input through handwritten characters or phrases. The features of
A. Shinde (&)
Sinhgad College of Engineering, Pune-41, Maharashtra, India
A. Shinde
Bhima Institute of Management and Technology, Kagal,
Kolhapur, Maharashtra 416216, India

https://doi.org/10.1007/978-981-10-7245-1_17
164 A. Shinde and A. Shinde
handwritten character make these tasks comparatively difficult due to high variance
in any person’s handwriting. The level of difficulty further increases when the
phrases or sentences are written without giving much importance to the separate-
ness of characters which causes them to overlap. A solution is to separate the
characters and recognize them separately. Hence, this research proposes a
person-dependent handwritten character recognition using Discriminant Hidden
Semi-Markov Model.
HMM was proposed by Stratonovich [1] in the field of mathematical modeling of
speech recognition. Later it was applied in DNA sequencing, character recognition,
and many other applications. The use of HMM in character recognition has been
really popular and successful hence. The switch to Discriminant Hidden
Semi-Markov model from HMM happened very recently due to the shortcomings of
HMM. The most prominent of them is hidden states of model. The important feature
of HsMM over HMM is the visibility of intermediate states. Figure 1 shows general
structure and observation scenario in HsMM. HMM is called hidden as the pro-
cessing states of the model for each input or event are hidden and hence are difficult
to understand and improve for a better result. HsMM overcomes this obstacle by
generating intermediate outputs which are observable and can also be processed on.
These observations are complex to understand directly and usually need processing
before the presentation. This implies that, unlike in HMM where a single state
produces one observation, a state in HsMM produces a sequence of characters.
In most cases of handwritten text, there is unreliability of recognition models due
to the incomplete, ambiguous, or imprecise contents of the handwritten text images.
So, performing handwritten text recognition aims toward capturing the organization
of the complexity and it has to deal with a high degree of tolerance toward
uncertainty. To summarize it is quite a challenging task to design a simplified
algorithm which may be optimally cost effective.
The appropriate feature extraction algorithm selection is the most important step,
because it decides the degree of recognition performance [2].
HMM is utilized for feature recognition of the words and the optimal paths
algorithm for classification. When tested on a group of 98 words, this system
achieved 72.3% success [3]. Another application of HMM in OCR was presented
by Huang et al. [4] with character segmentation-based word recognition approach
Fig. 1 General structure and

observation scenario in
HsMM
Overlapping Character Recognition for Handwritten Text … 165
(2001), who used CDVDHMM to build character models. This technique suffers
from slow speed to train and in operation.
Kamble et al. (2014) for the purpose of extraction of features have used the
“Rectangle Histogram Oriented Gradient representation” [5]. The recognition rate
for printed characters for this method nears at 95%.
The work by Deepjoy Das et al. has used SVM and HMM for the “Online
character recognition system for Assamese language”. The authors further per-
formed a recognition performance analysis for both the models [6]. After training,
181 different Assamese Stokes Recognition models were generated.
In a study, Yarman-Vural and Arica [7] have overviewed the present status of
CR research. The recognition results of different algorithms appear successful. Most
of the recognition accuracy rates reported are over 85%.
Most of the above algorithms work specifically for printed characters with less
variations. Consecutively, the problem of touching and overlapping characters are
not handled by many researchers and therein lies their shortcomings. We propose
an innovative perspective and solutions for this problem. The preprocessing
addresses the problem of touching characters, while the HsMM algorithm’s efficient
training module solves the problem of variation in handwriting pattern of a person
and provides a higher recognition rate.
2 Methodology
The proposed approach to recognize handwritten characters adopts a Discriminant

Hidden Semi-Markov field model for overlapping character segmentation and
recognition [6]. In the proposed problem, segmentations are modelled as the
duration of output or intermediates output state. The proposed methodology finds
segmentation and corresponding labeling for each of initial segments of image most
likely to be successful in separating characters.
2.1 Preprocessing
2.1.1 Thresholding
The proposed adaptive binarization methodology relies on the calculation of local

mean of block segments of the image. The technique is based calculation of local
threshold using formula

Tsub;P ¼ lsub;P 1 þ a * logNb;P ð1Þ
2.1.2 Skew Correction
The algorithm here uses entropy scoring measurement for the image. A simple way
to do it is, first decide constraints for the maximum skew angle. Measure least
entropy score for each angle change in previously decided step-size using formula.
A sum of entropies is calculated for each step-size. The smallest of entropy sum
is for least skewed text image.
2.1.3 Character Segmentation
Projection profiles and reverse water reservoir have been used for finding over-
lapping area of characters. Projection profiles are one-dimensional arrays used to
store the count of pixel each with value “1” in binary images which belong to each
character’s area when normal X–Y axis profiles are created. Each value in the
projection array is generated with the number of pixels above a predefined
threshold. The reverse water reservoirs are created by considering characters upside
down and imagining raining over them. The sections which look like lakes will
most probably hold water in them and hence these areas are called water reservoirs.
If the characters are isolated then the process is aborted, otherwise, for connected
characters the overlapping space is determined. The cut-set are selected based on
both the algorithms, where a bottom of the reverse reservoir and initial cut-sets are
used in the following algorithm.
Step 1. A segmented line sub-image is input

Step 2. A vertical Projection of Image is calculated
--------- (2)
Step 3. Binarization of vertical projections
Step 4. A profile of lowermost edge of whole image is generated
--------- (3)
Step 5. Using a reverse gravity water reservoir to find overlap
area and its local reverse bottom is chosen as area of
segmentation
Step 6. The point at which binarized ‘1’ of vertical projection
and reverse bottom is chosen as cut-point
Step 7. Search nearest –k and +k reverse bottom which is higher
than selected local reverse bottom as cut-point
a. If local reverse bottom with higher than selected
bottom is not found, then cut-point remains same
b. If local reverse bottom with higher than selected
bottom is found, then cut-point is made larger to cover
initial and final local bottoms and all pixels in
between.
Step 8. Pixel value of column near and at cut-point and between
later found bottom are set to background color
2.2 Character Feature Recognition Using HsMM
For recognition of each character, a given segmentation and corresponding labeling

is generated using similarity functions. The similarity function captures the
appearance of a character by means of gradient features. Additional functions
describing character properties such as bounding boxes overlap between characters
and gaps between characters are also used (Fig. 2).
A segmentation in this model will induce a set of unknowns y and a corre-
sponding setof similarity functions{CCR}C c. Each unknown yi Y(T) generates a
function Y CG, CSAr;t; , CSB , CGG
n;r . The sample in Fig. 3 shows one word of a
sentence. In that word, there exist five separated regions with the probability of
characters present which are given labels. Notice that character segment regions, y4,
which show the “blank space” character. Considering “blank spaces” between
words, a special character which will be recognized which will enable us to dif-
ferentiate between two sequential words and arrange the words in a defined index.
As shown in the figure, there are multiple overlaps (as for y1 and y2) and multiple
gaps (as for y2 and y3).
Fig. 2 Water reservoirs forming in a word with boundaries. Initial cut-points and final cut-points
of segmentation
Fig. 3 Example of input, reservoir formation and output of proposed projection profiles and
reverse water reservoir technique for character segmentation
Fig. 4 Segments of a word with 5 characters into 5 character segment regions (annotated by y1 to
y5), which can be segmented for recognition
( )
X
CGðyj x, CC Þa exp CSC ðyc , x; CC Þ ð4Þ
C2c
where x is a segment of the image being conditioned upon and C a feature function
for character learned with training. The notation yC will indicate sub-function of
character function CSC (Fig. 4).
2.3 Feature Function Extraction
Here, it is proposed to use five sub-functions for calculating the discriminant feature
of the character. Which are Appearance of Character, Segment Block, the overlap
between character segments, the gap between character segments, and character
gradients.
2.3.1 Appearance of Character
Each character segment, given by CS, is recognized using pretrained Feature

Function CC,
CSAr;t ðy; x; CC Þ ¼ CCr;t

A
ð yÞ Fr;t ð xÞ ð5Þ
for a character segment given by ‘y’ and the segment width from r to t.
Thus, the model will be trained that “w”s are usually wide while “i”s are narrow.
The features Fr,t(x) of each character image are calculated from r to t, which are
overlapping consecutive character spaces but the flexible design of the recognition
model enables researchers to use overlapping segments of the sub-image even if

they are extending the character segment.
2.3.2 Character Gradients
Feature function energies can be easily trained in the form of gradients. Each
neighboring series of character gives us a bigram score in form of overlapping or
nonoverlapping segment energy, CSB ðy0 ; x; CC Þ ¼ CC B ðy0 ; yÞ.
2.3.3 Overlap in Characters
A pair of neighboring segments may have an overlap or character gap between

them. In such case of neighboring character having overlap, an segment flexibility
o
is added, Un;r ðCC O Þ ¼ CCn;r
O
, of which’s value depends upon number of pixels in
overlapping region annotated by characters by n & n + 1. Thus, the resulting
bounding boxes may have an overlap to allow a soft boundary selection.
2.3.4 Gap Between Characters
As declared above, neighboring characters can also have a small gap (desirable for
seamless segmentation) between them. The gap is scored by a learning function,
G Pr
CGGn;r x;h ¼ i¼n CC G Fi ð xÞ which a count of number of horizontal pixel
between consecutive characters.
2.3.5 Sparse-Based Parsing
After creating a sub-image from the segmentation to have a best coverage for a
character, the optimal output of comparison function SLS(n, t, n + 1, t + 1) should
tend to infinity or be as large valued as possible, where the character t + 1 is fixed
and the segmentation boundaries n and n + 1 are varied. However, the function
only creates a competitive test between closely matching characters in possible
recognition scenarios. For example a given segment, the term SLS(n, t, n + 1,
k) may represent character “I” (capital I) which matches to feature functions of “i”
(small i), “t”, “l” (small L) or “1” (one).
CSðn þ 1, t þ 1jn, t, x, CC, I Þa expfCðn, t, n þ 1, t þ 1Þ/n t þ 1: ð6Þ
KNN used here as classifier makes the decision of selection of class based on the
entire training dataset or all subsets of them. The output class in our case is a
character. Training has been done with 180 samples of each character which makes
9360 sample dataset with 52 labels repeating 180 times each.
3 Experiment and Analysis
180 samples have been used for each character for the training of both small and
capital sets. The recognition was done on 15 different samples of each containing
80–120 samples of characters more than 50% of which were overlapped with one
other character and 20% overlapped with two other characters (both left and right
neighbor).
The segmentation of lines overlapping characters gives 96% average accuracy in
perfect segmentation while 3% shows undersegmentation and 1% shows overseg-
mentation which leaves very small segments which do not affect recognition rate as
they are not used during recognition (Fig. 5).
Fig. 5 Stage-wise results of methodology. a Initial overlapping characters. b Initial segmentation.

c Application of modified water reservoir technique for separation of overlapping characters.
d Segments of separated characters. e Skeletalisation of segmentation result. f Recognized text
Table 1 Change in recognition rate with increase in the number of training samples
Number of training sample/character Recognition rate for a sample of 84
characters
True positive (%) False positive (%)
5 48 48
12 51 45
48 63 33
60 71 28
120 92 6
180 96 1
The recognition rate for a sample containing 84 characters is used to extract values
described in Table 1. When classified with training dataset created using 5 training
samples for each character is recognition rate of 48% was achieved. Similarly, when
classified with training dataset created using 12 training samples for each character
is recognition rate achieved was 51%. After using samples of 180 different slant
varying from −5° to +15° for each character for training the recognition rate
reached to 98% true positive and 1% false positive for the same input sample.
The attained average recognition rate is 86% aggregated for all samples. All of
the samples contain 50% or more characters, which are overlapping and are sep-
arated using the proposed algorithm with almost perfect segmentation with very
small chance of over-segmentation.
5 Conclusion
The HsMM can be implemented efficiently for purpose of handwritten and over-
lapping or touching character recognition. The recognition rate depends upon
multiple things such as user’s handwriting consistency especially in text slants and
area of overlapping characters. It also depends upon no of training samples and
efficiency of preprocesses. The model can now be implemented in anybody’s
handwriting if samples of their handwriting are available in abundance for learn-
ingthe pattern. This model hence has a high applicability for handwritten CR in any
language but also has a requirement such as preprocessing and high training time.
References
1. Stratonovich, R. L. (1960). “Conditional Markov Processes”. Theory of Probability and its

Applications. 5 (2): 156–178.
2. Dasu, Nagendra Abhinav, “Implementation of hidden semi-Markov models” UNLV

Professional Papers. Paper 997, 2011.
3. Chen, X., & Yuille, A. L., “Detecting and reading text in natural scenes”, Proc. IEEE Conf. on
Comp. Vis. and Pat. Recog. 2004, pp. 366–373.
4. Huang, X., Hon, H. W., & Acero, A., “Spoken language processing: A guide to theory,
algorithm, and system development”, Prentice Hall PTR., 2001.
5. Ravinda S. Hegadib, Parshuram M. Kamblea, “Handwritten Marathi character recognition
using R-HOG Feature”, Procedia Computer Science 45, ICACTA, 266–274, 2015,.
6. Deepjoy Das, SRM Prasanna, Rituparna Devi, Subhankar Ghosh, Krishna Naik, “Performance
Comparison Of Online Handwriting Recognition System For Assamese Language Based
On HMM And SVM Modelling”, CSCP 2014, pp. 87–95, 2014.
7. Fatos T. Yarman-Vural, Nafiz Arica, “An Overview of Character Recognition Focused on
Off-Line Handwriting”, IEEE Transactions on Systems, Man and Cybernetics—Part C:
Applications and Reviews, Vol. 31, No. 2, 2001.
A Literature Survey on Authentication
Using Behavioural Biometric Techniques
Rajvardhan Oak
Abstract With technological advancements and the increasing use of computers

and internet in our day to day lives, the issue of security has become paramount.
The rate of cybercrime has increased tremendously in the internet era. Out of the
numerous crimes, identity theft is perhaps the one that poses the most dangers to an
individual. More and more voices strongly declare that the password is no longer a
reliable IT security measure and must be replaced by more efficient systems for
protecting the computer contents. Behavioural biometrics is an emerging technol-
ogy that resolves some of the major flaws of the previous scheme. This paper is the
first stage of a project which aims to develop a novel authentication system using
behavioural biometrics. It presents a comprehensive survey of various techniques
and recent works in the respective fields.
Keywords Information security Biometrics Authentication mechanism
1 Introduction
Biometrics can be characterized basically as the estimation of human qualities.

Biometric identifiers are then unmistakable, quantifiable qualities used to name and
depict people [1]. Famous cases of biometric confirmation are retina examinations,
fingerprint tests and DNA tests.
Biometrics verification is a strategy used by coordinating an individual’s
hereditary attributes or behavioural qualities with information that have already
been learned, enlisted into a layout and organized in a framework database or on a
token [2]. It can likewise be characterized as the idea of recognizing oneself by
something that you know, something that you have, or something that you are [3].
That is, it identifies the innate qualities.
R. Oak (&)
Department of Computer Engineering, Pune Institute of Computer Technology, Pune, India
e-mail: rvoak@acm.org

https://doi.org/10.1007/978-981-10-7245-1_18
174 R. Oak
For a parameter to be called a biometric identifier, it must satisfy some properties

[1, 4]:
(1) Universality: Every person possesses that particular characteristic. For example,
DNA can be called a biometric parameter, whereas birthmarks cannot.
(2) Uniqueness: The characteristic is different for every person. For example, while
a fingerprint is a biometric, eye colour and blood group are not.
(3) Permanence: The characteristic does not disappear or change with time. For
example, DNA is biometric but hormone levels are not.
(4) Collectability: It is possible to obtain readings of the characteristic using sen-
sors in a feasible, fast and highly accurate manner. For example, voice can be
called a biometric characteristic, but factors such as confidence and self-esteem
cannot.
(5) Circumvention: The parameter is forgery-proof and it is nearly impossible to
replicate it.
Biometric parameters may be classified into two types: Physiological and
behavioural biometrics [1]. Physiological characteristics are those that are anatomic
and biological properties of an individual. They include fingerprints, facial recog-
nition, iris scan, voice recognition, palm veins, DNA, etc. [5]. These are the tra-
ditional means by which an individual’s identity is verified.
Behavioural biometrics, on the other hand, refers to factors such as gait, GUI
interaction, Haptics [6], programming style, registry access, system call logs, mouse
dynamics, etc. [7, 8]. It depends on an individual’s inward qualities and attributes [9].
The advantages of this technique over other customary biometric approaches are
as follows [3, 5, 7]:
(1) It provides persistent security. The authentication process is not complete after
the login, but there is continuous monitoring.
(2) Behaviours can be collected surreptitiously without alerting the user.
(3) No special hardware is necessary to identify the behaviours.
(4) It is difficult to replicate the behaviour, hence making identity theft less likely.
2 Classification of Behavioural Biometrics
Depending on the nature of the parameters collected and evaluated, behavioural

biometric mechanisms can be subdivided into five types (Fig. 1).
(1) Authorship Based: It is based on the analysis of a work produced by the user.
The system identifies styles and characteristics particular to a user as he writes/
draws and verification is done based on the matches of these characteristics.
(2) HCI Based: It is based on the traits and mannerisms exhibited by the user while
interacting with the system such as mouse movements or touchscreen strokes.
A Literature Survey on Authentication Using Behavioural … 175
Behavioural
Biometrics
Authorship Indirect HCI Motor Skills Purely

HCI Based
Based Based Based Behaviourial
Fig. 1 Classification of behavioural biometrics
Every person applies different strategies, shortcuts and conventions while using
a computer. A collection of such interaction particularities serves as a base for
constructing a user identity and for authentication.
(3) Indirect HCI: It is very similar to HCI-based interaction. In indirect HCI-based
systems, the system monitors the effects of the normal HCI actions. All user
actions leave certain low-level digital evidence in the system in the form of
system call traces, audit logs, execution traces, call stack analysis, etc.
(4) Motor-Skills based: It refers to the way in which the user utilizes the muscles. It
is presumably the best inquired about of all behavioural biometric techniques.
Human movements are a combination of muscle action, bone action and
impulses travelling in the nervous system.
(5) Purely Behavioural: It is not based on the body part metrics or intrinsic
behaviour. Rather, it is based on the fact that human beings utilize different
strategies, innovative ideas, critical thinking and creative thinking. It attempts
to quantify such traits and use them for authentication.
3 Evaluation Metrics
The following parameters are used as a measure of the effectiveness of the

biometric systems:
(1) FAR: It is the false acceptance rate [5]. It is the ratio of the amount of attack
instances incorrectly labelled as authentic to the total number of attack
instances [5, 10]. Value of FAR must be as small as possible. It gives a measure
of the percentage of attacks that could not be identified by the biometric system.
(2) FRR: It is the false rejection rate. It is the ratio of the amount of authentic
interactions that were incorrectly classified as attacks and the total number of
authentic, valid instances [10]. It gives a measure of the percentage of authentic
interactions that were incorrectly classified as intrusive.
(3) ROC: It is the Receiver/Relative operating characteristic. It is a plot which
represents a compromise between FAR and FRR. The matching algorithm of a
biometrics mechanism has an established threshold which evaluates how close
to the learned format an instance must be so that it qualifies as authentic [5].
Increasing the threshold reduces the FAR but increases the FRR. Decreasing
176 R. Oak
the threshold will reduce FRR but lead to higher FAR. The ROC helps to
identify the optimum threshold so that both FAR and FRR are minimized.
(4) EER/CER: At this rate, both acceptance and rejection rates are numerically the
same [5, 10]. ROC gives the value of EER. Ideally, the value of EER should be
small [5].
4 Survey of Techniques
4.1 Keystroke Dynamics
It is a type of HCI-based behavioural biometric parameter. The keyboard is the

primary input device used by humans to interact with a system. Different indi-
viduals have varied characteristics with respect to the speed of typing, error rate, use
of certain key combinations, use of the touch typing method, etc. [10]. As a result of
these differences, it is possible to verify the identity of the user. For example, the
touch typing method is used by veteran typists, whereas novices use a
hunt-and-peck technique which utilizes just two fingers. [7, 10]. Keystroke
dynamics are based on the two important parameters: Flight Time and Dwell Time
[8, 10]. Flight time is the time gap between releasing a key and pressing the next
one, and the latter is the time for which a key is pressed. A large amount of research
regarding keystroke dynamics for verification has been carried out. Bartolacci in
2005 and Curtin in 2006 have studied keystroke dynamics for long text analysis
[4, 7, 10]. A study for email authorship identification was carried out by Gupta et al.
in 2005. In [11], the author constructs digraphs (consisting of two adjacent char-
acters) which he classifies into seven different categories and calculates mean
latency for each category. This gives a measure of the programming experience of
the user [11]. The authors in [12] have deduced that keystroke on integrating it with
accelerometer biometrics, has a false acceptance rate of just 7%.
4.2 Mouse Dynamics
The mouse is probably the most important device after the keyboard which aids
humans in interacting with the Graphical User Interface (GUI). In general, mouse
dynamics refer to the characteristics of different individuals to use different pointing
devices like mouse and light pen in different manners [13]. A number of different
mouse gestures can be analyzed, such as single clicks, double clicks, scrolling, drag
and drop and stillness [7]. To acquire the features, a software program intercepts the
low-level events occurring because of mouse dynamics, along with associated
attributes such as timestamps and cursor coordinates [10, 13]. At the high level, the
gathered information would incorporate abnormal state itemized data about the
GUI-related activities of the client, for example, left the tap on the begin menu,
double tap on explorer.exe, close notepad.exe window, and so forth [7]. Other
factors such as velocities in a horizontal, vertical and tangential direction along with
the angular direction, tangential acceleration and jerk also form a part of the cap-
tured data [10]. From this event log, various statistical and kinematic features are
extracted which are used to build a user profile. Pusara and Brodley in 2004 have
proposed an approach in which split the mouse event data is classified into
movements of the mouse wheel and clicks on different entities on the screen [7].
Gamboa and Fred in 2003 have described an approach in which identification and
authentication of humans are carried out by analyzing the human–mouse interaction
in online gaming. In [14], mouse movements were captured as functions of
timestamps and graphical coordinate values and analysis was done using support
vector machines with an error rate of 1.3%.
4.3 Haptics
It identifies with the feeling of touch. Nowadays, intelligent cell phones are
equipped with various sensing elements such as an accelerometer, gyroscope,
computerized compass and high-resolution camera [6, 9]. As a result, it is possible
to measure several physical quantities during use. The parameters measured are
direction, pressure, force, angle and position of the user’s interactions [7]. In [15], a
biometric authentication system based on the haptics was built by integrating it with
fuzzy logic. A combination of three factors such as hold-time, inter-key behaviour
and finger pressure was proposed in [16]. As much as 30 behavioural features were
proposed by Frank et al. (2012) [17]. Xu et al. (2014) have proposed a continuous
and passive mechanism which achieved an error of less than 1% [15]. Sitova et al.
(2015) introduced Hand Movement Orientation and Grasp (HMOG)-based system.
Furthermore, Buriro et al. (2016) have developed a system which analyses
micro-movements of a phone, and the exact points on the screen which are pressed
[15]. In [18], a characteristic and consistent metric is known as Index of
Individuality has been proposed which uses Gaussian Process Models to capture
data.
4.4 Gait
It refers to a person’s way of walking about. It is a muscle controlled biometric

parameter. In gait-based biometrics, parameters such as kinematic patterns, knee
ankle movements, moments, angles, hunch, etc. [19, 20]. It is a complex
spatio-temporal activity which permits biometric identification of individuals at a
distance generally via recorded video [7, 19]. The capture of data may be carried
out by floor sensors, machine vision systems, or wearable sensors [20]. Gait is a
178 R. Oak
factor which is subject to several variations from person to person depending on

age, gender, bone density, waddling, muscle strength, fat percentage and energy
level [7]. As a result, this is one of the biometric parameters which is nearly
impossible to replicate. Typical features may include arm swing, walking speed,
stooping of the back, step size, head-foot distance and head-pelvis distance [7, 19].
By using three different approaches of signal correlation, frequency domain and
data distribution statistics [15, 21], it was found by Mantyjarvi et al. (2005) that the
lowest error of 7% was achieved with signal correlation method [21]. Gafurov et al.
(2006) developed a method to identify the person using an accelerometer attached
to the leg at an error rate of 10%. Derawi et al. (2010) attached the accelerometer to
the hip and an error of 20% was seen. In [22], a gait-based WiFi signature system
was proposed using a simple Naïve–Bayes classifier with a correct identification
rate of 87%. Cola et al. [23] a device is worn on the wrist and identification was
done with an error rate of 2.9%.
4.5 Log Files
Operating systems generally maintain exhaustive log files which contain records of
every small activity initiated by the user. Such log file entries contain fields such as
the identity of the user who fired a command, the timestamp, CPU usage and other
associated parameters [7]. In a system based on the audit files, there is a high chance
of false positives due to routine, legal activities such as adding new users, changing
network settings or change in permissions [7]. Hence, in these systems, there is an
overhead of informing the authentication program of such possibilities. Network
level logs which maintain traffic and various attributes such as protocol, sequence
number, length, correction checksum, etc. serve to identify intruders in the system.
In the training phase, a profile is built which identifies certain behaviours as normal.
A field known as ‘alert flag’ is set if any abnormal activity is detected. In [24], the
authors describe a five-step process: (i) Formatting data, (ii) Compare degree of
similarity, (iii) Clustering, (iv) Retranslation and (v) Detection, in order to identify
insider threats using log entries. [25] suggests the creation of a distributed Control
Flow Graph (CFG) by extracting template sequences from log files This is the
normal, expected behaviour. All activities are analyzed by comparison with this
CFG.
5 Comparative Study
(Table 1)
Table 1 A comparison of major behavioural biometric approaches

Behavioural approach Typical parameters analysed Results obtained
Keystroke dynamics • Flight Time FAR of 7% [11]
• Dwell Time
• Error rate
• Typing technique
Mouse dynamics • Mouse wheel velocity FAR of 1.3% [14]
• Types of clicks
Haptics • Direction FER of 1% [7]
• Force
• Pressure of touch
Gait • Arm swing FER of 2.9% [23]
• Head-foot distance
• Knee angle
• Head-pelvis distance
Log Files • Network traffic No quantified results yet
• CPU usage
• Dump files
6 Conclusion
Behavioural biometrics analyses various parameters of a person’s behaviour such as

gait, stride, typing patterns, mouse patterns, etc. As it is very difficult to replicate,
these systems have a high-security. systems with very low errors have been
developed. Clearly, behavioural biometrics promises to usher in a new era in the
domains of computer and information security.
References
1. L. Wang, X. Geng, L. Wang, and X. Geng. Behavioral Biometrics For Human Identification:
Intelligent Applications. Information Science Reference—Imprint of: IGI Publishing,
Hershey, PA, 2009
2. Michelle Boatwright, Xin Luo, “What Do We Know About Biometrics Authentication?”,
InfoSecCD ‘07: Proceedings of the 4th annual conference on Information security curriculum
development 2007
3. James L. Wayman, Anil K. Jain, Davide Maltoni, and Dario Maio,” Biometric Systems:
Technology, Design and Performance Evaluation”, Springer, ISBN: 1852335963
4. Tarik Mustafić, Arik Messerman, Seyit Ahmet Camtepe, Aubrey-Derrick Schmidt, Sahin
Albayrak, “Behavioral Biometrics for Persistent Single Sign-On”, Proceedings of the 7th
ACM workshop on Digital identity management, 2011
5. K P Tripathi, “A Comparative Study of Biometric Technologies with Reference to Human
Interface”, International Journal of Computer Applications (0975 – 8887) Volume 14– No.5,
January 2011
6. Wolff, Matt. Behavioral Biometric Identification on Mobile Devices. Foundations of
Augmented Cognition. Springer Berlin Heidelberg, 783–791, 2013
180 R. Oak
7. Yampolskiy, R.V. and Govindaraju, V. (2008) ‘Behavioural biometrics: a survey and

classification’, Int. J. Biometrics, Vol. 1, No. 1, pp. 81–113
8. Ioannis C. Stylios, Olga Thanou, Iosif Androulidakis, Elena Zaitseva, “A Review of
Continuous Authentication Using Behavioral Biometrics”, Proceedings of the SouthEast
European Design Automation, Computer Engineering, Computer Networks and Social Media
Conference 2016
9. Esther Vasiete, Vishal Patel, Yan Chen, Larry Davis, Ian Char, Rama Chellappa, Tom Yeh,
“Toward a Non-Intrusive, PhysioBehavioral Biometric for Smartphones”, MobileHCI 2014,
Sept. 23–26, 2014, Toronto, ON, CA
10. Monika Bhatnagar, Raina K Jain, Nilam S Khairnar, “A Survey on Behavioral Biometric
Techniques: Mouse vs Keyboard Dynamics”, Proceedings of the International Conference on
Recent Trends in engineering & Technology - 2013(ICRTET’2013)
11. Juho Leinonen, Krista Longi, Arto Klami, Arto Vihavainen, “Automatic Inference of
Programming Performance and Experience from Typing Patterns”, Proceedings of the 47th
ACM Technical Symposium on Computing Science Education, 2016
12. Kyle R. Corpus, Ralph Joseph DL. Gonzales, Alvin Scott Morada, Larry A. Vea, “Mobile
User Identification through Authentication using Keystroke Dynamics and Accelerometer
Biometrics”, Proceedings of the IEEE International Conference on Mobile Software
Engineering and Systems, 2016
13. Zach Jorgensen, Ting Yu, “On mouse dynamics as a behavioral biometric for authentication”,
Proceedings of the 6th ACM Symposium on Information, Computer and Communications
Security, 2011
14. Nan Zheng, Aaron Paloski, Haining Wang, “An Efficient User Verification System Using
Angle-Based Mouse Movement Biometrics”, ACM Transactions on Information and System
Security, Vol 18, Issue 3
15. Andrea Kanneh, Ziad Sakr, “Biometric user verification using haptics and fuzzy logic”,
Proceedings of the 16th ACM international conference on Multimedia, 2008
16. Saevanee H., Bhatarakosol, P., (2008). User Authentication Using Combination of Behavioral
Biometrics over the Touchpad Acting Like Touch Screen of Mobile Device. International
Conference on Computer and Electrical Engineering, 2008. Page(s): 82–86
17. Frank, M., Biedert, R., Ma, E., Martinovic, I., Song, D., (2012). Touchalytics: On the
Applicability of Touchscreen Input as a Behavioral Biometric for Continuous Authentication.
IEEE Transactions on Information Forensics and Security. 2012. (Volume:8, Issue: 1). Page
(s): 136–148
18. Daniel Buschek, Alexander De Luca, Florian Alt, “Evaluating the influence of Targets and
Hand Postures on Touch-based behavioural biometrics”, Proceedings of the ACM Conference
on Human Factors in Computing Systems, 2016
19. Benabdelkader, C., Cutler, R., and Davis, L. ‘Person identification using automatic height and
stride estimation’, IEEE International Conference on Pattern Recognition 2002
20. S Laxmi, Tata A S K Ishwarya, S Sreeja, “Exploring Behavioural Type Biometrics: Typing
Rhythm, Gait, Voice”, International Journal of Innovative Research in Computer and
Communication Engineering, Vol.4, Issue 10, October 2016
21. Mantyjarvi, J., Lindholm, M., Vildjiounaite E., Makela, S.M., Ailisto, H. A. (2005).
Identifying users of portable devices from gait pattern with accelerometers. IEEE International
Conference on Acoustics, Speech, and Signal Processing, 2005. (Volume:2). Page(s): ii/973–
ii/976 Vol. 2
22. Yan Li, Ting Zhu, “Gait-Based WiFi signatures for Privacy-Preserving”, Proceedings of the
11th ACM on Asia Conference on Computer and Communications Security, 2016
23. Gugliemlo Cola, Marco Awenuti, Fabio Musso, Alessio Vecchio, “Gait-based Authentication
using a wrist-worn device”, Proceedings of the 13th International Conference on Mobile and
Ubiquitous Systems: Computing, Networking and Services, 2016
24. Markus Wurzenberger, Florian Skopik, Roman Fiedler, Wolfgang Kastner, “Discovering
Insider Threats from Log Data with High-Performance Bioinformatics Tools”, Proceedings of
the 8th ACM CCS International Workshop on Managing Insider Security Threats, 2016
25. Animesh Nandi, Atri Mandal, Shubham Atreja, Gargi B. Dasgupta, Subhrajit Bhattacharya,
“Anomaly Detection Using Program Control Flow Graph Mining from Execution Logs”,
Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, 2016
Number System Oriented Text
Steganography in English Language
for Short Messages: A Decimal Approach
Kunal Kumar Mandal, Santanu Koley and Saptarshi Mondal
Abstract Information or data transfer between nodes are the necessary part of
computing but the biggest need is to secure the data from evil sources. There are
many technique but they are insufficient in case of information larceny. Our
approach with number system oriented text steganography for short messages is the
outcome for preventing such purpose. The combination of mathematical, compu-
tation rules, and knack of us hides the data from view. The approach used to hide
data is innovative among data transfer protocols as word to word or rather alphabet
to alphabet including numbers and special characters is derived. This approach
makes data invisible when moving in any one of the ways such as SMS, WhatsApp,
Email, or Facebook messenger. Good mixtures of any number, characters form a
pair of sets that can be used to hide information. Security agencies like navy, army
or air force can use such kind of techniques transferring data from one node to
another for sake of setting aside their native soil.
Keywords Text steganography Pair of number Information hiding

Secret message
K. K. Mandal
Mankar College, Mankar, West Bengal, India
e-mail: kunal@turiyan.com
S. Koley (&)
Budge Budge Institute of Technology, Bardhaman, West Bengal, India
e-mail: santanukoley@yahoo.com
S. Mondal
Indian Railways, Asansol, West Bengal, India
e-mail: saptarshi.crj@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_19
184 K. K. Mandal et al.
1 Introduction
Today, the data/information transfer is done using e-media right through the globe.
Short data/information transfer between different nodes (especially text data) has
grown up exponentially in the last decade, which is nearly about 561 billion per
month [1]. WhatsApp and Facebook messenger are responsible for transferring text
data size of 60 billion practically [2]. Therefore, this enormous dimension of data
makes us to do something for transmitting data safely. Now, it is a trial on English
language but not limited to it. In case of producing language other than English, the
alphabets of that language will be processed as an image. Steganography is the best
approach so far for securing a message from external obstruction. Here, the exis-
tence of the message will be acquainted with the addressee barely.
This technique provides a way that anticipated recipient recognizes the subsis-
tence of the message(s). This is just a obscuring a file, message, image, or video
within another. In other way, steganography is a way of covering a message sur-
rounded by an extra message, so that nobody can have a notion of being its
presence [3]. Messages can be perceived by the predetermined addressee only.
Polygraphia and Steganographia are a grand inscription by Trithemius. Trithemius
is the person to start off the expression “Steganography”. It is originated from two
Greek words steganos, meaning “covered”, and graphein, meaning “to write”. We
may find the black and white facts on the subjects of steganography by uncovering
legendary Herodotus story of sending messages on shaved heads by their slaves.
This world of internetworking transfers data and information through worldwide
network, besides the security threats growing exponentially. This is need of
invention for securing data as much as possible. The protection of data or infor-
mation on network plays a significant role as data transfer increases on Internet.
Privacy, secrecy, and integrity of data are mandatory for defending in opposition to
legitimate access by intruders. We say need is the father of inventions; this results in
a massive development on the ground of data/information hiding. They followed
numerous techniques resembling cryptography, steganography, and coding for
safety measures. Conversely, it can be said without any doubt that steganography
portrays further concentration than others. Steganography is a combination of fine
art of putting sensitive data out of sight and science of covering information so that
its existence cannot be identified [4]. Encoding is performed on furtive messages
which is a technique where the subsistence of the information is out of sight for
unintended users. In comparison to different existing data transmission schemes,
steganography is the finest approach on secreted data communication through
internet. Developing a protected messaging in an entirely untraceable mode is the
main purpose of steganography. In addition to set up and to stay away from
portrayal of uncertainty to the diffusion of a secret data is through digital com-
munication channels [4]. This technique is not to ensure others to be acquainted
with the hidden information, but to make certain further thoughts of the existence of
information. Suspicion of surreptitious message in a shipper standard for a
steganographic means fall short the entire procedure at once. Categorical of
Number System Oriented Text Steganography in English Language … 185
submission like digital image can be purposeful [5] together with copyright pro-
tection, feature tagging, and undisclosed communication. Copyright notice or
watermark can be entrenched and delimited by an image to be acquainted with it as
intellectual possessions [6]. Several illicit efforts to use this image can be
acknowledged by digging out the watermark.
2 Steganography
Steganography is the procedure of hiding messages by which the dispatcher of that

memorandum and his beneficial counterpart recognize the subsistence of the
message. The beauty of this process is that it does not create a center of consid-
eration for the hackers. Steganography is not a new process, it was devised even in
primeval era and these olden techniques are identified physical steganography [7].
There are a number of illustrations for such processes which are messages con-
cealed in messages in various supplementary modes, messages carved in secret
inks, messages shaped on envelopes in areas covered by stamps, etc. This elderly
process is termed as digital steganography. These contemporary schemes comprise
put out of sight of messages contained by piercing pictures, implanting messages
surrounded by arbitrary data, inserting images among the message contained by
video files, etc. [8]. In addition, network steganography is drawn on the telecom-
munication networks. These include the procedures resembling steganophony
(covering a message in Voice-over-IP exchanges) and WLAN steganography
(techniques for conveying steganograms in wireless local area networks) [9, 10].
Steganography can be classified according to its importance and goals. Various
types of steganography are shown in the diagram (Fig. 1).
The study of “hide from view” of any information is known as cryptography. On
the other hand, steganography deals with arranging hidden messages so that only
the sender and the receiver are acquainted with the presence of the message. In
steganography, only the sender and the receiver know the reality of the message,
while in cryptography the survival of the encrypted message is able to be seen to the
world [11]. For the above reason, steganography gets rid of the unnecessary con-
centration to the hidden message. It can be clearly understood that steganography
exercises on hiding both the message as well as the content whereas cryptographic
methods try to protect the content of a message only [12]. By coalescing cryp-
tography and steganography, one can accomplish improved safekeeping.
3 Previous Work
Several efforts are associated here to illustrate in this segment on information hiding
by the side of their benefits and drawbacks.
Steganography
Technical Linguistic
Steganography Steganography
Method Sema grams
Spacial Frequency Visual

Domain Domain
Least Discrete Text
Significant Bit Cosine
transform Open codes
Pixel Value
Differencing Disctrete Wevelet
Transform
Pixel Indicator
Cover
Image
Audio
Video
Text
Fig. 1 Classification of steganography
3.1 Use of Markup Languages in Text Steganography
To hide any data or information using text steganography, the markup languages
are used as a part of its existing services it provides [12]. The case inattentiveness of
tags of any HTML document may be an example of the characteristics of it. Now as
an illustration, the <BR> tag ought to be used as <br> or it may be as <Br>.
3.2 Text Steganography in Explicit Characters in Words
Here in this mode, one or a few of the word or a group of characters (words) are
chosen from all or some paragraphs [13]. For example, a hidden or secret message
shapes by the manner as the initial words of every paragraph are chosen. This is a
process by placing the first characters of a particular word side by side. An
unwanted receiver may find discrepancy in the text. Similarly, a series of text after
some intervals can be decided for hiding with steganography.
3.3 Line Shifting Technique
In this system, the lines of the text are vertically shifted to some extent of degrees
[14, 15]. For example, some selected appearances are being altered. They can be
1/695 inch up or down in the text and information are concealed by generating a
secreted distinctive nature of the text. This could be unknown to others. Bare eyes
cannot find this altered text. It can be assured that printed texts are sufficient for this
technique.
3.4 The Altering of Words
The techniques of horizontally shifting words and by changing distance among

words, information are hidden in the text [14, 16]. The distance between words
keeps changing in a technique that is acceptable for texts only. Filling of a line is
quite common due to change of distance, as a cause of it can be identified less.
Security is constrained in this method.
3.5 Syntactic Scheme
This technique can be used to hide information by changing the punctuations likely
full stop (.) and comma (,) in appropriate positions. There is an example: “When I
sing well, ladies feel sick”, now altering the punctuation and we could see “When I
sing, well ladies feel sick”. One can hide information in a text file [17, 18]. These
methods are inconsequential and also have need of recognizing appropriate posi-
tions of setting punctuation symbols.
3.6 Semantic Schemes
The semantic scheme hides information within the text using the synonym of words
for definite words [19]. The foremost improvement and significance of this scheme
is the protection of information. It cannot be violated in case of retyping or using
OCR programs (contrary to methods listed under Sects. 3.3 and 3.4). However, the
significance of the transcript may change.
4 Recommended Method
Steganography for the text communication can have quite a lot of approach in
Bengali text, Hindi text, or several languages. The text written with English
alphabets (may be in any language) is converted to ASCII number system. A model
number system is being defined that can cover the given text from external view.
Our representation is different than others. This decimal representation of ASCII
can be applied on any combination of text A to Z, a to z, 0 to 9 and other special
characters but we have restricted numbers and special characters. In this demon-
stration, any number is represented by two digits ordered pair. Now these two digits
can add 1 individually and thus send these pairs. At the moment of extraction using
these two added digits are subtracted by 1 and apply the formula to get the final
value. This concept will be clearer by looking at the following example.
Let us assume a secret message—“BOMB IN MARKET”
The counting is done by taking any quantity of the commodity as (Table 1).
Such type of text message is very common mixed with several languages like
Bengali, Hindi, and English. They are written using English alphabets. Now, the
first word “BOMB” represented in ASCII is 66 (B), 79 (O) 77 (M) and 66 (B). At
this point, the number 66 will be written as (10, 11). As per proposed algorithm, we
add 1 to each number in the pair and finally this will be (11, 12), which can be sent
to other node. Here, the pair of number is again subtracted as it was added and
becomes (10, 11). The computation is done by the formula [{M * (M + 1)}/2] + N
for an ordered pair (M, N) [20]. Now the value of (10, 11) is given by
½f10 ð10 þ 1Þg=2 þ 11

¼ ½f10 11g=2 þ 11
¼ ½110=2 þ 11
¼ 55 þ 11
¼ 66
This concept pursues a different way of counting, which means in ancient times
some counted commodities are exchanged by the same and equal number of some
different commodity (Figs. 2 and 3).
The numbers are divided the group in such a manner that the first group can
contain only one digit/number, i.e., 1 (as a single number). In the same way, second
group can contain two digits/numbers, i.e., 2, 3 and the nth group can contain “n”
numbers. Thus, if we want to represent the number given by the second position in
the 11th group then it is represented as (10, 2). The formula gives us the value as
(10 * 11)/2 + 2 = 110/2 + 2 = 57. This can be verified by counting the second
position from left side of the 11th group.
Table 1 Grouping applicable in proposed number system

Group number Group elements
1 (1)
2 (2, 3)
3 (4, 5, 6)
4 (7, 8, 9, 10)
5 (11, 12, 13, 14, 15)
6 (16, 17, 18, 19, 20, 21)
7 (22, 23, 24, 25, 26, 27, 28)
8 (29, 30, 31, 32, 33, 34, 35, 36)
9 (37, 38, 39, 40, 41, 42, 43, 44, 45)
10 (46, 47, 48, 49, 50, 51, 52, 53, 54, 55)
11 (56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66)
12 (67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78)
… …
Fig. 2 Proposed encryption technique
5 Technical Background
In this technique, the cover media is a pair of number with text. Here, it is hiding
one message inside the couple. Sender only knows what the message is, how to
hide the message, how to extract and all. It also knows what encoding technique is
applied to hide the message inside the pair of number. Receivers end must know
what the decoding technique is and how to extract the message within the pair. An
equation used in this paper is [{M (M + 1)}/2] + N. There some exceptions as in
the case of spaces (Hex value 32), no such pair would be created. Obtain an
illustration as the secret message (Figs. 4 and 5).
Fig. 3 Proposed decryption technique
Fig. 4 Original text
Fig. 5 Steganographic message with cover text
6 Proposed Algorithm
We have presented a newly developed algorithm for encryption as well as

decryption technique. These algorithms will help us to determine the techniques for
hiding data in different ways.
6.1 Encoding Algorithm
Step 1: Start.
Step 2: Initiation of the short message.
Step 3: Characters from the message are taken from the beginning and proceed
for the next character.
Step 4: If the character is a blank go to Step 8. If it is a special character, it
remains the same and concatenated with the set produced in Step 7. Else
it is converted it into corresponding ASCII value.
Step 5: ASCII value is converted into a pair of number (M, N) where
M = Group number of a fully completed cluster. N = Position from the
left most digit of the next group. They can be found by comparing with
the given table for the proposed number system.
Step 6: This pair of number (M, N) is added by 1 individually as (M + 1,
N + 1).
Step 7: Added pairs are kept one by one as a set.
Step 8: The set concludes when a blank space (ASCII code—32) is
accomplished.
Step 9: If the last character is not reached go to Step 3.
Step 10: Before sending, the data will be prepared as a number of sets depending
upon the words within the message.
Step 11: End.
6.2 Decoding Algorithm
Step 1: Start.
Step 2: Initiation of the sets as cipher text.
Step 3: Sets from the encoded message are taken from the beginning and pro-
ceed for the next. If there are no sets left, go to Step 10.
Step 4: Numbers within sets are taken as pair from the beginning one by one. If
there is no pair within the set, go to Step 3.
Step 5: The pair of number (M, N) is subtracted by 1 individually as (M − 1,
N − 1).
Step 6: This pair of number (M − 1, N − 1) will be applied with a proposed
formula (((M * (M + 1))/2) + N).
Step 7: Output of the formula is an ASCII number.
Step 8: ASCII will be converted into its corresponding character.
Step 9: When last pair of the set will reach go to Step 3, else go to Step 4.
Step 10: Before sending, the data will be prepared as a number of sets depending
upon the words within the message.
Step 11: End.
7 Conclusion
Exploring an ancient art is steganography for hiding information from past to

future. On a large scale, it is used in military, diplomatic, personal, and intellectual
property applications. A technically different way of decimal number system is
introduced in this paper where we are using steganography and a pair of numbers
with text as a cover media. Here, we have tried to hide a secret message to develop
a number system and the proposed formula is itself an invention. Our encryption
and decryption of the original message from cover are also a problematic task for
intruders as another up-to-the-minute algorithm we are using. We will try to explore
steganography from its most primitive instances through probable future
application.
Future Scope
The work presented in this paper is optimistically within the defined scope, but
research work in no way ends, therefore, future research work is expected to survey
horizons beyond the scope of this paper. It is tricky to acquire a steganography that
satisfies both security and robustness. Therefore, we will make an effort to enhance
the security and robustness for steganography.
References
1. HOW MANY TEXTS DO PEOPLE SEND EVERY DAY? Source:https:// www.textrequest.

com/blog/many-texts-people-send-per-day/, Last Access: May (2017).
2. Messenger and WhatsApp process 60 billion messages a day, three times more than SMS,
Source: http://www.theverge.com/2016/4/12/11415198/facebook-messenger-whatsapp-
number-messages-vs-sms-f8-2016, Last Access: May (2017).
3. K. Tutuncu, A.A. Hassan, New approach in E-mail based text steganography, Int. J. Intell.
Syst. Appl. Eng., 3 (2015), pp. 54–56.
4. A.A. Mohamed, ―An Improved Algorithm for Information Hiding Based on Features of
Arabic Text: A Unicode Approach‖ A Egyptian Informatics Journal 15, 79–87, 2014.
5. Staff, CACM. 2014. Know your Steganographic Enemy. Communications of the ACM, Vol.
57 No. 5, Page 8.
6. Kaur, S., Bansal, S., Bansal, R.K. Steganography and classification of image steganography
techniques. International Conference on Computing for Sustainable Global Development
(INDIACom), 2014. pp. 870–875.
7. Zielińska, E., Mazurczyk, W. and Szczypiorski, K. 2014. Trends in Steganography.
Communications of the ACM, Vol. 57 No. 3, Pages 86–95.
8. Abdelmgeid Amin Ali and Al - Hussien Seddik Saad, “New Text Steganography Technique
by using Mixed-Case Font”, International Journal of Computer Applications (0975 – 8887)
Volume 62– No.3, January 2013.
9. Chen, T. S. Chen, M. W. Cheng, A New Data Hiding Scheme in Binary Image, in Proc. Fifth
Int. Symp. on Multimedia Software Engineering. Proceedings, pp. 88–93 (2003).
10. B. Feng, W. Lu and W. Sun, “Secure Binary Image Steganography Based on Minimizing the
Distortion on the Texture”, Information Forensics and Security, IEEE Transactions, vol. 10,
(2015), pp. 243–255. (2001).
11. Z. Pan, X. Ma and X. Deng, “New reversible full-embeddable information hiding method for
vector quantisation indices based on locally adaptive complete coding list”, Image Processing,
IET, vol. 9, no. 1, (2015), pp. 22–30.
12. G. Doërr and J.L. Dugelay, A Guide Tour of Video Watermarking, Signal Processing: Im-age
Communication, vol. 18, Issue 4, pp. 263–282, (2003).
13. G. Doërr and J.L. Dugelay, Security Pitfalls of Frame by Frame Approaches to Video
Wa-termarking, IEEE Transactions on Signal Processing, Supplement on Secure Media, vol.
52, Issue 10, pp. 2955–2964, (2004).
14. K. Gopalan, Audio Steganography using bit modification, Proceedings of the IEEE
Inter-national Conference on Acoustics, Speech, and Signal Processing, (ICASSP ‘03), vol.
2, 6–10 April 2003, pp. 421–424, (2003).
15. M. Nosrati, A. Hanani and R. Karimi, “Steganography in Image Segments Using Genetic
Algorithm”, Advanced Computing & Communication Technologies (ACCT), (2015),
pp. 102–107.
16. R. Aqsa, M. Muhammad, M. Saas and S. Nadwwm, “Analysis of Steganography Technique
using Least Significant Bit in Graysclae Images and their ectension to Colour Images”,
Journal of Scientific Research and Report, vol. 9, no. 3, (2016), pp. 1–14.
17. N. Provos and P. Honeyman, Hide and Seek- An introduction to Steganography, IEEE
Security and Privacy, pp 32–44, May/June (2003).
18. T. Moerland, Steganography and Steganalysis, www.liacs.nl/home/tmoerlan/privtech.pdf,
May15 (2003).
19. K. Bennett, Linguistic Steganography: Survey, Analysis, and Robustness Concerns for Hiding
Information in Text, Purdue University, CERIAS Tech. Report (2004–13).
20. K. K. Mandal, A. Jana, V. Agarwal, A new approach of text Steganography based on
ma-thematical model of number system, Circuit, Power and Computing Technologies
(ICCPCT), IEEE, pp. 1737–1741, 20–21 March (2014).
Novel Robust Design for Reversible Code
Converters and Binary Incrementer
with Quantum-Dot Cellular Automata
Bandan Kumar Bhoi, Neeraj Kumar Misra

and Manoranjan Pradhan
Abstract This work, we employ computing around quantum-dot automata to

construct the architecture of the reversible code converters and binary incrementer.
The code converter and binary incrementer are made up of Feynman gate and Peres
gate, respectively. We have presented the robust design of Ex-OR in QCA, which is
used for the construction of code converters and binary incrementer. The layouts of
proposed circuits were made using the primary elements such as majority gate,
inverter, and binary wire. A novel binary-to-gray converter design offers 59% cell
count reduction and 36% area reduction in primitives improvement from the
benchmark designs. Being pipeline of PG gate to construct the 1-bit, 2-bit, and 3-bit
binary incrementer, we can use this robust layout in the QCA implementation of
binary incrementer. By the comparative result, it is visualized that the binary
incrementer such as 1-bit, 2-bit, and 3-bit achieved 60.82, 60.72, and 64.79%
improvement regarding cell count from the counterpart.

Keywords Reversible logic Quantum-dot cellular automata Nanotechnology
Code converter, binary incrementer
1 Introduction
The current microelectronics has grown for the past years, but it faces problems in
miniature, power, and design cost. Nanoelectronics have an alternative to tackle
these problems [1]. The challenges task today is the construction of robust design.
B. K. Bhoi (&) M. Pradhan

Veer Surendra Sai University of Technology, Burla, India
e-mail: bkbhoi_etc@vssut.ac.in
M. Pradhan
e-mail: manoranjan66@rediffmail.com
N. K. Misra
Institute of Engineering and Technology, Lucknow, India
e-mail: neeraj.mishra@ietlucknow.ac.in

https://doi.org/10.1007/978-981-10-7245-1_20
196 B. K. Bhoi et al.
QCA computing is suitable for the efficient computing around nanometer scale [2].
This paper introduces the robust design with emerging QCA technology which has
a potential to the difficulty of conventional CMOS technology [3].
The essence of reversible circuits is accessible to the new generation designer
regarding no loss of information. Landauer [4] proved that conventional irreversible
logic gates have information loss, i.e., energy dissipate. Bennett [5] proved that for
a digital circuit construct from reversible logic gates no information loss also
negligible energy dissipation. Moreover, the quantum circuit is to be developed by
reversible gates [6].
This paper proposes a robust design of Ex-OR which uses the QCA technology
are presented to be used in the construction of reversible code converter and binary
incrementer. The design was driven toward less latency, less cell count, minimum
clock zones, and no crossover. QCADesigner tool was adapted toward achieving
these QCA primitives’ results. However, more cell complexity means more power
dissipation similarly low complexity increase the computation speed [7].
The complete quantitative analysis of converters and binary incrementer are
explored to less complexity means faster computational speed. The results show
reduction of QCA primitives when compared with previous work reported herewith.
The reversible binary incrementer introduced in this paper will be useful in various
fields, including digital signal processing and ALU. Whereas, reversible code con-
verter will be applicable in a change of information from one format to another.
Moreover, reversible code converter and binary incrementer will be useful in digital
electronics computing where dedicated code converter and binary incrementer
modules are demanded. The proposed idea of this article is described as follows.
• A novel two-input Exclusive-OR (Ex-OR) gate is proposed which is used to
design an efficient Feynman gate (FG) in QCA.
• The efficient Feynman gates are used to construct the basic designs of a code
converter and incrementer.
• Using proposed two-input Ex-OR design of QCA, the 1-bit, 2-bit, and 3-bit
binary incrementers have been constructed and reported better QCA primitives
as compared to previous designs.
• Half adder and n-bit binary incremental are designed using the proposed Peres
gate.
The rest of the article is organized as follows: Sect. 2 discusses previous work.
Section 3 presents novel design architecture and the results are discussed and in
Sect. 4 comparative statistics is presented. In Sect. 5, conclusion is presented.
2 Previous Work
The architecture of reversible code converter and binary incremental in QCA

framework are drawing attention in the previous works. Many works of these
designs have been proposed. In 2016, Das et al. [8] proposed binary incrementer in
Novel Robust Design for Reversible Code Converters … 197
QCA. The 1-bit design consists of 97 cells, 0.075 µm2 area, seven majority voter,
four inverter, and four clocking zones. The 2-bit incrementer design consists of 14
majority voter, 8 inverters, 196 cell count, 0.272 µm2 area, and 4 clocking zones.
Several reversible code converters in QCA have been proposed in [9, 10]. Das et al.
[9] present the code converters in QCA. In these designs, multilayer design is used
for layout, with suffers from high latency and complexity. Misra et al. [10] presents
the 3-bit reversible binary-to-gray and gray-to-binary code converter in QCA. The
design consists of 118 and 112 cell count in binary-to-gray and gray-to-binary code
converter, respectively. Several efforts have been made in optimizing the code
converter and binary incrementer in QCA technology for efficient reversible logic
computing. In this work, the robust QCA design of the proposed reversible circuits
is introduced first by using Ex-OR. Such a proposed QCA design shows the robust
in the reliability as a whole.
3 Building Blocks of Code Converter and Binary

Incremental
This section introduces robust architecture of reversible code converter and binary
incrementer. Toward the QCA design of code converter and binary incrementer,
one robust design of Ex-OR is presented which incurs zero latency.
3.1 Ex-OR Gate in QCA
Ex-OR operation is commonly used in many digital logic applications such as

designing arithmetic circuits, pseudorandom number generation, correlation, and
sequence detection. In this work, we are presenting a new two-input Ex-OR gate.
Figure 1a, b illustrates the gate symbol and QCA layout of this gate respectively.
Figure 2c illustrates the simulation result which indicates that there is a zero
latency.
Our proposed QCA layout of two-input Ex-OR gate requires 13 cell count with
an overall area of 0.02 µm2. This gate will be highly cost efficient in designing
complex digital circuits. The comparison result is presented in Table 1. This is the
first Ex-OR gate that utilizes less cell count and zero latency as compared to
previous.
3.2 Reversible Code Converter Implementation in QCA
To design a code converter, we used the two-input Ex-OR gate in logic output
construction. In the first type, code converter is binary-to-gray. In this design, three
(a) (c)
(b)
Fig. 1 Ex-OR gate a functional diagram b quantum-dot cell layout c timing waveform
(a) (b) (c)
FG
FG
FG
Fig. 2 Reversible four-input binary-to-gray code converter a functional diagram b quantum-dot

cell layout c timing waveform
Table 1 Comparison of two-input Ex-OR gate QCA layouts

Design Cell count Latency Area (lm2)
[11] 67 1.25 0.06
[12] 29 0.75 –
[13] 18 0.75 0.01
[14] 14 0.5 0.02
New 13 0.5 0.02
(a) (b) (c)
FG
FG
FG
Fig. 3 Reversible four-input gray-to-binary code converter a Functional diagram b quantum-dot

cell layout c timing waveform
Ex-OR gates were settled in the specified logic manner. The function diagram and cell
layout of binary-to-gray are exhibited in Fig. 2a, b. Figure 2c presents the result of
binary-to-gray converter. In this result, settled outputs are correctly with zero latency.
The second code converter is gray-to-binary, which was drawn in Fig. 3a. The
gray-to-binary design in QCA is presented in Fig. 3b, which was utilized by the less
cell count, latency, and two clock zones. Figure 3c presents simulation result of
proposed gray-to-binary. In this result attained correctly logic signal, zero latency is
achieved in output B3, B2, and B1, whereas B0 after one clock cycle delay.
3.3 Reversible Binary Incrementer in QCA
In the proposed scheme, we have shown pipeline design of binary incrementer in

size of 1-bit, 2-bit, and 3-bit in QCA. With a pipeline design in QCA of binary
incrementer, we utilize Peres gate as a building block. Implementation of Peres gate
in QCA has been presented in Fig. 4b. It is noted that the latency of the design in
large size order of binary incrementer has been diminished. Peres gate design in
QCA has been achieved by use of Ex-OR. Peres gate is a reversible gate having
three inputs and three outputs as shown in Fig. 4a. Here, the input to output
mapping is P = A, Q = AB, R = ABC and the quantum cost of this gate is four.
The simulation result of the Peres gate is shown in Fig. 4c. In simulation result, it is
shown that P and Q outputs have zero latency, whereas R output has one latency
(clock delay). This quantum-dot cell of PG is achieved by taking three majority
gates and two inverters.
3.4 Reversible 1-Bit Binary Incrementer
In the quantum-cell design of reversible binary incrementer only by normal cell,

and fix polarization cell are utilized. Most of the existing binary incrementers are
based on coplanar or multilayer [8]. In this work, we are targeting the normal and
fix cells based design. Figure 5a shows the binary incrementer block for one bit.
The proposed Peres gate implemented reversible binary incrementer circuit is
shown in Fig. 5b. Figure 5c, d shows the cell layout and simulation results related
to binary incrementer, respectively.
(b) (c)
(a)
PG
Fig. 4 Reversible PG a functional diagram b quantum-dot cell layout c timing waveform

(a) (c) (d)
HA
(b)
PG
Fig. 5 Binary incrementer a Half adder as Binary incremental b Peres gate as an incremental
c quantum-dot cell of PG as half adder d timing waveform
(c) (d)
(a)
HA HA
(b)
PG PG
Fig. 6 2-bit reversible binary incrementer a functional diagram b quantum-dot cell c timing
waveform
3.5 2-Bit Reversible Binary Incrementer
A 2-bit binary incrementer is implemented with cascading two half adders as shown
in Fig. 6a. Here A1, A0 is 2-bit input binary number and S1, S0 represents incre-
mented output. A structure of 2-bit binary incrementer using Peres gates is shown in
Fig. 6b. QCA schematic and simulation results of the proposed 2-bit binary
incrementer are drawn in Fig. 6c and Fig. 6d respectively. In timing waveform
result, arrow sign indicates the output signal that has arrived at that point after two
latency.
(d)
(c)
(a)
HA HA HA
(b)
PG PG PG
Fig. 7 Reversible 3-bit binary incrementer a functional diagram b quantum-dot cell c timing
waveform
3.6 3-Bit Reversible Binary Incrementer
The 3-bit binary incrementer is implemented using three number of cascaded HA as

shown in Fig. 7a. Here, A2A1A0 represents 3-bit input number and S2S1S0 repre-
sents 3-bit output number. The 3-bit reversible binary incrementer with the PG’s is
shown in Fig. 7b. Its proposed QCA layout and simulation results are drawn in
Fig. 7c, d, respectively. The QCADesigner has two type of engines that are
available for simulation. Coherence vector and the bistable approximation are two
engines [15, 16]. In this work, we use a bistable engine which is a default set of
parameters.
3.7 Reversible n-Bit Binary Incrementer
The structure of 1-bit reversible binary incrementer requires one Peres gate as
shown in Fig. 5 and it has one garbage output. Two numbers of Peres gates are used
in designing 2-bit binary incrementer having two garbage outputs as shown in
Fig. 6. Figure 7 shows that three Peres gates are used in the design of 3-bit
reversible binary incrementer, and it has three garbage outputs. Similarly to design
n-bit reversible binary incrementer n-number of Peres gates are used with n-number
of garbage outputs. The structure of n-bit binary incrementer using PG’s is shown
in Fig. 8.
4 Comparative Statistics
Table 1 explores the comparison of the quantum-dot cell circuits of reversible

binary-to-gray and gray-to-binary converters. Table 2 shows that the new
binary-to-gray converter has an improvement of 59% in cell count and 36%
PG PG PG PG
Fig. 8 Functional diagram of n-bit reversible binary incrementer
Table 2 Comparison of QCA layouts of reversible binary-to-gray code converters

Design Capability Cell count Area Crossover Latency
Binary-to-gray [10] 3-bit 90 0.95 0 2
[17] 4-bit 192 0.34 0 2
[9] 3-bit 118 0.38 1 0.75
[13] 4-bit 108 0.11 1 0.75
New 4-bit 44 0.07 0 0.5
Gray-to-binary [10] 3-bit 194 0.28 0 4
[17] 4-bit 269 0.69 0 6
[9] 3-bit 99 – 0 0.75
[13] 4-bit 69 0.10 0 0.75
New 4-bit 50 0.09 0 1
improvement in the total area over the previous design [9]. This circuit also has no
crossover and 33% improvement in latency [9]. The proposed reversible
gray-to-binary code converter has 27% improvement in cell count and 10%
improvement in the area over the previous work [9].
The proposed quantum-dot cell of PG and 1-bit binary incrementer have 60 and
6% improvement in terms of cell count and area over [8], respectively. Our pro-
posed reversible 2-bit binary incrementer has 60 and 52% improvement in cell
count and area over existing QCA layout [8], respectively. Similarly, the proposed
reversible 3-bit binary incrementer has 64 and 34% improvement in cell count and
area compared to existing designs [8], respectively. The overall evaluation is drawn
in Table 3. It is realized that there is a significant enhancement of results in the
proposed QCA layout designs.
Table 3 Comparison of Peres gate and reversible binary incrementer QCA layouts
MV INV CC Area Crossover Latency
No No. (µm2)
Peres gate [8] 7 4 97 0.075 1 1
New 3 2 38 0.070 0 1
1-bit reversible [8] 7 4 97 0.075 1 1
incrementer New 3 2 38 0.070 0 1
2-bit reversible [8] 14 8 196 0.272 2 2
3-bit reversible [8] 21 12 321 0.274 3 3
5 Conclusion
This paper has reported the novel designs of reversible code converters and binary
incrementer using QCA technology. All layouts are implemented in QCADesigner
tool. The simulation results of all the QCA layouts have confirmed the functionality
of the designs. All these designs have achieved a higher efficiency compared to
previous works available in the literature. We also presented optimized designs of
Feynman gate and the Peres gate in QCA. These two are basic gates having a low
quantum cost which can be used for designing reversible circuits in QCA tech-
nology. Therefore, all our proposed QCA layouts can be used as an effective
architecture for designing efficient digital circuits with the power of reversibility
and QCA technology.
References
1. Lent, C.S., Taugaw, P.D.: A device architecture for computing with quantum dots. Proc. IEEE
85, 541–557 (1997).
2. Misra, N.K., Sen, B. and Wairya, S.: Designing conservative reversible n-bit binary
comparator for emerging quantum-dot cellular automata nano circuits. Journal of
Nanoengineering and Nanomanufacturing, 6(3), 201–216 (2016).
3. Allan, A., Edenfeld, D., Joyner, W.H., Kahng, A.B., Rodgers, M. and Zorian, Y.: 2001
technology roadmap for semiconductors. Computer, 35(1), 42–53 (2002).
4. Landauer, R.: Irreversibility and heat generation in the computating process. IBM J. Res. Dev.
5(3), 183–191 (1961).
5. Bennett, C.H.: Logical reversibility of computation. IBM J. Res. Dev. 17, 525–532 (1973).
6. Misra, N.K., Sen, B. and Wairya, S.: Towards designing efficient reversible binary code
converters and a dual-rail checker for emerging nanocircuits. Journal of Computational
Electronics, 16(2), 1–17 (2017).
7. Misra, N.K., Sen, B., Wairya, S. and Bhoi, B.: Testable Novel Parity-Preserving Reversible
Gate and Low-Cost Quantum Decoder Design in 1D Molecular-QCA. Journal of Circuits,
Systems and Computers, 26(9), 1750145 (2017).
8. Das, J. C., & De, D.: Novel low power reversible binary incrementer design using
quantum-dot cellular automata. Microprocessors and Microsystems. 42, 10–23 (2016).
9. Das, J.C., De, D.: Reversible Binary to Grey and Grey to Binary Code Converter using QCA.
IETE Journal of Research. 61(3), 223–229 (2015).
10. Misra, N.K., Wairya, S. and Singh, V.K.: Optimized Approach for Reversible Code
Converters Using Quantum Dot Cellular Automata. In Proceedings of the 4th International
Conference on Frontiers in Intelligent Computing: Theory and Applications (FICTA)
Springer India, 367–378 (2016).
11. Angizi, S., Alkaldy, E., Bagherzadeh, N., Navi, K.: Novel robust single layer wire crossing
approach for exclusive or sum of products logic design with quantum-dot cellular automata.
Journal of Low Power Electronics. 10, 259–271 (2014).
12. Singh, G., Sarin, R.K., Raj, B.: A novel robust exclusive-OR function implementation in
QCA nanotechnology with energy dissipation analysis. Journal of Computational Electronics.
15(2), 455–465 (2016).
13. Karkaj, E. T., Heikalabad, S.R.: Binary to gray and gray to binary converter in quantum-dot
cellular automata. Optik-International Journal for Light and Electron Optics. 130, 981–989
(2017).
14. Sasamal, T.N., Singh, A.K., Ghanekar, U.: Design of non-restoring binary array divider in
majority logic-based QCA. Electronics Letters. 52(24), (2016).
15. Sen, B., Dutta, M., Goswami, M., Sikdar, B.K.: Modular Design of testable reversible ALU
by QCA multiplexer with increase in programmability. Microelectronics Journal. 45(11),
1522–32 (2014).
16. Sen, B., Dutta, M., Some, S., Sikdar, B.K.: Realizing reversible computing in QCA
framework resulting in efficient design of testable ALU. ACM Journal on Emerging
Technologies in Computing Systems (JETC). 11(3), 30 (2014) https://doi.org/10.1145/
2629538.
17. Rafiq, B.M., Mustafa, M.: Design and Simulation of Efficient Code Converter Circuits for
Quantum-Dot Cellular Automata. Journal of Computational and Theoretical Nanoscience.
11(12), 2564–2569 (2014).
Routing with Secure Alternate Path
Selection for Limiting the Sink Relocation
and Enhanced Network Lifetime
Renuka Suryawanshi, Kajal Kapoor and Aboli Patil
Abstract The main challenge of wireless sensor network is its lifetime. In this type
of network, single static sink node is present; a sensor device node requires more
energy for estimating information packet specifically those that are available in the
area of the sink node. Such nodes separate the energy so fast due to the numerous
tone traffic patterns and at the end they die. This uneven event is named as hotspot
issue which gets more real as the numbers of sensor nodes increase. Generally,
replacement of such energy sources is not a feasible and cost-effective solution. For
this problem, there is one solution regarding to distance. If the distance among
sensor and sink node is minimized, the energy consumption will be effectively
reduced. This paper presents the solution for enhancing network lifetime with
energy saving of sensor nodes. Here, we also discuss on the limitations and
advantages of previous methods. The sensors nodes consume more battery power
which is at minimum distance from sink node. Therefore, energy of sensor nodes in
network will quickly consume their energy. So that, the lifetime of a sensor nodes
will be produced. To overcome this drawback of this system, we propose alternate
shortest path technique. To enhance the efficiency of energy along with network
lifespan, this approach is used. Furthermore, we developed a novel technique
known as Energy Aware Sink Relocation (EASR) for remote base station in WSN if
the energy of alternate path is going to die. This system exploits information
recognized with the remaining energy of sensor nodes battery for increasing the
range of transmission of sensor node and relocation technique for the sink node in
network. Some calculated numerical and theoretical calculations are given to
demonstrate that the EASR strategy is used to increase the network energy of the
remote system essentially. Our system proposes secure data sending using ECC
algorithm and increases more network lifetime.
R. Suryawanshi (&) K. Kapoor A. Patil

Department of Computer Engineering, MITCOE, Pune, India
e-mail: renukasuryawanshi743@yahoo.com
K. Kapoor
e-mail: kajal.kapoor@mitcoe.edu.in
A. Patil
e-mail: aboli.patil@mitcoe.edu.in

https://doi.org/10.1007/978-981-10-7245-1_21
208 R. Suryawanshi et al.
Keywords Wireless sensor networks Cluster head Base station

Cache-based system Sensor nodes
1 Introduction
In the wireless network of sensor, nodes are outlined by heavy deployment of the
huge types of sensor nodes in a particular geographical area. The information
captured from sensor nodes is transmitted to monitor the station. These monitoring
stations worked as a sink or base station. This base station is placed far from actual
sensing area. To transfer this sensed information from source to base station,
concept of multi-hop routing and flooding is used. With the multiple numbers of
base stations, the total number of hops can get minimized. This will result in
minimized energy consumption by sensor node. The minimum energy consumption
of sensor nodes will improve lifetime of sensor network along with high rate of
packet transmission to base station. So the communication nodes deployment and
the different sink nodes are treated as most important components in the lifetime in
wireless sensor network.
The WSN have various applications such as climate observing, battlefield
investigation and inventory, manufacturing progressions. For the maximum amount
of time, the sensor environment needs to be intolerant. In the remote network
system, the sensor devices are not present to replace when their batteries get
drained. The battery exhausted from nodes can be brought several issues, for
example, take coverage hollow space moreover, communication hollow space
issues. Consequently, some WSN systems are busy in planning proficient approach
to keep the energy of sensor nodes, as an instance, drawing schedule of cycle for
sensor nodes, which is used to permit some of nodes and enter into the die state to
moderate power of energy. Now sensor node does not damage the running sensing
process of the wireless network. The efficient design of energy algorithms aims at
balancing the depletion of the battery exploitation strength of every sensor node or
consuming a limited data aggregation technique for mixture of sensory information
into a unit to decrease the number of message transmitted to prolong the wireless
network lifetime. The enormous majority of such system scans coincide in the
network system work. The another methodology is used for the purpose of storing
energy as well as utilizing remote sensors to maintain the locations of region with
aggregating lifetime network energy of nodes.
For enhancing the lifespan of network as well as efficiency of energy, we pro-
pose a shortest alternate path mechanism in this paper. For transmitting data from
sender node to sink node, safely use alternative route and ECC algorithm. When
energy level is less than given threshold for alternate path, we trigger the relocation
of sink. Scheme for sink relocation is explored here, which decides when and where
to relocate the position of sink. The mathematical performance evaluations are
calculated to determine the proposed sink relocating scheme which is beneficial for
enhancing the network lifetime of a system. The simulation of technique is used in
Routing with Secure Alternate Path Selection … 209
project to check out the precaution of the EASR technique. This type of approach
can work to improve the lifetime of a WSN system. The sink node relocation will
be prolonging the battery usage of nodes.
Section 2 illustrates the related work studied for our new topic. Section 3
demonstrates the details of project implementation, definitions of terms and in
addition the documentation can express the proposed system undertakings in this
paper. Section 4 includes conclusions and represents future work of project.
2 Literature Survey
In this part, we illustrate the previous techniques proposed by the authors for WSN
system.
Sara and Sridharan [2] represent a survey of routing schemes in remote sensor
nodes networks. In WSN’s author review the challenges for routing protocol
designs were discussed. The comprehensive research of individual routing tech-
nique is classified into three stages depending upon the structure of network such as
flat, hierarchical, location-based routing, etc.
Somasundara et al. [3] investigate a network system which depends on the
utilization of mobile components and reduces the utilization of the energy con-
strained nodes at the time of communication and enhances useful network.
Similarly, their approach gains the advantages in sensor networks and inadequately
deployed sensors in network. They demonstrate how their procedure supports to
reduce energy utilization at energy controlled nodes. After that, for enhancing the
performance of energy, author illustrates their framework model which uses their
proposed way.
Sensor network deployment is highly challenging because of the aggressive as
well as volatile nature of consumption environments. Mousavi et al. [4] imple-
mented two routines for the self-deployment of mobile sensors. Basically, author
developed a randomized way that offers both simplicity and applicability to dif-
ferent environments.
Akyildiz et al. [5] describe idea of network formed by sensors. These sensors
have combined microelectromechanical technology, wireless communication and
digital physics. First, the sensing tasks and applications of sensor networks are
examined, and a comparative analysis of things influencing the look of sensor
networks is given.
The main benefit of sensor node networks is their self-organizing nature as well
as autonomous process and possible architectural alternatives suitable for a different
types of data-centric driven applications. During this article, Jain and Agrawal [6]
deal with the presenting an outline of this state of the art inside the field of wireless
sensor networks.
Tian and Georganas [7] have introduced inclinations to a node scheduling
scheme, which can reduce the energy utilization of complete system, therefore
growing network time period, by characteristic redundant nodes with respect to
sensing coverage of network, moreover distributing them an offline operation mode.

This offline mode has minimum energy consumption than the online mode.
Hong et al. [8] implemented a capable route setup for distributed sensing ele-
ment network utilizing the similar process between the wireless and multi-hop
communications network concerning instruments and rovers and therefore the
packet radio network is utilized as a typical ad hoc networking surroundings.
Huang and Jan [9], for enlarging the lifetime of network’s, presented an Energy
Aware Cluster Based Routing Algorithm (ECRA) in WSN’s. This algorithm
chooses some nodes as cluster heads to construct Voronoi outlines and move the
cluster head load balancing in every cluster of nodes. To improve the execution of
the ECRA, a two-tier architecture (ECRA-2T) is designed. The reproductions
demonstrate that both the ECRA-2T as well as ECRA algorithm perform better than
other routing schemes such as direct communication, static clustering, and LEACH.
Shah and Rabaey [10] developed an energy aware routing mechanism that
depends upon sub-optimal paths which give substantial gain. Additionally, the
experimental results show the increment in lifetime of network over practically
similar plans like directed diffusion routing. The more elegant degradation of service
with time in a fairer manner to overcome the burning energy of nodes were showed.
In planning sensor networks, sensor deployment is a primary problem. Wang
et al. [11], they review and make use of disseminated self-deployment protocols for
mobile sensors. The protocols are proposed to estimate the target positions of the
sensors after finding coverage of hole in network where the sink is ready to move.
3 Implementation Details
In this field, we illustrate the overview of system, algorithmic steps of system, and
mathematical formulation of the proposed system.
3.1 System Overview
Figure 1 represents the architecture of the proposed system. System works as

described in the following:
• Network Generation
In this phase, user can generate vertices or nodes. These nodes are connected by
edges.
• Path Generation:
It creates all possible routes from source to sink node after creating source as
well as sink.
• Shortest Path Selection:
Fig. 1 Proposed system

architecture
After generation of all possible paths from source to destination node, shortest
path is selected on the basis of minimum weight of edge.
• Generation and distribution of Keys:
At key generation center, keys are generated and distributed to all nodes belongs
to shortest path.
• Encryption of data:
At every node, collected data is encrypted by using ECC algorithm for security
purpose.
• Estimation of Energy Consumption:
After collection of data or sending of data or any type of action, consumed
energy is calculated at each node.
• Data Authentication:
Sink node checks the authenticated data after determining the hash value at
source node.
• Data Decryption:
After receiving the data from source node, sink node decrypts the data for
further processing. For decryption, sink node has the decryption key.
3.2 Mathematical Model
The proposed system (PS) is represented as

PS= {DN, SN, SiD, PA, SpS, AK, DS}
1. Deployment of nodes
DN = {dn1, dn2,…,dnn}
where
DN is the set of number of deployed nodes.
2. Source node selection
SN = {sn1, sn2,…,snn}
where SN is a set of sources selected at each runtime.
3. Sink node selection
SiD = {Sid1, Sid2,….,Sidn}
Where, SiD is a set of sink nodes selected at each runtime.
4. All paths from source to destination
PA = {pa1, pa2,….,pan}
where PA is a set of all n number of paths from source to destination.
5. Shortest path selection
SpS = {sps1, sps2, sps3,….,spsn}
where SpS is the set of all possible shortest path from source to destination at
each runtime.
6. Authentication with keys
AK = {ak1, ak2,….,akn}
where AK is a set of n number of keys generated and distributed to each node
for authentication.
7. Data sending from source to destination.
DS = {ds1, ds2, ds3,….,dsn}
where DS is the set of all data packets securely routing through shortest path.
3.3 Algorithm
The proposed scheme works as follows:

• Algorithm 1: Proposed algorithm description
1. The network graph is generated such as Gi(Vi, Ei)
where Vi is the set of vertices and Ei is the set of all connecting edges to
vertices.
2. Select sender and destination node from all sensor nodes.
3. Produce all possible paths from selected source to destination node.
4. Among all generated possible path, select the one shortest path based on
weight factor.
5. Generate and distribute public–private key pair for source and destination
node.
6. Perform data sending at source node through selected shortest path.
7. Encrypt the data with the private key before actually sending.
8. Estimate energy consumed by each node belongs to shortest path.
9. Decrypt the private key and authenticate received data at destination.
10. If energy node in path is going to die then select alternate path among
shortest path.
11. Resend the data from source to destination node through alternate path and
also calculate energy consumed by path.
12. Again energy may expire of alternate path.
13. When energy is minimized, use Energy Aware Sink Relocation (EASR)
technique to relocate sink node at other place.
Explanation: Algorithm 1 describes the levels of proposed system. Primarily,
with sensor nodes, source and sink node network is created. Then generate all
routes from source to sink node and for data sending purpose choose the shortest
path. Sensor nodes are not working properly if energy utilization is greater.
Therefore, systems must choose optional communication path between source and
destination node and also estimate energy consumed by each node in network. By
using the ECC algorithm, encrypt the data with the secret key. With the help of its
hash value, data is validated. Only verified data is accepted by sink node. Decrypt
the received data with the appropriate public key. If again energy is evacuated and
path is expired, then repeat the procedure of sink relocation.
• Algorithm 2: ECC Algorithm
1. Sender and receiver node calculate edB = S = (s1, s2).
2. Sender node sends a message M E to receiver node as follows:
3. Compute L such that, (s1 * s2) mod N = L.
4. Compute L * M = C.
5. Send C to sender node.
6. Receiver node receives C and decrypts as follows:
7. Compute (s1 * s2)modN = L.
8. Compute (L-1)mod N, Where N = E.
9. L-1*C = L-1*L*M = M.
Basically, the system is constructed with the help of java framework with version
JDK 1.8 on window platform. IDE is used for developing the system, Netbeans
with version 1.8 is used as a development tool. For generating the network, Jung
simulation tool is used, by using this tool, network is created with the number of
sensor node.
4 Results Discussion
Figure 2 describes the evaluation graph among ratio of energy consumption for
existing as well as proposed system. In the designed system, energy utilization is
less than the energy utilization in the existing one (Table 1).
Figure 2 represents the evaluation of designed system and existing system on the
basis of accuracy. The accuracy of designed system is higher than the existing
system.
Table 2 illustrates the outcomes for time consumption for developed site clas-
sifier and train classifier for designed and existing system.
Figure 3 exhibits the comparison between designed system and existing system.
The time required for implementing the proposed system is less than the existing
system.
Fig. 2 Energy consumption graph comparison
Table 1 Energy Existing system Proposed system

consumption
Energy consumption 484 345
Table 2 Network lifetime Existing system Proposed system

Total energy 3645 5068
Fig. 3 Network lifetime comparison
5 Conclusion
This paper implemented the different methods to improve the lifetime of network.
A relocatable sink is one approach to enhance the lifetime of network but still it has
its own limitations as sink relocation involves more energy so we have proposed
alternate shortest path technique which optimizes all nodes in the network system
and also enhances lifetime of network by limiting the number of sink relocating
actions. In addition, we also proposed secure data sending and node authentication
for communication purpose. In future, we can increase lifetime of a network and
secure the network by providing security.
References
1. Chu-Fu Wang, Jau-Der Shih, Bo-Han Pan, and Tin-Yu Wu, “A Network Lifetime
Enhancement Method for Sink Relocation and Its Analysis in Wireless Sensor Networks,
IEEE sensors journal, vol. 14, no. 6, June 2014.
2. G. S. Sara and D. Sridharan, Routing in mobile wireless sensor network: A survey,
Telecommun. Syst., Aug. 2013.
3. A.A. Somasundara, A. Kansal, D. D. Jea, D. Estrin, and M. B. Srivastavam, Controllably
mobile infrastructure for low energy embedded networks, IEEE Trans. Mobile Comput., vol.
5, no. 8, pp. 958973, Aug. 2006.
4. H. Mousavi, A. Nayyeri, N. Yazani, and C. Lucas, Energy conserving movement-assisted
deployment of ad hoc sensor networks, IEEE Commun. Lett., vol. 10, no. 4, pp. 269271, Apr.
2006.
5. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayiric, Wireless sensor networks: A
survey, Comput. Netw., vol. 38, no. 4, pp. 393422, Mar. 2002.
6. N. Jain and D. P. Agrawal, Current trends in wireless sensor network design, Int. J. Distrib.
Sensor Netw., vol. 1, no. 1, pp. 101122, 2005.
7. D. Tian and N. D. Georganas, A node scheduling scheme for energy conservation in large
wireless sensor networks, Wireless Commun. Mobile Comput., vol. 3, no. 2, pp. 271290,
Mar. 2003.
8. X. Hong, M. Gerla, W. Hanbiao, and L. Clare, Load balanced energyaware communications
for Mars sensor networks, in Proc. IEEE Aerosp. Conf., vol. 3. May 2002, pp. 11091115.
9. S. C. Huang and R. H. Jan, Energy-aware, load balanced routing schemes for sensor
networks, in Proc. 10th Int. Conf. Parallel Distrib. Syst., Jul. 2004, pp. 419425.
10. R. C. Shah and J. Rabaey, Energy aware routing for low energy ad hoc sensor networks, in
Proc. IEEE Wireless Commun. Netw. Conf., vol. 1. Mar. 2002, pp. 350355.
11. G. L. Wang, G. H. Cao, and T. L. Porta, Movement-assisted sensor deployment, in Proc.
IEEE Inf. Commun. Conf., Aug. 2004, pp. 24692479.
12. Chu-Fu Wang, Jau-Der Shih, Bo-Han Pan, and Tin-Yu Wu, A Network Lifetime
Enhancement Method for Sink Relocation and Its Analysis in Wireless Sensor Networks,
ieee sensors journal, vol. 14, no. 6, June 2014.
Gene Presence and Absence in Genomic
Big Data for Precision Medicine
Mohamood Adhil, Mahima Agarwal, Krittika Ghosh, Manas Sule

and Asoke K. Talukder
Abstract The twenty–first-century precision medicine aims at using a

systems-oriented approach to find the root cause of disease specific to an individual
by including molecular pathology tests. The challenges of genomic data analysis for
precision medicine are multifold, they are a combination of big data, high
dimensionality, and with often multimodal distributions. Advanced investigations
use techniques such as Next Generation Sequencing (NGS) which rely on complex
statistical methods for gaining useful insights. Analysis of the exome and tran-
scriptome data allow for in-depth study of the 22 thousand genes in the human
body, many of which relate to phenotype and disease state. Not all genes are
expressed in all tissues. In disease state, some genes are even deleted in the genome.
Therefore, as part of knowledge discovery, exome and transcriptome big data needs
to be analyzed to determine whether a gene is actually absent (deleted/not
expressed) or present. In this paper, we present a statistical technique to identify the
genes that are present or absent in exome or transcriptome data (big data) to
improve the accuracy for precision medicine.

Keywords Big data Algorithms Genomics Multimodal distribution

Exome analysis Transcriptomics analysis Gaussian mixture model
Precision medicine
M. Adhil M. Agarwal K. Ghosh M. Sule A. K. Talukder (&)

Interpretomics, 5th Floor, Shezan Lavelle, 15 Walton Road, Bangalore 560001, India
e-mail: asoke.talukder@interpretomics.co
M. Adhil
e-mail: mohamood.adhil@interpretomics.co
M. Agarwal
e-mail: mahima.agarwal@interpretomics.co
K. Ghosh
e-mail: krittika.sasmal@interpretomics.co
M. Sule
e-mail: manas.sule@interpretomics.co

https://doi.org/10.1007/978-981-10-7245-1_22
218 M. Adhil et al.
1 Introduction
Biology and medicine used to be a data-poor science. With the completion of the
Human Genome Project (HGP), biology has entered the quantitative data-rich era.
This has led the implosion in the amount of data being generated. In precision,
medicine of twenty-first-century genetic tests of body fluid (blood) or tissues (for
cancer or various neoplastic tissues) are being included in evidence-based medicine
(EBM) as “molecular pathology” tests. The analysis of molecular pathology tests
needs to overcome complex big data algorithmic challenges. The whole genome
data for a human ranges upward of 300 GB (Giga Bytes) based on the sequencing
depth. The volume of data generated by the protein coding region (exome) is of the
order of 18–20 GB while the transcriptome (RNA-Seq) data measuring the protein
expression of the coding region is of the same order as well [1]. To discover the
actionable insight or a biomarker from this data, it is necessary to perform
extensive statistical and exploratory data analysis [2], Although this “Big data” in
form of DNA, gene expression, or the pathway data has been collected and
recorded in a wide variety of databases, the intricate understanding as a holistic
process to address the an individual at a system level still remains elusive. This is
due to lack of explicit correlations amongst these different data forms. Key ques-
tions that still confront biologists include (a) the functional roles of different genes
and the cellular processes in which they participate; (b) which genes are func-
tionally present or absent within a sample and how they interact to trigger a phe-
notype; (c) how gene expression levels differ in various cell types and states, as well
as how gene expression is changed by various disease pathologies and their
treatments.
A phenotype is any observable characteristic of an organism related to its
external appearance or behavior, both physical and biochemical; whereas, the
genotype is the state of the DNA inside the body, which is transcribed and mani-
fests in RNA and subsequently in proteins; and finally how the nutrients and
environmental conditions make the cells in the body function in harmony. The
linking of genotype or mutations in the gene and the gene expression to phenotype
is the fundamental aim of modern genetics [3]. In other words, what is the genetic
root of the abnormal manifestation (disease state) of an individual? The focus
currently lies on study of interactions amongst genes supported through the
expression data and phenotype data through integrative analysis. The inherent
complexity of phenotypes makes high-throughput phenotype profiling a very dif-
ficult and laborious process. Phenotype–genotype association analysis problems are
complicated by small sample size of data coupled with high dimensionality. Many
complex phenotypes, such as cancer, are the product of not only gene expression,
but also gene interactions [4].
The central dogma of molecular biology explains how the genetic information
flows within the biological system starting from DNA to protein expression and
functional implications of the same. Changes in this flow of information at the gene
level, transcript level, or protein level can lead to the genetic causation of diseases.
Gene Presence and Absence in Genomic Big Data for Precision Medicine 219
The monitoring of gene functions and their interactions can provide an overall
picture of the genes being studied, including their expression level and the activities
of the corresponding protein under certain conditions. To accomplish this, func-
tional genomics provides techniques characterized by high-throughput or
large-scale experimental methodologies combined with statistical and computa-
tional analysis of the results [5]. Current methods in evidence-based medicine
require advanced technologies for evidence search and discovery. Next generation
sequencing provides convenient and high-throughput means for accomplishing
systemic investigation and functional genomics compared to more traditional
methods as q-PCR and Sanger based methods for mutation detection. Genes
expressed in cells in different organs (tissues) are different. This implies that though
all cells were born from one mother cell (with genetic information carried down half
from father and half from mother), their functions differ vividly due to a very
sophisticated process of cellular differentiation and gene expression. Apart from
this, environmental cues, epigenetics play an important role in gene regulation.
A core challenge in molecular biology therefore is to identify the genes that are
expressed in a cell within a tissue (or organ) and to measure their expression levels
[6]. The major challenge is to determine, if a gene is not expressed in a tissue—is it
due to normal phenotype or a disease state phenotype. Therefore, the analytics
software must be able to determine which genes are expressed (present) and not
expressed (absent) in the tissue of interest from the experiment data. In order to do
this, it must be able to identify and eliminate the noise in the expression values. In
microarray experiments, special negative control probes are designed based on the
knowledge of the nucleotide sequence of a gene. During the analysis of microarray
data, these negative control probe measures are used in the normalization process to
eliminate the noise. Gene level presence/absence calls in microarrays are made
using these negative probe sets (no hybridization) and cutoffs generated using
p-values to eliminate the noise [7].
The RNA-Seq procedures in NGS have been widely used to estimate the
expression of genes in specific tissues and under particular condition. In this
technique, the RNA is extracted from the human tissue of interest, libraries of
sequences are prepared and then sequenced using NGS. The data generated in such
experiments are in the order of 18–30 GB, generally consisting of around 30
million reads. In principle, the NGS derived reads are statistically mapped across a
genome to estimate the regions mapped. However, the mapping of reads is statis-
tical and not devoid of system bias. Since, genomic regions differ in terms of
sequence complexity, the regions with lower complexity present lower potential for
mapping than others. Moreover, there are repetitive regions which are considered
“unmappable” owing to the absence of unique reads originating from those regions
which can map them back. Such regions extend bias in alignment by mapping
algorithms leading to noise in mapping across the genome [8]. This leads to some
regions having over represented coverage whereas other regions may be less
covered.
Transcriptome data presents an added challenge due to the phenomenon of
splicing. The alignment requires considering “splice aware” alignment, wherein
220 M. Adhil et al.
reads from transcripts can be fragmented and aligned to the reference genome from
where it had originated [9]. These exert a selective bias and sometimes give rise to
nonspecific alignments in regions and transcripts which are not expressed under a
particular condition thus leading to incorrect assessment. Similar, features can be
expected in any enriched library being sequenced out from the human genome.
However, accuracy is a must when application space is health care.
In case of RNA-Seq and NGS gene expression data, the absence of negative
control probe-set as in microarray studies, there is a need to discover which genes
are expressed. Therefore, the need to algorithmically eliminate the noise generated
holds prime importance. In this paper, we present a novel statistical algorithm that
does this job accurately. Here, we improve the accuracy of this calling of genes by
introducing the algorithm that acts as an additional filter to the initial results pro-
vided by transcriptomic experiments. The algorithm works to identify mapping
artifacts thereby refining the genomic coverage for RNA sequencing analysis by
estimating the presence or absence of alignment of reads and hence gene expres-
sion. The presence/absence call algorithm presented here is part of the reference
based RNA-Seq analysis of the iOMICS software [10] and uses statistical tech-
niques to discover the genes that are not explicitly expressed in the sample under
investigation.
2 Methods
The following key points have been considered while developing the algorithm for
identification of presence/absence of genes in RNA-Seq data:
1. Parallel computing technique has been used to calculate the Reads Per
Nucleotide (RPN) to reduce the time taken by many folds.
2. The algorithm works both on single sample (n = 1) and cohort data (n = N) of
assembled RNA-Seq data to identify the sample level presence/absence.
3. Genes are also optionally annotated with tissue level baseline gene expression
values from external databases.
Because the sequence reads generated by sequencing machine are few hundreds
of nucleotides (generally 300 nucleotides from two 150 nucleotide reads), these
sequence reads often align to nonspecific regions. Also, reads that are erroneous but
not filtered by the cleaning up procedure will align to regions as noise. Due to
fundamental properties of cells, a gene may be expressed low, generating lesser
number of reads. Therefore, the challenge is to detect whether less number of read
is due to lower expression of the gene or due to miss-alignment. Here, our challenge
is to devise a mathematical model that will be able to detect the various expression
profiles within the experimental data. Before calculating RPN value for each gene,
the raw sequence quality and alignment statistics are verified for data integrity.
Gaussian mixture model (GMM) which is a parametric model is used to divide

the whole population into subpopulation. Here, the whole population refers to all
genes in the sample with logarithmic reads per nucleotide and subpopulation refers
to presence, absence, and marginal expression [11]. GMM is preferred mainly due
to its nature of distinguishing clusters with definite boundaries using p-value and
with no preassignment of data point into a cluster. GMM is an unsupervised model
similar to k-means, but unlike k-means we do not need to assign a data point (gene)
to a single cluster. GMM parameters are estimated using the iterative Expectation–
Maximization (EM) algorithm [12]. GMM is a weighted sum of M components’
Gaussian densities as given in Eq. 1. The components are predefined before the
model fitting.
X
M X
pðxjkÞ ¼ wi g xjli ; i
; ð1Þ
i¼1
where
1. x is a D-dimensional vector of continuous value
2. wi, i = 1,…, M, are the mixture weights,
3. g(x|µi, Ri), i = 1,…, M, are the component Gaussian densities. Each component
density is a D-variate Gaussian function given in Eq. (2)
4. µ is the mean vector, Ri covariance matrix and mixture weight satisfy the
constraint Ri=1 to M wi = 1

1 1 0 1
gðxjli ; Ri Þ ¼ exp ðx li Þ Ri ð x l i Þ ; ð2Þ
ð2pÞD=2 jRi j1=2 2
Here, the number of components is two (presence and absence). µ (mean) ini-
tially will be zero and x will be the vector of RPN values. Model fit is performed
with two components; mean and covariance are calculated for those two compo-
nents. After model fitting, RPN values are iterated and assigned into two distri-
butions. The GMM is parameterized by the mean vectors, covariance matrices, and
mixture weights from all component densities.
The noise distribution is identified using the RPN value calculated from inter-
genic reads aligned to reference genome due to by chance or error. The RPN for
noise distribution is calculated using the coordinates of intergenic regions
throughout the genome. Once we have the noise distribution, it is used as a
threshold or boundary for labeling the gene as present, absent, and having marginal
expression. This is done by expectation–maximization to find the maximum like-
lihood, where the mean and covariance of noise distribution is compared with the
mean and covariance of the genes calculated from the GMM model. Values of
genes where a presence/absence label cannot be assigned with high confidence
using the given p-value cutoff are labeled as marginally expressed. The p-value and
222 M. Adhil et al.
confidence level is calculated for three labels (Presence, Absence and Marginal).
The significance level (p-value) depends on the user, where it can be less than 0.05.
The RNA-Seq reads are aligned to the exonic regions (coding regions) of genes.
The number of reads aligning to the exonic regions of a gene is proportional to the
expression level of the gene and the length of the exonic region. However, due to
sequencing errors and the statistical nature of reference assembly, errors are
incorporated into the read alignment process. This results in leakage or spillover in
the data. There will be a few reads which will align to genes as alignment artifacts
(noise). Similarly, there will also be some reads that will align to the intergenic
regions also. In order to identify the noise, the algorithm uses the feature Reads Per
Nucleotide (RPN) statistic (Eq. 3). RPN is defined as
RPN ¼ R=L where R is the number of reads in a region of length L nucleotides ð3Þ
It may be noted that while the read alignment at a whole genome level is
modeled using Poisson distribution, the RPN distribution is log-normal.
The RPN statistic is calculated for the genic as well as intergenic regions to
identify the noise. The algorithm models the RPN for both exonic (collapsed to
genic level) and noncoding/intergenic regions, and uses it to generate p-values for
each gene and obtain an RPN cutoff (Fig. 1). Using this RPN cutoff, the sample
level presence/absence of a gene expression is estimated.
Fig. 1 RPN distribution across the genome for genic and intergenic (noise) along with regions of
absence, marginal, and presence
3.1 Case Study (RNA)
In order to test and validate our algorithm, we have used RNA-Seq data from 20
samples of the lymphoblastoid cell lines generated by Lappalainen et al. [13]
available from ArrayExpress (E-GEUV-1). The RPN statistical model for each
sample results in two distributions, viz., genic (red) and intergenic (blue). The gene
level presence/absence expression for each sample can be estimated using the genic
distribution showing two expression profiles through p-value cutoffs. Figure 1
shows the distribution of log RPN for the genic and intergenic regions for the
sample HG00096. Table 1 shows a few records from the final PA-call output for
this example dataset. This table gives log (RPN), presence/absence call, and the call
confidence. Baseline tissue gene expression values (Fragments Per Kilobase per
Million mapped reads (FPKM) shown in last column of Table 1) from Expression
Atlas [14] have been used to cross validate the results.
In the above table, “-Inf” in log (RPN) column represents that no reads were aligned
to the particular gene. Using the log (RPN) cutoff value obtained from the genic and
intergenic distributions, Presence (P)/Absence (A)/Marginal (M) expression calls are
identified as shown in Sample PA Call (column 3). The confidence levels (ranging
from 0 to 1) for these calls are given in column 4 (Call Confidence). The results can be
compared with the tissue baseline expression values (Column 5) based on the tissue
of origin.
3.2 Estimating Presence/Absence of Genes Using Exome

Data
Variations in the DNA are responsible for various phenotypes. When these varia-
tions become pathogenic, they manifest into cancer and many other diseases. These
Table 1 The gene level presence/absence of expression for 10 genes as estimated by the
algorithm
Gene log (RPN) Sample PA Call confidence Tissue baseline
call (1 − p) (FPKM)
ACTL9 -Inf A 1 –
C140R64 −10.2146419812 A 1 –
ACTB 5.7318105034 P 0.9999999902 3439
RPS29 3.7241632036 P 0.9999984735 63
MBD2 1.9461681249 P 0.9999332766 31
PARN 0.7630505519 P 0.9994225476 14
CARD11 0.1472238208 P 0.9984111854 23
OR2T5 −3.7027823593 M 0.8505041566 0.1
OR2T3 −3.7802292327 M 0.8738975872 0.3
HIST1H3I −3.8649313979 M 0.8996086247 0.2
224 M. Adhil et al.
Fig. 2 Mixture of two populations in the DNA data. a Distribution plot of localized CNV in
exome data. b Density plot of localized CNV in exome data
variations may include single base substitutions (single nucleotide polymorphism—

SNP), small insertions, and deletions (indels), or rearrangements and copy number
variations (CNV) together with epigenetic changes [15]. All these variations are
heterogeneous in nature and represented through categorical variables. For quan-
titative analysis of CNV, however, we used quantitative matrix of the read abun-
dance at each exonic location. We use exonic limits as breakpoints to quantify these
CNV. The measure of DNA variation is presented as SPKMG (Sequence Per
Kilobase of exon, per Megabase of the mappable Genome) and already reported in
our earlier publication [16]. We showed that this quantitative measure of DNA
demonstrates a normative pattern with higher level of statistical evidence with
phenotypic association. We also showed that localized CNV works very well to
discover DNA biomarkers [16] for both single patient and a population. The
deletion of a gene in exome can be assumed similar to low expression as in RNA.
The amplification of genes is similar to high expression. Therefore, in principle the
localized CNV will show similar characteristics like the RNA. The frequency
distribution and the density plot of localized CNV are shown in Figs. 2a, b,
respectively. The localized CNV for this graph are taken from the familial breast
cancer data generated by Gracia-Aznarez et al. [17]. This also indicates that the
same model described earlier for RNA-Seq will be applicable to identify the genes
that are amplified or deleted in a DNA-Seq data.
4 Conclusion
Genomic data processing faces combinations of complex challenges related to big

data, high dimensionality, and multimodal distribution. Knowledge discovery in big
data genomics is often top-down. Therefore, for genomics-based molecular
pathology tests for clinical that can lead towards precision medicine, we must
address all these challenges algorithmically. Here in this paper, we have presented
the presence–absence algorithm for genomic big data analysis. We have shown that
RNA data from RNA-Seq experiments and the localized CNV data from exome are
multimodal. The limited data samples and high dimensionality problems can be
addressed efficiently using statistical methods. We have presented here an analytic
algorithm for processing such genomic data and identified accurately the presence
or absence of genes in both RNA and DNA data from NGS experiments, where the
volume of a single patient data is upward of 10 GB. We also showed that the
multimodal challenges in genomic data can be addressed using the Gaussian
mixture model. We compared the result of our model with the experimental results
from other experiments and bibliomic data. Results from our model match with the
experiments performed by other independent studies. This shows that our algo-
rithms we designed are able to solve all these challenges of precision medicine
accurately and can be used in clinical setups.
References
1. Eisenstein, Michael. “Big data: the power of petabytes.” Nature 527.7576 (2015): S2–S4.
2. Bock, Hans-Hermann, and Edwin Diday, eds. Analysis of symbolic data: exploratory methods
for extracting statistical information from complex data. Springer Science & Business Media,
2012.
3. Morley, Michael, et al. “Genetic analysis of genome-wide variation in human gene
expression.” Nature 430.7001 (2004): 743–747.
4. Ried, Thomas, et al. “Genomic changes defining the genesis, progression, and malignancy
potential in solid human tumors: a phenotype/genotype correlation.” Genes, Chromosomes
and Cancer 25.3 (1999): 195–204.
5. Kitano, Hiroaki. “Computational systems biology.” Nature 420.6912 (2002): 206–210.
6. Maniatis, Tom, Stephen Goodbourn, and Janice A. Fischer. “Regulation of inducible and
tissue-specific gene expression.” Science 236 (1987): 1237–1246.
7. Komura, Daisuke, et al. “Noise reduction from genotyping microarrays using probe level
information.” In silico biology 6.1, 2 (2006): 79–92.
8. Schwartz, Schraga, Ram Oren, and Gil Ast. “Detection and removal of biases in the analysis
of next-generation sequencing reads.” PloS one 6.1 (2011): e16685.
9. Trapnell, Cole, et al. “Differential gene and transcript expression analysis of RNA-seq
experiments with TopHat and Cufflinks.” Nature protocols 7.3 (2012): 562–578.
10. iOMICS-Research Version 4.0.
11. Reynolds, Douglas. “Gaussian mixture models.” Encyclopedia of biometrics (2015):
827–832.
12. Moon, Todd K. “The expectation-maximization algorithm.” IEEE Signal processing
magazine 13.6 (1996): 47–60.
13. Lappalainen, Tuuli, et al. “Transcriptome and genome sequencing uncovers functional
variation in humans.” Nature 501.7468 (2013): 506–511.
14. Petryszak, Robert, et al. “Expression Atlas update—an integrated database of gene and
protein expression in humans, animals and plants.” Nucleic acids research (2015): gkv1045.
15. Pleasance, Erin D., et al. “A comprehensive catalogue of somatic mutations from a human
cancer genome.” Nature 463.7278 (2010): 191–196.
226 M. Adhil et al.
16. Talukder, Asoke K., et al. “Tracking Cancer Genetic Evolution using OncoTrack.” Scientific
Reports 6 (2016).
17. Gracia-Aznarez, Francisco Javier, et al. “Whole exome sequencing suggests much of
non-BRCA1/BRCA2 familial breast cancer is due to moderate and low penetrance
susceptibility alleles.” PloS one 8.2 (2013): e55681.
A Survey on Service Discovery Mechanism
Gitanjali Shinde and Henning Olesen
Abstract A wide range of web services are available, with a deployment of a huge
number of the Internet of Things (IoT) applications, e.g., smart homes, smart cities,
intelligent transport, e-health and many more. In the IoT paradigm, almost every
device around us will be on the Internet to provide specific services to end users.
A large number of services will be available to help end users, however, providing
and selecting appropriate service from a massive number of available services is a
challenge of IoT. Service discovery plays a vital role in web service architecture.
Traditionally, keyword search has been used for service discovery. The traditional
approaches of service discovery cannot be applied as it is for web services in the
IoT environment as devices providing, and accessing services are resource con-
strained in nature, e.g., limited size, limited energy source, limited computational
capability, and limited memory. Hence, there is need of a lightweight and semantic
service discovery approach for the IoT. This work analyzes different service dis-
covery approaches. This analysis will focus on the key requirements of service
discovery mechanism needed for the IoT environment.

Keywords Internet of Things (IoT) Service description Service discovery

Semantic service discovery User preference
1 Introduction
In the IoT environment, devices around us are connected to the Internet to provide
services to the end users. IoT devices can be anything around us, i.e., sensors, RFID
tag, RFID reader, actuators, etc. The different IoT devices may use different
G. Shinde (&) H. Olesen

Center for Communication, Media and Information Technologies (CMI),
Aalborg University Copenhagen, 2450 Copenhagen SV, Denmark
e-mail: gis@es.aau.dk
H. Olesen
e-mail: olesen@cmi.aau.dk

https://doi.org/10.1007/978-981-10-7245-1_23
228 G. Shinde and H. Olesen
wireless/wired technology to transmit data to cluster head, i.e., WLAN, ZigBee,

NFC, Zwave, WiFi [1]. These devices may be different in sizes, computational
capability; hence heterogeneity will be the major challenge to design WS. IoT will
be deployed in large number and almost in all application domains, i.e., e-health,
smart home, the smart city [2], intelligent transport, logistics and many more.
Hence, a large number of devices will be working with each other over the Internet
to provide services to the end users. End users access a web service
(WS) depending on their requirements, hence selecting appropriate service from a
large number of available services is an essential design parameter for WS.
The remainder of this paper is structured as follows: the Motivation is discussed
in Sect. 2, the Detailed Literature of Service Discovery Mechanisms is discussed in
Sect. 3, Comparative Analysis of literature work is presented in Sect. 4, and the
Discussion is summarized in Sect. 5.
2 Motivation
Predicting and providing appropriate service depending on context and user pref-
erence is the necessity of IoT environment. The user may be mobile and will require
accessing services via his/her smartphone; therefore the lightweight WS discovery
process is an important design parameter of WS architecture for IoT. In the IoT
environment, devices around us sense information and depending on this infor-
mation services are built. Every IoT device is not capable of building and providing
WS as IoT devices are resource constrained in nature. Hence, few devices are
grouped together as a cluster, and one computationally strong device is selected as
cluster head to perform complex task [3], i.e., building and providing WS. The
information senses by the IoT devices are raw and unstructured in nature, so there is
need of describing services depending on the raw and unstructured information by
converting it into a structured format. The need of clustering IoT devices and
lightweight WS discovery is shown in Fig. 1.
The traditional WS discovery methods are not applicable to IoT environment
due to its different requirements, hence it is important to review different WS
discovery mechanisms. WS are the description of operations that perform a certain
task. WS are language and platform independent. WS can perform a simple task or
complex task. A WS has three components, which include service provider, service
registry, and service requestor. Service provider describes services using a standard
Extensible Markup Language (XML), i.e., Web Service Description Language
(WSDL). A service is advertised by publishing it to a registry. Universal
Description Discovery and Integration (UDDI) is the registry used to publish web
services. End users are the service requestors, requesting for web service using find
operation. WS has a four-phase life cycle, which includes Build, Deploy, Run, and
Manage.
The build phase of life cycle plays very important role in the WS life cycle to
provide appropriate service to end user when the description of WS is matched with
A Survey on Service Discovery Mechanism 229
Fig. 1 Requirement of service discovery mechanisms for IoT environment
requirements of the end user. WS uses the Simple Object Access Protocol (SOAP)
to access WS, and it is not suitable for the WS architectures in the IoT environment
as a SOAP-based approach are heavyweight in nature. Traditionally, syntactic
service discovery mechanism is used to discover the WS. A syntactic service dis-
covery mechanism results in a list of services among that user has to select
appropriate service depending on his/her requirement, and this is not feasible
approach for IoT environment. To provide appropriate service to end user, the
service description should be done in a semantic way, so that in the service dis-
covery phase, a service can be searched very efficiently.
The WSDL, SOAP, and UDDI standards are used to build service architecture.
However, the WSDL is not designed for semantic description. WSDL describes WS
in the form of a port, port type, and bindings. Hence, it is very difficult to specify
semantic details about WS, i.e., nonfunctional requirements, QoS requirements,
availability, reliability, throughput, response time, and mean time before failure.
This invokes the need for a semantic description of services. To provide the
semantic descriptions, different approaches are invoked, i.e., OWL-S, WSDL-S.
However, these methods have few limitations, i.e., a huge number of services exist
are described using WSDL format and it is not feasible to add semantic in the
existing services, every new service cannot have a semantic description with it and
service requestor may be unaware of terms related to the service. UDDI is used to
publish a WS, however, the runtime parameters cannot be stored in the UDDI and
there is no universal registry for storing WS. Hence, it is very difficult to store
runtime parameters to improve WS description [4].
A few challenges of WS architectures for IoT environment are as follows:
• Information sensed by IoT device is heterogeneous, raw, and unstructured
• IoT devices are resource constrained in nature, i.e., limited energy source,
limited computational capability, limited memory, and limited size
• Traditional WS discovery mechanism is heavyweight and keyword search based
• User-centric approach is not considered in the traditional service discovery
mechanisms
• The semantic service description is required for WS.
In the near future, IoT devices around us will offer a huge number of services to
end users for betterment and comfortable life. The end user must get services
depending on their requirement. However, the IoT comes with challenges
mentioned above. This invokes the need for a lightweight and semantic service
architecture for IoT. In the next section, review of different approaches of service
discovery is done. This analysis helps find out design parameters of service archi-
tecture for IoT environment.
3 Related Work
The service description is the first and important step of service architecture as the
efficiency of service discovery is dependent on the service description. In the lit-
erature, service description is done by three different approaches, such as func-
tional, nonfunctional, and data structural [5]. To provide appropriate service
depending on a user’s requirement, the need of betterment in a service description is
invoked. Semantic service description is the step to fulfill the need mentioned
above. The semantic service description can be done using the following two
approaches: annotation based and semantic language based. The annotation-based
approach improves the expressivity of service description by providing an anno-
tative description for data types, inputs, and output of WS. In the IoT paradigm,
devices around us will provide services to end users. A large number of services
will be available as IoT applications will be deployed in all domains. In order to
provide appropriate and useful service to the end user, efficient service discovery
mechanisms are needed. Traditionally, services are registered in the UDDI registry;
the services are searched using the keyword search mechanism. The UDDI-based
service architecture faces the limitation due to a centralized approach, hence
different ways of deploying service architectures are discussed by Demian et al. [6],
i.e., centralized and distributed. Centralized approach is categorized into the portal
and UDDI-based architecture. UDDI-based architecture is further categorized into
hybrid and portal-based architecture. Distributed architecture approach is catego-
rized into three different ways, such as agent-based, Internet-based, and point to
point (P2P) based architecture. Among these different ways of WS architecture,
appropriate architecture is used depending on the requirement of the application.
The ontology-based services discovery is proposed by Paliwal et al. [7]. Service
discovery is achieved by functional level service categorization and semantic
enhancement of service request. The service categorization is done at UDDI by
performing the following two steps: parameter based service refinement and
semantic similarity based matching. Service refinement is done by combining
semantic characteristics with syntactic characteristics in WSDL description. In this
approach, each service is represented as vector, i.e., Service Description Vector
(SVD). Service refinement is done by improving semantic contents of the SVD.
Services having similar functionality may be grouped in different categories at
UDDI, such services from different categories are grouped together in the cluster.
These clusters are labeled depending on their functionality for efficient service
discovery. To label the cluster, an ontology concept is used. Services from the
cluster are mined to find the pattern to satisfy the user’s request. Hyperclique
patterns are used for this purpose. A user’s request is enhanced using Latent
Semantic Indexing (LSI) technique. In this approach, the service refinement and
discovery is dependent on the WSDL file. It is impractical that every existing
service will have a semantic description with WSDL file.
Logic-based and knowledge-based service search and discovery are proposed by
Cassar et al. [8]. The latent factor based approach is used to describe the service.
Latent factors are recovered from service description by probabilistic machine
learning approach. A latent factor is related to the semantic description of web
service and can be expressed as probability distribution function. In this, Ontology
Web Language for Services (OWL-S) is used for service description. To retrieve
latent factor from the description, it is converted into Service Transaction Matrix
(STM). The STM provides a training set, and it is served as input to probabilistic
machine learning technique to extract latent factors. Clustering of relevant services
is done on the basis of latent factors using Latent Dirichlet Allocation (LDA).
Service discovery process works efficiently on clustered services.
Meditskos et al. [9] proposed service profile ontology-based service description.
Object-based structural matching technique and logic reasoning methods are used
for match making process of service discovery. Ontology roles are combined with
description logic reasoning for improved service discovery process. The compar-
ison between the Complex Concept (CC) and Service Profile (SP) description of
WS description is proposed in this work. The four ontology levels are used to
describe WS, such as Service, Service Profile, Service Process, and Service
Grounding. In the profile-based discovery mechanisms, the service request and
service advertisement are both represented as the profile. The profile-based dis-
covery gives discovery results very efficiently.
Bootstrapping ontology mechanism for service discovery is proposed by Segev

et al. [10]. The Term Frequency and Inverse Document Frequency (TF/IDF)
methods are used to evaluate the WSDL description. These methods review the
service description from internal and external perspective. Lastly, free text
description method is used to resolve the inconsistencies in the service description
ontology. It works in four steps: token extraction, token evaluation, concept evo-
cation, and ontology evolution. Initial filtering and extraction of the labels from
WSDL document are done in the token extraction step. In the second step, the
tokens are evaluated by TF/IDF and Web Context Extraction (WCE) method. The
TF/IDF method searches for the token that presents most frequently in its WSDL
document, but least present in WSDL documents of other services, i.e., token that
can represent service in a unique way. In a WCE method, group of tokens are
extracted from WSDL file of service that can represent the context of service. In the
concept evocation, step tokens extracted from both methods are evaluated, and
tokens present in both methods are selected to describe the service. In the last step,
current ontology is compared with newly extracted descriptors. In this work, the
description of WS is improved for efficient discovery.
Chakraborty et al. [11] proposed service discovery for pervasive computing. In
the pervasive environment, usually services are advertised in a broadcast way.
However, the global broadcasting has limitations, such as broadcasting of services
cannot be applied to the large-scale network as it works poorly with increasing
network size and it uses the more resources and power. In the proposed method,
solution to overcome the drawbacks of global broadcasting is discussed. It works in
two steps: service caching and group-based advertisement of the service request.
Services are cached at every peer and service requests are forwarded in the
group-based approach. Every group of nodes advertises the services provided by
the group in an abstract way. Hence, service request can be easily forwarded to the
appropriate group. The group-based approach minimizes the efforts to search
appropriate node by providing the service demanded by the user.
WS description, advertisement, execution, and discovery using DARPA Agent
Markup Language (DAML) are discussed by McIlraith et al. [12]. Descriptive API
is used to provide service specification, service capabilities, interfaces, and pre-
requisites for service execution. DMAL Agent-based technology is used in this
work to discover WS depending on the user preferences and constraints. Zisman
et al. [13] proposed the need and solution of runtime service discovery. Service
replacement is needed due to various reasons, such as unavailability or malfunc-
tioning of service, changes in the structure of interface or functionality of service,
and better service is available to complete the associated task. The pull and push
model of runtime service discovery is used to replace the service that is no longer
needed. The requirements of Quality of Service (QoS) parameters for WS
description and discovery are discussed by Kritikos et al. [14]. QoS parameters are
categorized into two types, such as domain dependent and domain independent.
QoS broker service architecture is proposed in this work. The broker plays the
middle layer between user and WS. Analysis of different QoS parameters helps to
describe WS semantically with nonfunctional qualities.
Guinard et al. [15] proposed the RESTful service architecture. Proposed

approach is targeted for the IoT environment. The IoT devices are resource con-
straints in nature, hence RESTful approach is used for WS. Device profile (DP) is
generated for each IoT device that provides services. DP provides services to the
users with minimal user interaction. Minimal user interaction is the need of the IoT
environment to achieve lightweight service architecture. The context awareness is
used as a key parameter for service discovery as in future network large number of
services will be available. Among such large number of services, only appropriate
service should be provided to the user, hence user preference and context are the
parameters for service discovery. In the proposed work, these requirements of the
future network are focused, e.g., RESTful services, context, and user preferences.
However, the access methodology for WS is not addressed in this work. Universal
Presence Server (UPS) based lightweight RESTful WS architecture is discussed by
Chunyan et al. [16]. The proposed architecture includes the following three entities:
Request Manager (RM), Publication Manager (PM), and Subscription Manager
(SM). The PM publishes the service description in the publication database.
The SM retrieves the list of services present from the UPS database. The SM
dynamically decides the interface agent to interact with the end user. The dynamic
decision of message interface makes this approach lightweight for accessing and
interacting with WS.
A vast literature work is done in the service architecture field to achieve the
dynamic, on-demand, lightweight, and context-aware service discovery. Thirteen
different approaches of service discovery are analyzed by Debajyoti et al. [17], such
as context-aware WS discovery, publish-subscribe model, keyword clustering,
service request expansion method, graph ranking, layer-based semantic discovery,
WS discovery techniques for heterogeneous networks, WS indexing, structural case
based reasoning, agent and QoS based discovery, collaborative tagging, using finite
automaton, and hybrid approach for WS discovery.
When the service discovery method is not capable of discovering service based
on a user’s requirement, then approximate WS is suggested. For such type of
process, the behavioral matching of WS is required. The user’s requirements are
described using the behavioral graph and the WS discovery process is converted
into graph matching process. From the analysis of different techniques of WS
discovery in this work, the importance of QoS parameters is focused. Using dif-
ferent semantic and syntactic discovery approach, a large number of services can be
discovered for the same type of function. Selecting one WS for a user’s requirement
will be more dependent on the nonfunctional parameters, such as QoS parameters.
A survey of QoS-based WS discovery approaches is done by Phalnikar et al.
[18]. The detailed categorization of QoS parameters is discussed in this work.
Ontology and syntactic approaches of WS Discovery with QoS parameters are
compared, and the requirements of ontology-based WS discovery are proposed in
the work, which include the automatic WS Discovery, WS invocation, and WS
composition. WS description is the base of WS discovery process, the efficiency of
discovery process increases with the semantic description of WS. Bitar et al. [19]
reviewed different ways of WS description methods. The need of lightweight
semantic description is proposed in this work. WS described using semantic method

then that can be efficiently discovered through semantic discovery. In the near
future, huge number of WS will be present around us, hence a need of semantic WS
discovery is invoked; tremendous research work is going in the direction to dis-
cover efficient protocol for WS discovery.
Review of different semantic WS discovery approaches is discussed in the
Ibrahim et al. [20] categorized the service matching processes into three types
depending on the heterogeneity level in the matching process, i.e., algebraic,
deductive, and hybrid approach. In an algebraic approach, the degree of similarity
between WS descriptions is taken as a parameter to discover appropriate service for
user’s requirement. Structured graphs are used to find out the degree of similarity
between WS descriptions. In a deductive approach, the WS discovery is based on
logic, such as description logic and first-order logic. In this approach instead of
semantic different, other elements of WS description are considered for service
matching process, which include Preconditions and Effect matching (PE matching),
Input and Output matching (IO matching), and Input, Output, Precondition and
Effect matching (IOPE).
4 Comparative Analysis
In the above section, different prominent techniques of WS discovery are discussed.

In view of realizing IoT application, there is need of dynamic WS architecture,
hence this literature survey will help to figure out the designing parameters of WS
architecture for IoT environment. In the IoT environment, end users may be mobile
and may not be aware of services provided at that location. In each location, there
may be a large number of services present. Among these services, end users must
be made aware of only a few services that are required for them. Therefore, the
context-aware and user preference aware WS framework is required in the IoT
environment. Thinking ahead, the IoT devices that will provide and access WS are
resource constraint in nature. Thus, there is a requirement of lightweight WS
architecture. That is why user preference, context awareness, and lightweight
framework parameters are taken for comparative analysis.
Table 1 shows whether approach support these three parameters and summarizes
the applicability of WS discovery approaches for IoT environment.
UPerf—User Preferences, CAW—Context Awareness and LW—Lightweight.
Requirements of WS architecture for IoT environment are as follows:
• Semantic description of WS
• Automatic on-demand WS discovery
• User-centric approach
• Context awareness approach
• Lightweight registration process for WS
• RESTful approach.
Table 1 Analysis of WS discovery approaches

Approaches Discovery technique UPref CAW LW
[7] Semantic with Latent Semantic Indexing No Yes No
[8] Semantic with Service Transaction Matrix and No Yes No
Latent Dirichlet Allocation
[9] Service profile ontology-based No Yes No
[10] Term Frequency and Inverse Document Frequency No Yes No
and web context generation
[11] Group-based service request forward approach No No Yes
[12] DARPA agent Markup Language Yes Yes No
[13] pull and push model No Yes No
[14] Quality of Service Yes Yes No
[15] RESTful Approach Yes Yes Yes
[16] Universal Presence Server No No Yes
5 Conclusions and Future Work
In order to provide WS framework for IoT environment, it is important to determine

the requirements of WS architecture for IoT network. Hence, we have reviewed the
literature for service description and service discovery methods. In the literature,
tremendous work is done for service architecture, however, these approaches are
not appropriate to apply for IoT settings as devices that will provide and access
services in IoT paradigm are resource constrained in nature. This survey work will
help to determine the designing parameters for WS architecture for IoT network.
The user-centric approach is required in the near future as a large number of
services will be available, among such huge number of services user must get
appropriate service depending on a user’s requirements. The future plan is to
explore a semantic service description method and lightweight, user-centric, and
context-aware service discovery methods for IoT network.
References
1. J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internet of Things (IoT): A vision,

architectural elements, and future directions,” Futur. Gener. Comput. Syst., vol. 29, no. 7,
pp. 1645–1660, Sep. 2013.
2. A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi, “Internet of Things for Smart
Cities,” IEEE Internet Things J., vol. 1, no. 1, pp. 22–32, Feb. 2014.
3. G. Shinde and H. Olesen, “Interaction between users and IoT clusters : Moving towards an
Internet of People, Things and Services (IoPTS),” in World wireless Research Forum Meeting
34, 2015.
4. M. K. Nair and V. Gopalakrishna, “Look Before You Leap: A Survey of Web Service
Discovery,” Int. J. Comput. Appl., vol. 7, no. 5, pp. 975–8887, 2010.
5. H. Omrana, I. El Bitar, F. -Z. Belouadha, and O. Roudies, “A Comparative Evaluation of Web

Services Description Approaches,” in 2013 10th International Conference on Information
Technology: New Generations, 2013, pp. 60–64.
6. D. A. D’Mello and V. S. Ananthanarayana, “A Review of Dynamic Web Service Description
and Discovery Techniques,” in 2010 First International Conference on Integrated Intelligent
Computing, 2010, pp. 246–251.
7. A. V. Paliwal, B. Shafiq, J. Vaidya, Hui Xiong, and N. Adam, “Semantics-Based Automated
Service Discovery,” IEEE Trans. Serv. Comput., vol. 5, no. 2, pp. 260–275, Apr. 2012.
8. G. Cassar, P. Barnaghi, and K. Moessner, “Probabilistic Matchmaking Methods for
Automated Service Discovery,” IEEE Trans. Serv. Comput., vol. 7, no. 4, pp. 654–666,
Oct. 2014.
9. G. Meditskos and N. Bassiliades, “Structural and Role-Oriented Web Service Discovery with
Taxonomies in OWL-S,” IEEE Trans. Knowl. Data Eng., vol. 22, no. 2, pp. 278–290, Feb.
2010.
10. A. Segev and Q. Z. Sheng, “Bootstrapping Ontologies for Web Services,” IEEE Trans. Serv.
Comput., vol. 5, no. 1, pp. 33–44, Jan. 2012.
11. D. Chakraborty, A. Joshi, Y. Yesha, and T. Finin, “Toward Distributed service discovery in
pervasive computing environments,” IEEE Trans. Mob. Comput., vol. 5, no. 2, pp. 97–112,
Feb. 2006.
12. S. A. McIlraith, T. C. Son, and Honglei Zeng, “Semantic Web services,” IEEE Intell. Syst.,
vol. 16, no. 2, pp. 46–53, Mar. 2001.
13. A. Zisman, G. Spanoudakis, J. Dooley, and I. Siveroni, “Proactive and Reactive Runtime
Service Discovery: A Framework and Its Evaluation,” IEEE Trans. Softw. Eng., vol. 39, no.
7, pp. 954–974, Jul. 2013.
14. K. Kritikos and D. Plexousakis, “Requirements for QoS-Based Web Service Description and
Discovery,” IEEE Trans. Serv. Comput., vol. 2, no. 4, pp. 320–337, Oct. 2009.
15. D. Guinard, V. Trifa, S. Karnouskos, P. Spiess, and D. Savio, “Interacting with the
SOA-Based Internet of Things: Discovery, Query, Selection, and On-Demand Provisioning of
Web Services,” IEEE Trans. Serv. Comput., vol. 3, no. 3, pp. 223–235, Jul. 2010.
16. C. Fu, F. Belqasmi, and R. Glitho, “RESTful web services for bridging presence service
across technologies and domains: an early feasibility prototype,” IEEE Commun. Mag., vol.
48, no. 12, pp. 92–100, Dec. 2010.
17. D. Mukhopadhyay and A. Chougule, “A survey on web service discovery approaches,” in
Advances in Intelligent and Soft Computing, 2012, vol. 166 AISC, no. VOL. 1,
pp. 1001–1012.
18. R. Phalnikar and P. A. Khutade, “Survey of QoS based web service discovery,” in 2012
World Congress on Information and Communication Technologies, 2012, pp. 657–661.
19. I. El Bitar, F. -Z. Belouadha, and O. Roudies, “Review of Web services description
approaches,” in 2013 8th International Conference on Intelligent Systems: Theories and
Applications (SITA), 2013, pp. 1–5.
20. I. El Bitar, F. -Z. Belouadha, and O. Roudies, “Semantic web service discovery approaches:
overview and limitations,” Sep. 2014.
Analysis of Multiple Features
and Classifier Techniques Combination
for Image Pattern Recognition
Abstract Automatic visual pattern recognition is complex and highly researched

area of image processing. This research aims to study various pattern recognition
algorithms, cloth pattern recognition is presented as research problem and to find
out best combination suited for the cloth pattern recognition problem. The dataset is
collected from CCNY clothing pattern dataset and contains 150 samples of each
category (Patternless, Striped, Plaid, and Irregular). The presented study compares
all combinations of three different feature extraction techniques and three classifier
techniques. Feature extraction techniques used here are Radon Feature Extraction,
projection of rotated gradient, and quantized histogram of gradients. The classifiers
used are KNN, neural network, and SVM classifier. The highest recognition rate is
achieved using Radon Signature feature and KNN classifier combination which
reaches to 93.7% of accuracy.
Keywords Clothing pattern recognition Texture analysis Image feature

extraction Classifier
1 Introduction
Image pattern recognition is a tool for machine or a computer system to understand

the pattern which makes it possible humans to recognize objects. Pattern recogni-
tion is also a simplest way of distinguishing between multiple objects depending on
a single property which is unique to that object. Patterns are repetitive visual
A. Shinde (&)
Sinhgad College of Engineering, Pune 41, Maharashtra, India
A. Shinde
Bhima Institute of Management and Technology, Kagal,
Kolhapur 416216, Maharashtra, India

https://doi.org/10.1007/978-981-10-7245-1_24
markers present on an object’s surface. Here, a comparative study is presented to

show compatibility between a feature recognition system and classifier.
Feature extraction techniques are basis of smart machine vision algorithms. The
understanding of data by machine usually comes from a property of data type by
which it can be distinguished from other image. For image pattern recognition, this
is the most important step too. Features can be of various types. Features such as
color, edges, size, etc., which are primary visual features while histogram, pro-
jections, spatial texture, etc., are much complex types of feature [1].
Multiple combinations of feature extraction and classifier are compared in this
research to ensure a broader understanding of machine vision algorithms.
A lot of efforts have already been taken in terms of pattern recognition for cloth
classification. Hidayati et al. [2] proposed a methodology for classification of
clothes based on genres. Davis et al. [3] proposed a new method for “texture
analysis based on the spatial distribution of local features” for unsegmented image
textures. Generalized co-occurrence matrices (GCM) are used for texture descrip-
tion and feature extraction.
Eisa et al. [4] proposed the use of the LBP feature for analyzing texture feature.
This feature extraction technique presents an effective way for both structural and
statistical texture analysis. In approach proposed by Kalantidis et al. [5], they adopt
the locality sensitive hashing (LSH), a very popular and effective high-dimensional
indexing technique. Here, it is applied specifically for cloth pattern classification.
2 Feature Extraction for Pattern Recognition
In this study, three different feature extraction methods have been considered.
Radon Signature characterizes the directionality feature of clothing patterns.
Projection of Rotated Gradient takes x-projection of images which are rotated
stepwise at a given angle to create a multidirectional projection array. Histogram of
Oriented Gradient takes histogram of gradient taken one by one from multiple
direction of original image.
2.1 Radon Signature Features
Radon Signature is feature extracted using the Radon transform [6] which is
commonly used to extract information of the principle orientation of an image. The
image is then rotated to find dominant direction represented by projection of the
image intensity along a radial line oriented at a specific angle. A two-dimensional
Radon transform of function is given as [6]:
Analysis of Multiple Features and Classifier Techniques … 239
Z1 Z1
Rðr; ;Þ ¼ I ðx; yÞdðr x cos ; y sin ;Þdxdy ð1Þ
1 1
where ‘U’ is the projection line angle and ‘r’ is the perpendicular distance of a
projection line to the origin.
To ensure the consistency in variance of Radon transform for multiple projection
orientations, Radon transform has been applied on the image selected by maximum
radius area instead of the complete image. The large intra-pattern variations of
clothing are also reflected as shown in Fig. 1.
The probable orientation of image is represented by variance in radon signature
calculated by equation [6]
1XN 1
^ ðr; ;i Þ 2
varðr; ;i Þ ¼ R r j ; ;i R ð2Þ
N j¼1
where
X
N 1
^ ðr; ;i Þ ¼ 1
R R r j ; ;i ; ð3Þ
N j¼1
Fig. 1 Radon signature feature extraction. a Input image. b Disk area selected for calculating
radon signature. c Radon signature of image calculated at 0–180°. d Radon signature variance used
as extracted image feature
Fig. 2 Different cloth pattern images (Irregular, Plaid, Striped and Patternless) and their
corresponding radon signature variance
where R(r, U) is value of perpendicular projection at distance r and angle U and

Ȓ(r, U) is its expected value.
The Radon Signature feature array is extracted by concating all variances sig-
natures acquired at all projection directions
½varðr; ;1 Þ; varðr; ;2 Þ; varðr; ;3 Þ; varðr; ;4 Þ; . . .varðr; ;N1 Þ
To enhance the large variations in image intensity, image has been filtered using
the X/Y-Sobel operator to create the gradient image as in Eq. (1). Figure 1b illus-
trates the area selected for Radon transform and Fig. 1c demonstrates the radon
signature. The selected feature is shown in Fig. 1d which is variance in Radon
Signature (Fig. 2).
2.2 Projection of Rotated Gradient Features
Rotated gradient images are edge enhanced versions of multidirectional projection

of an image [7]. An image can be enhanced by taking its gradient. However,
gradient taken from a single direction will only enhance its limited edges. By taking
gradient from multiple directions ensures image is collectively enhanced and hence
gives higher degree of information about the features of that image. The gradient
enhancement of an image is done using Sobel operator given by [1, 2, 1; 0, 0, 0; −1,
−2, −1] which is x-directional Sobel operator and its rotated version [1, 0–1; 2, 0,
−2; 1, 0, −1] is a y-directional Sobel operator. The resulting image is given by
Gðx; yÞ ¼ ff ðx; yÞ sðx; yÞg ^

Sðx; yÞ; ð4Þ
where s(x, y) and ŝ(x, y) are x-direction and y-direction sobel operator.
The rotation of image here is by 90° each time hence there will four concated
gradient images in total. The projections of gradient are
X
no:ofcolumns
Pð xÞi ¼ G x; yj i ð5Þ
j¼1
2.3 Quantized Histogram of Orientated Gradient Features
Quantized oriented gradient is an enhanced version of HOG features. HOG has

robustness to change of illumination and attains high computational accuracy in
detection of variously textured objects [8]. HOG features are calculated by taking
histogram of image created in by generating gradient images with different orien-
tation. Here too gradients were calculated using Sobel operator given by [1, 2, 1; 0,
0, 0; −1, −2, −1] which is x-directional Sobel operator and its rotated version [1, 0–
1; 2, 0, −2; 1, 0, −1] which is a y-directional Sobel operator (Figs. 3, 4 and 5).
The histograms were however quantized at bin of 8, i.e., each histogram array
consisted of eight values. From this histogram, a single gradient descriptor was
calculated by calculating highest value of each bin. This value describes the average
orientation of that sub-image. Each of such histograms was calculated for whole
image by creating a block of 8 8, which gives 64 values for an angle theta. The
theta was varied seven times further to calculate the whole feature vector (Fig. 6).
The calculated HOG features from each block are as follows (where length of
each vector is 8 and n = 64):
Fig. 3 Sample pattern image and its projection of gradient feature. a Input image. b Rotated
version of input image. c Gradient image of each rotated image. d Projection of gradient images
Fig. 4 Different cloth pattern images (Irregular, Plaid, Striped and Patternless) and their
corresponding projection of gradient feature
Fig. 5 Flow of calculation of Q-HOG feature
HOGf8; n; mg ¼ ½HOG1 ; HOG2 ; HOG3 ; . . .; HOGn m¼1;2...8 ð6Þ
QHOGfn; mg ¼ ½QHOG1 ; QHOG2 ; QHOG3 ; . . .; QHOGn m¼1;2...8 ð7Þ
3 Comparative Implementation and Analysis
The example to showcase this study uses dataset of cloth patterns with four labels.
The evaluations are done using CCNY dataset of cloth patterns. It contains 150
samples each belonging to Striped, Plaid, Random, and Patternless category. The
feature extraction is done using Radon Signature, Projection of Rotated Gradients
(PRG), and Histogram of Oriented Gradients (HOG).
The research aims toward comparison of compatibility between feature extraction
and feature classifier for given example of pattern dataset of cloths. CCNY Clothing
Fig. 6 Example image of each pattern in first row. Second row is histograms of all 8 directed
gradient of each image in row one. Third row is final feature vector for each image in row one
Fig. 7 Example images for each pattern. Irregular, Plaid, Striped and patternless in order
Pattern Dataset includes four different typical clothing pattern designs: Plaid,
Striped, Patternless, and Irregular with 150 images in each category. The resolution
of each image is 140 140. Figure 7 shows example images in each category.
Each of the dataset is split in training set and recognition set with varying
percentage of split. The training dataset images are processed using all pattern
recognition techniques described in Sect. 2. Here, three different classifiers are used
namely:
1. Feedforward neural network with 10 hidden layers and Levenberg–Marquardt
optimization,
2. KNN classifier with neighborhood-1 and
3. Multilabel SVM classifier.
Different combinations of training data set splitting as well as multiple feature

extraction-classifier pairing ensured thorough analysis and study of both feature
extraction technique and classifier techniques.
150 samples have been considered belonging to each label, i.e., pattern types. The
data was split for Training Data-Test Data in various combinations of 10–90, 30–
70, 50–50, and 70–30% of dataset.
The recognition for all combinations shows a consistent growth except for
Q-HOG and NN combination which do not show consistent growth. NN also
performs poorly in combination with PRG. SVM however less accurate than KNN
is quite consistent with varying size of training data when used in combination
RadSig and Q-HOG (Fig. 8; Table 1).
Fig. 8 Comparison of recognition accuracy of all feature extraction and classifier combinations at
a 70% Training data b 50% Training data c 30% Training data d 10% Training data
Table 1 Change in recognition rate in percent with increase in number of training samples
Training 70% 50% 30% 10%
sample
used
NN KNN SVM NN KNN SVM NN KNN SVM NN KNN SVM
RADON 85.8 93.7 88.1 78.6 90.5 87.4 80.2 90.5 82.6 78.6 82.6 88.9
PRG 50 82.6 49.3 46.9 69.1 38.1 42.1 65.9 48.5 46.1 44.5 25.4
Q-HOG 51.6 88.1 90.5 53.2 80.2 77.8 56.4 75.4 81.8 52.4 59.6 69.9
5 Conclusion
The most consistent results were achieved with Radon Signature Feature-SVM
classifier. Highest results however were obtained at 70% training data using Radon
Signature Feature-KNN features. Also KNN performed comparatively better than
all other classifiers in all combinations consistently.
In terms of operation speed training of neural network took longest during
training. KNN do not need special training while its recognition time was shortest
making it the fastest classifier technique. Q-HOG took longest for feature training
while Radon Signature feature took shortest. In terms of the algorithm complexity
represented mathematical calculations, the PRG proved simplest while Q-HOG
hardest.
All considerations resulted in conclusion that RadSig-KNN combination is best
feature extraction-classifier combination for presented scenario.
The aim of research is to find out which combination of feature
extraction-classifier produces best results for a given problem. The same scenario
can be recreated using a different dataset of patterns of other types, which may
include wood pattern, wall color pattern, skin color pattern, etc. The presented
research statement is flexible and can be introduced as a methodology of com-
parative analysis of a number of image processing problems.
References
1. Van Gool L, Dewaele P, Osterlinck A (1983). Texture analysis anno 1983. Computer Vision
Graphics Image Processing 29(12):336–57.
2. S. Hidayati. W. Cheng, and K. Hua, “Clothing Genre Classification by Exploiting the Style
Elements,” In Proc. ACM Multimedia, 2012.
3. Larry S. Davis, Steven A. Johns, J. K. Aggarwal, “Texture Analysis Using Generalized
Co-Occurrence Matrices”, IEEE Transactions on Pattern Analysis and Machine Intelligence
(Volume: PAMI-1, Issue: 3, July 1979), pp. 251–259.
4. M. Eisa, A. ElGamal, R. Ghoneim and A. Bahey, “Local Binary Patterns as Texture
Descriptors for User Attitude Recognition,” International Journal of Computer Science and
Network Security, vol. 10 No. 6, June 2010.
5. Y. Kalantidis, L. Kennedy, L. Li, “Getting the Look: Clothing Recognition and Segmentation
for Automatic Product Suggestions in Everyday Photos,” ICMR, ACM 978-1-4503-2033-7/13/
04, April, 2013.
6. K. Khouzani and H. Zaden, “Radon Transform Orientation Estimation for Rotation Invariant
Texture Analysis,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 27, no. 6,
pp. 1004–1008, 2005.
7. Lin-Lin Huanga, Akinobu Shimizua, Yoshihoro Hagiharab, Hidefumi Kobatakea, “Gradient
feature extraction for classification-based face detection”, Pattern Recognition Volume 36,
Issue 11, November 2003, Pages 2501–2511.
8. S.K. Uma, Srujana B.J., “Feature Extraction for Human Detection using HOG and CS-LBP
methods”, International Journal of Computer Applications (0975–8887), 2015, pp. 11–14.
Smart and Precision Polyhouse Farming
Using Visible Light Communication
and Internet of Things
Krishna Kadam, G. T. Chavan, Umesh Chavan, Rohan Shah

and Pawan Kumar
Abstract Recently, Polyhouse farming has taken the place of traditional farming.
This technique reduces dependency on rainfall and dramatically changing envi-
ronmental conditions and makes the optimum use of land and water resources in a
controlled environment. Manually controlled system requires a lot of attention and
care. To overcome this limitation, novel technologies like Internet of Things
(IoT) and sensor network (SN) can be used. Case studies show that use of elec-
tromagnetic or radio spectrum for communication degrades the growth and quality
of crops. The visible light communication (VLC) provides an effective solution to
this problem using visible light spectrum for communication. VLC outperforms
when compared with traditional communication techniques in terms of available
frequency spectrum, reliability, security, and energy efficiency. LED source used in
VLC acts as communication element as well as is used for illumination. This paper
proposes use of sensor network to sense the data like temperature, humidity, soil
moisture, and luminosity along with VLC as a medium of communication from
node to the network gateway. IoT performs its part in automating polyhouse
countermeasures and fertilizer suggestion based on the data processed in cloud for
Business Intelligence and Analytics. Thus, the proposed system will ensure auto-
mated and sustainable polyhouse farming using VLC and IoT.

Keywords Internet of Things Sensor network Fertilizer suggestion

Cloud Visible light communication Business intelligence Analytics
K. Kadam (&) G. T. Chavan U. Chavan R. Shah P. Kumar

Sinhgad College of Engineering, Pune, India
e-mail: kjk12895@gmail.com
G. T. Chavan
e-mail: gtchavan.scoe@sinhgad.edu

https://doi.org/10.1007/978-981-10-7245-1_25
248 K. Kadam et al.
1 Introduction
Globally, traditional agriculture is one of the most practiced occupations. It heavily

depends on human labor and is surviving on the mercy of environment that is
deteriorating day by day. Technological advances in agriculture are providing much
easier ways to farm efficiently. Automation of various farming devices in open
farms like a sprinkler and drip irrigation has gained a lot of popularity in recent
times. Despite its popularity, it fails to address the issue of unpredictable envi-
ronmental conditions. Unpredictable environmental conditions make it difficult to
take up farming as primary occupation resulting in less agricultural yield forcing
farmers to commit suicide. Hence, farmers are moving to a novel environment for
farming, viz., Polyhouse farming (a.k.a. Greenhouse Farming). Polyhouse provides
a healthy environment with controlled environmental conditions to grow high
quality crops. It provides suitable temperature, humidity, soil moisture, and lumi-
nosity for the entitled crop. Polyhouse, seemingly a perfect solution falls short to
resolve all the limitations. Major limitations include the harmful effect of chemical
fertilizers and pesticides on living beings and in turn animal husbandry, effect of
radio frequency on the health of crops when radio communication is used for sensor
data transfer and no accumulation of sensor and statistical data. These limitations
propose a threat when devising a green technology for agricultural industry.
The chemical fertilizers are used to gain extra yield, but malignant effects of
them are now surfacing out. Chemical fertilizers reduce the fertility of soil and
causes harmful effect on human beings who apply it in the field. Fertilizer sug-
gestion ability is descended down the generations and requires expertise.
PLC based parameter regulation was devised to automate activities based on
sensor data. Wireless communication technologies like Wi-Fi and ZigBee are used
to transfer data from sensor nodes to sink node. This nearly crunched and unsafe RF
spectrum proves to be an obstacle as it affects the health of crops when exposed for
a longer duration [1].
The sensor network deployed gathers very important data which is used to
regulate environmental conditions inside polyhouse. This data is not accumulated to
a nonvolatile storage and thus loses its significance. Thus, it can only be used for
local automation and no knowledge is extracted out of it.
Dan et al. [2] uses the sensor network that is designed for collecting greenhouse
parameters like temperature, pressure, light, humidity, and CO2 and a control
system that contains fan control, curtain control, fertilizer control, and sprinkler.
Nakutis et al. [3] propose a methodology for separation of data collection and
automated control system through IoT gateway using OPC UA server for con-
trolling remote agricultural processes. Chaudhary et al. [4] propose IoT-based case
study that uses GPRS for communication and GPS for remote monitoring. FuBing
[5] provides agriculture Intelligent System based on IoT. Pavithra et al. [6] show the
other applications of IoT such as home automation. Chavan et al. [7] propose a
system with PLC and WSN for monitoring environmental parameters of polyhouse.
Chaudhary et al. [8] use Programmable System on Chip for precision agriculture.
Smart and Precision Polyhouse Farming … 249
Sachdeva et al. [9] design a farmer friendly software for polyhouse automation
using LabVIEW. Khandekar et al. [10] use mobile telephony for monitoring and
control of polyhouse through Internet. Purnima et al. [11] provide a system for
automatic irrigation using GSM Bluetooth. Borade et al. [12] provide a system with
star network applied for polyhouse automation.
Nowadays, Internet of Things (IoT) and cloud computing are gaining a lot of
popularity. IoT is one of the very few novel technologies that enables things to
communicate via the Internet and gives a new dimension to automation. The data
can now be processed and coupled with cloud platforms. Cloud allows to store and
process big data. The knowledge, when extracted out of this data, can be used to
analyze it and make predictions on future strategies using business intelligence (BI).
So, we propose a system that takes the benefit out of these novel technologies to
automate polyhouse activities using IoT, data processing and analytics using cloud,
and green communication using Visible Light Communication (VLC). The pro-
posed system provides precise organic fertilizer suggestion based on NPK
parameters of analyzed soil.
The rest of the paper is organized as follows: Sect. 2 describes the methodology
for the implementation of the system. Section 3 describes experimental results and
analysis. Section 4 concludes the paper and Sect. 5 describes the scope for future
work.
2 Methodology
When combining novel technologies for the betterment of agriculture, various

factors such as cost-effectiveness, scalability, and analytics capability should be
considered. Embedded systems allow the use of compact computation units to be
deployed on field making the system fairly cost-effective. Use of cloud technology
delivers modularity and scalability. Cloud also allows to store big data and to
operate on that data and get knowledge extracted. The environmental data gathered
by using sensors needs to be communicated with the central gateway and here VLC
comes into the picture. It enables green communication between sensor nodes to the
network gateway. Figure 1 shows system architecture of the proposed system. The
overall system can thus be divided into four modules, viz., sensor network (SN),
VLC, cloud computing and analytics, and fertilizer suggestion.
2.1 Sensor Network
Polyhouse farming is focused on environmental parameter control. The major

parameters which favor the healthy growth of crops are temperature, humidity,
luminosity, and soil moisture. The appropriate and durable sensors for the above
parameters are available in the market. The sensors used are LM35 (Temperature
250 K. Kadam et al.
Fig. 1 System architecture diagram
Sensor), SY-HS 220 (Humidity sensor), bh1750fvi-e (Light Intensity Sensor) and
Soil moisture sensor.
To control the environment inside the polyhouse, the corresponding actuators
need to be controlled based on sensed data. These sensor–actuator pairs are ven-
tilation fans and curtains for temperature control, Fogging System for humidity
control, Shade Nets for Light intensity control, and Drip Irrigation System for soil
moisture control.
Beaglebone Black (BBB) is used to attach sensors. Figure 2 shows the circuit
diagram to attach sensors and actuators to the BBB. The LM35 temperature sensor
is used since it has a 0.5 °C ensured accuracy and has low cost due to wafer-level
trimming. While, SY-HS 220 humidity sensor is used because it has an accuracy of
5%RH and it operates at 30–90% RH. Bh1750fvi is used since it has an inbuilt
amplifier and an ADC that can be interfaced with BBB’s I2C interface. It gives wide
range and high resolution for sensing light intensity. Soil moisture sensor used has
to be interfaced with the analog input of BBB and this reading can be mapped to the
soil moisture. For controlling actuators thresholds are used. These are crop specific
and can be found on NIC website. MCT2E Optocoupler is used to separate the relay
driver circuit from BBB. ULN2803A is used as a relay driver IC. It allows to drive
eight relays at a time. To control any actuators, 12 V DC relay is used. For a
polyhouse of size 60 m 40 m, soil moisture sensors are placed at every 15 m
along the length and 10 m along the width. Temperature sensors are placed at an
interval of 30 m along the length. A humidity sensor and a light intensity sensor are
placed at the center of the polyhouse. User interface provided by the web appli-
cation allows the user to control any on-site actuators remotely. This gives the
control in hands of the farmer, when system misbehaves due to any reason.
Fig. 2 Sensor attachment circuit diagram
2.2 Visible Light Communication (VLC)
Communicating sensed data from the sensor node to the network gateway is
quintessential. Various communication technologies can be used and all of them
have their own advantages and limitations. The wired network helps with higher
data rate while lacks when concerned with scalability. Wireless networks such as
Wi-Fi and ZigBee, on the other hand, are suitable for scaling the system as required.
Dan et al. [2] provide an IoT-based automation mechanism which works using
ZigBee for communication. These wireless technologies are based on radio fre-
quency and can have harmful effects on growth of crops after prolonged exposure
[13]. To enable green communication that uses natural visible light frequency
spectrum, VLC can be used.
VLC works on the principle of transmitting data using LEDs and receiving that
data using photodetectors. Various techniques such as On-Off Keying (OOK), Pulse
Width Modulation (PWM), and NRZ-OOK are suitable for the transmission [14].
On the receiver side, photodiodes are used to sense the light intensity. These
intensity values are then demodulated. The proposed system uses NRZ-OOK.
Figure 3 shows the block diagram of VLC module. BBB contains the python script
for transmitter. For a zero in binary, MOSFET 2 and MOSFET 3 are turned off
while for a one in binary, MOSFET 2 and MOSFET 3 are turned on. On the
receiver side, light intensity sensor with a photodiode detects the data sent from
LED. The receiver python script running on receiver BBB demodulates the
intensity values by mapping them in binary. For this, reference intensity value is
sensed before the start of each packet.
252 K. Kadam et al.
Fig. 3 VLC block diagram
2.3 Cloud Computing and Analytics
The cloud platform is the backbone of IoT as it enables remote connectivity and
monitoring through the Internet. Even latest polyhouse automation technologies
lack in providing a mechanism to store big data generated by IoT and process that
data for business intelligence. The traditional system just allows to control and
regulate a single polyhouse. As agriculture is an industry based on experience and
knowledge of farmers, why not to use data analytics and business intelligence to
share and plan agricultural activities based on the success of a polyhouse and data
gathered from farming cycle. Among various available cloud platforms, Amazon
Web Services (AWS) was chosen since it provides special flavor for AWS IoT and
a cloud instance to host the web application. A web application in python is
developed to provide user interface where the live environmental parameters along
with the status of actuators are displayed. Data from network gateway is commu-
nicated to a thing on AWS IoT using MQTT protocol with the help of Node-RED.
This data is then made available on Amazon EC2 instance where the web appli-
cation is running. This data then can be provided to a BI system to analyze and
devise future business plans and to share the recipe of success to other polyhouses.
2.4 Fertilizer Suggestion
The essential part of any farming cycle is providing a proper fertilizer in a precise
amount. To provide an appropriate fertilizer, farmers depend on inherited knowl-
edge and experience. A novice farmer when lacks the both, gets into a paralyzed
situation. The government has come up with lots of soil testing laboratories where
farmers can visit with their soil samples to get them tested. The basic report of any
soil test shows the nutrient values that are present and a fertilization amount of
predefined fertilizers. But the traditional fertilizer suggestion methods are not
precise and can sometimes have an adverse effect on the produce. Often, suggested
fertilizers are chemical or inorganic fertilizers. These fertilizers have proved to have
an adverse effect on fertility of soil and on health of living beings that come in
contact with them.
Proposed system provides precise organic fertilizer suggestion. Organic fertilizer

has very less amount of fillers and is not harmful to any form of living beings. The
algorithm for fertilizer suggestion is as follows:
Program Fertilizer_Suggestion (N, P, K, crop, age)

Begin
[N0, P0, K0]:= get_optimum_npk_values (crop, age)
[Nr, Pr, Kr]:= [N0 - N, P0 - P, K0 - K]
If ([Nr, Pr, Kr] in lookup_table)
Return lookup_table.suggested_fertilizer
Else
Var min := 999
Repeat for each f in fertilizer_DB
Euclidean_distance:= sqrt ((f.N - Nr) 2
+ (f.P –
2 2
Pr) + (f.K - Kr) )
If (Euclidean_distance < min)
min := Euclidean_distance
Suggested_fertilizer:= f
Add_to_lookup_table (f)
Return Suggested_fertilizer
End.
The above algorithm takes nitrogen (N), phosphorus (P), potassium (K), selected
crop and age of the crop as input. Farmer inputs these values obtained from soil
analysis on the web page hosted on the cloud. First, the optimum values required
for the crop at the particular age are taken from database present for all of the crops.
According to this, required macro-nutrient values are calculated. If these are already
present in lookup table then suggested fertilizer is returned from lookup table. If it is
not in the lookup table, Euclidean distance is calculated from the database of
nutrient values of organic fertilizers present in the market. The fertilizer with the
minimum Euclidean distance is hence returned and displayed on web page.
3 Experimental Results and Analysis
The proposed system is fully implemented and is ready to be deployed on field. The
prototype of the system is shown in Fig. 4.
The UI for the system is shown in Fig. 5. It shows the appropriate status of
actuators and live temperature parameters. The web application is hosted on
Amazon EC-2 instance. Figure 6 shows the fertilizer suggestion module for the
crop tomato. The module gives 98% accuracy in suggesting fertilizer. It is ensured
that the suggested fertilizer is the nearest one that fulfills the nutrient requirement.
254 K. Kadam et al.
Fig. 4 System prototype (A: Sensor network with VLC Transmitter, B: Display of Transceiver
python script running on BBB, C: Cloud Server running remotely, D: Web Application UI)
Fig. 5 Screenshot of working UI of the web application
Fig. 6 Screenshot of working fertilizer suggestion module

Table 1 Sensor–actuator pairs and operating thresholds for tomato crop

Sensor Actuator Threshold for control
Temperature sensor Ventilation fans and curtains 15–29 °C
Humidity sensors Fogging system 60–90% RH
Light intensity sensor Shade nets 10,000–140,000 lx
Soil moisture sensor Drip irrigation system 50–90% SM
Table 1 shows sensor–actuator pairs along with their operating thresholds for
sample crop Tomato.
4 Conclusion
The system proposed in this paper automates activities inside the polyhouse and
reduces the human requirement for monitoring environmental parameters. The
system gathers real-time data for analytics and for decision-making using Business
Intelligence. It also provides organic fertilizer suggestions based on N-P-K values
of analyzed soil, thus reduces harmful effects of chemical fertilizers on the health of
land, nature, and living beings. Use of VLC eliminates the effect of radio frequency
on crop health.
5 Future Scope
The proposed system will be implemented on site for a couple of years thus pro-
viding a bigger picture considering BI and Analytics. The data gathered from longer
time span allows to map success rate to analyzed data and ensure a well-defined
success recipe. The further cost reduction can be achieved by developing a custom
system on chip (SOC) with optimum processing power. Higher data rates will be
achieved in VLC module by using a high quality LED and a fast sensing
photodiode.
References
1. Special article: Effect of ultra-short radio waves on plant growth: Science 17 Nov 1939: Vol.
90, Issue 2342, pp. 470–471 (1939).
2. LIU Dan, Cao Xin, Huang Chongwei, JI Liangliang: Intelligent Agriculture Greenhouse
Environment Monitoring System Based on IOT Technology: 2015 International Conference
on Intelligent Transportation, Big Data & Smart City, IEEE (2015).
3. Ž. Nakutis, V. Deksnys, I. Jauruševiius, E. Marcinkeviius, A. Ronkainen, P. Suomi,
J. Nikander, T. Blaszczyk, B. Andersen: Remote Agriculture Automation using Wireless Link
256 K. Kadam et al.
and IoT Gateway Infrastructure: 26th International Workshop on Database and Expert
Systems Applications, IEEE (2015).
4. Rashi Chaudhary, Jalaj Ranjan Pandey, Prakhar Pandey, Prakhar Pandey: Case study of
Internet of Things in area of Agriculture, ‘AGCO’s Fuse Technology’s ‘Connected Farm
Services, IEEE (2015).
5. FuBing: Research on the Agriculture Intelligent System Based on IoT: 2012 International
Conference on Image Analysis and Signal Processing (IASP), IEEE (2012).
6. Pavithra. D, Ranjith Balakrishnan: IoT based Monitoring and Control System for Home
Automation: 2015 Global Conference on Communication Technologies, IEEE (2015).
7. Sachin. V. Chavan, Sumayaa. C. Pathan, Suhas. N. Patil and Bhimrao. P. Ladgaonkar: design
of LM4F120H5QR based node for wireless sensor network to monitor environmental
parameters of polyhouse: International Journal of Advances in Engineering & Technology
(2015).
8. D.D. Chaudhary, S.P. Nayse, L.M. Waghmare: application of wireless sensor networks for
greenhouse parameter control in precision agriculture: International Journal of Wireless &
Mobile Networks (IJWMN) Vol. 3, No. 1 (2011).
9. Kanwal Sachdeva, Neeraj Kumar, Rajesh Kumar, Rohit Verma: Monitoring and Controlling
of Environmental Parameters of Polyhouse Based On Lab VIEW: International Journal of
Innovative Research in Science, Engineering and Technology Vol. 4, Issue 9 (2015).
10. Yogesh R. Sonawane, Sameer Khandekar, Bipin Kumar Mishra, K. K. Soundra Pandian:
Environment Monitoring and Control of a Polyhouse Farm through Internet: IIT Kanpur
(2016).
11. Purnima, S.R.N. Reddy: Design of Remote Monitoring and Control System with Automatic
Irrigation System using GSM-Bluetooth: International Journal of Computer Applications
Volume 47– No. 12 (2012).
12. Kiran E. Borade, Prof. C.S. Patil, Prof. R.R. Karhe: Polyhouse Automation System:
International Journal of Advanced Research in Computer Science and Software Engineering
Volume 3, Issue 8 (2013).
13. Natural News: http://www.naturalnews.com/043238_Wi-Fi_routers_radiation_plant_growth.
html.
14. Krishna Kadam, Prof. Manisha R. Dhage: Visible Light Communication for IoT: iCATccT,
pp. 275–278, IEEE (2016).
Acceleration of CNN-Based Facial
Emotion Detection Using NVIDIA GPU
Bhakti Sonawane and Priyanka Sharma
Abstract Emotions often mediate and facilitate interactions among human beings
and are conveyed by speech, gesture, face, and physiological signal. Facial
expression is a form of nonverbal communication. Failure of correct interpretation
of emotion may cause for interpersonal and social conflict. Automatic FER is an
active research area and has extensive scope in medical field, crime investigation,
marketing, etc. Performance of classical machine learning techniques used for
emotion detection is not well when applied directly to images, as they do not
consider the structure and composition of the image. In order to address the gaps in
traditional machine learning techniques, convolutional neural networks (CNNs)
which are a deep Learning algorithm are used. This paper comprises of results and
analysis of facial expression for seven basic emotion detection using multiscale
feature extractors which are CNNs. Maximum accuracy got using one CNN as
96.5% on JAFFE database. Implementation exploited Graphics Processing Unit
(GPU) computation in order to expedite the training process of CNN using GeForce
920 M. In future scope, detection of nonbasic expression can be done using CNN
and GPU processing.
Keywords CBIR CNN GPU processing Emotion detection
1 Introduction
Facial expression which is a form of nonverbal communication plays an important

role in interpersonal relations. Facial expressions represent the changes of facial
appearance in reaction to persons inside emotional states, social communications, or
intentions. Recently, active research is going on in the area of Automatic Facial
B. Sonawane (&) P. Sharma

Nirma University, Ahmedabad, India
e-mail: bhakti.sonawane@sakec.ac.in
P. Sharma
e-mail: priyanka.sharma@nirmauni.ac.in

https://doi.org/10.1007/978-981-10-7245-1_26
258 B. Sonawane and P. Sharma
Expression Recognition (FER). It is useful in the field of medicine, human emotion

analysis, etc., and will be one of the best steps for improving Human Machine
Interaction (HMI) systems. Many factors make emotion detection as challenging
problem. Among them is the problem of an unavailability of the standardized
database for FER. Benchmark database that can fulfill the various requirements of
the problem domain in order to become standard database for future research is a
tough and challenging exercise [1]. Most of existing databases expressions is posed
and not spontaneous. Biggest challenge is to capture spontaneous expressions on
images and video. Like different subjects express the same emotions at different
intensities and sometimes laboratory conditions become hurdle for the subject to
display spontaneous expressions. Another major challenge is labeling of the data
which is a time-consuming process and possibly error prone also. Challenges
involved in capturing and recognizing spontaneous nonbasic expression are more
than basic expressions. Most of the FER has lack of rotational movement freedom
[1]. Here, our aim is to present an approach based on CNNs for FER and a
systematic comparison of five different 12-layer CNNs. The input to CNN is an
image to predict the facial expression label which should be one of these labels:
anger, happiness, fear, sadness, disgust, and neutral. Among various database,
JAFFE database is chosen for implementation as this is the most commonly used in
other automatic FER systems.
2 Background
Generally, face detection, facial feature extraction, and facial expression classifi-
cation are the parts of an automatic FER system using traditional machine learning
techniques. In the face detection step, given an input image system performs some
image processing techniques on it in order to locate the face region. In feature
extraction step, from located face, geometric features and appearance features are
the two types of features that are generally extracted to represent facial expression.
Geometric features describe shape of face and its components like lips, nose or
mouth corners, etc. Whereas appearance features depict the changes in texture of
face when expression is performed. Classification is the last part of the FER system
which based on machine learning theory. Output of previous stage which is a set of
features retrieved from face region is given as an input to the classifier like Support
Vector Machines, K-Nearest Neighbors, Hidden Markov Models, Artificial Neural
Networks, Bayesian Networks, or Boosting Techniques [2]. Some expression
recognition systems classify the face into a set of standard emotions. Other system
aims to find out movements of the individual muscle that the face can produce. In
[1], author provided an extensive list of researches between 2001 and 2008 on FER
and analysis. Since past few years, there were several advances to perform FER
using traditional machine learning methods which involve different techniques of
face tracking and detection, feature extraction, training classifier, and classification.
In [3], to provide a solution for low resolution images, framework for expression
Acceleration of CNN-Based Facial Emotion … 259
recognition based on appearance features of selected few prominent facial patches

which are active when emotion are expressed is proposed. In order to get dis-
criminative features for classification, salient patches are obtained after processing
selected patches further. One-against-one classification task is performed using
these features and recognition of the expression is done based on majority vote.
Experimentation of the proposed method is carried on CK and JAFFE facial
expression databases. In [4], feature extraction is done using PCA along with LBP
and SVM classifier used to obtain results. Database used are JAFFE database and
MUFE database and obtained results show that both PCA and LBP gave high
performance together. In [5], live video stream frames are extracted containing face
using Gabor feature extraction method and neural network and modified k means
with PCA is used for classification of emotion. JAFFE database is used for simu-
lation of framework. In [6] Active Appearance Models (AAM) were used to
identify the face and extract its graphic features. For expression prediction, HMM is
used and to identify the person in the image, K-NN is used. In [7], improved
Directional Ternary Patterns (DTP) feature extraction and SVM classifier are used
for real-time purpose emotion detection by facial expressions on JAFFE database.
In FER using traditional machine learning techniques, programmer has to be
very specific about what he is interested which involves laborious process of feature
extraction. Domain knowledge is expected for feature extraction. Thus, success rate
of system depends on programmer’s ability to accurately define a feature set. In
addition, whenever the problem domain changes the whole system needs to change
requiring a redesign of the algorithm from the start [8]. In [9], authors proposed a
novel FER system based on features resulting from principal component analysis
(PCA) which are fine-tuned by applying particle swarm. The best classification
result achieved was 97% for CK database.
Most of researcher also used neural networks as its ability to extract undefined
features from the training database. Most of the time it is observed that if neural
networks that are trained on large amounts of data are able to extract generalized
features well to scenarios that the network has not been trained on. In [10], con-
structive training algorithm for MLP neural networks has been proposed as clas-
sification step for the FER system. Experiments carried on three well-known
databases show that the best recognition rate has been obtained using the con-
structive training algorithm as compared to the fixed MLP architecture. In [11],
authors proposed Neural Network and K-NN based model for facial expression
classification. For extraction of facial features on JAFEE database, ICA is used.
Recent approaches include increased use of deep neural networks (neural networks
with many numbers of hidden layers) for automatic FER problem. With growing
computing power, for finding complex patterns in images, sound, and text, deep
neural network architectures provide learning architecture similar to the develop-
ment of brain-like structures which can learn multiple levels of representation and
abstraction. Extreme variability patterns with robustness to distortions and simple
geometric transformations are recognized by CNNs which are deep neural net-
works. It has been proven by a wide range of applications that are using CNN such
as face detection, face recognition, gender recognition, and so forth that minimal
domain knowledge of the problem at hand is sufficient to perform efficient pattern

recognition tasks [8, 12–16]. CNNs have become the traditional approach for
researchers examining vision and deep learning. Starting with LeNet-5 [17], vari-
ations of this basic design are prevalent in the image
classification literature with the best results. The recent trend is to increase the
number of layers and layer size for larger datasets such as ImageNet and use of
dropout in order to deal with the problem of overfitting [18, 19].
In [20], authors proposed network consists of two convolutional layers each
followed by max pooling with next four inception layers and conducted experi-
ments on seven publicly available facial expression databases. In [21], two different
deep network models are proposed, for extraction temporal appearance features
from image sequences and for extraction temporal geometry features from temporal
facial landmark points. A new integration method for combining these two models
is required in order to boost the performance of the FER. For Emotions in the Wild
[22] contest for static images in [23], multiple deep convolutional neural networks
are trained as committee members and combine their decisions, generating up to
62% test accuracy.
3 JAFFE Database and Proposed CNN Architecture
Proposed CNN architectures are tested on JAFFE database set of facial expression
images for posed emotions (six different emotions and neutral face displays) of 10
Japanese female subjects. These six expressed emotions are the basic emotions
given by Ekman and Friesen [24]. Figure 2 shows sample images from JAFFE
database [25]. Expressed emotion seems to be universal across human ethnicities
and cultures which are happiness, sadness, fear, disgust, surprise, and anger. The
grayscale images are 256 256 pixels size. The images were labeled into
6 + 1 = 7 emotion classes. Some head pose variations can be featured by these
images [26].
Five different 12-Layer CNNs architectures are proposed for facial expression
classification up to seven different basic emotions. In all five CNNs, input layer is
followed by convolutional layer with different filter size and number of filters. This
layer is followed by relu layer and max pooling layer with pool filter size as 2 which
outputs maximum among the four values. Max pooling layers are trailed by con-
volutional layer with different filter sizes and number of filters for different CNNs
again followed with relu and max pooling layer. Next layer is a fully connected layer
with a number of output neurons which varies in different CNNs and followed by
relu layer. And last fully connected layer is with seven output neurons and output of
this layer is given to final softmax and classification layer. Table 1 shows detail
regarding proposed 5 different CNNs. For CNN_1, CNN_2, CNN_3, and CNN_4,
number of training images are 164 and testing images are 26. In CNN_1 and
CNN_2, training images are repeated to increase the total number of training image
set. For CNN_5, number of training images are 178 and testing images are 35.
Table 1 Details for proposed CNNs for emotion detection

Convolution Convolution Fully connected Fully connected
Sr. No. CNNS Layer 1 Layer 2 Layer 1 Layer 2
01 CNN 1 3(3 3) 9(3 3 3) 512 7
02 CNN 2 3(12 12) 6(9 9 3) 16 7
03 CNN 3 3(12 12) 6(9 9 3) 16 7
04 CNN 4 3(15 15) 6(9 9 3) 16 7
05 CNN 5 3(15 15) 6(9 9 3) 16 7
Fig. 1 Proposed CNN_5 architecture
In Table 1, numbers in convolutional layer column give number of feature maps

generated with filter size and numbers in fully connected layer column gives number
of output neurons. Thus in CNN_1, number of training images is 164 and first
convolutional layer generates 3 feature map using 3 3 filter size. Second con-
volutional layer generates 9 feature map using 3 3 filter size. Fully connected
layer 1 has output neurons 512 and last fully connected layer has number of output
neurons as 7 representing basic emotion.
Proposed 12 layers architecture are shown in Fig. 1. Similar architecture is
(referring Table 1) for CNN_1 CNN_2, CNN_3, and CNN_4 only with different
filter size, feature map, and number of output neurons in fully connected layer.
The proposed design is tested on a 2.40 GHz Intel i7-5500U quad core processor;
8 GB RAM, with windows 10, 64 bit system. As CNN is computationally intensive
which requires GPU processing for faster computation, thus system used was
CUDA-enabled NVIDIA GPU with compute capability higher than 3.0, DirectX
Runtime Version 12 (graphics card GeForce 920 M). MATLAB used for
Fig. 2 JAFFE sample images [25]
Table 2 Comparison for Sr. No. CNNS Accuracy (%)

CNNs
1 CNN 1 80.76
2 CNN 2 88.46
3 CNN 3 73.07
4 CNN 4 88.46
5 CNN 5 96.15
implementation and system tested on JAFFE database From Tables 1 and 2 it is

observed that more image detail gets captured using 12 12 or 15 15 filters. In
CNN_1 and CNN_2, number of training set is increased by repeating that set but its
effect is similar to increasing the epoch during training. If first layer captures good
detail from the input image of 256 256 size and number of epoch is more this
leads final accuracy. Maximum accuracy achieved in CNN_5 in which total number
of images to be trained is more as compared to other CNNs.
Minimal preprocessing involved as CNNs are designed to recognize visual patterns

directly from pixels of images. This is completely in contrast with the conventional
pattern recognition tasks in which prior knowledge of the problem at hand is needed
in order to apply a suitable algorithm to extract the right features. In this paper, five
new CNN architectures have been proposed for automatic facial expression
recognition. Among them, CNN_5 resulted in the highest accuracy with a classi-
fication accuracy of 96.15% achieved on test samples of JAFFE database using
2.5 GHz i7-5500U quad core processor, 8 GB RAM with GeForce 920 M.
Proposed architecture can be extended for detection of nonbasic expression in
future.
References
1. Face Expression Recognition and Analysis: The State of the Art Vinay Kumar Bettadapura
Columbia University.
2. Nazia Perveen, Nazir Ahmad, M. Abdul Qadoos Bilal Khan, Rizwan Khalid, Salman Qadri,
Facial expression recognition Through Machine Learning, International Journal of Scientific
& Technology Research, Volume 5, Issue 03, March 2016, ISSN 2277-8616.
3. S L Happy, Aurobinda Routray Automatic Facial expression recognition Using Features of
Salient Facial Patches, IEEE transactions on Affective Computing, VOL. 6, NO. 1, January–
March 2015.
4. Muzammil Abdulrahman, Alaa Eleyan, Facial expression recognition Using Support Vector
Machines, in Proceeding of 23nd Signal Processing and Communications Applications
Conference, PP. 276–279, May 2015.
5. Debishree Dagar, Abir Hudait, H. K. Tripathy, M. N. Das Automatic Emotion Detection
Model from Facial Expression, International Conference on Advanced Communication
Control and Computing Technologies (ICACCCT), ISBN No. 978-1-4673-9545-8,2016.
6. G. Ramkumar E. Logashanmugam, An Effectual Facial expression recognition using HMM,
International Conference on Advanced Communication Control and Computing Technologies
(ICACCCT), ISBN No. 978-1-4673-9545-8,2016.
7. S. Tivatansakul, S. Puangpontip, T. Achalakul, and M. Ohkura, “Emotional healthcare
system: Emotion detection by facial expressions using Japanese database,” in Proceeding of
6th Computer Science and Electronic Engineering Conference (CEEC), PP. 41–46,
Colchester, UK, Sept. 2014.
8. A.R. Syafeeza, M. Khalil-Hani, S.S. Liew, and R. Bakhteri, Convolutional neural network for
face recognition with pose and illumination variation, Int. J. of Eng. and Technology, vol. 6
(1), pp. 44–57, 2014.
9. Vedantham Ramachandran, E Srinivasa Reddy, Facial Expression Recognition with enhanced
feature extraction using PSO & EBPNN. International Journal of Applied Engineering
Research, 11(10):69116915, 2016.
10. Boughrara, Hayet, et al. “Facial expression recognition based on a mlp neural network using
constructive training algorithm.” Multimedia Tools and Applications 75.2 (2016): 709–731.
11. Hai, Tran Son, and Nguyen Thanh Thuy. “Facial expression classification using artificial
neural network and k-nearest neighbor.” International Journal of Information Technology and
Computer Science (IJITCS) 7.3 (2015): 27.
12. Shih, Frank Y., Chao-Fa Chuang, and Patrick SP Wang. “Performance comparisons of facial
expression recognition in JAFFE database”. International Journal of Pattern Recognition and
Artificial Intelligence 22.03 (2008): 445–459.
13. C. Garcia and M. Delakis, “Convolutional face finder: a neural architecture for fast and robust
face detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 26,
pp. 1408–1423, 2004.
14. S. Chopra, R. Hadsell, and Y. LeCun, “Learning a Similarity Metric Discriminatively, with
Application to Face Verification,” in In Proceedings of CVPR (1) 2005, 2005, pp. 539–546.
15. T. Fok Hing Chi and A. Bouzerdoum, “A Gender Recognition System using Shunting
Inhibitory Convolutional Neural Networks,” in International Joint Conference on Neural
Networks, 2006, pp. 5336–5341.
16. Caifeng Shan, Shaogang Gong, Peter W. McOwanb, Facial expression recognition based on
Local Binary Patterns: A comprehensive study, Image and Vision Computing, V. 27 n. 6,
pp. 803–816, May 2009. https://doi.org/10.1016/j.imavis.2008.08.005].
17. Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L.
D. Jackel. Backpropagation applied to handwritten zip code recognition. Neural Comput., 1
(4):541551, Dec. 1989.
18. O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A.
Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. arXiv preprint
arXiv:1409.0575, 2014.
19. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and
A. Rabinovich. Going deeper with convolutions. arXiv preprint arXiv:1409.4842, 2014.
20. Mollahosseini, Ali, David Chan, and Mohammad H. Mahoor. “Going deeper in Facial
expression recognition using deep neural networks.” Applications of Computer Vision
(WACV), 2016 IEEE Winter Conference on. IEEE, 2016.
21. Jung, Heechul, et al. “Joint fine-tuning in deep neural networks for facial expression
recognition.” Proceedings of the IEEE International Conference on Computer Vision. 2015.
22. Dhall A, Murthy OVR, Goecke R, Joshi J, Gedeon T (2015) Video and image based emotion
recognition challenges in the wild: Emotiw 2015. In: Proceedings of the 2015 ACM on
International Conference on Multimodal Interaction, ACM, pp 423–426.
23. B. Kim, J. Roh, S. Dong, and S. Lee, Hierarchical committee of deep convolutional neural
networks for robust facial expression recognition, Journal on Multimodal User Interfaces,
pp. 117, 2016.
24. P. Ekman and W. Friesen. Constants Across Cultures in the Face and Emotion. Journal of
Personality and Social Psychology, 17(2):124129, 1971.
25. Michael J. Lyons, Shigeru Akemastu, Miyuki Kamachi, Jiro Gyoba. Coding Facial
Expressions with Gabor Wavelets, 3rd IEEE International Conference on Automatic Face
and Gesture Recognition, pp. 200–205 (1998).
26. Lyons M., Akamatsu S., Kamachi M., and Gyoba J. Coding Facial Expressions with Gabor
Wavelets. In Third IEEE International Conference on Automatic Face and Gesture
Recognition, pages 200205, April 1998.
Research Issues for Energy-Efficient Cloud
Computing
Nitin S. More and Rajesh B. Ingle
Abstract Everyone is using highly computing devices nowadays and indirectly

getting part of cloud computing, IOT, and virtual machines being either as a client
or server. All this was possible because of virtualization techniques and upgradation
of technologies from networking to ubiquitous computing. However, today, there is
a need of considering the cost of computing devices versus the cost of power
consumed by computing devices. Everybody required mobile of more computing
capacity and along with that more battery backup. This also is needed to be thought
about cloud computing, as energy consumption by data centers was increased year
to year which also causing footprints of CO2 behind. According to Gartner, worlds
2% CO2 emission was only due to IT industry. This paper deals with investigating
such research issues for energy-efficient cloud computing.

Keywords Energy efficient cloud computing Research issues

Virtualization Data centers Green computing CO2
1 Introduction
Cloud computing has changed the information technology industry by many ways
and thus also contributed in the environment hazards indirectly. In cloud com-
puting, client request resource may be infrastructure, platform or software resource
and servers situated across the datacentres will serve the requested resource as per
the software level agreement (SLA) existing between client and server. On the other
hand, cloud data centers end up with consuming lots of power required for cooling,
operating, and maintaining data centers and emitting ample amount of CO2. In
N. S. More (&) R. B. Ingle

Pune Institute of Computer Technology, Savitribai Phule Pune University, Pune, India
e-mail: nsmore27@gmail.com
R. B. Ingle
e-mail: rbingle@pict.edu

https://doi.org/10.1007/978-981-10-7245-1_27
266 N. S. More and R. B. Ingle
2020, energy consumption by data centers worldwide was predicted to be 4% of the

worldwide electricity use and is likely to cultivate in the upcoming years [1].
This paper tries to deal with presenting novel research issues for energy-efficient
cloud computing. As new researchers are coming out to work on cloud computing
they indeed go to find the issues to be work on, this paper will suggest some issues
for them. Energy is the parameter of concern mostly for wireless or battery operated
device but now the time has come to consider it for cloud computing as well.
Energy utilization in cloud data centers can be optimized by reducing the need for
physical servers present in data centers and there are many ways for it, but along
with better quality of service (QOS) and without violating software level agreement
(SLA). The planned methodology should be enough scalable, distributed, and
effectual for managing the power utilization trade-off. Figure 1 shows how IT
industries paying on Servers, Power, and Cooling Management. Industries are
acquiring cloud computing for their companies instead of going for actual invest-
ments in servers, and other computing devices. It is clearly predictable that cost on
server spending is decreasing due to virtualization techniques but power and
cooling cost is increasing day by day which is concerned a lot and cannot be
underestimated as it will create disaster for all human beings with issues like the
greenhouse effect, global warming, etc. This motivates for further vital research
issues
• pointing energy efficient virtual machine (VM) topologies
• identifying energy efficient attributes and finding relations between them
• investigate energy efficient software design for cloud
• optimized VM migration algorithms
• verifying energy efficient techniques against QOS and SLA.
Fig. 1 Cost of data center versus cost of power consumption [3]

Research Issues for Energy-Efficient Cloud Computing 267
2 Survey of Work Done
Further, Table 1 shows the gap analysis of energy efficient cloud computing
techniques referred during survey, which includes 123 Springer journal/conference
papers, 78 IEEE, 5 ACM, and 12 other papers published in various journals or
conferences. The survey also shows that the regular power load consumption of the
companies surveyed increased from 2.1 MW in 2011 to 2.9 MW in 2013 [1]. In the
age of weather change and global warming, it is unacceptable to avoid power
consumption and hence, it is essential to review existing body of knowledge.
Table 1 shows comparison of various power saving strategies in the cloud.
As tabulated in Table 1, Beloglazov et al. [3] studied various hardware and
software techniques and proposed VM consolidation algorithms which do not affect
SLA. Bergmann et al. [4] used swarm intelligence and dynamic voltage scaling
(DVS), scheduling jobs technique, and achieving energy efficiency without dis-
turbing QOS. Nguyen et al. [5] proposed a genetic algorithm for VM placement.
Kulseitova et al. [6] observed many techniques with server and network as a scope
and strategies like virtualization, dynamic voltage frequency scaling (DVFS),
dynamic power management, and sleep mode for the servers which are not active,
etc., but without any secondary objective.
Priya et al. [7] used a technique sleep mode for network as a scope and by
monitoring link load as a metric they do not consider any another second objective
other than energy efficiency.
Strunk et al. [8] proposed fault tolerant switching off strategy based on CPU load
metric with vm migration, they also consider QOS. Rahman et al. [9] proposed
virtual network embedding with vm migration they consider bandwidth metric.
Cavdar et al. [10] used green routing strategy for switch and link load metrics along
with QOS, SLA and CO2 emission parameters are considered.
Mahadevan et al. [11] used server load consolidation with network sleep mode
by considering CPU and link load metrics for server load migration with QOS as a
secondary objective. No one has considered vm topologies effect on power con-
sumption. The above-discussed strategies are mainly focusing on saving power and
making datacenters more efficient by considering server and network as its scope,
however in future due to the requirement of high bandwidth and network con-
nectivity of datacenters, the power requirement of datacenter will go beyond our
imagination.
To overcome the above-mentioned concern, Koutitas et al. [2] mention how
actually power consumption happens in datacenters, most of the part of power are
consumed by computing servers 70%, CPU 43% and switches 25%, i.e., network
devices. There still are lots of research need to be done to design such modern
algorithm and techniques which achieve better energy efficiency along with con-
sideration of QOS, SLA, and VM consolidation by defining VM topologies which
are fault tolerant and also helps nature, mankind by minimizing CO2 emission.
268
Table 1 Gap analysis of energy efficient techniques in cloud computing

Reference Scope Strategy Metrics Migration Secondary Fault VM CO2
objective tolerant topologies Emission
Beloglazov Server Virtualization, VM migration Physical machine VM SLA – – –
et al. [3] utilization
Bergmann Server DVFS, Swarm intelligence Task execution – QOS – – –
et al. [4] time
Nguyen et al. Server VM Migration Server load VM – – – –
[5]
Kulseitova Network Sleep mode Queue length at – – – – –
et al. [6] switch
Priya et al. [7] Network Sleep mode Link load – – – – –
Strunk et al. Server Switched off CPU load VM QOS C – –
[8]
Rahman et al. Network Virtual network embedding Bandwidth VM – – – –
[9]
Cavdar et al. Network Green routing Switch and link – QOS & SLA – – C
[10] load
Mahadevan Server and Server load consolidation with CPU and link load Server QOS – – –
et al. [11] Network network sleep mode load
N. S. More and R. B. Ingle
3 Mathematical Model for Energy Efficient Cloud
For achieving mathematical model for energy consumption in cloud, there are
ample of parameters for concentrating on power consumption of cloud, let us
consider energy of virtual machine (vm) while running, migrating and when vm
idle, type of task submitted to vm, energy utilization of vm at time of reading data
as well writing data, energy consumption of resources, all these technical issues and
their power usage are important to consider. The above-mentioned content can be
placed in mathematical equation as follows. Let us consider total energy consumed
by cloud is denoted by Ec then
Ec ¼ Evmrunning þ Evmmigrating þ Evmidle ; ð1Þ
where
Nomenclature
Evmrunning ¼ Nt :et þ Rt :ert ð2Þ
Evmmigrating ¼ f ðnu; pmÞ ð3Þ
Evmidle ¼ Evm ð4Þ
Ec The energy consumption of cloud

Evmrunning The energy consumption of vm when vm is running
Evmmigrating The energy consumption of vm while migrating
Evmidle The energy consumption of vm when vm is idle
Nt The number of task t executing at vm
et The energy consumed by task t
Rt The number of resources used during task t
ert The energy consumed by resources during task t
nu The network utilization
pm The physical machine
Evm The energy of vm when task t = 0.
Equation 1 says that energy of the cloud is equal to the sum of energy consumed
by vm in three states, i.e., vm running, vm migrating and when vm is idle.
Equation 2 says energy consumed by vm during running is the sum of product of
two things, one is the product of number of task and energy required by that task
and second is the product of number of resources required for task and energy
required by resources like disk, memory, etc., which can be more expressed with
different equations. Equation 3 says the energy of vm which is migrating is a
function of energy used by network utilization and physical machine on which vm
running. Lastly, Eq. 4 stating the energy of vm when the task is zero and vm is idle.
Fig. 2 Cloud entities relationship with energy consumption
4 Simulation Result with MATLAB
For identifying energy-efficient attributes in the cloud and finding relationship of

them classification of cloud entities like a physical machine, vm, hypervisor,
memory, hard disk, motherboard, switches, etc., done in three sets as
1. Size = [large, medium, small]
2. Processing power = [good, average, minimum]
3. Energy consumption = [high, medium, low]
Based on these three sets, 27 rules are derived in MATLAB fuzzy toolbox by
considering size and processing power of cloud entities as input and energy con-
sumption as output. Then membership functions are created by considering
appropriate values for all three sets. By processing all 27 rules in rule viewer an
output is obtained in surface viewer as given in Fig. 2, showing on x-axis size of
cloud entities, y-axis processing power of cloud entities by considering zero to ten
numbering for small, medium, large for size and minimum, average, good for
processing power as input and output on z-axis, i.e., energy consumption of cloud.
The result shows that as the size of cloud entities increasing energy consumption
also increased and it will vary with processing power.
5 Research Issues
Shortening all of the above studies, stating some of the research issues which can be
considered for further enhancements in the green cloud computing. They are stated
as follows:
• To invest more for renewable energy use in cloud data centers.

• To design algorithms and techniques for energy efficient cloud without affecting
QOS, SLA.
• To investigate and propose fault tolerant heuristics approaches for live virtual
machine migration.
• To design energy efficient software for cloud applications without hurting
security and privacy of the user.
• To inspect each and every entity in the cloud for energy efficiency.
Methodologies and techniques to be adapted in order to address the research
issues
1. Study of existing energy efficient techniques.
2. Gap Analysis.
3. Propose innovative Algorithms and Techniques.
4. Experimenting and Evaluating proposed Algorithm and Techniques.
5. Come up with new dynamic, scalable, and fault-tolerant energy-efficient com-
puting techniques.
6 Conclusion
This paper briefly analyzes and reviews the different strategies for energy-efficient
cloud computing along with the motivation of it. Finding research issues and doing
a gap analysis of existing energy efficiency techniques. A mathematical model for
energy efficient cloud which gives the basic equation of calculating energy con-
sumed at cloud by considering different activities of the virtual machine was dis-
cussed. Experiment with MATLAB simulator for investigating interrelationships
between cloud entities and energy consumption using fuzzy logic was carried out.
Lastly, different research issues along with methodologies and techniques to be
adapted for achieving them were discussed. There are still other relevant issues that
need to be comprehensively inspected for energy-efficient cloud computing.
References
1. Infoworld, http://www.infoworld.comtgreen-itkoomeys-law-computing-efficiency keeps-pace-

moores-law-172681 access date 12 December 13.
2. G. D. Koutitas, P. Demestichas, Challenges for Energy Efficiency in Local and Regional Data
Centers, Journal of Green Engineering, pp. 1– 32, 2010.
3. Anton Beloglazov, Energy-efficient management of virtual machines in data centers for cloud
computing, Ph.D. thesis submitted to Melbourne university, February 2013.
4. Neil Bergmann, Yuk Ying Chung, Xiangrui Yang, Zhe Chen, Wei Chang Yeh, Xi- angjian
He, and Raja Jurdak. 2013. “Using swarm intelligence to optimize the energy consumption
for distributed systems”. Modern Applied Science 7, 6 (2013), 5966. DOI:http://dx.doi.org/

10.5539/mas.v7n6p59.
5. Quang-Hung Nguyen, Pham Dac Nien, Nguyen Hoai Nam, Nguyen Huynh Tuong, and Nam
Thoai. 2013. “A genetic algorithm for power-aware virtual machine allocation in private
cloud. In Information and Communicatiaon Technology”, Lecture Notes in Computer
Science, Vol. 7804. Springer, Berlin, 183191.
6. Kulseitova, A., Ang Tan Fong, “A survey of energy-efficient techniques in cloud data
centers,” ICT for Smart Society (ICISS), 2013 International Conference on, vol., no., pp. 1, 5,
13–14 June 2013.
7. Priya, B., Pilli, E.S., Joshi, R.C., “A survey on energy and power consumption models for
Greener Cloud,” Advance Computing Conference (IACC), 2013 IEEE 3rd International, vol.,
no., pp. 76, 82, 22–23 Feb. 2013.
8. Strunk A., “Costs of Virtual Machine Live Migration: A Survey,” Services (SER- VICES),
2012 IEEE Eighth World Congress on, vol., no., pp. 323, 329, 24–29 June 2012.
9. Rahman, A., Xue Liu, Fanxin Kong, “A Survey on Geographic Load Balancing Based Data
Center Power Management in the Smart Grid Environment,” Communications Surveys &
Tutorials, IEEE, vol. 16, no. 1, pp. 214, 233, First Quarter 2014.
10. Cavdar, D., Alagoz, F., “A survey of research on greening data centers,” Global
Communications Conference (GLOBECOM), 2012 IEEE, vol., no., pp. 3242, 3–7 Dec. 2012.
11. P. Mahadevan, P. Sharma, S. Banerjee and P. Ranganathan, Energy aware network
operations, in Proceedings of the 28th IEEE international conference on Computer
Communications Workshops, Rio de Janeiro, Brazil, 2009, pp. 25–30.
12. T. Kaur and I. Chana, “Energy Efficiency Techniques in Cloud Computing: A Survey and
Taxonomy” ACM Computing Surveys, Vol. 48, No. 2, Article 22, Publication date: October
2015.
13. Ching-Hsien Hsu, Kenn Slagter, Shih-Chang Chen, and Yeh-Ching Chung. 2014. Optimizing
energy consumption with task consolidation in clouds. Information Sciences 258 (Feb. 2014)
452462.
14. L. Wang, S. U. Khan, D. Chen, J. Koodziej, R. Ranjan, C.-z. Xu, and A. Zomaya,
“Energy-aware parallel task scheduling in a cluster,” Future Generation Computer Systems,
2013.
15. White paper by T systems intel, “DataCenter 2020: first results for energy-optimization at
existing data centers”, july 2010.
16. RichaSinha et al, Int. J. Comp. Tech. Appl., Vol 2 (6), ISSN: 2229-6093, 2012,
pp. 2041-2046.
17. Dzmitry Kliazovich, Pascal Bouvry, and Samee U. Khan. 2013. DENS: Data center
energy-efficient networkaware scheduling. Cluster Computing 16, 1 (2013), 6575, 2013.
Implementation of REST API Automation
for Interaction Center
Prajakta S. Marale and Anjali A. Chandavale
Abstract AVAYA Interaction Center is developed for communication between

customer and the company. The Interaction Center provides business-class control
of contact center communications across several channels like voice, email, web
chat. Chat is a progressively popular communication channel due to growth in the
era of Internet. Existing chat generator tool supports only IE browser. So, there is a
need that the tool should be browser independent. The paper proposes the tool
based on REST API. In the proposed system, the REST API of Chat media of
CSPortal is captured through SOAPUI tool. The request–response of REST API is
captured for different resources like /csportal/cometd/of CSPortal Chat Media. The
result shows the proposed approach is more efficient than existing tool and chat is
live for more than 2 min.
Keywords AVAYA interaction center SoapUI REST API, etc.
1 Introduction
A REST (Representational State Transfer) API (Application Programming Interface)

explicitly captures supremacy of HTTP (Hypertext Transfer Protocol) methodolo-
gies. A RESTful API [1] is application program interface which uses HTTP requests
to GET, POST, PUT, and DELETE data. Mostly browser uses REST technology can
be thought of as the language of the Internet. REST allows users to connect and
interact with cloud services. Social media or E-commerce websites nowadays mostly
uses RESTful APIs. Transactions break down for creating sequences of small
modules by using REST APIs. Each module directs a particular part of the trans-
action. This module provides developers a lot of flexibility, but it can be challenging
P. S. Marale (&) A. A. Chandavale

MIT College of Engineering Pune, Pune, India
e-mail: prajaktamarale@gmail.com
A. A. Chandavale
e-mail: Anjali.chandavale@mitcoe.edu.in

https://doi.org/10.1007/978-981-10-7245-1_28
274 P. S. Marale and A. A. Chandavale
for developers. For communication in between client and agent of Interaction center
of AVAYA, CSPortal webpage is available. Here, our main focus is on the
CSPortal API. We are going to capture different responses from providing request
and resource parameters of API through SOAPUI tool. Rest of the paper is divided
into following sections. Section 2 presents the literature survey. Section 3 describes
the proposed system. In Sect. 4, actual result is outlined, Sect. 5 describes conclu-
sion, and finally acknowledgment.
2 Related Work
Hamad et al. [2] evaluated a RESTful web service for mobile devices. Author
developed string concatenation and float number addition web services by using
RESTful and conventional SOAP web service. The performance evaluation results
show the advantages of using RESTful web services over SOAP web services for
mobile devices are higher. RESTful web service offers a perfectly good solution for
the majority of implementations, with higher flexibility and lower overhead.
Belqasmi et al. [3] present a case study which is based on a comparison in between
SOAP-based web services and RESTful web services. This case study is related to
the end-to-end delays for executing the different conferencing operations like cre-
ating a conference, add participant, etc. The delays via the REST interface are three
to five times less than when using SOAP-based web service.
Chou et al. [4] presented algorithms and methods to materialize the extensibility
of REST API under hypertext-driven navigation. First, the author described a
Petri-Net based REST Chart framework. Second, the author described and char-
acterized a set of important hypertext-driven design patterns. Third, the author
presented a novel differential cache mechanism that can significantly reduce the
overhead of clients in hypertext-driven navigation.
We studied that SOAP-based services are much harder to scale than RESTful
services. SOAP protocol uses XML for all messages, which make the message size
much larger and less efficient than REST protocol. RESTful web service provides
efficient, flexible, better performance than SOAP-based web services.
3 Proposed Work
CSPortal Webpage is for customer for interaction between customer and agent of
Interaction Center. The CSPortalWebAPI is a client-side JavaScript API included in
the CSPortalWebAPI SDK. The high-level architecture for the custom client and
various other components involved. Client and agent can communicate through
REST API Façade. For communication purposes, REST Protocol is used. SoapUI
tool can test this REST protocol communication effectively. The API enables
web-based clients to integrate various collaboration options available in Interaction
Implementation of REST API Automation for Interaction Center 275
Center, such as Email, LiveChat, and Schedule Call Back. The Live Chat API is
required for initiating the chat between client and agent. The actual end-to-end flow
of CSPortal chat media is depicted in Fig. 1 which is described below.
1. Customer escalates the chat by calling escalatemedia API through SOAPUI tool.
The chat lands on the Apache server.
2. CSPortal Client tomcat redirector escalates the incoming chat to CSPortal Server
via HTTP protocol.
3. CSPortal Server redirects the chat to remote ICM server. ICM sends makecall
request to ICMBridge and sends the chat request to the attribute server.
4. Attribute server creates eDU making a request to EDU Server. It sends the
incoming event to WebACD. WebACD sends onhold to ICMBridge.
5. ICMBridge forward onhold messaging to ICM. ICM sends onhold messages to
HTMLClient Servlet. HTMLClient receives the onhold messages. The messages
are displayed in the transcript.
6. WebACD contacts workflow server to run the Qualify Chat flow. WebACD
sends task assignment message to agent via paging server.
7. Agent accepts the task assignment. Agent enters room. ICM notifies the
CSPortal Server on Tomcat Server which in turn notifies the CSPortal client and
connected message is sent to Customer. CommetD polling takes place to publish
messages on the server to the CSPortal Javascript running on SoapUI as events
from the server.
SoapUI is the most useful open-source, cross-platform API Testing tool. SoapUI
captures different API’s required for the chat media of CSPortal. Some API’s
resources are listed below.
Fig. 1 End-to-end flow of CSPortal API [1]

276 P. S. Marale and A. A. Chandavale
(1) Csportal/cometd/handshake—CometD handshake request to establish con-

nection. (2) Csportal/cometd—CometD request to send commands to AIC chat
server. (3) Csportal/cometd/connect—CometD request to periodically poll and fetch
the Chat Events from AIC Chat Server.
For implementation of automation of REST API, we provide URI, resources and
request parameter as input to SOAPUI tool and get response. All the requests to
servers go through SoapUI Tool instead of browsers. So for establishing chat, we
are not dependent on the browsers.
4 Result
A transaction is breaking down into sequences of small modules. Single chat

scenario considers different modules which considered as a different API’s. To
capture chat scenario through SOAPUI, we need to call some set of API’s. For Chat
purpose, we need to call listed API’s in Sect. 3. Request and response of some
API’s is depicted in the Fig. 2.
The response we have captured through SOAPUI tool has been compared with
logs at server side. The request goes to the server side, we can check through logs.
Through SOAPUI, we are getting higher chat duration as compared to existing chat
generator tool. Existing chat generator tool works on only IE instance. Here by
using SoapUI tool, we eliminate the browser dependency (Fig. 3).
Fig. 2 Request–response of csportal/cometd/(Command: open) API

Implementation of REST API Automation for Interaction Center 277
Fig. 3 Logs that shows correct response for request through SOAPUI tool
Table 1 Chat duration Channel Existing chat duration Proposed chat

duration
Live chat 1–2 min 3–4 min
Sometimes, our request might be misplaced or it may not go to appropriate

server so through SoapUI we can get accurate response and we get clear under-
standing of the request (Table 1).
5 Conclusion
The paper has concentrated on REST API’s. Different responses are captured for
different resources like /csportal/cometd/connect of CSPortal REST API. We have
not attempted to compare the effectiveness of tools related to the REST API. We
believe that the automating REST API through SOAPUI tool improved the accuracy
and chat is alive for more than 2 min. We have included a list of references enough to
deliver a more detailed understanding of the approaches described. We ask for pardon
to researchers whose important contributions may have been neglected.
Acknowledgements The author would like to thank Mr. Harshad Apshankar for his guidance.
The work presented in this paper is supported by AVAYA PVT LTD.
References
1. http://searchcloudstorage.techtarget.com/definition/RESTful-API.
2. Hatem Hamad, Motaz Saad, and Ramzi Abed “Performance Evaluation of RESTful Web
Services for Mobile Devices” in International Arab Journal of e-Technology, Vol. 1, No. 3,
January 2010.
3. Fatna Belqasmi, Jagdeep Singh, SuhibBanimelhem, Roch H. Glitho “SOAP-Based Web
Services vs. RESTful Web Services for Multimedia Conferencing Applications: A Case Study”
in IEEE Internet Computing, Volume 16, Issue 4, July–Aug. 2012.
4. Li Li, Wu Chou, Wei Zhou and Min Luo “Design Patterns and Extensibility of REST API for
Networking Applications”, in IEEE Transactions on Network and Service Management, 2016.
Predictive Analysis of E-Commerce
Products
Jagatjyoti G. Tuladhar, Ashish Gupta, Sachit Shrestha,

Ujjen Man Bania and K. Bhargavi
Abstract For the past few years, there has been increasing trend for people to buy
products online through e-commerce sites. With the user-friendly platform, there is
loop hole which does not guarantee satisfaction of the customers. The customers
have the habit of reading the reviews given by other customers in order to choose
the right product. Due to high number of reviews with mixture of good and bad
reviews, it is confusing and time-consuming to determine the quality of the product.
Through these reviews, the vendors would also want to know the future trend of the
product. In this paper, a predictive analysis scheme is implemented to detect the
hidden sentiments in customer reviews of the particular product from e-commerce
site in real-time basis. This serves as a feedback to draw inferences about the quality
of the product with the help of various graphs and charts generated by the scheme.
Later, an opinion will be drawn about the product on the basis of the polarity
exhibited by the reviews. Finally, prediction over the success or failure of the
product in the regular interval of the timestamp is done using time series forecasting
method. A case study for iPhone 5s is also presented in this paper highlighting the
results of rating generation, sentiment classification, and rating prediction.

Keywords Predictive analysis Product rating Amazon Review

Polarity Preprocessor Sentiment identifier
J. G. Tuladhar (&) A. Gupta S. Shrestha U. M. Bania K. Bhargavi

Department of Computer Science and Engineering, Siddaganga Institute of Technology,
Tumkur, Karnataka 572103, India
e-mail: jagat.tula@gmail.com
A. Gupta
e-mail: 1foraashish@gmail.com
S. Shrestha
e-mail: ssachit3761@gmail.com
U. M. Bania
e-mail: ujjenms.bania@gmail.com
K. Bhargavi
e-mail: bhargavi.tumkur@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_29
280 J. G. Tuladhar et al.
1 Introduction
Predictive analysis is the advanced analysis technique that is used to predict

unspecific future events. It uses techniques like machine learning, statistical algo-
rithm, artificial intelligence, and data mining in order to analyze present-day data so
that the predictions on future outcome are based on historical data. The goal of the
predictive analysis is to do an in-depth analysis and to go further to know what
would happen in order to provide the best assessment of what may happen in the
future [1–3].
Predictive analysis has become a greater tool which is used to establish customer
responses and purchases and promote multiple opportunities. This model helps to
attract business, retain it, and increase their most valuable customers. A number of
the companies use this model to predict and manage their resources. These analytics
help organizations to function in a more efficient manner [4, 5].
In this paper, the reviews given by the customers over a particular product are
extracted. Sentiment analysis is performed on the reviews to generate a product
rating on a weekly basis. Predictive analysis is used to predict the rating for the nth
week using the records of the previous weeks. This results in the prediction of the
success and failure of the product.
The rest of the paper is organized as follows: Sect. 2 gives some of the related
works, Sect. 3 provides proposed scheme to be used in this paper, Sect. 4 discusses
the proposed algorithm in detail, Sect. 5 deals with the case study and finally
Sect. 6 draws the conclusions.
2 Related Works
Mining and summarizing customer reviews were discussed in [6–8]. This sums up
an array of positive words and negative words focuses on user reviews. The neg-
ative array of words has 4783 and the positive array of words has 2006. Both arrays
also contain various words which are spelled incorrectly that are frequently used in
social media posts and tweets. Sentiment taxonomy deals with categorization
problem, and features containing sentiment, views or opinions information need to
be identified before the categorization. Here, they considered only the polarization
of the reviews, i.e., whether it is positive or negative. However, the rating of the
product was not generated by extracting the sentiments in the preprocessed reviews
[9].
Sentimental analysis using subjectivity summarization was discussed in [10]. In
this, a unique machine learning methodology was applied that applies sentence
categorization techniques to just the subjective part of the sentence. The sentences
in the document can be either objective or subjective and that apply a novel
machine learning classification tool to extract results. This can avert the polarity
classifier from taking into consideration trivial or ambiguous text. Subjectivity
Predictive Analysis of E-Commerce Products 281
contents can be extracted from the document that are provided to the users as sum
up of the sentiment content of that document. The result represent the subjectivity
extracted and created sentiment information of that documents in a very compact
way with discussed on sentence-level categorization [11].
Sentiment analysis and opinion mining were discussed in [12]. Here in the text
or the sentence, the problem was to classify the text or sentence into a particular
sentimental polarity, negative, or positive. It mainly focuses on text or sentence
which has three levels of sentimental polarity classification, and they are the entity
and aspect level, the sentence level, and document level. The entity and aspect level
focuses on what exactly users dislike or like from their reviews and opinions.
Likewise, the sentence level focuses on each and every sentence with sentiment
classification. While, the document level focuses on the content of the whole
document and expresses the positive or negative sentiment. Here, the written text
categorization and their polarity are discussed however the rating of the product
was not generated and predictive analysis was not taken into consideration [13].
3 Proposed Scheme
The scheme for predictive analysis of e-commerce products is given in Fig. 1. The
customer reviews on electronic gadgets are scrapped from various e-commerce
websites like Amazon, Flipkart, E-bay, and Snapdeal using Python-based Web
Scrapper and stored in review log file. The scheme mainly consists of four func-
tional modules. They are Preprocessor, Sentiment Identifier, Rating Generator, and
Rating Predictor.
The Preprocessor filters the reviews based on the part of speech exhibited by
each and every token in the review. The Sentiment Identifier classifies the review as
positives, negatives, or neutrals based on their polarity. The Rating Generator rates
the review based on the weekly sentiment provided by the customer over the
product. Finally, the Rating Predictor aggregates the previous rating on the monthly
scale to predict the success and failure of the product.
Fig. 1 Scheme for predictive analysis of e-commerce products

4 Algorithm
In this section, the detailed discussion on every functional module of the proposed
scheme along with their algorithm is given.
Preprocessor
In the Preprocessor module, Tokenization is a process which breaks down the
review into tokens of words separated by spaces between them. Then appropriate
part of speech like nouns, pronouns, verbs, adjectives, adverbs, and prepositions
will be tagged to each of the tokens in the review using NLTK (Natural Language
Tool Kit) Tagging process. From the tagged tokens, the based words are converted
using NLTK Stemming process is carried out on the review log via Suffix-stripping
algorithm shown in Fig. 5. The reviews are then Filtered based on the parts of
speech. The stop words in the reviews are then removed using stop word remover
process. The working of the Preprocessor is given in Algorithm 1 and functioning
of the Preprocessor is given in Fig. 2. A sample example depicting the process of
tokenization and tagging is given in Figs. 3 and 4.
Fig. 2 Preprocessor
Fig. 3 Example of tokenization
Fig. 4 Example of tagging

Fig. 5 Example of stemming
Fig. 6 Sentiment identifier
Algorithm 1: Preprocessor
Step 1: Start
Step 2: for each review Ri from the review log RL
use word_tokenize to generate the list of
tokens T
tag Part-Of-Speech to each ti ϵ T {(ti,
POSi), (tj, POSj), (tk, POSk), …}
convert to base words using stemming
process
filter the words based on Part-Of-Speech
remove stop words in the review
end for
Step 3: Output the pre-processed review PR
Step 4: End
Sentiment Identifier
The Sentiment Identifier takes the preprocessed list as input and compares each
element in the list with the Bag of Words with the reference from Naïve Bayes
classifier. If the words in the preprocessed list match with the Bag of Words, the
sentiment value of those bag of words is extracted. The reviews are classified into
positive, negative or neutral based on their sentiment value using Naïve Bayes
Fig. 7 A sample of
sentiment identifier output
classifier. The working of Sentiment Identifier is given in Algorithm 2, the func-

tioning of the Sentiment Identifier is given in Fig. 6 and a sample example is
demonstrated in Fig. 7.
Algorithm 2: Sentiment Identifier
Step 1: Start
Step 2: Intialize a bag of words in sentiment
database
Step 3: for each PRi ϵ PR
match PR[i] in PR with the bag of words
if match is not found
discard the item
else
sentiment_value[] = Extract the value
from the bag of words
end if
calculate ∑sentiment_value[]
end for
Step 4: if ∑sentiment_value > 0
PR is "positive"
else if ∑sentiment_value < 0
PR is "negative"
else
PR is "neutral"
end if
Step 5: Store PR in Sentiment Database
Step 6: End
Rating Generator
The Rating Generator assigns a rank to each of the preprocessed sentiment database
based on the mean of the sentiment values exhibited by every review over a week.
The working of Rating Generator is given in Algorithm 3, the functioning of the
Rating Generator is given in Fig. 8 and a sample example is demonstrated in Fig. 9.
Fig. 8 Rating generator
Fig. 9 A sample of rating

generator output
Algorithm 3: Rating Generator

Step 1: Start
Step 2: Accept pre-processed PR review along with the
sentiment value.
Step 3: for each PRi ϵ PR
sum= ∑sentiment value {PR1, PR2, PR3, …}
average = sum / total no. of reviews with
sentiment value in that
time period
end for
Step 4: Rate the product according the average
Step 5: Output product rating of every review
Step 6: End
Rating Predictor
The Rating Predictor forecasts the rating of the product in the upcoming time using
time series forecasting technique. The calculation of the rating for the nth week uses
the records of the previous weeks. This results in the prediction of the success and
failure of the product. The working of Rating Predictor is given in Algorithm 4, the
functioning of the Rating Predictor is given in Fig. 10.
Fig. 10 Rating predictor
Algorithm 4: Rating Generator

Step 1: Start
Step 2: Accept the product rating
Step 3: for each product rating at time ∆t
if product rating ∆t-1<=product rating ∆t
Product will be a success
else
Product will fail
end if
end for
Step 4: End
5 Result
In this section, the implementation details of the proposed work along with the
results obtained are discussed. The tools used are Python-based Web Scrapper, used
to extraction of reviews from the Amazon Website real time, NLTK for prepro-
cessing, text processing API for Bag of Words comparison, and finally ARIMA
model for Trend prediction modeling.
1000 reviews on Iphone 5s have been taken into consideration during imple-
mentation of the proposed work, the most common keywords used while
expressing the reviews were battery life, touch screen and price of the Iphone 5s.
Negative sentiment was expressed in reviews by using adjectives like bad, poor,
worst, and horrible whereas positive sentiments were expressed by using sentiments
like nice, good, best, excellent, and amazing.
The Rating Generator output over Iphone 5s is shown in Fig. 11. It was found
out that majority of the reviews were above average (5). The Sentiment Identifier
output over those reviews is depicted in Fig. 12 where around 450 reviews
exhibited positive sentiment, 250 reviews exhibited negative sentiments and
remaining were found to be neutrals. The Rating Predictor output of Iphone 5s over
next 2 months is shown in Fig. 13. The ARIMA modeling output is shown in
Fig. 14. It was found that Iphone 5s receives high rating in starting month and later
the rating will remain saturated.
RaƟng per review (Range 0 -10)
RaƟng 10
RaƟng 9
RaƟng 8
RaƟng 7
RaƟng 6
RaƟng 5
RaƟng 4
RaƟng 3
RaƟng 2
RaƟng 1
0 2 4 6 8 10
Value
Fig. 11 Rating generator output
Fig. 12 Sentiment identifier

output
9
8
RaƟng of Iphone 5s 7
6
5
4
3
2
1
0
Jan 1st Feb 1st Mar 1st Apr 1st May 1st June 1st
Months
Fig. 13 Rating prediction output
Fig. 14 Prediction output
6 Conclusion
This paper presents a novel scheme for predicting the rating of the product in
E-commerce websites. The aim of this paper is to provide the vendor a framework
to quickly gain insight about what customer in their field think about the product. It
brings out the fact that there are sentiment values hidden in the reviews which can
be used to generate a rating for the product. With the help of time series algorithm,
the future market trend of the electronic products is predicted. This is a practical
way to use unstructured data in the form of review to inform a vendor on how
customers think about and react to product.
References
1. Mohammad Salehan and Dan J. Kim, “Predicting the Performance of Online Consumer
Reviews: A Sentiment Mining Approach”, Elsevier Journal on Information Processing and
Management, 2016.
2. ShahriarAkter and Samuel FossoWamba, “Big data analytics in e-commerce: A systematic

review and agenda for future research”, Electronic Markets 26 173–194, 2016.
3. Wikipedia, https://en.wikipedia.org/wiki/Predictive_analytics.
4. Pavithra B, Dr. Niranjanmurthy M, Kamal Shaker J and Martien Sylvester Mani F, “The
Study of Big Data Analytics in E-Commerce”, International Journal of Advanced Research in
Computer and Communication Engineering, 2016, Vol. 5, Special Issue 2.
5. Galit Shmueli and O. Koppius, “Predictive Analytics in Information Systems Research”, MIS
Quarterly, 2011, Vol 35 No. 3 pp. 553–571.
6. Hu M and Liu B, “Mining and summarizing customer reviews”, In Proceedings of the tenth
ACM SIGKDD international conference on Knowledge discovery and data mining, 2004,
pp 168–177.
7. Rui Xia, FengXu Jianfei Yu, YongQi and ErikCambria, “Information Processing and
Management Polarity shift detection, elimination and ensemble: A three-stage model for
document-level sentiment analysis”, Elsevier Journal on Information Processing and
Management, 2016, vol 52, Issue 1, pp 36–45.
8. Hu M and Liu B, “Mining and summarizing customer reviews”, In Proceedings of the tenth
ACM SIGKDD international conference on Knowledge discovery and data mining, 2004,
pp 168–177.
9. Yorick Wilks and Mark Stevenson. “The grammar of sense: Using part-of-speech tags as a
first step in semantic disambiguation”. Journal of Natural Language Engineering, 1998,
pp 135–144.
10. Pang B and Lee L, “A sentimental education: Sentiment analysis using subjectivity
summarization based on minimum cuts”. In: Proceedings of the 42nd Annual Meeting on
Association for Computational Linguistics, Stroudsburg, PA, USA, 2004, pp 1–8.
11. Jeonghee Yi, Tetsuya Nasukawa, Razvan Bunescu, and Wayne Niblack, “Sentiment analyzer:
Extracting sentiments about a given topic using natural language processing techniques”, In
Proceedings of the IEEE International Conference on Data Mining (ICDM), 2003.
12. Liu B, “Sentiment Analysis, and Opinion Mining. Synthesis Lectures on Human Language
Technologies”, Morgan & Claypool Publishers, 2012, vol. 5, no. 1, pp 1–67.
13. Ellen Riloff and Janyce Wiebe, “Learning extraction patterns for subjective expressions”,
Proceedings of conference on Empirical methods in natural language processing, 2003,
pp. 105–112.
User Privacy and Empowerment: Trends,
Challenges, and Opportunities
Prashant S. Dhotre, Henning Olesen and Samant Khajuria
Abstract Today, the service providers are capable of assembling a huge measure
of user information using big data techniques. For service providers, user infor-
mation has become a vital asset. The present business models are attentive to collect
extensive users’ information to extract useful knowledge to the service providers.
Considering business models that are slanted towards service providers, privacy has
become a crucial issue in today’s fast-growing digital world. Hence, this paper
elaborates personal information flow between users, service providers, and data
brokers. We also discussed the significant privacy issues like present business
models, user awareness about privacy, and user control over personal data. To
address such issues, this paper also identified challenges that comprise unavail-
ability of effective privacy awareness or protection tools and the effortless way to
study and see the flow of personal information and its management. Thus,
empowering users and enhancing awareness are essential to comprehending the
value of secrecy. This paper also introduced latest advances in the domain of
privacy issues like User-Managed Access (UMA) that can state suitable require-
ments for user empowerment and will cater to redefine the trustworthy relationship
between service providers and users. Subsequently, this paper concludes with
suggestions for providing empowerment to the user and developing user-centric,
transparent business models.
Keywords Big data Personal information Privacy Privacy by design

Awareness Trust
P. S. Dhotre (&) H. Olesen S. Khajuria

CMI, Aalborg University Copenhagen, Copenhagen, Denmark
e-mail: psd@es.aau.dk
H. Olesen
e-mail: olesen@cmi.aau.dk
S. Khajuria
e-mail: skh@cmi.aau.dk

https://doi.org/10.1007/978-981-10-7245-1_30
292 P. S. Dhotre et al.
1 Introduction
Every day, Internet users send ten billion text messages and make one billion posts.
This sharing of messages and post is said as “mass-self-communication” by Manuel
Castells [1]. This communication contains personal information. Hence, along with
protection of personal information, there is a strong need to notify users about data
collection, its use and what happens next to their data.
The developed privacy protection principles like W3C P3P [2], Privacy by
Design (PbD) [3], “Laws of Identity” from Kim Cameron [4] emphasize on noti-
fication of data collection, user control over data, minimum data disclosure, and
user-defined consent for data use. However, today user has no control over their
personal information. Service providers have already assigned a value to personal
data. They apply it to build accurate customer profiles, do target advertising, sell
and share collected information or analyze information to other entities without user
notification and consent [5–7].
Personal data is becoming an integral voice of service use on the network. The
present business models focused less on user privacy. The current framework for
the protection of user privacy includes ENISA [8], OpenID [9], UMA [10] focused
on minimum data disclosure to have fine-grained control.
Data brokers make their revenue by selling user information to publishers and
advertisers [11]. Hence, big data is now the big business. Practices of data brokers’
raise privacy issues. The end users are unaware of the involvement of data brokers
in the bargain. There is no notification to end user about the user information
collection, manipulation, and sharing. Besides, a recent study taken in India
highlighted the fact that the user privacy concerns could be noteworthy [7]. 65% of
respondents are unaware of privacy issues, rights, and knowledge. Moreover, 80%
of respondents are concerned about the solicitation of personal data. This indicates
that they want to experience more about the personal information aggregation.
Therefore, there is a strong need to empower users in understanding the value of
privacy and flow of personal information between service providers and user. This
will offer an opportunity to both users and service provider to come together on a
common platform and to focus on accessing, sharing, and managing personal
information in a manipulated way. Without user awareness and empowerment, it is
exceedingly hard to protect users’ privacy. From the user’s perspective, the new
business model should consider the minimum information collection, users’ rights,
consent, and choice. Likewise, from the service providers’ point of opinion, it is as
significant to identify valuable consumers/users and building a long relationship
that establishes trust between them.
In this report, we introduced some of the principal challenges for user
empowerment like how to visualize information flow, controlled bargain, and user
consent gain. Also, we presented research guidelines that address challenges like
the formation of a visualization tool, semiautomated negotiation process where both
(users and service providers) will receive an equal return in information bargain.
User Privacy and Empowerment: Trends, Challenges, and … 293
2 Users, Data Brokers, and Privacy Concerns
2.1 Internet User
A user has different attributes and is widely diffused. As shown in Fig. 1, the user
interacts with social media, online subscribers, etc., by disclosing a set of user
attributes. Depending on context, the set of attributes forms the partial identity of
the user. Hence, an identity comprised multiple identities [12]. The user must share
their identity with the separate service provider. At the other end, data brokers
collect, store, manage, manipulate, and share extensive information about end users
without direct interaction with them.
In many instances, it is shared and sold to data brokers [13]. Still, worries about
privacy have arisen with the increasing flow of personal information in many
directions to different companies. Thus, it gets more and more hard to protect the
user’s data privacy.
2.2 How Do Data Brokers Work?
Data brokers collect the user data from the discrete sources like governments
(central and state government), openly available media (social media, the Internet,
blogs), and commercial sources (phone companies, banks, car companies, and so
on). After analysis of gathered data, they separate the users as sports loving,
interested in cancer solutions, trekking enthusiasts, blended with political posts,
food choices, and hence along. Besides, they organize collected information into
Fig. 1 Information flow between user, service provider, and data broker
different segments, which are individually identifiable dossiers that containing, e.g.,
marital status, location, income, shopping, hobbies, job, and travel plan. Then,
segmented information is ready to sell to other data brokers, advertisers, or gov-
ernment sectors without the user’s direct awareness [6].
2.3 Who Are the Data Brokers?
Acxiom, the largest data broker, has gathered information relating to 700 million
customers worldwide and 3000 data segments for nearly every U.S. user [14]. In the
Indian setting, “Zifzi” company provides IT and marketing solutions by providing
12 categories of Indian databases (electronic mail addresses, mobile numbers,
business seekers, credit card holders, statewide and citywide databases, and so on)
[15]. People can buy information like profiles, company, religion, usernames, salary
and income sources, medical history, drug abuses, and sexual orientation from
many information brokers. Internet users do not recognize that the greatest menace
to personal data privacy comes from them.
3 State-of-the-Art Technologies and Frameworks
In the field of privacy awareness and protection, there have been a long practice and
activities. Some of the research includes:
The users do not really know to what extent their data are collected, used, and
dealt. Consent and privacy conditions (terms and conditions, agreements) are not
effective for reading because of the length and complexity of language. According
to the Futuresight [16], 50% of mobile users are giving their consent without
reading the agreement. Hence, there is a requirement to raise the awareness of the
users by analyzing these documents and implementing an easy-to-understand way
to visualize the content of the agreement.
Lorrie Cranor said that “It is difficult to protect your privacy even if you know
how” [17]. The conclusion of a survey conducted by her is that the respondents
were unfamiliar with the tools, struggled, and failed to protect their privacy. The
article on “The Economics of Personal Data and Privacy: 30 years after the OECD
Privacy Guidelines” [11] revisits the Fair Information Practice principles including
transparency, purpose specification, and use limitation. Looking at the major
reclamation of the European Union legal framework on the security of personal data
[18], there is a demand to strengthen individual rights of protection by raising user
awareness. In the Indian context, despite the Indian Information Technology Act
2000 [19], there is a lack of consumer organization to address privacy issues.
The DSCI [20] focuses on a privacy awareness program for employees of the
governing body.
3.1 Tools to Support Privacy
“Disconnect me” is a software program which can offer you with a list of third
parties like ad networks, analytics, and marketing companies, who is tracking your
clicks, your interests, page views, etc. This information goes to data brokers [6].
Online Permission Technologies have provided an online privacy guard called
MyPermission for Android and Apple devices [21]. MyPermission provides a menu
for users to handle all service permissions. So, the depth of these companies and
darkest corners has only been investigated to a very limited extent. There is a need
to exploit the concepts of such tools and provide new enhancement for them that
will allow the user to decide what to do with their personal data.
3.2 Privacy Enhancing Technologies
The European Network and Information Security Agency (ENISA) carried out a
study and survey on security, privacy, and usability of online seals [8]. Their
findings reaffirmed: “At the moment, it is too difficult for consumers to compare
different information practices of online service providers” and also identified five
main challenges of the seals: awareness, standards, validity checks, usability, and
presence.
A framework of User-Managed Access (UMA) [10] helps the users to find out
the value of their personal data and its flow control. Different scenarios were
discussed to show fine-grained control of user attributes [12]. The OAuth protocol
for delegation of authorization is becoming the standard mode to gain access to
protected resources for websites and apps using tokens [22]. Also, ongoing research
addresses Big Privacy and the Personal Data Ecosystem (PDE) based on PbD [3].
Seven architectural elements are a personal cloud, semantic data exchange, trust
framework, identity and data portability, data by reference or subscription, pseu-
donyms, and contractual data anonymization.
4 Privacy Issues from Users and Service Providers View
The initial research of the granted project from the National Science Foundation’s
Secure and Trustworthy Computing initiative program [SaTC] represents online
privacy issues filed in FTC enforcement actions [23]. The complaints are examined
and separated as established in Table 1. Most of the descriptive claims relate to data
disclosure without user consent and partial implementation measures for the user
privacy protection.
Table 1 Summary of privacy issues filed in FTC enforcement actions
296
Classification grouping # of Examples of complaint Remark/Issues

(based on privacy claims
violation claims)
Unauthorized disclosure 51 Use of personal information for financial gain Use, disclosure, processing of personal information
of personal information (identity theft)
Use of personal information for unspecified purposes
Unfair personal information collection practices
Unauthorized use and sale of information
Unauthorized third party access without protection
Disclosed, offered/sold personal information
Surreptitious collection of 111 No notification of privacy changes Notice, information collection, opt-in and opt-out
data Sharing of online shopping carts to third parties
Failed to reveal types of information collection
Unnecessary information collection
Failed to disclose the information collection sources.
Obtain customers financial information
Without user consent, gathering of personal
information through false way
Failed to notify the software installation that pass on
personal information
Continued to collect and display information even
after opt-out
(continued)
P. S. Dhotre et al.
Table 1 (continued)
Classification grouping # of Examples of complaint Remark/Issues
(based on privacy claims
violation claims)
Inadequate 275 Failed to develop, implement, and sustain a complete Threat detection and prevention, employees and training,
information security program for protection of access credentials and authorization, Breach response,
information Information disposal
Failed to work on security in the design of software
delivered to consumer
Failed to implement acceptable security policies for
sensitive information
Failed to implement satisfactory measures to detect,
monitor and prevent unauthorized access to
information
Failed to train/inform employees regarding consumer
information privacy
Failed to monitor outbound network traffic to
recognize and block unauthorized transfer sensitive
information
User Privacy and Empowerment: Trends, Challenges, and …
Failed to implement complete policies regarding

appropriate collection, handling, and removal of
personal information
Failed to evaluate information security program in
light on the possible risks
Wrongful retention of 6 Storage of personal information for a long time Information store
personal information unnecessary
297
4.1 User Understanding and Awareness About Privacy
There is a simple ignorance and user’s apathy toward their privacy because privacy
is something whose value is completely understood only after it has been taken
away. Users are unaware of the repercussions of exposing personal information [24,
25]. The users are unaware of online practices and do not recognize what happens
to their data once it is picked up by the service suppliers. It is found that many
consumers claim that they do not like anyone knowing anything about themselves
or their habits unless they decide themselves to share that information [23, 24].
4.2 Privacy Policy Understanding
Privacy policies are convoluted and are characterized by a lot of jargons which are
difficult for the naive users to understand. Privacy policies are complicated, take too
long to understand and read [24]. This is a blindly accepted item on websites
despite the user express increasing concern about personal data collection practices
[26]. This document is important and acts as a single source of information to know
how service providers collect, use, and share their data. The challenge is to be
capable of studying and visualizing privacy policy in an easygoing manner.
4.3 User Control
Most web service providers do not have selective data submission, i.e., the user
does not have complete control over the amount of personal information they want
to enter. It is often mandatory to enter, birthday, phone number, organization, and
so on while signing up for any service [12].
4.4 Service Provider’s Issues
The current approach of the service provider relies on the formation of the business
model that requires end users to provide maximum information from the initial
moment of dealing. Thus, the challenge for service providers is to get valuable
customers from a large pool of available customers by asking least/minimum
information from them, providing privacy assurance for maintaining a long-term
relationship with them. Therefore, privacy issues from the users and service pro-
vider’s view are shown in Table 2.
Table 2 Summary of privacy issues

Entity Issues
User side Users disclose more information than needed
Users are unaware of online practices and privacy and privacy policy
Users are unaware of privacy protection tools
Users are not aware of what happens to their data
Users are not notified for collection and sharing personal information in a
visible way
Users are unable to control and manage their data
Users do not give consent or no option for opt-out
Service provider Maintaining the privacy of customers’ information
side Binding a customer connection (trust establishment)
5 New Challenges and Recommendations
To empower the users and protect user’s privacy, the new research challenges are
summarized in Table 3.
5.1 User Awareness and Personal Flow Visualization
To improve end user privacy decisions, a new visualization tool for measuring
privacy can be developed using visual cues schemes. Visual cues about the privacy
Table 3 New challenges for privacy awareness and protection

Type New challenges
User awareness and How do we create awareness among the end users about
visualization of personal data privacy and online practices?
How to enhance end users’ understanding and control over
their personal data
How do we visualize the flow of personal data?
How do we create awareness among end users about data
collection sources, methods, purposes of service providers,
and their relations with other entities?
Assurance and truth How do we ensure that service provider will not use
information for any other purposes than what user have
given consent for?
Privacy How do we protect user’s privacy?
User consent gaining How to increase the general acceptability of the user–
provider liaison?
User-centric businesses with How do we provide user-centric business cases to the
transparency service provider?
How do we allow transparency of the personal data
management? How will the user be enabled to see where
their data are flowing and for what purposes?
intrusiveness of services and mobile applications must communicate using a

visualization tool. This should focus on the completeness of privacy policy and
development of a dashboard to assist users in making consent judgments. Machine
learning-based technologies will facilitate the users to figure out the terms and
privacy policies.
A bargaining mechanism must be implemented to make an informed decision on
how much users accept to disclose for what in return. It should include easy-to-use
and easy-to-understand semiautomated negotiation process, where the user can
determine the benefits from service providers and be given whether they care to
furnish this information.
5.2 Assurance and Truth
The “accept or reject” strategy of service providers should change in future. The
service provider must notify the purpose of data information to end user and should
have a panel to see the data collected from the user, its purpose, and a list of the
third party involved. They should not share the information without notification and
proper consent of Internet user.
5.3 Privacy Protection
PbD [3] is a major initiative from Canada, providing a set of principles for
incorporating privacy protection in applications and services from the beginning,
without sacrificing essential functionality. One may think of the analogy with a
personal privacy firewall, where all ports are held in private by default, and you
only open the ports that are needed.
Hence, on that point is a substantial demand to consider privacy-by-design
principles in the fundamental concept of the privacy protection solution. The
solution should be implemented that must comply with the existing regulations as
per data protection directives and the framework proposed by the county
commission.
5.4 User Consent Gaining
End user consent is gaining the desired level of permission from them while pro-
viding services. User consent is received through long, complicated text mentioned
in privacy policies presented to end users in an ineffective way. Thus, a new method
must furnish an effective and efficient means of requesting consent from users. This
is a very challenging task because knowledge, reading time, and location of the user
change from person to person. Hence, cognitive science and human–computer

interface principles will assist in this research.
5.5 User-Centric Businesses with Transparency
The marketing strategy of the service provider should focus on specific and needy
users’ instead of looking at every user. This will facilitate the service providers to
know their valuable consumers. User choices and interests must be protected from
other parties. A novel mechanism must allow for transparency of the personal
information management process through only transactions. It must assist end users
and service providers in their trustworthiness to build better relations. History-based
reputation will help users in building trust between user and service supplier.
6 Privacy Protection Parties and Responsibilities
Figure 2 represents interested and concerned entities for the protection of personal
information privacy: Individual users, service providers/website developers, data
brokers/advertisers/third party app developers, and government and governing
bodies.
6.1 Individual Users
The user needs to be more mindful of their personal information disclosure when
they are online. The minimum information should be shared with the other com-
municating parties. Before availing any services, the user must read privacy policies
set by the service suppliers. They must be cognizant of the online activity tracking
entities. A user must use privacy protection tools on a veritable base. Besides, they
should render a thought to the consequences of information sharing when they are
online.
Fig. 2 Privacy protection

parties
6.2 Service Providers
Service providers are responsible for providing a panel where an end user can see
what information has been collected, the purpose of collection and if shared, the
entity names. A new mechanism must include grant, revoke and change their
consent to have control over their information before availing a service. A new
user-centric business model must be established to provide services and hide
information from additional and unnecessary entities. Likewise, a new platform
may be provided where two partied can strongly negate on information interchange
to have a win-win position.
6.3 Government and Governing Bodies
The government and governing bodies must play a leading role and enforce stan-
dardization and legitimate efforts. Data storage and transfer standardization must
include secure communication, fine-grained access control, selected attribute-based
authentication and authorization, an agreeable user consent mechanism to reduce
barriers for illegal entries. Also, government bodies establish a common access
point where an end user can file complaints of privacy breaches.
6.4 Data Brokers/Advertisers/Third Party App Developers
To gain explicit user consent, application developers are responsible for asking the
required information from end users in a clean way. The list of services/features
must be presented along with the authorization they wish to activate each service.
Also, they must be able to convey the data collection methods, its purpose, and list
the entities to which the data is shared.
7 Conclusion
Aggregation and analysis of user data through big data techniques are a gift to both
service providers and private users. Nonetheless, present techniques are not satis-
factory enough to support fine-grained access control and privacy-guaranteed data
management. In today’s information society, there is a compelling need to empower
the user and enhance user awareness about privacy by building up and deploying
novel privacy mechanism based on digital technologies. In this paper, we provided
guidelines that will help to implement privacy awareness mechanisms to enhance
the end users’ understanding and control of their personal data during online
transactions by the use of visualization tools. Also, innovative research method

should enable service providers and app developers to engage in explaining their
requests for, treatment and security of end users’ personal data, ultimately
increasing the general acceptability of the user–provider liaison. Introducing the
possibility of assigning a value for users to pay and service providers to charge for
privacy at desired levels will benefit both the parties.
We also identified a need to redefine the relationship between users’ personal
data and service providers’ business models on data analytics by focusing on
economic values of end users’ data. This relation should be ethically corrected,
comply with regulations and profitable to all. The responsibilities of privacy pro-
tection parties must lead to a solution that must consider user’s awareness about
privacy, empowering users, effective access control and authentication mechanisms
to the users’ personal data, understanding of the terms of services and build a
comprehensive trust.
In the future work, the authors will do an analysis of privacy policy for different
service providers and develop an improved privacy awareness tool to assist users in
managing their personal information and trust toward service providers.
References
1. M. Kuneva, “European Commission - PRESS RELEASES - Press release - Meglena Kuneva

European Consumer Commissioner KeynoteSpeech Roundtable on Online Data Collection,
Targeting and Profiling Brussels, 31 March 2009.” [Online]. Available: http://europa.eu/rapid/
press-release_SPEECH-09-156_en.htm. [Accessed: 17-Dec-2016].
2. “P3P: The Platform for Privacy Preferences.” [Online]. Available: http://www.w3.org/P3P/.
[Accessed: 17-Oct-2016].
3. A. Cavokian and D. Reed, “Big Privacy:Bridging Big Data andthe Personal Data Ecosystem
Through Privacy by Design,” 2013. [Online]. Available: https://www.ipc.on.ca/images/
Resources/pbd-big_privacy.pdf. [Accessed: 28-Jan-2017].
4. Kim Cameron, “The Laws of Identity -.” [Online]. Available: http://www.identityblog.com/
stories/2005/05/13/TheLawsOfIdentity.pdf. [Accessed: 21-Sept-2016].
5. “Sharing Information: A Day in Your Life| Consumer Information.” [Online]. Available:
http://www.consumer.ftc.gov/media/video-0022-sharing-information-day-your-life.
[Accessed: 18-Dec-2016].
6. “The Data Brokers: Selling your personal information.” [Online]. Available: http://www.
cbsnews.com/news/data-brokers-selling-personal-information-60-minutes/.
7. “Getting to know you| The Economist.” [Online]. Available: http://www.economist.com/
news/special-report/21615871-everything-people-do-online-avidly-followed-advertisers-and-
third-party.
8. “Deliverables — ENISA.” [Online]. Available: https://www.enisa.europa.eu/activities/risk-
management/emerging-and-future-risk/deliverables. [Accessed: 12-Jun-2015].
9. “OpenID Connect | OpenID.” [Online]. Available: http://openid.net/connect/. [Accessed:
29-Dec-2016].
10. “Home - WG - User Managed Access - Kantara Initiative.” [Online]. Available: https://
kantarainitiative.org/confluence/display/uma/Home. [Accessed: 29-Jan-2017].
11. A. Acquisti, “The economics of Personal Data and the Economics of Privacy,” Econ. Pers.
Data Priv. 30 years after OECD Priv. Guidel.
12. H. Olesen and S. Khajuria, “Accessing and Disclosing Protected Resources: A User-Centric
View,” in 2015 IEEE 81st Vehicular Technology Conference (VTC Spring), 2015, pp. 1–5.
13. “Sharing Information: A Day in Your Life, Fed. Trade Comm’n,.” [Online]. Available: http://
www.consumer.ftc.gov/media/video-0022-sharing-information-day-your-life. [Accessed: 19-
May-2016].
14. “Acxiom, Annual Report 2013.” [Online]. Available: http://d3u9yejw7h244g.cloudfront.net/
wp-content/uploads/2013/09/2013-Annual-Report.pdf.
15. “Indian Database.” [Online]. Available: http://www.zifzi.com/index.php?route=product/
category&path=20. [Accessed: 18-May-2016].
16. Futuresight, “Futuresight:, ‘User perspectives on mobile privacy, Summary of research
findings.’” [Online]. Available: http://www.gsma.com/publicpolicy/wp-content/uploads/
2012/03/futuresightuserperspectivesonuserprivacy.pdf.
17. “Why Privacy Is So Hard to Get | MIT Technology Review.” [Online]. Available: http://
www.technologyreview.com/view/526421/self-defense/. [Accessed: 12-Jun-2016].
18. “Commission proposes a comprehensive reform of the data protection rules.” [Online].
Available: http://ec.europa.eu/justice/newsroom/data-protection/news/120125_en.htm. [Accessed:
12-Nov-2016].
19. “Ministry of Law, Justice and Company Affairs (Legislative Department) New Delhi, the 11,”
pp. 1–13, 1994.
20. “Frameworks | Data Security Council of India.” [Online]. Available: https://www.dsci.in/dsci-
framework. [Accessed: 02-May-2016].
21. “MyPermissions - About.” [Online]. Available: http://mypermissions.com/whoweare.
[Accessed: 17-Aug-2016].
22. “OAuth - The Big Picture by Greg Brail & Sam Ramji.” [Online]. Available: https://itunes.
apple.com/us/book/oauth-the-big-picture/id515116202?mt=11. [Accessed: 29-Apr-2016].
23. “CLIP- Center on Law and Information Policy,” 2014. [Online]. Available: www.fordham.
edu/download/…/id/…/privacy_enforcement_actions.pdf.
24. P. S. Dhotre and H. Olesen, “A Survey of Privacy Awareness and Current Online Practices of
Indian Users,” WWRF34, St. Clara, California, US, Apr. 2015.
25. X. Liang, K. Zhang, X. Shen, and X. Lin, “Security and privacy in mobile social networks:
challenges and solutions,” IEEE Wirel. Commun., vol. 21, no. 1, pp. 33–41, Feb. 2014.
26. J. R. Reidenberg, T. Breaux, L. F. Cranor, B. French, A. Grannis, J. T. Graves, F. Liu,
A. M. McDonald, T. B. Norton, R. Ramanath, N. C. Russell, N. Sadeh, and F. Schaub,
“Disagreeable Privacy Policies: Mismatches between Meaning and Users’ Understanding,”
Aug. 2014.
Evolution Metrics for a BPEL Process
N. Parimala and Rachna Kohar
Abstract A Business Process Execution Language (BPEL) process in

Service-Oriented Architecture (SOA) evolves over time. The study of evolution
helps in analyzing the development and enhances maintenance of a BPEL process.
In this paper, we study the evolution using metrics which measure the changes and
the quantity in which the changes have occurred. A process could evolve on its own
or because of the evolution in the partner services. In both the cases, changes could
be to the internal logic of the process or involve changes in the interaction with the
partner services. The evolution metrics proposed in this paper are—BPEL Internal
Evolution Metric (BEMI) and BPEL External Evolution Metric (BEME). The time
complexity for computing the metrics is linear. The metrics are theoretically vali-
dated using Zuse framework and are found to be above the ordinal scale. In our
previous work, metrics were proposed for a single service under evolution. The
cohesiveness, between the changes of an evolving service and an evolving BPEL
process which uses this service, is demonstrated using metrics.
Keywords BPEL process Service Evolution Metrics
Electronic supplementary material The online version of this chapter (doi:10.1007/978-981-

10-7245-1_31) contains supplementary material, which is available to authorized users.
N. Parimala (&) R. Kohar

School of Computer and Systems Sciences, Jawaharlal Nehru University, New Delhi, India
e-mail: dr.parimala.n@gmail.com
R. Kohar
e-mail: rachnajnu20@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_31
306 N. Parimala and R. Kohar
1 Introduction
In SOA, different services are composed together to realize a business process. It is

used to achieve specific consumer business requirements. BPEL is considered as a
standard language to represent these processes. In the current fast growing business
market, a BPEL process evolves quite frequently. This evolution is driven by
different factors. One of the factors is the changing demands of today’s market.
Other factor is improving the process in a way so that its quality or performance
may be increased.
Evolution in software has been studied to understand the nature and amount of
changes in its source code [1, 2]. Similarly, in SOA, the evolution of a BPEL
process can be studied to analyze the nature of evolution. In this paper, we propose
to study the evolution along two axes. These are “what” and “by how much” in its
code. Here, “what” refers to nature of changes made in BPEL process and “by how
much” refers to the amount of these changes. The answers to these questions help to
have an idea of changes in the offered functionalities of BPEL process under
evolution.
The changes that can occur in a BPEL process could be addition, deletion,
modification, split and merge [3–6] in its different activities. These changes may be
concerning the interactions with external services or with the internal logic.
Therefore, there are two categories of changes in a BPEL process—Internal and
External changes.
As an example, consider a Travel Booking BPEL process, TB version1 which
coordinates with the consumer, Employee and the Airline web service and provides
the functionality of booking the flight for an Employee. Activities in this version are
‘receive TB request’, ‘invoke Employee service’ (to retrieve Employee Travel
details), ‘receive Employee service’ (to receive travel status), ‘invoke Airline
service (for travel booking), ‘receive Airline service’ (receive booking details),
‘invoke consumer’ (to reply travel booking confirmation)’. Let us say that the
Airline service has added a Privilege functionality which provides the Employee an
opportunity to book a Hotel, rent a Car or subscribe for a Magazine along with the
discount given by the Airline. The BPEL Process accommodates the newly offered
functionality by the Airline service for the Employee. Therefore, in TB version2,
‘invoke Airline service’ (for availing Privileges) activity and ‘invoke consumer’
(for returning Privilege confirmation details) activity are added. These changes
involve interactions with the external partner services. Let us say that a new version
TB version3 is made in which a wait activity is added in the process to wait for
some duration to perform functionalities of TB version2. This is an internal change
in the process. Clearly, there are two types of BPEL evolution, one is external in
nature and other is internal in nature.
The changes need to be measured to give us a measure of evolution of BPEL
process. Metrics provide a quantitative measure of quality of software [7, 8]. They
are computed by considering its successive releases [9]. The evolution of a BPEL
process is measured by taking into account its different versions. When a BPEL
Evolution Metrics for a BPEL Process 307
process evolves, the evolution can be measured form the consumer as well as the
producer point of view. In [10], we have proposed metrics from the consumer point
of view. In this paper, the metrics show the evolution for the producer of the BPEL
process. As brought out above, two categories of changes—Internal and External
changes are proposed. Metrics for these categories, namely, BPEL Internal
Evolution Metric (BEMI) and BPEL External Evolution Metric (BEME) are
proposed.
Changes in a service may demand changes in a BPEL process. We measure
changes using metrics. The cohesiveness of these changes is shown using metrics
proposed for a service [11] and those proposed in this paper for a BPEL process.
The paper is organized as follows. In Sect. 2, related work is discussed. BPEL
process is described in Sect. 3. Evolution Metrics are defined in Sect. 4. In Sect. 5,
the formal validation of metrics using Zuse [12] framework is given. In Sect. 6,
metrics computation algorithm is presented. In Sect. 7, the cohesiveness of changes
between a BPEL process and its partner service is demonstrated. Finally, the paper
is concluded in Sect. 8.
2 Related Work
Metrics are a standard of measurement of a process or a product [7, 8]. In SOA,

metrics have been proposed for a composite service to measure its performance,
reusability, granularity, coupling etc. [13–18].
In [13], a set of metrics is proposed for the composite service providers taking
into account quality features such as availability in terms of time, reliability in terms
of process requests and performance to measure throughput, discoverability etc.
Metrics are presented in [14] for a composite service to measure its resource quality
in terms of performance behavior, utilization of network and management of its
versions. The metrics proposed in [15] measures its performance and evaluate its
efficiency and effectiveness based on runtime data. Metrics proposed in [16] con-
siders Service level agreement between consumer and provider of the composite
service to measure how much performance, reliability and availability are actually
met. Four metrics are defined in [17] to quantitatively measure the granularity
appropriateness of a BPEL process considering attributes such as business value,
reusability, context independency etc. The work in [18] measures decoupling
considering black-box parameters of service stateness, service interface required,
service interface provided invocation modes, self-containment, implicit invocation
and binding modes.
Whereas the above are concerned with measuring QoS attributes, the work in
this paper measures structural changes of a BPEL process when it evolves. To the
best of our knowledge, metrics have not been proposed for measuring evolution of a
process.
3 BPEL Process
In SOA, different services may be composed together through orchestration using

BPEL. This process defines the activities and order of execution of these activities
to attain a specific business objective. In this paper, BPEL 2.0 standard is used to
categorize changes as well as to compute metrics.
4 Evolution Metrics
The evolution of a BPEL process is analyzed along internal and external changes.
These changes are defined below.
1. Internal Changes: Changes in BPEL activities such as If, while, assign etc. may
occur. These changes are internal to the BPEL process itself i.e. they are not
concerned with the interaction with the services. These changes are categorized
as internal changes. For example, addition of a wait activity is an internal
change.
2. External Changes: A BPEL process uses services to accomplish the required
business functionalities. It interacts with the services via invoke, receive, reply
activities. Any change in this interaction is classified as external change. For
example, addition of an invoke activity is an external change.
Table 1 provides detailed list of changes in the activities under both categories.
Now specification of metrics for measuring the evolution of a BPEL process is
considered. The metrics are defined for both the categories of changes. When a
process evolves, its new version is created. Metrics are computed for changes
across different versions of a BPEL process. Now, to define the metrics, let there be
two different versions of a BPEL process i.e. x and x + 1. Let, the table, which
contains all the evolution data between these changes for an activity i stored in
Tx,x+1. As an example, number of modification, deletion may occur in a flow
activity. Let there be 3 wait activities added in Tx,x+1. This will make value for Ci as
3. When the context of the changes is unambiguous Tx,x+1 is not mentioned while
defining the metrics.
Table 1 Category of changes

Category Type of change: Activities in the BPEL process
External Add/Delete/Modify/Split/Merge: invoke, receive, reply
Internal Add/Delete/Modify/Split/Merge: rethrow, wait, if, pick, forEach, while,
repeatUntil, flow, assign, throw, exit, sequence, empty
1. Internal Evolution Metric (BEMI):
P
n
w i ci
i
BEMI ¼ ð1Þ
n
where n is the count of types of changes which occur in a process under the
category “Internal”, Ci denotes the total number of changes for ith type of
change within this category and wi is the weight of ith type of change. The
weight is computed as a proportion of the number of changes for each type of
change within the category to the total number of changes in the category. For
example, if in a BPEL process, two wait activities are added (C1) and one wait
activity is deleted (C2) then count of types of changes is 2. C1 is 2 and w1 is 2/3;
C2 is 1 and w2 is 1/3. So, BEMI = (2/3) * 2 + (1/3) * 1 = 1.67
2. External Evolution Metric (BEME):
P
m
wj cj
j
BEME ¼ ð2Þ
m
where m is the count of types of changes which occur in a process under the
category “External”, Cj denotes the total number of changes for jth type of
change within this category and wj is the weight of jth type of change. For
example, if invoke activities are added and modified then there are two types of
changes i.e. n = 2. If 3 invoke activities are added and 2 are modified then C1 is
3 & w1 is 3/5 and C2 is 2 & w2 is 2/5. So, BEME = (3/5) * 3 + (2/5) * 2 = 2.2
5 Metrics Formal Validation
Theoretical validation of the proposed metrics is presented in this section using a

software measurement i.e. Zuse framework [12]. Axiomatic approach is used in this
framework to determine the scale of a metric. Measurement scale of a metric is used
to analyze the metric values as well as its empirical properties. The four scales are
nominal, ordinal, ratio and absolute [12]. Zuse framework is summarized in Table 2.
5.1 BEMI Metric Formal Validation
Let, there be a process P having n versions. Between any two process versions,
evolution data is computed and stored in a file. This file is referred to as a change
Table 2 Summarized Zuse Modified extensive structure

framework [12]
Axiom1: (A, >=) (weak order)
Axiom2: A1 o A2 > = A1 (positivity)
Axiom3: A1 o (A2 o A3) (A1 o A2) o A3 (weak
associativity)
Axiom4: A1 o A2 A2 o A1 (weak commutativity)
Axiom5: A1 > = A2 ) A1 o A > = A2 o A (weak
monotonicity)
Axiom6: If A3 > A4 then for any A1, A2, then there exists a
natural number n, such that A1 o nA3 > A2 o nA4
(Archimedean axiom)
Independence conditions
C1: A1 A2 ) A1 o A A2 o A and
A1 A2 ) A o A1 A o A2
C2: A1 A2 , A1 o A A2 o A and A1 A2
,A o A1 A o A2
C3: A1 > = A2 ) A1 o A > = A2 o A, and
A1 > = A2 ) A o A1 > = A o A2
C4: A1 > = A2 , A1 o A > = A2 o A, and A1 > = A2 ,
A o A1 > = A o A2
Modified relation of belief
MRB1: 8 A, B e Ӟ: A > = B or B > = A (completeness)
MRB2: 8 A, B, C e Ӟ: A > = B and B > = C ) A > = C
(transitivity)
MRB3: 8 A B ) A = < B (dominance axiom)
MRB4: 8 (A B,
A \ C = ∅) ) (A > = B ) A [ C > B [ C)
MRB5: 8 A e Ӟ Á: A > = 0 (positivity)
file. Let, there be two process versions i.e. x and x + 1. Let, their change file is
denoted by Fx,x+1. Let F be the set of all change files.
Consider BEMI. The measure BEMI is a mapping: BEMI: F ! R such that
the following holds for all change files Fx,x+1, Fx,x+1 e F: Fx,x+1 • >= Fx,x+1,
BEMI(Fx,x+1) >= BEMI(Fx,x+1).
Concatenation operation is used to combine the values of two metrics.
A combination rule is used to determine the behavior of metric. This behavior is
required to validate the metric for the different axioms. In the proposed metrics,
concatenation operation is denoted as follows for the combination rule.
BEMI(Fx,x+1 o Fy,y+1) = BEMI(Fx,x+1 [ Fy,y+1)
where Fx,x+1 [ Fy,y+1 is a file which contains changes (distinct) in Fx,x+1 and
Fy,y+1. This means that a common change in two different files appears once in their
concatenated file.
5.2 BEMI and the Modified Extensive Structure
Axiom 1: The binary relation • >= is known to be of weak order when it is tran-
sitive and complete. Let F1,2, F3,4 and F5,6 be the three change files where F1,2, F3,4,
F5,6 e F. It must be true that either BEMI(F1,2) >= BEMI(F3,4) or
BEMI(F3,4) >= BEMI(F1,2). Therefore, completeness property is fulfilled.
Next, is the transitivity property. If BEMI(F1,2) >= BEMI(F3,4) and
BEMI(F3,4) >= BEMI(F5,6) then it is obvious that BEMI(F1,2) >= BEMI(F5,6).
Thus, transitive property is also accomplished. Therefore, BEMI fulfills Axiom 1.
Axiom 2: BEMI(F1,2 o F3,4) >= BEMI(F1,2). It can be seen that when two files are
combined then the value of the metric BEMI is larger than the value of the metric
for each of those files. This proves Axiom 2 for BEMI.
Axiom 3: When weak associativity rule is applied to metric BEMI, the formulation
of the rule becomes, BEMI(T1,2 o (T3,4 o T5,6)) = BEMI((T1,2 o T3,4) o T5,6).
In this paper, the concatenation operation for the metric is Union operation.
It is known that the union operation is associative, therefore,
BEMI(T1,2 [ (T3,4 [ T5,6)) = BEMI((T1,2 [ T3,4) [ T5,6). Axiom 3 is satisfied.
Axiom 4: The weak commutative axiom for the metric BEMI is stated as
BEMI(F1,2 o F3,4) = BEMI((F3,4 o F1,2). It is known that the union operation is
commutative. Hence, BEMI fulfills Axiom 4.
Axiom 5: The property of weak monotonicity is stated as
BEMI(F1,2) >= BEMI(F3,4) ) BEMI(F1,2 o F5,6) >= BEMI(F3,4 o F5,6). To prove
BEMI(F1,2 [ F5,6) >= BEMI(F3,4 [ F5,6) (given BEMI(F1,2) >= BEMI(F3,4)), let
the count of common changes between F1,2 and F5,6 be more than the count of
common changes between F3,4 and F5,6. Since these common changes appear once
after applying concatenated operation, then the resultant metric computed based on
their concatenated files be BEMI(F3,4 [ F5,6) >= BEMI(F1,2 [ F5,6). Therefore,
BEMI does not fulfill this axiom.
Axiom 6: Idempotent property is considered here to prove this axiom. A metric
is idempotent going by definition of concatenation operation i.e.
BEMI(F1,2 o F1,2) = BEMI(F1,2). Therefore, BEMI does not fulfill this axiom.
It is concluded that modified extensive structure is not fulfilled by BEMI.
5.3 BEMI and the Independence Conditions
C1: To prove this condition, it has to be shown that BEMI(F1,2 o

F5,6) = BEMI(F3,4 o F5,6) and BEMI(F5,6 o F1,2) = BEMI(F5,6 o F3,4) given
BEMI(F1,2) = BEMI(F3,4). Now, BEMI(F1,2 [ F5,6) may or may not be equal to
BEMI(F3,4 [ F5,6) because the changes (which are common) in F1,2 [ F5,6 and
F3,4 [ F5,6 may not be the same. This makes BEMI(F1,2 o F5,6) 6¼ BEMI(F3,4 o
F5,6). Similarly, BEMI(F5,6 o F1,2) 6¼ BEMI(F5,6 o F3,4). Therefore, BEMI does
not fulfill this condition.
C2: When a metric does not accomplish C1, it will also not fulfill C2. The metric
BEMI does not fulfill C1 and therefore does not fulfill C2.
C3: When a metric does not accomplish fifth axiom of the modified extensive
structure, it will also not fulfill this condition which is the case with BEMI.
C4: A metric not satisfying the condition C3 cannot accomplish the condition C4.
Hence, BEMI does not accomplish C4.
It can be concluded that BEMI does not fulfill the independence conditions.
5.4 BEMI and the Modified Relation of Belief
MRB1: When Axiom 1 of modified extensive structure is fulfilled by a metric, then

it also satisfies MRB1. BEMI fulfills Axiom 1 of modified extensive structure
(proved above) and therefore, BEMI satisfies MRB1.
MRB2: If Axiom 1 of modified extensive structure is satisfied by a metric then that
metric satisfies MRB2. BEMI fulfills Axiom 1 of modified extensive structure and
therefore, it satisfies MRB2.
MRB3: In order to prove MRB3, let all the changes of the change file F3,4 are
included in F1,2, then it is obvious that BEMI(F1,2) >= BEMI(F3,4). Therefore, this
axiom is satisfied.
MRB4: In order to prove MRB4, let all the changes of the file F3,4 are included
in F1,2 and F1,2 \ F5,6 = ∅. Then it needs to be proved that
BEMI(T3,4) >= BEMI(T1,2) ) BEMI(T3,4 [ T5,6) >= BEMI(T1,2 [ T5,6). Due to
the fact that BEMI(F3,4) >= BEMI(F1,2) and that there are no common changes
between F3,4 and F5,6, the value of BEMI(F3,4 [ F5,6) will be more than
BEMI(F1,2 [ F5,6). This proves that the metric BEMI satisfies MRB4.
MRB5: This axiom is satisfied because changes in a process cannot be less than 0.
Therefore, modified relation of belief is fulfilled by BEMI. Thus, BEMI is a
measure above the level of the ordinal scale.
BEME has also been validated using Zuse framework. Both metrics fulfill
Axiom 1 to Axiom 4 but do not fulfill Axiom 4 and Axiom 5 of the modified
extensive structure. They do not fulfill any of the independence conditions C1 to C5
but fulfill all the axioms MRB 1 to MRB5. Therefore, as per [19], both metrics are
above ordinal scale.
6 Metrics Computation Algorithm
All the evolution data between any two versions of a process is stored in database
table. It lists the activities, number of changes and change category. Whenever a
process evolves, a new table is created and the evolution data is recorded. The
metrics are computed using Algorithm 1.
Algorithm 1: Input: Table from the Database Output: Metric Values
Read each row from the table
Compute BEMI and BEME using each row’s data
Time complexity of metric’s computation is O(n) when number of rows are n.
7 Cohesiveness of Changes in a BPEL Process Vis-à-Vis

Changes in a Service
A web service is invoked using client code. When a service undergoes changes, its
corresponding client code may also undergo changes. In [11], changes in a web
service have been classified into three categories as explained below.
1. Mandatory changes: The client has to include these changes in the code. E.g.,
corresponding invocation for the deleted operation needs to be removed from
client.
2. Optional changes: The client may opt to include them in the code. An example
may be addition of service operation.
3. Trivial changes: Changes which are immaterial to the client are known as trivial
changes. E.g., addition of documentation which has no effect on client code.
This is because while writing client code, this is not used.
Corresponding to these categories, SCEMM (Mandatory changes), SCEMO
(Optional changes) and SCEMT (Trivial changes) metrics have been proposed in [11].
A BPEL Process is the consumer of web services. When a service changes, the
BPEL process may have to accommodate the corresponding changes—depending
upon the type of changes. The metrics proposed in this paper and in [11] are shown
to be cohesive. For example, when service client code metrics reflect the mandatory
changes [11] in service, then BEMI and BEME metrics of a BPEL process must
exhibit a value indicating that a change has occurred for the successful execution of
the BPEL process. This cohesiveness is demonstrated with the help of an example.
The example of Booking BPEL Process from Oracle Technology Networks
http://www.oracle.com/technetwork/articles/matjaz-bpel1-090575.html is taken. It
has two partner services: Airline and Employee services. Employee service is used
to give travel status of employee to the process and then based on this status Airline
service returns airline booking details to process. Service and BPEL process code
Table 3 Description of the changes in the service and BPEL process

Versions Changes in the Service Changes in the BPEL BPEL version
airline service version process
1 Addition: Travel Addition of activities
Airline WSDL Version 1.wsdl
update & cancel for travel update & Travel BPEL Version 1.bpel
and refund cancel and refund

functionality functionality
2 Deletion: Travel Deletion of activities
update & cancel for travel update Travel BPEL Version 2.bpel
and refund functionality

functionality Deletion of activities
for cancel and refund
functionality
3 Addition: Client Addition of activities
Travel BPEL Version 3.bpel
privilege for client privilege
functionality functionality
4 Addition of flight No change
Airline WSDL Version 4.wsdl Travel BPEL Version 4.bpel
schedule
functionality
5 Addition of No change
documentation Travel BPEL Version 5.bpel
Table 4 Metrics for the airline service and the travel booking process
S. No. Service table Service metrics BPEL process table BPEL process metrics
1 SV1,2 SCEMM = 2.44 BV1,2 BEMI = 3.00
SCEMO = 0.00 BEME = 3.33
SCEMT = 0.00
2 SV1,3 SCEMM = 0.00 BV1,3 BEMI = 1.80
SCEMO = 3.31 BEME = 0.80
SCEMT = 0.00
3 SV1,4 SCEMM = 0.00 BV1,4 BEMI = 0.00
SCEMO = 2.44 BEME = 0.00
SCEMT = 0.00
4 SV1,5 SCEMM = 0.00 BV1,5 BEMI = 0.00
SCEMO = 0.00 BEME = 0.00
SCEMT = 1.86
taken from the reference cited above are modified. The modified versions are shown
in Table 3.
Service version Vi (i 2) is compared with service version V1 and the changes
are stored in the table SV1,i in the database. The second column of Table 4 lists
these tables. Similarly, tables for BPEL process are listed in the fourth column of
Table 4.
Metrics as given in [11] and the metrics proposed in this paper are shown in
Table 4. Next, we analyze the metrics.
(1) Mandatory changes: SCEMM > 0 and BEMI and BEME have positive values
for SV1,2 and BV1,2. Therefore, there is a clear synchronization between
mandatory changes in service client code vis-à-vis changes in the process.
(2) Optional changes: Metrics for SV1,3 and SV1,4 show optional changes. The
business process may accommodate (as in BV1,3) or may not (as in BV1,4).
Again, clearly, the changes are synchronized.
(3) Trivial changes: The last row shows that the BPEL process is unaffected by the
changes in SV1,5.
8 Conclusion
A BPEL process may evolve over time. Evolution is analyzed in terms of what
types of changes have occurred and by how much. In order to understand the type
of changes, two categories of changes are proposed: Internal and External changes.
Subsequently to estimate the amount of changes, metrics are defined for each of
these categories. The corresponding metrics are Internal Evolution Metric (BEMI)
and External Evolution Metric (BEME). A high BEME could indicate that services
used by a BPEL process are not available anymore or the services themselves are
evolving. A low BEMI could indicate that the BPEL process doesn’t need any more
correction and is mature. Complexity of the proposed metrics is linear. Theoretical
validation of the metrics is done using Zuse framework. The proposed metrics are
found to be above the ordinal scale.
A BPEL process is composed of web services. When a web service evolves,
BPEL process may also evolve. Proposed metrics truly reflect the cohesiveness of
changes in a process vis-a-vis changes in services.
As future work, our plan is to propose evolution metrics for choreography (via
WS-CDL Process) of services when they evolve.
References
1. Mockus, A., Votta, L. G.: Identifying reasons for software changes using historic databases.
In: International Conference on Software Maintenance, IEEE, pp. 120–130 (2000)
2. Lehman, M. M., Ramil, J. F., Wernick, P. D., Perry, D. E., Turski, W. M.: Metrics and laws of
software evolution-the nineties view, In: Fourth International Software Metrics Symposium,
IEEE, pp. 20–32 (1997)
3. Fdhila, W., Rinderle-Ma, S., Reichert, M.: Change propagation in collaborative processes
scenarios, In: 8th International Conference on Collaborative Computing: Networking,
Applications and Worksharing (CollaborateCom), IEEE, pp. 452–461 (2012)
4. Dongsoo, K., Minsoo, K., Hoontae, K.: Dynamic business process management based on
process change patterns, In: Convergence Information Technology, International Conference
on, IEEE, pp. 1154–1161 (2007)
5. Fdhila, W., Baouab, A.., Dahman, K., Godart, C., Perrin, O., Charoy, F.: Change propagation
in decentralized composite web services. Collaborative Computing: Networking, Applications
and Worksharing, 7th International Conference on. IEEE, 508–511 (2011)
6. Slominski A.: Adapting BPEL to scientific workflows. In: Workflows for e-Science, Springer
London, pp. 208–226 (2007)
7. Fenton, N., James, B.: Software metrics: a rigorous and practical approach, CRC Press, (2014)
8. Curtis, B.: Measurement and experimentation in software engineering, Proceedings of the
IEEE 68, no. 9, 1144–1157 (1980)
9. Mens, T., Demeyer, S.: Future trends in software evolution metrics, In: 4th international
workshop on Principles of software evolution ACM: 83–86 (2001)
10. Parimala N., Rachna Kohar: A Quality Metric for a BPEL Process under Evolution. Eleventh
International Conference on Digital Information Management (ICDIM), pp. 197–202 (2016)
11. Kohar, R., Parimala N. A Metrics Framework for measuring Quality of a Web Service as it
Evolves. International Journal of System Assurance Engineering and Management, Springer,
doi:https://doi.org/10.1007/s13198-017-0591-y, pp. 1–15 (2017)
12. Zuse, H.: A framework of software measurement. Walter de Gruyter. (1998)
13. Choi, S. W., Her, J. S., Kim, S. D.: QoS metrics for evaluating services from the perspective
of service providers. e-Business Engineering, ICEBE IEEE International Conference on.
IEEE, (2007)
14. Rud, D., Schmietendorf, A., Dumke, R.: Resource metrics for service-oriented infrastructures.
In: Proc. SEMSOA, pp. 90–98 (2007)
15. Wetzstein, B., Strauch, S., Leymann, F.: Measuring performance metrics of WS-BPEL
service compositions, In: 5th IEEE International Conference on Networking and Services,
ICNS’09, IEEE, pp. 49–56 (2009)
16. Dyachuk, D., Deters, R.: Using SLA context to ensure evolution of service for composite
services. In: IEEE International Conference on Pervasive Services, IEEE, pp. 64–67 (2007)
17. Khoshkbarforoushha, A., Tabein, R., Jamshidi, P., Shams, F.: Towards a metrics suite for
measuring composite service granularity level appropriateness. In: 6th World Congress on
Services. IEEE, (2010)
18. Qian, K., Jigang, L., Frank, T.: Decoupling metrics for services composition. In: 5th IEEE/
ACIS International Conference on Computer and Information Science and 1st IEEE/ACIS
International Workshop on Component-Based Software Engineering, Software Architecture
and Reuse (ICIS-COMSAR’06), IEEE, pp. 44–47 (2006)
19. Calero, C., Piattini, M., Pascual, C., Serrano, M. A.: Towards Data Warehouse Quality
Metrics. In: International Workshop on Design and Management of Data Warehouses.
pp. 2.1–2.10 Switzerland (2001)
Development of Performance Testing Suite
Using Apache JMeter
Jidnyasa Agnihotri and Rashmi Phalnikar
Abstract Testing a product has become one of the most important tasks for any
organization (Be it small scale or large scale). Without testing the product, it is not
delivered to the customer. Testing is an ongoing activity from the beginning of a
product’s development. A performance testing suite shall be developed using
Apache JMeter for the purpose of testing a product. To perform performance testing
on client- and server-type softwares, a 100% pure Java application named Apache
JMeter is used. Apache JMeter is not a browser, it works at protocol level. Static
and dynamic resources performance testing can be done using JMeter. A high level
performance testing suite will be developed in capturing aspects of performance at
UI and System level. Developing the testing suite helps in saving the time and cost
of the organization. The discussion follows and describes benefits of performance
testing and the performance testing suite.
Keywords Silk performer Performance testing JMeter, etc.
1 Introduction
Testing of any application, product, web applications, etc., has become very
important to ensure that the product works efficiently as per the requirement of the
customer. Testing department has gained a lot of importance for this reason. Every
organization and company (large scale or small scale) does have testing department.
The fast-growing world is leading to the demands of updating and implementing
their software’s as soon as possible to meet the requirements of the market and also
J. Agnihotri (&) R. Phalnikar

Department of IT, MITCOE, Pune, India
e-mail: jigyasathakurdas@gmail.com
R. Phalnikar
e-mail: rashmiphalnikar@yahoo.co.in

https://doi.org/10.1007/978-981-10-7245-1_32
318 J. Agnihotri and R. Phalnikar
to stay at lead in the race. Software should run correctly even if hundreds of people
are trying to access it at one time. To stay away from bugs, software must undergo
software testing. SDLC has software testing as one of its phase. Testing in itself is
very vast and lengthy process. Testing ensures that the software will work properly
as per the demands of the customer. Testing leads to successful completion and
working of a product.
In this paper, performance testing suite shall be developed using Apache JMeter.
The tool shall be developed for a particular product, describing in detail the method
to build the testing suite for any product. Also, the importance of testing and the
testing suite for a particular product shall be discussed in detail.
Rest of the paper is divided into the following section. Section 2 presents
existing system. Section 3 Literature Survey Sect. 4 Architecture Sect. 5 Results
and Analysis and Sect. 6 Conclusion.
2 Existing System
Performance testing is a nonfunctional testing and is carried out to test the system
under load conditions. A user can do performance testing on a product using
various tools. Many tools are available online/in market for the purpose of per-
formance testing. A tool which is easy to handle and is available will always be
preferred by any organization.
Silk performer tool [1] helps creating practical load and tests all the environ-
ments by creating huge stress. Performance testing on internet applications can be
done using WebLOAD [2], a commercial tool for performance testing. A user can
create practical load on websites and applications on web using LoadComplete [3].
A demo of original transactions is created for load testing by Rational Performance
Tester [4]. Performance of your web application can be improvised using
NeoLOAD [5]. Apache JMeter [6] is a Java platform application and can also create
a functional test plan.
Silk performer [1] can create huge load and requires very less size, because silk
performer is a UI-based performance testing. WebLOAD [2] is the generated load
from cloud and the machine present in the environment. Virtual users are created by
LoadComplete [3] by first recording the actions performed by the user. These
virtual users are created from the cloud or the machines. Thousands of users can be
created and handled by HP LoadRunner [7]. It is a combination of more than 2–3
different machines. So, it is complex to handle. When user interaction is involved,
only then Rational Performance Tester [8] can be used. NeoLOAD [5] can be of use
only for testing web applications.
Apache JMeter is a tool widely used for performance testing. Apart from testing
the web applications, JMeter can also be used for testing the Servlets, JMS, LDAP,
Web Services, Java classes, FTP, Pearl Scripts, and JAVA objects.
Development of Performance Testing Suite Using Apache JMeter 319
3 Literature Survey
Software testing of a product is done to make sure that the software is of a good
quality. Software testing also makes sure that the software is reliable. Performance
testing is performed on any application especially web application to understand
how it behaves under heavy load. Cloud computing can be used in the field of
testing for the purpose of overcoming the space issues [9].
Web services require testing on large scale. Extensive Human Intervention is
required for calculating the performance of web service. Enhancing the components
and creating new components also needs the obvious human interference. In paper
[10], JMeter is used for web services testing which can be done without human
intervention and dependencies are reduced so, in return, the speed of execution is
increased. For measuring the reasons for the decline in the performance, in the
paper [11], JMeter-based distributed test structure and point to point test structure
are used to set up a computing environment. A fully automated testing service is
made available recently for the ease of the testers. This service is available com-
mercially. This service, known as Testing as a Service (TaaS), reduces the time and
increases the efficiency. The paper [12] proposes a framework for performance
testing. Problems generally faced by the testers are addressed quite neatly in this
framework. Additionally generating and executing the test cases is completely
automated. In this paper [12], JMeter is also one of the TaaS moreover it is easily
available (open source).
In [11], point to point test structure is developed, it helps in leading to the
conclusion that the main reason of decline of performance is memory loss. This loss
happens because of Tomcat server. JMeter is a very flexible tool and a tool with
large number of possible variations which helps in testing of data from various
different sectors [13].
Apache JMeter is a freely available feature rich tool and easy to set up and use.
Flexibility of JMeter makes it use more compatible (Fig. 1).
Apache JMeter executes its actions as shown in the flow above. Apache JMeter
initially generates load on the server under test. It then collects the responses from
these servers. The obtained responses are then analyzed and calculated. Report
generation is the last step in the workflow which takes place after the response data
is ready. Looking at this workflow and all the other factors form installing JMeter to
generating the reports, it is understood that JMeter is the best tool to test the load as
it is free, great load generation, and easy user interface.
4 Architecture
Performance testing or load testing is a process of evaluating the quality or capa-

bility of a product (Fig. 2).
Fig. 1 Workflow for JMeter (Source [14])
Fig. 2 Performance testing process (Source [15])
Figure above describes how the performance testing process is actually carried
out. The tests are identified again and again to ensure that the system under test is
giving appropriate output every time. After the test is complete, benchmark is noted
and the same procedure is carried out again to note all the possible results. Also, no
action should be left which does not undergo testing.
Doing performance testing is a tedious task if it has to be done in the traditional
way. Developing a testing suite will lead to time deduction. As in, the tester will
have the user interface ready, only the test should be started. So, from the point of
view of an organization, it is a necessity to have a performance testing suite. The
performance testing suite is developed in five modules. Each module is explained in
detail ahead.
Module 1: Learning the product
Input—Setup of the product
Output—Understanding the product
Procedure—In this module, the expert of the product will give the training.
Understanding the product will lead to the clear idea of what has to be done. The
detailed study of the product and all the flows in it are necessary to prepare the
performance testing suite. This module will lead to finalization of the product for
which the performance testing suite has to be developed.
Module 2: Understanding the Scenarios for Performance Testing
Input—Scripts if available from previous test and different scenarios for Testing
Output—Various scenarios studied
Procedure—In this module, the possible scenarios will be studied. The performance
suite developer will understand all the possible scenarios that are available and
necessary at the moment. The scenarios which make any difference to the product’s
operations should be considered for script preparation. All the flows must be
thoroughly studied by the developer to record the scripts correctly.
Module 3: JMeter
Input—The product setup and the JMeter Setup
Output—Scripts for the Performance Testing Suite
Procedure—In this module, the possible scenarios are taken into consideration and
those scenarios are recorded in JMeter. These scripts will be the important portion
of the performance testing suite. Plugins required for the testing are in the devel-
opment phase. The actual operations performed for the development of scripts
include the following steps.
(1) Initially, change the port number of the browser to be used.
(2) Using the JMeter GUI’s test script recorder start the recording.
(3) Stop the recorder of JMeter when the necessary navigation is completed.
(4) Change the recorded script and remove all the random generated values and
hard coded values.
(5) To check the results, add the appropriate listener.
(6) Now, the initial script should be checked.
(7) Enhancement of script wherever possible shall be done in this step.
(8) Start the performance test, evaluate the results, repeat these actions for the
application.
The above steps are performed for every script/action of the product. The script
with all the changes and modification which gives expected results is included in
the suite.
Module 4: Performance Testing

Input—Performance Testing Suite and the product setup
Output—Performance Test results
Procedure—At the end of this module, we will have performance results of how the
system is performing under load conditions. The performance runs are usually
carried out for 1 h, 4 h, 1 day, 4 days, 10 days, etc. The performance tests are
performed on the product according to the requirement. The load generation is
specified accordingly.
Module 5: Verifying the Performance Results
Input—Results of the Performance tests
Output—Check the achieved accuracy
Procedure—At the end of this module, the results achieved are verified for their
correctness. It is checked whether the system is behaving according to the expec-
tations. If yes how to enhance is the next step. If there are certain problems in the
system behavior, the changes to be done in the system are reported. For example, if
a page should load in 5 s and if it taking 55 s, the issue should be reported and the
necessary actions should be taken.
Figures 3 and 4 are the visual description of the steps that are to be carried out in
the development of performance testing suite. Various steps are combined together
to put up in two figures.
Figure 5 represents the graphical user interface of the performance testing suite
developed. The load can be managed changing the parameters which appear in the
thread group section. The graph represents the presence of threads on the particular
timings. The scripts are present on the left-hand side. The performance run must be
done through the non-gui mode for the run to be completed without causing any
failure or problem to the system. Nineteen scripts are developed so far that are
included in the suite.
Fig. 3 Test script generation

Fig. 4 Test script execution
Fig. 5 Performance testing Suite GUI
5 Results and Analysis
Results obtained by the JMeter performance run are very simplified and can be
viewed as a CSV file or statistics (in table format) or even as separate graphs for
every operation performed. JMeter provides results by writing the results to the
CSV file and to the folder which are provided in the command line of the non-gui
mode.
Fig. 6 Latencies over time
Fig. 7 Statistics
Figure 6 includes a graph which shows the latencies of the operations over time.
Different color indicates different operation. Every operations graph can be viewed
individually by disabling the other operations. The average time responses can be
viewed using the statistics table, using graphs an approximate estimate of the times
of the requests can be observed. The responses having incorrect reply can be
pointed out easily using the graphs.
Figure 7 shows the table which is produced as the output of the performance
testing. The statistics are very precise, accurate, and give detailed information about
the error, minimum time, maximum time, average time, throughput, etc. The graphs
and the statistics tables are equally important to monitor the system.
6 Conclusion
The performance testing is an important task to be done for every product or

application. Having convenient means for this testing provides various benefits to
the organization. Time and cost savings are the two most important benefits that can
be noted for having performance testing suite. Here, the tool used is also free so the
cost reduction is to a high level.
In this paper, we discuss the performance testing suite developed for SMGR
(System Manager). For the testing of this product, initially silk performer was used
which has high cost. The suite developed, gives accurate results when compared
with the results of the silk performer. It is noticeable that the results are accurate up
to 90%. The reports generated are also very detailed, precise, and give information
about every single action performed during the run. Apache JMeter performance
testing suite has a GUI which is more user friendly when compared to silk per-
former. The suite developed in here is specific to SMGR similarly it can be
developed for any application (static or dynamic).
Acknowledgements The work of researchers and publishers is meant a lot to the author during
the whole process. Author expresses regards for the same. Also the help by the professors proved
to be of great use. Author would also like to thank Avaya India Pvt. Ltd. for giving this oppor-
tunity of working for Avaya. The suite developed during this project is specific to SMGR (product
of Avaya India Pvt. Ltd.) as the project is sponsored by Avaya.
References
1. http://community.microfocus.com/borland/test/silk_performer__application_performance_testing/
w/silk_performer_trial/23695.tutorial-1-part-1-getting-started-with-silk-performer.aspx.
2. 2016, Sharma, M., Iyer, V. S., Subramanian, S., & Shetty, A. Comparison of Load Testing
Tools.
3. Vinayak Hegde, Pallavi. “Web Performance Testing: Methodologies, Tools and Challenges.”
4. 2006, Vol. 45(3), pp. 463–480, Brown AW, Iyengar S, Johnston S. A Rational approach to
model-driven development. IBM Systems Journal. 2006.
5. 2016, Vol. 6(5), Bhatia R, Ganpati A. In Depth Analysis of Web Performance Testing Tools.
2016 Sep.
6. http://jmeter.apache.org/usermanual/get-started.html.
7. 2013, pp. 429–434, Gao Q, Wang W, Wu G, Li X, Wei J, Zhong H. Migrating load testing to
the cloud: A case study. InService Oriented System Engineering (SOSE), 2013 IEEE 7th
International Symposium on 2013 Mar 25. IEEE.
8. 2003, Vol. 26, Issue 8, pp. 888–898, Apte V, Hansen T, Reeser P. Performance comparison of
dynamic web platforms. Computer Communications. 2003 May 20.
9. http://www.softwaretestinghelp.com/getting-started-withcloud-testing/.
10. 2014, pp. 314–318, Shenoy S, Bakar NA, Swamy R. An adaptive framework for web services
testing automation using JMeter. InService-Oriented Computing and Applications (SOCA),
2014 IEEE 7th International Conference on 2014 Nov 17. IEEE.
11. 2010, Vol. 5, pp. 282–285, Jing Y, Lan Z, Hongyuan W, Yuqiang S, Guizhen C.
JMeter-based aging simulation of computing system. In Computer, Mechatronics, Control and
Electronic Engineering (CMCE), 2010 International Conference on 2010 Aug 24. IEEE.
12. 2015, pp. 356–361, Ali A, Badr N. Performance testing as a service for web applications.
InIntelligent Computing and Information Systems (ICICIS), 2015 IEEE Seventh International
Conference on 2015 Dec 12. IEEE.
13. 2016, pp. 1–5, Harikrishna P, Amuthan A. A survey of testing as a service in cloud
computing. InComputer Communication and Informatics (ICCCI), 2016 International
Conference on 2016 Jan 7. IEEE.
14. http://www.testingjournals.com/jmeter-introduction/.
15. https://www.tutorialspoint.com/software_testing_dictionary/performance_testing.htm.
Characterizing Network Flows
for Detecting DNS, NTP, and SNMP
Anomalies
Rohini Sharma, Ajay Guleria and R. K. Singla
Abstract Network security can never be assured fully as new attacks are reported
every day. Characterizing such new attacks is a challenging task. For detecting
anomalies based on specific services, it is desirable to find characteristic features for
those service specific anomalies. In this paper, real-time flow-based network traffic
captured from a university campus is studied to find if the traditional volume-based
analysis of aggregated flows and service specific aggregated flows is useful in
detecting service specific anomalies or not. Two existing techniques are also
evaluated to find characteristic features of these anomalies. The service specific
anomalies: DNS, NTP, and SNMP are considered for study in this paper.
Keywords Network flows DNS tunnel DNS amplification reflection

NTP SNMP
1 Introduction
Network security is an area which never becomes old. Everyday forth, new tech-
niques and attacks are invented by attackers to breach the security of a network.
Tackling the security issue is of utmost importance as hardware and software
resources are kept on network for sharing. Intrusion detection is an important aspect
of network security where intrusions from unauthorized sources are avoided, pre-
vented, and detected. Intrusion Detection Systems (IDS) are of two types:
Network IDS and Host IDS. In network IDS, network data is taken by the system
and analyzed whereas in Host IDS, resources local to host like logs are taken for
R. Sharma (&) A. Guleria R. K. Singla

Department of Computer Science and Applications, Panjab University, Chandigarh, India
e-mail: rohini@pu.ac.in
A. Guleria
e-mail: ag@pu.ac.in
R. K. Singla
e-mail: rksingla@pu.ac.in

https://doi.org/10.1007/978-981-10-7245-1_33
328 R. Sharma et al.
analysis and detection. Host IDS are effective in securing host but its effect is not
much as other hosts on the network are not involved. Network IDS (NIDS) is
further divided into signature-based NIDS and anomaly-based NIDS. In
signature-based NIDS, signatures of known attacks are stored in a database which
are used to detect attacks in real time. It is a widely used NIDS but the main
problem is not being able to detect zero day attacks. In anomaly-based NIDS, the
normal behavior of the system is modeled and deviation from normal behavior is
termed as anomalous. Anomaly-based NIDS are able to detect zero day attacks. In
the current scenario, every day a new kind of attack is generated and a new service
is target of an attack in a new way, therefore, detecting zero day attacks has become
very important. Anomaly-based NIDS can use Deep Packet Inspection for detecting
anomalies in the network which is a reliable approach as large amount of detailed
data is available but at the same time it is quite time consuming. For detection in
real time, deep packet inspection is a costly and time consuming solution.
Using Network Flows for detecting anomalies is less costly than deep packet
inspection as these flows are based on packet headers only. A number of packets are
combined into a flow if they share five attributes: SourceIP, DestinationIP,
SourcePort, DestinationPort, and Protocol. It decreases the amount of data to be
analyzed and helps in speeding up the process of detection as the flow-based data is
less than 1% of the packet data [1]. With many advantages, flows have certain
limitations also. As only header information is available in flows, many kinds of
anomalies are not detectable using this. Therefore, it cannot be used as an alter-
native to Deep Packet Inspection [2]. However, it can be used as first step in
detection process followed by Deep Packet Inspection wherever required. Anomaly
Detection Mechanism can be implemented in supervised mode where labeled
dataset is there to train the system or unsupervised mode where labeled dataset is
not available. In the absence of labeled dataset, data can be categorized into certain
classes such that anomalous data gets confined to these class(es).
2 Problem Definition
Generally for flows based detection, the volume of traffic is taken into considera-
tion. The volume of flows, bytes, and packets per time bin is used in [3–8] for
detecting anomalies. Time bin is decided according to the volume of data available.
In [4], time bins of 1, 2, and 10 s are used. In [9], time bin of 1 min and in [1] time
bin of 10 min are used, respectively. As the size of time bin increases, the infor-
mation gets hidden in the aggregation of flows. If sum or average of flows, bytes,
and packets is taken, a large number of flow records of the order of thousands and
lakhs in a single time bin do not provide an accurate picture. Moreover, the
problems detected in networks are service specific and can be detected if a service is
monitored properly. The research goals of this paper are
Characterizing Network Flows for Detecting DNS … 329
• Analyzing time series data for flows, bytes, and packets for different time bin
sizes in order to study the effect of size of time bin over anomalous detections.
• Breaking down the flows according to services and analyze the effectiveness of
resultant time series data for detecting service specific anomalies.
• Generating time series for DNS service using different attributes in order to
study the effect of attribute selection and to characterize flows for effective
detection of service specific anomalies.
In this paper, focus is on amplification and tunneling anomalies. The specific
services used for analysis are Domain Name System (DNS), Network Time
Protocol (NTP), and Simple Network Management Protocol (SNMP) services.
3 Dataset Used
The real-time data is collected from the computer center of Panjab University,
Chandigarh which is handling the networking of the whole university.
3.1 Network Architecture
The network is divided into two parts: access network and server network.
Dedicated switches are there on both the networks such that the requests for servers
and the replies from servers are moved through server network and the other
outgoing data moves through access network. UTM is installed on the dedicated
switches to check any kind of attack inward or outward. To capture flow data of
access network, a machine is attached to the switch installed for outgoing traffic.
The port on the dedicated switch is mirrored and the whole data is captured on the
machine attached in parallel. The process of data collection is shown in Fig. 1.
Fig. 1 Data collection architecture

3.2 Data Collection
The machine installed on the network to capture flows is made ready to capture
packets on an interface using softflowd that converts the captured packets into
flows. These flows are stored in binary nfcapd files by NfDump Collector. The
interval for capturing flows in nfcapd files is 5 min by default which was kept
unchanged. The captured flows are extracted with and without aggregation and
analysis is performed on these flows according to the research goals.
4 Analysis of Flows
Flows are collected for one day, i.e., 24 h from 03-10-2016 1100 to 04-10-2016
1100 h. Non-aggregated flows amount to 453,476,798 and aggregated flows
amount to 40,031,076 which is roughly 9% of non-aggregated flows. Hence, if
aggregated flows are used for analysis, it will result in huge savings in terms of time
and space. However, non-aggregated flows give detailed data as compared to
aggregated flows for analysis. Analyzing such a large amount of flows is very much
time consuming and requires high power and high speed machine. The machine
used for collection and analysis uses i7-5500v CPU@2.4 GHz with 8 GB RAM
and with CentOS 7 installed over it.
4.1 Research Goal 1
For generating time series based on number of bytes, packets, and flows, the flows
are aggregated using—a option with NfDump that aggregates flows over five
parameters SourceIP, DestinationIP, SourcePort, DestinationPort and Protocol.
Flows, bytes, and packets are added while aggregating whereas an average is
calculated for bytes per packet, bits per second, and packets per second.
Aggregation is done over 3 h. In order to get the insight little earlier, we create time
series of the aggregated flow records over time bins. There is a tradeoff between the
size of the time bin, analysis time, and accuracy of the time series. As the size of
time bin increases, the analysis time decreases as there will be lesser number of
records to analyze but it may affect the accuracy of the time series. In this paper, it
has been studied whether size of time bin affects the accuracy of time series or not.
Sperotto [1] used time bin of 10 min for 2 days resulting in 288 records. In this
study, time bins of 1, 5, and 10 min are taken for 24 h resulting in 1440, 288, and
144 records, respectively. The graphs generated are shown in Fig. 2. It is quite
evident from the graphs that the shapes of the curves are quite similar in all the time
bins. A clear pattern which is visible in the curves representing sum of packets,
bytes, and flows is the day/night pattern. The volume of packets, bytes, and flows
decreases gradually around 2 am till 8 am. These curves also have some spikes
which show increase in volume of packets, bytes, and flows on that particular
instant of time. As shown in Fig. 2a, spikes show increase in number of packets
around 14:00, 17:00, 20:00, and 23:00. If we consider bits per second, packets per
second, and bytes per packet, no clear pattern is visible in figures with different time
bins but shapes of curves for all the three time bins are same. It has also been shown
in Fig. 2d that transfer speed was quite less from 23:00 to 02:00 in the night. The
spikes shown in the graphs are very less as compared to the service specific
anomalies in the data which may be due to the granularity of the data which is quite
low. As the spikes are not coinciding with the anomalies, volume-based analysis
cannot detect service specific anomalies. Hence following conclusions can be
drawn.
• Size of the time bin does not affect the shape of the curve and hence the
accuracy of the time series. For different time bins, same spikes have been
shown. Hence, larger time bins can be used to decrease response time. For
localization of anomalies, the anomalous interval can be detailed out further.
• Volume-based analysis of flows, packets, bytes, bits per second, packets per
second, and packet size cannot detect service specific anomalies.
• Aggregating flows over 3 h results in lesser number of records but do not give
clear indications of anomalies. Aggregating over lesser time comparatively may
give better results.
4.2 Research Goal 2
For analyzing the data services wise, three services are selected: Domain Name
System (DNS) that works on port 53, Network Time Protocol (NTP) that works on
port 123, and Simple Network Management Protocol (SNMP) that works on port
161. Most common anomalies for these services/protocols are amplification
reflection attacks and data tunneling which are quite common these days. In
amplification reflection attacks, an attacker, with spoofed SourceIP address, sends
request to legitimate servers and the servers send replies to the target victim as
shown in Fig. 3a. Servers involved in the attacks generally send large packets as
replies for a small packet of request. For example, a DNS request packets is
maximum 80 bytes in size whereas a DNS reply can be as large as 16 K bytes
which can be fragmented due to network limitations up to 1500 bytes. As cited by
United States Computer Emergency Readiness Team (US-CERT) [10], the
amplification factor for DNS protocol is 28–54. If attacker sends requests with
spoofed SourceIP address to number of servers and all the servers, behaving as
reflectors, send amplified replies to the victim, the victim will get bombarded with
data leading to Denial of Service at victim side. NTP and SNMP are other key
players because the amplification factor for these two protocols is also high as given
in Table 1.
Fig. 2 a Sum of packets, b Sum of bytes, c Sum of flows, d Average of bits per second, e Average
of packets per second and f Average of packet size for Time Bin of 1, 5, and 10 min
Fig. 3 a Reflection amplification attack, b DNS Tunneling attack
Table 1 Amplification factors

Protocol Amplification Protocol Amplification Protocol Amplification
factor factor factor
DNS 28–54 NTP 556.9 SNMPv2 6.3
In DNS Tunneling, data is hidden in the request and reply packets and trans-
mitted through the network without detection. DNS traffic is made to pass through
network boundaries as it is meant for name resolution and does not require any
checking. Because of this feature, DNS is most exploited for tunneling attacks.
Basic working of DNS Tunneling is shown in Fig. 3b.
For analyzing anomalies specific to DNS, again aggregated flows over 3 h have
been taken, filtered for destination port number 53,123, and 161 and total number of
packets, bytes, and flows are analyzed. The graphs generated for DNS are shown in
Fig. 4 for time bin of 1 min. Time bins of 5 and 10 min are not considered as that
will also produce same spikes as concluded in Sect. 4.1. As shown in Fig. 4a, b, c,
spikes are shown in each figure at different times. Many other spikes are also shown
in packets, bytes, and flows time series but are not much prominent. Average
packets per second metric is not effective in detecting anomalies as it is displaying a
very regular pattern without any spike as shown in Fig. 4e. Bytes per packet and
bits per second, however, are able to detect a few anomalies which are shown in the
form of spikes clearly distinct from rest of the curve in Fig. 4d, f but these spikes
are just 20% of the anomalies present in the data. Hence for DNS flows, bits/s and
bytes/packet based time series are also not much effective.
Figures 5 and 6 show the graphs of aggregated flows for destination port 123
and 161, respectively. In Fig. 5, the curves representing total packets, bytes, and
flows show spikes indicating anomalous behavior. Bytes per packet and bits/s
curves are showing very few spikes as compared to the anomalies present. It can be
concluded that leaving packet/s, all the other attributes are able to show some of the
anomalies using service specific volume based analysis. In Fig. 6, the anomalies are
clearly visible in the curves representing total flows, bytes, and packets. Bits per
Fig. 4 a Sum of Packets, b Sum of bytes, c Sum of flows, d Average of bits per second,
e Average of packets per second and f Average of bytes per packet for DNS aggregated flows
e Average of packets per second and f Average of bytes per packet for NTP aggregated flows
e Average of packets per second and f Average of bytes per packet for SNMP aggregated flows
second and packet per second show a single spike which is quite prominent but
many anomalies are left untraced. Hence, volume-based analysis using packets,
flows, and bytes can be useful in detecting some SNMP anomalies only.
4.3 Research Goal 3
It is found in Sects. 4.1 and 4.2 that service specific packets, bytes, flows, bits/
second, and packet size time series can detect very few DNS anomalies. So it is
decided to use specific techniques. In this section, two different techniques have
been employed specifically for amplification of reflection and tunneling attacks. For
detecting DNS amplification reflection attacks, the technique proposed by Huistra
[11], shown in Algorithm 1 has been used. For the study, non-aggregated flow
records have been taken which are collected in NfCapd files each of 5 min duration.
After applying the technique for DNS, it has been found that on an average 13%
source IPs are suspicious in each set of flow records of duration 5 min and the
duration during which the suspicious IP addresses are sending request spans
complete 5 min approximately. This might be the reason that these anomalies were
not much visible in the curves representing total packets, bytes, and flows.
It has been found that out of the listed sourceIPs, 10% of the source IPs are not
involved in anomalous activity. Hence, false positive rate is quite high in this
technique. For NTP anomalies, same technique is applied with minor modification.
As in case of NTP, generally every minute, a host sends requests to all NTP servers
for synchronization. So the technique is modified by grouping all the flows in 5 min
file by 1 min with same source IP to find out that one IP sends how many requests
to NTP servers in a minute. Maximum requests sent by an IP in a minute are 108.
As no means are available to find out threshold, top 10% of the sourceIPs can be
identified and monitored to find out if they are being used as bots. After applying
the technique over NTP flows, it has been found that some hosts are sending only 4
requests per minute while others are sending 108 requests to outside NTP Servers.
Top 10% of the sourceIPs are sending approximately 70 or more NTP requests to
outside NTP Servers which can be analyzed further using deep packet inspection.
Algorithm 1 Technique for detecting DNS amplification reflection attacks [11]
1. Input: Flow records.
2. Output: Suspicious IP addresses.
3. Filter Flow records with destination port = 53 and combine flow records with
same source
4. IPs
5. if Total requests generated by an IP > 1000 then
6. IP Address is suspicious.
7. else if Total requests generated by an IP > 100 then
8. Calculate Standard Deviation of request size in non-aggregated flows
where source ports are different.
9. if Standard Deviation < 1 then
10. IP Address is suspicious.
11. end if
end if
For detecting SNMP amplification reflection attacks, same technique has been
used with threshold set to 1 as in a University network, a single request is also
suspicious. If a host is sending an SNMP request outside the enterprise network and
the number of requests to a particular IP is large enough, it means that host is
compromised and has been directed for initiating SNMP amplification attack. It has
been found while analyzing data that all the hosts are involved in sending SNMP
requests to other hosts in outer enterprise networks specifically in US.
For Detecting DNS Tunneling attacks, Karasaridis et al. [12] have proposed
Tunneling Attack Detector (TUNAD) for detecting DNS Tunnels in near real time.
Working of TUNAD is shown in Algorithm 2. The technique has been applied over
the non-aggregated flow records. Relative entropy converges to Chi-Square dis-
tribution under null hypothesis [12], so threshold was taken as the table value of
Chi-Square with degree of freedom equal to number of time bins and false positive
rate taken as 10−3 as the level of significance. However, the results are very sen-
sitive to the threshold taken.
The graph in Fig. 7 shows the results for 6 h from 1100 to 1800 h on
03-10-2016 and 04-10-2016. It shows the presence of DNS Tunneling anomalies
but the results are quite sensitive to the selection of threshold. The UTM installed in
the university network also shows that anomalous requests are there in the network
in each hour. Hence, the results shown are somewhat in conformance with what the
existing security measures show.
Algorithm 2 Technique for detecting DNS Tunneling [12]
1. Input: Flow records
2. Output: Suspicious Flows
3. Combine all 5 min NfCapd files to form hourly flows.
4. For each request packet size, calculate request size frequency.
5. If Request packet size > 300 then
6. Mark flows suspicious.
7. else
Fig. 7 Relative entropy of

hourly request size histograms
8. Apply Cross Entropy on packet size histograms.

9. If Relative Entropy > Threshold then
10. Mark flow suspicious.
11. end if
12. end if
5 Conclusion and Future Work
From the above study, it is concluded that not all types of anomalies can be detected
using volume-based analysis for packets, bytes and flows. For detecting service
specific anomalies like amplification reflection attacks and tunneling of data,
specific attributes are required to be identified in the given network. Normal profile
of the network can be generated based on the identified attributes that can be
compared with real-time traffic characteristics to detect anomalies in real time.
Detecting service specific anomalies with 100% accuracy and localization of the
root cause are challenging tasks to accomplish.
References
1. Sperotto, A., Sadre, R., Pras, A.: Anomaly characterization in flow-based traffic time series.
In: International Workshop on IP Operations and Management, pp. 15–27. Springer,
Heidelberg. (2008).
2. Sperotto, A., Pras, A.: Flow-based intrusion detection. In: IFIP/IEEE International
Symposium on Integrated Network Management (IM), pp. 958–963. (2011).
3. Fernandes, G., Rodrigues, J. J., Proena, M. L.: Autonomous profile-based anomaly detection
system using principal component analysis and flow analysis. Applied Soft Computing, 34,
513–525. (2015).
4. Ellens, W., Żuraniewski, P., Sperotto, A., Schotanus, H., Mandjes, M., Meeuwissen, E.: Flow
based detection of DNS tunnels. In: IFIP International Conference on Autonomous
Infrastructure, Management and Security, pp. 124–135. Springer, Heidelberg. (2013).
5. Johnson, T., Lazos, L.: Network anomaly detection using autonomous system flow
aggregates. In: IEEE Global Communications Conference (GLOBECOM), pp. 544–550.
(2014).
6. Pena, E. H., Barbon, S., Rodrigues, J. J., Proena, M. L.: Anomaly detection using digital
signature of network segment with adaptive ARIMA model and Paraconsistent Logic. In:
IEEE Symposium on Computers and Communication (ISCC), pp. 1–6. (2014).
7. Carvalho, L. F., Rodrigues, J. J., Barbon, S., Proenca, M. L.: Using ant colony optimization
metaheuristic and dynamic time warping for anomaly detection. In: International Conference
on Software, Telecommunications and Computer Networks (SoftCOM), pp. 1–5. (2017).
8. Li, Y., Luo, X., Li, B.: Detecting network-wide traffic anomalies based on robust multivariate
probabilistic calibration model. In: IEEE Military Communications Conference, MILCOM,
pp. 1323–1328. (2015).
9. Hellemons, L., Hendriks, L., Hofstede, R., Sperotto, A., Sadre, R., Pras, A.: SSHCure: a
flowbased SSH intrusion detection system. In: IFIP International Conference on Autonomous
Infrastructure, Management and Security, pp. 86–97. Springer, Heidelberg. (2012).
10. US-CERT United States Computers Emergency Readiness Team, https://www.uscert.gov/
ncas/alerts/TA14-017A.
11. Huistra, D.: Detecting reflection attacks in DNS flows. In: 19th Twente Student Conference
on IT. (2013).
12. Karasaridis, A., Meier-Hellstern, K., Hoeflin, D.: Nis04-2: Detection of dns anomalies using
flow data analysis. In: IEEE Global Telecommunications Conference, GLOBECOM’06,
pp. 1–6. (2006).
Periocular Region Based Biometric
Identification Using the Local Descriptors
K. Kishore Kumar and P. Trinatha Rao
Abstract Biometric systems have become a vital part of our present day automated
systems. Every individual has its unique biometric features in terms of face, iris and
periocular regions. Identification/recognition of a person by using these biometric
features is significantly studied over the last decade to build robust systems. The
periocular region has become the powerful alternative for unconstrained biometrics
with better robustness and high discrimination ability. In the proposed paper,
various local descriptors are used for the feature extraction of discriminative fea-
tures from the regions of full face, periocular and city block distance is used as a
classifier. Local descriptors used in the present work are Local Binary Patterns
(LBP), Local Phase Quantization (LPQ) and Histogram of Oriented Gradients
(HOG) and Weber Local Descriptor (WLD). FRGC database is used for the
experimentation to compare the performance of both periocular and face biometric
modalities and it showed that the periocular region has a similar level of perfor-
mance of the face region using only 25% data of the complete face.

Keywords Periocular region Local feature extraction methods

Local appearance-based approaches Equal Error Rate Rank-1 recognition rates
1 Introduction
The traditional biometric systems [1], such as face and iris systems reaching the
state of maturity with almost having high performances of 100% accuracy. These
systems work very well under controlled and ideal circumstances like
high-resolution images taken from the short distance, well-illuminated conditions
K. Kishore Kumar (&)

Department of ECE, IcfaiTech School, IFHE University, Hyderabad, India
e-mail: kkishore@ifheindia.org
P. Trinatha Rao
Department of ECE, GITAM School of Technology, GITAM University, Hyderabad, India
e-mail: trinath@gitam.in

https://doi.org/10.1007/978-981-10-7245-1_34
342 K. Kishore Kumar and P. Trinatha Rao
and from a cooperative subject. The performance of these systems will degrade
when they operate in nonideal conditions, i.e. when the images are captured from
longer distances when the subject is on the move or from the non-cooperative
subject. Now, the research is concentrated on the design of robust biometric sys-
tems under nonideal condition with better accuracies. The periocular region has
become the best alternative for unconstrained biometrics with better robustness and
high discrimination ability.
The periocular region is the sub-portion of the face in the vicinity of an eye with
eyebrows, eyelids and eye folds as shown in Fig. 1.
Periocular region [2] is the most discriminating portion of the face, gaining
significance as a useful biometric modality under the nonideal conditions, supple-
menting to the traditional face and iris biometric systems. Periocular region based
recognition is still a comparatively new area which is cost-effective [3, 4] and does
not require any extra storage space. The periocular region is captured along with the
face and iris during the biometric acquisition process. For obtaining the better
recognition accuracies, the periocular region can be used as in fusion with face/iris
modalities. In the design of age-invariant face recognition systems [5, 6], the
periocular region is less affected by the ageing process than compared to the face.
Periocular biometrics uses small periocular region [7] templates which use only
25% of the full face image when compared with the large face models makes them
faster in the recognition process.
Sato et al. [8] presented the first partial face recognition using the subregions of
the face which showed that the eye region achieved better recognition rates com-
pared to others regions suggesting us that the periocular region could be a major
area of the face with better discriminative features. Savvides et al. [9] performed
partial face recognition on the Facial Recognition Grand Challenge (FRGC) data set
and the results showed the eye region achieved better recognition rates compared to
others. Teo et al. [10] compared the performance of full face recognition with the
eye-based partial face recognition. Park et al. [2] presented the periocular-based
recognition system using the local appearance-based approaches for the feature
extraction of the periocular region and the performance is compared to the face.
Lyle et al. and Merkow et al. presented the use of the periocular biometric for the
Fig. 1 Periocular region

Periocular Region Based Biometric Identification … 343
gender classification [7]. Santos and Hoyle [11] showed the fusion of iris and
periocular modalities. Miller et al. explored the effects of data quality in a periocular
biometric system [12].
Proposed methodology for the periocular biometric system [13] is depicted in

Fig. 2 which consists of the following steps: Data pre-processing, partitioning
testing/training sets, feature extraction and comparison methods.
2.1 Face Recognition Grand Challenge (FRGC)
Face Recognition Grand Challenge (FRGC) database consists of large face images,
and it is significantly used for extracting the periocular regions from the face
images. It consists of 16,029 still frontal face images, high-quality resolution faces
of size 1200 1400 with different sessions and variable expressions.
Training Phase Testing Phase
FRGC FRGC
Pre-processing of Face/ Pre-processing

Periocular Images
Feature Extraction Using Feature Extraction

Local Descriptors
Comparison
Unknown Known Face/Periocular
Fig. 2 Block diagram of proposed approach

2.2 Pre-processing of Face/Periocular Images
Pre-processing of the face/periocular images involves the following steps as shown

in Fig. 3. (i) Converting a raw colour, facial image into the preprocessed particular
images (ii) Geometric normalization (iii) Histogram equalization.
2.3 Periocular Region Extraction
Periocular region is extracted from normalized and equalized facial images, which
is accomplished by placing a square bounding box around each eye, centred on the
post-geometric normalization eye centre locations.
2.4 Feature Extraction
Each feature extraction technique transforms a two-dimensional image into a

one-dimensional feature vector through its unique process. Feature extraction
techniques employed in the periocular region are local appearance-based feature
representations are LBP, HOG, LPQ and WLD. These multiple descriptors are used
for deriving the discriminating features of the periocular region and city block
distance is used to compute the similarity between the feature vectors.
2.4.1 Local Binary Pattern (LBP)
Local appearance-based approaches [14] are the class of feature extraction tech-
niques which collects statistics within local neighbourhoods around each pixel of an
image providing the information related to the occurrence of certain textures and
patterns. The outcome of these approaches is one-dimensional feature vectors.
A Local Binary Pattern (LBP) is a texture classification method that was developed
by Ojala [15]. LBP collects the texture information from an image into a feature
Fig. 3 Process flow for the periocular region extraction

vector by labelling pixels with a binary number by placing a threshold on the

neighbourhood around each pixel. A histogram of these values forms the output
feature vector. LBP is used extensively for both facial recognition [16] and peri-
ocular recognition [17].
The LBP value of pixel of concern Pk, is a function of intensity changes in the
neighbourhood of M sampling points on a circle of radius r, then the LBP operator
is given by
X
M 1
LBPM;r ¼ sðgn gc Þ2n ; ð1Þ
n¼0
where
Sð pÞ ¼1 if p [ 0
¼ 0 if p\0
gc intensity of the pixel of concern at the centre of the pixels on the circum-
ference of a circle with values of gn, where n = 0, …, M − 1. In the proposed work,
all LBP calculations are made from a circle of radius 1 pixel with 8 pixels along the
circumference of the circle.
2.4.2 Local Phase Quantization
LPQ proposed by Ojansivu et al. [18] is a texture descriptor which quantizes the
phase information of a discrete Fourier transform (DFT) in patch-sized neigh-
bourhoods of an image. LPQ is robust to image blurring which has been used for
face recognition [8]. Like LBP, the resulting LPQ codes are formed into a
histogram.
In LPQ, the local spectra at a pixel p is calculated from a short-term Fourier
transform and is given by
X
F ðu; pÞ ¼ f ðp lÞej2puTl ; ð2Þ
l2px
where Pu is a pixel in a M M neighbourhood around u. At frequency points

u1 = [a, 0]T, u2 = [0, a]T, u3 = [a, a]T, u4 = [a, −a]T, local Fourier coefficients are
computed where a is 1/M. The phase portion of the Fourier coefficients is defined as
the sign of the real and imaginary components of F (u, x) given by
qj ðuÞ ¼ 1; if qj ð uÞ [ 0
ð3Þ
¼ 0; otherwise
The LPQ score is the binary coding of the eight binary coefficients qj(u) and all
LPQ calculations were made on a 9 9 pixel window.
2.4.3 Histogram of Oriented Gradients (HOG)
Histogram of Oriented Gradients (HOG) is an edge- and gradient-based feature

descriptor [19] originally developed by Dalal and Triggs to detect humans in
images. HOG is a local appearance-based approach that counts the occurrences of
different gradient orientations in localized portions of a picture. Even though HOG
was originally intended for object detection, it has been used for both facial and
periocular recognition. HOG is a simple technique which is invariant to geometric
and photometric transformation. A modified HOG algorithm is used for extracting
features from the periocular region. The first step is computing the gradient of the
image using a Prewitt convolution kernel. The gradient magnitude, Gmag and gra-
dient angle, Gangle, are calculated from the image gradient, Gx in the horizontal
direction and Gy in the vertical direction, as defined by
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Gmag ¼ G2x þ G2y ð4Þ
Gangle ¼ atan2ðGx ; Gy Þ ð5Þ
The values of Gmag and Gangle at each pixel location are accumulated into a
histogram.
2.4.4 Weber Local Descriptor (WLD)
Weber Local Descriptor (WLD) is a texture descriptor developed by Chen et al. The
law states that the change in a signal that will be just noticeable is proportional to
the magnitude of the original signal. WLD is concerned with the ratio between the
intensity value of a pixel and the relative intensity differences of the pixel to its
neighbours, also called the differential excitation and the gradient orientation of a
pixel. The WLD feature vector is a histogram of the occurrences of each excitation
and direction.
2.5 Classification Using City Block Distance Metric
City Block Distance Metric is used to determine the closeness between two feature
vectors obtained from the feature extraction methods and City Block Distance
Metric is given by
X
n
d ðk; lÞ ¼ jki li j; ð6Þ
j¼0
where k and l are feature vectors extracted from two images that have a feature
dimensionality of n.
3 Performance Measures
Following performance measures define the robustness of the designed system

using face/iris modalities:
Rank 1 recognition rates illustrate the successfulness of the scheme in identifying
the best match for a subject.
The rate at which the False Rejection Ratio (FRR) and False Acceptance Ratio
(FAR) are equal is defined as Equal Error Rate (EER).
FRGC protocol advocates the verification rate at 0.1% false accept rate be used to
compare the performance of two methods.
D shows the separability between the similarity score distributions of the set of true
matches and false matches.
4 Results
In the proposed paper, four local feature extraction methods are applied to three
image regions, i.e. left eye, right eye and full face images of the FRGC Experiment
1 dataset. Tables 1, 2 and 3 show the performance statistics [20] for the four local
feature extraction methods in terms of Rank-1 Accuracy, Equal Error Rate (EER),
VR at 0.1% FAR, and D. In most of the cases, face region performs better when
compared with the periocular regions as it has more information in terms of nodal
points. Particular region showed significant accuracies compared to face with only
using 25% of the full face information. LPQ scheme produced the best performance
results when extracted from face images (Figs. 4, 5, 6, 7 and Table 4).
Table 1 Results obtained Region Local Binary Pattern (LBP)

from Local Binary Pattern
Rank-1 EER VR @ 0.1% D
FAR
Left eye 99.7057 8.8312 64.8837 2.7330
Right 99.7003 8.2005 69.7148 2.8245
eye
Face 99.9178 7.0837 69.9876 2.9423
Table 2 Results obtained Region Local Phase Quantization (LPQ)

from Local Phase
Quantization
FAR
Left eye 99.7682 7.1179 75.9181 2.8653
Right 99.7814 6.7217 76.6564 2.9434
eye
Face 99.9424 5.4145 79.9201 2.9911
Table 3 Results obtained Region Histogram of Oriented Gradients (HOG)

from Histogram of Oriented
Gradients
FAR
Left eye 99.6069 8.0829 69.6951 2.8350
Right 99.6444 7.5245 72.2473 2.9378
eye
Face 99.8815 6.7912 70.3877 2.9692
Rank-1 Recognition Rates

100
Rank-1 Recognition
99.9
99.8 Left Eye
99.7 Right Eye
Rate
99.6 Face
99.5
99.4
99.3
LBP LPQ HOG WLD
Fig. 4 Rank-1 recognition rates for the LBP, LPQ, HOG and WLD methods
Equal Error Rate

14
Equal Error Rate
12
Left Eye
10
Right Eye
8
Face
6
4
LBP LPQ HOG WLD
Fig. 5 Equal Error Rates (EER) for LBP, LPQ, HOG and WLD methods
Verification Rate at 0.1% False Accept Rate

85
(VR @ 0.1% FAR)

80
75
70
Left Eye
65
Right Eye
60
Face
55
50
LBP LPQ HOG WLD
Fig. 6 Verification rate at 0.1% false accept rate (VR @ 0.1% FAR) for the LBP, LPQ, HOG and
WLD methods
Seperability between the Similarity Score D

3.2
Seperability between the
Similarity Score D
2.8
Left Eye
2.6
Right Eye
2.4 Face
2.2
2
LBP LPQ HOG WLD
Fig. 7 D shows the separability between the similarity score for the LBP, LPQ, HOG and WLD
methods
Table 4 Results obtained Region Weber Local Descriptor (WLD)

from Weber Local Descriptor
FAR
Left 99.3699 10.5697 56.5907 2.5122
eye
Right 99.4010 9.3778 61.5161 2.6390
eye
Face 99.6007 14.5565 41.9101 2.1189
5 Conclusion
In the proposed paper, various local descriptors are used for the feature extraction of
discriminative features from the regions of the full face, periocular regions and city
block distance is used as a classifier. Local descriptors used in the present work are
Local Binary Patterns (LBP) [16], Local Phase Quantization (LPQ) and Histogram
of Oriented Gradients (HOG) and Weber Local Descriptor (WLD). FRGC database
is used for the experimentation to compare the performance of both periocular and
face biometric modalities and it showed that the periocular region has a similar level
of performance of the face region using only 25% data of the complete face.
References
1. A. Jain, A. Ross, and S. Prabhakar. An introduction to biometric recognition. IEEE

Transactions on Circuits and Systems for Video Technology, 14(1):4–20, January 2004.
2. U. Park, R. Jillela, A. Ross, and A. Jain. Periocular Biometrics in the visible spectrum. IEEE
Transactions on Information Forensics and Security, 6(1):96–106, March 2011.
3. F. Juefei-Xu and M. Savvides. Unconstrained periocular biometric acquisition and
recognition using COTS PTZ camera for uncooperative and non-cooperative subjects. In
Proceedings of the IEEE Workshop on ACV, pages 20 1–208, January 2012.
4. F. Juefei-Xu, K. Luu, M. Savvides, T. Bui, and C. Suen. Investigating age invariant face
recognition based on periocular biometrics. In Proceedings of the International Joint
Conference on Biometrics, pages 1–7, October 2011.
5. Haibin Ling, S. Soatto, N. Ramanathan, and D.W. Jacobs. “Face Verification Across Age
Progression Using Discriminative Methods”, IEEE Transactions on Information Forensics
and Security, 2010.
6. N. Ramanathan. “Face Verification Across Age Progression”, IEEE Transactions on Image
Processing, Nov 2006.
7. J. Merkow, B. Jou, and M. Savvides. An exploration of gender identification using only the
periocular region. In Proceedings of the IEEE International Conference on Biometrics:
Theory, Applications, and Systems, pages 1–5, September 2010.
8. K. Sato, S. Shah, and J. Aggarwal. Partial face recognition using radial basis function
networks. In Proceedings of the IEEE International Conference on Automatic Face and
Gesture Recognition, pages 288–293, April 1998.
9. M. Savvides, R. Abiantun, J. Heo, S. Park, C. Xie, and B. Vijayakumar. Partial holistic face
recognition on FRGC-II data using support vector machine. In Proceedings of the IEEE
Conference on Computer Vision & Pattern Recognition, page 48, June 2006.
10. C. Teo, H. Neo, and A. Teoh. A study on partial face recognition of eye region. In
Proceedings of the International Conference on MV pages 46–49, December 2007.
11. G. Santos and E. Hoyle. A fusion approach to unconstrained iris recognition. Pattern
Recognition Letters, 33(8):984–990, 2012.
12. P. Miller, J. Lyle, S. Pundlik, and D. Woodard. Performance evaluation of local appearance
based periocular recognition. In Proceedings of the IEEE International Conference on
Biometrics: Theory, Applications, and Systems pages 1–6, Sept 2010.
13. K. Hollingsworth, K. Bowyer, and P. Flynn. Identifying useful features for recognition in
near-infrared periocular images. In Proceedings of the IEEE International Conference on
Biometrics: Theory Applications and Systems, pages 1–8, September 2010.
14. A. Joshi, A. Gangwar, R. Sharma, and Z. Saquib. Periocular feature extraction based on LBP
and DLDA. In Advances in Computer Science, Engineering & Applications, Volume 166 of
Advances in Intelligent and Soft Computing, pages 1023–1033. Springer, 2012.
15. T. Ojala, M. Pietikainen, and T. Maenpaa. A generalised local binary pattern operator for
Multiresolution gray-scale and rotation invariant texture classification. Second International
Conference on Advances in Pattern Recognition, pages 397–406, 2001.
16. Kishore K Kumar and P. Trinatha Rao “Face Verification across Ages using the
Discriminative Methods and See 5 Classifier”. In Proceedings of First International
Conference on ICTIS: Volume 2, Springer Smart Innovation, Systems and Technologies
51, Pages 439–448.
17. Mahalingam, Gayathri, and Chandra Kambhamettu. “Face verification with ageing using
AdaBoost and local binary patterns”, Proceedings of the Seventh Indian Conference on
Computer Vision Graphics and Image Processing - ICVGIP 10 ICVGIP 10, 2010.
18. T. Ahonen, E. Rahtu, V. Ojansivu, and J. Heikkilä. Recognition of blurred faces using local
phase quantization. In Proceedings of the International Conference on Pattern Recognition,
pages 1–4, December 2008.
19. Ramanathan. N “Computational methods for modelling facial ageing: A survey”, Journal of
Visual Languages and Computing, June 2009.
20. S. Bharadwaj, H. Bhatt, M. Vatsa, and R. Singh. Periocular Biometrics: When iris recognition
fails. In Proceedings of the IEEE International Conference on Biometrics: Theory,
Applications, and Systems, pages 1–6, September 2010.
Model-Based Design Approach
for Software Verification Using
Hardware-in-Loop Simulation
Pranoti Joshi and N. B. Chopade
Abstract Increasing demand in high quality products with high safety require-
ments and reduced time-to-market are the challenges faced during the development
of embedded products. These products irrespective of different domains (consumer,
automotive, medical, aerospace) incorporate multidisciplinary (electrical, mechan-
ical, electronic) systems as a part of hardware along with complex software that
controls the hardware. Late integration of these multidisciplinary systems in the
development cycle followed by the software verification and validation may lead to
expensive redesigns and delayed time-to-market. Model-based design
(MBD) approach can be used to overcome these challenges. Hardware-in-loop
(HIL) verification is an effective method that can be used to verify the control
software. Plant modeling is the crucial part for HIL verification. This paper will
provide a review of one of the steps of model-based design (plant modeling) that
can be used for software testing along with the impact of fidelity of model on the
verification.

Keywords Model-based design (MBD) Plant modeling Hardware in loop

(HIL) Model fidelity Software verification
1 Introduction
Embedded systems are nothing but control systems that include sensors to sense the
parameter of interest. These inputs are taken by the control system, processing is
done and the loads are driven accordingly. In modern systems (e.g., cell phones,
P. Joshi (&) N. B. Chopade

Pimpri Chinchwad College of Engineering, Akurdi, Pune, India
e-mail: jpranu1234@gmail.com
N. B. Chopade
e-mail: nbchopade@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_35
354 P. Joshi and N. B. Chopade
ovens, cars, etc.), the control part is taken care of by the software. As the users
demand more and smarter systems with new features, the software complexity has
increased to about millions of lines. This makes the testing process more chal-
lenging. Stringent time-to-market requirements with no compromise in quality add
to the testing challenges. In such a scenario, manual testing is obviously not a
correct way of testing. Hence, new approaches need to be used to sustain the current
demands and scenarios. Socci [1] has mentioned that at a frequency of 1000 lines of
code, there are 10–20 defects encountered in an embedded system.
2 Literature Survey
Testing control software in open loop may lead to time as well as cost expenses
when it will be integrated with actual plant because of potential errors. Hence, in
[2], an approach of closed loop modeling for verification and validation has been
suggested. To have an assurance about correctness of the control software, the
controller and the plant have to be taken into account that together form a
closed-loop system. Plant modeling is one of the phases of model-based design.
A model-based design workflow for embedded control systems can accelerate
product development, improve performance, increase reliability, and reduce engi-
neering cost [1].
In [3], authors suggest that data-driven models could serve as a substitution for
highly complex physics-based models with an insignificant loss of prediction
accuracy for many applications. Thus to assess the quality of models, model fidelity
comes into picture.
In [4], analysis of behavioral abstraction relation to the type of behaviors that can
be validated and how different model fidelity levels can be related to each other is
done. However, several studies have shown that high fidelity simulators may not be
necessary to produce effective training results. The fidelity of the model repre-
sentation also affects the analysis capabilities of the early design method.
Nowadays, many vendors of hardware-in-loop simulation systems are available
in the market, but each system has its own pros and cons. A system on chip (SoC) is
an integrated circuit (IC) that integrates all components of a computer or other
electronic system into a single chip [4]. For power electronics related applications,
system-on-chip solution can provide many combinations of digital, analog, or
mixed signal architecture inside one die. Having reconfigurable logic along with the
processor on a single die provides high processing power in less computation time
(up to resolution in nanoseconds). In [5], authors have developed a Xilinx Zynq
based HIL platform that can be suitable for power electronics applications. The
below sections will include limitations of traditional software testing methods,
overview of model-based design, hardware-in-loop simulation, and model fidelity
impact on HIL.
Model-Based Design Approach for Software Verification … 355
3 Traditional Design Approach of Application Control

Units (ACU)
In a traditional workflow, requirements analysis and high level design tradeoffs are
limited to paper or done using expensive prototype hardware. Control software is
written manually based on these framed requirements, as is the testing and verifi-
cation [6]. Embedded products contain interdisciplinary subsystems whose design
and development process is carried out separately using specific tool and methods.
After design and development of each subsystem, integration of the subsystems is
carried out followed by validation process.
Thus, the errors/faults in the system can be detected very late in the development
cycle; after which the entire process needs to be repeated till validation stage which
incurs more cost, design time, and efforts. The testing of the embedded system
(ACU) is also dependent on the availability of the hardware components even if a
prototype or actual system is used. This approach may cause damage to the system
hardware affecting the cost of failure.
An embedded system includes an ACU (controller) and the system plant which
includes the sensors and the loads are shown in Fig. 1. The “process” part of the
control system forms the plant model that is used to replicate the behavior of actual
hardware. It involves the system behavior modeling, sensor, and load modeling. In
traditional method for ACU testing, actual hardware is used. The sensor outputs are
read by the ACU and depending on the control signals sent to the plant, the loads
are driven. Using model-based design methodology, this hardware is replaced by
the models. The sensor models will sense the system parameters from the plant
model which will be sent to ACU. Necessary control signals will be sent from ACU
to the model which will drive the corresponding loads.
Fig. 1 Embedded System as

a controller and a process
3.1 Concept of Model-Based Design (MBD)
Model-based design is a model-centric approach to the development of control,

signal processing, communications, and other dynamic systems. Rather than relying
on physical prototypes and textual specifications, model-based design uses a model
throughout development [6]. A model includes every component relevant to system
behavior: algorithms, control logic, and physical components. Model-based design
has an advantage of early error detection in the software during the development
cycle even before the hardware is available. Simulation helps designers spot
problems that would require hardware changes. It is very important as hardware
changes are more expensive than software fixes [6].
Requirements analysis and research is the first and most important phase of
MBD. The purpose and application define the scope/boundaries for the plant model
development as well as the controller design. Plant modeling can be used for
controller design as well as verification.
Once the controller is designed, both the plant model and controller are simu-
lated together. Code generation then takes place that can be dumped into the actual
controller [7]. Verification is done after code generation to check that the algorithm
still confines to the requirements. Various tools, viz., MATLAB/Simulink,
LabVIEW, Dymola are used for modeling [8].
4 Hardware-in-Loop Simulation
Hardware-in-loop simulation is the technique used in the development and testing

of embedded systems. An HIL simulation must also include the models of sensors
and actuators. These models act as a bridge between the plant model and the control
hardware under test. The value of each sensor is controlled by the plant simulation
and is read by the embedded system under test. Likewise, the embedded system
under test implements its control algorithms by outputting actuator control signals.
Changes in the control signals result in changes to variable values in the plant
simulation [1]. Once the plant model is developed, code is generated out of it and
dumped into a real-time PC. Thus, real-time response can be obtained. The block
diagram can be seen in Fig. 2. ETAS, dSpace, OPAL-RT technologies, and
National Instruments are some of the suppliers of HIL hardware platforms.
Hardware-in-loop simulation is advantageous when the hardware system is yet to
be built, or the failure conditions cannot be replicated physically. It also ensures
safety of the people performing the testing. Software testing can be started early in
development cycle as there remains no dependency on the actual hardware greatly
facilitating reduction in development time. Automated tests can be run that increase
test coverage and shorten testing times by running complete test suites and over-
night tests. HIL systems testing 24 h, 7 days per week are not fiction but reality
[9, 10].
Fig. 2 HIL process flow
4.1 Plant Modeling
Plant model is a description of the physical hardware that is to be controlled by the

embedded system. Plant model may include interdisciplinary systems. A plant
model can be developed using experimental data or mathematically where different
types of systems whether mechanical, thermal, and electrical are described using
differential equations based on physical laws. Such kind of modeling is also known
as modeling using first principles. The problem with this type of approach is that
certain physical phenomena are very difficult to be described using physics [9].
The values of constants that need to be defined while deriving the equations may
not be available in all cases. System identification proves fruitful in such cases. The
model and controller forming the loop can be seen in Fig. 3.
Fig. 3 Plant model (includes

sensors, actuators, and system
behavior)
When the input–output relationship of a component or a part of hardware is

known, then instead of following a detailed approach system identification or curve
fitting can be used. System identification includes black box, gray box, and white
box modeling.
5 Model Fidelity
Model fidelity defines the accuracy of the representation of a model when compared
to the real world. Fidelity must be quantified with respect to a referent (set of
metrics against which a model can be compared), i.e., metrics should exist to
determine if a model resembles a referent. The metrics can include the level of
detail, resolution, accuracy, precision, and error. Efficiency, cost-effectiveness, and
schedules should dictate the appropriate level of modeling to meet the needs of the
customer.
Quality of modeling is very application specific. It will not be a case always that
the same component model is valid for different applications. In a broader sense,
models can have high or low fidelity. Low fidelity models are the ones built on
statistical fits on experimental data. High fidelity models are the ones constructed
using physics theory.
5.1 Scaling-Based Approach to Determine Model Fidelity
Scaling method can be used to determine fidelity based on the type of method used
to construct the model and the model response. Lowest level defined as: no
abstraction used high fidelity. Highest level defined as: highly abstract, low fidelity.
The more precise description about fidelity can be given as
Case 1: If the model is constructed using the experimental data and the response is
as intended, but the model response for different set of data is deviating from its
intended response then the fidelity is low.
Case 2: If the model is constructed using the physics theory and the response is as
intended, and is the same for different sets of data then the fidelity is high.
The desired level of fidelity is one that provides analysis results which are
complete and effective in aiding the decision-making process. Fidelity level of the
model impacts on the verification of the system.
5.2 HIL and Model Fidelity
The level of fidelity directly impacts the execution time required for the model and
also the development time and cost. The main purpose of HIL is to verify the
control algorithm and not the hardware design. The main focus is to get the inputs
from the ACU and provide corresponding feedback signals to the ACU. Hence,
behavioral kind of modeling can be leveraged instead of doing detailed modeling as
far as the desired response is obtained from the model.
6 Case Study: Temperature Sensor Modeling
Consider embedded software has been released that includes a feature of sensing
water temperature inside a tank. To test this feature, the controller will require
feedback from a temperature sensor. The data flow is described in Fig. 4. If the
traditional testing method is followed, a temperature sensor, a tank with water or a
prototype including these will be required to proceed with testing. Two types of
testing are needed to be carried out to ensure maximum test coverage, test-to-pass,
and test-to-fail. Generating failure conditions using real hardware is time con-
suming as well as cost adding.
Generating fault conditions in software is really easy, less time consuming and
can be done using automation as well, avoiding the presence of test engineer to
perform the testing.
The boundary conditions for temperature sensor can be checked such as how the
software responds if the temperature lies beyond the normal operating range.
Conditions such as sensor damage, accurate control of heater according to the
sensor feedback can also be tested.
Figure 5 shows the heater control done by ACU based on the feedback obtained
from the temperature sensor. When the desired temperature level is met, heater is
Fig. 4 HIL data flow

Fig. 5 Heater control using temperature sensor feedback
turned off. If the desired value is less than the required value, then the heater is
turned on to maintain the water temperature.
7 Conclusion
Traditional software verification methods that have been employed in organizations

are not able to cope up with the current embedded system trends. Reduced
time-to-market, high quality, and safety are the major concerns while designing any
product. Model-based design that was dominant in automotive domain is also being
employed in other domains. In current scenario, MBD can make up to the expec-
tations by the use of models throughout the development cycle and also as a
communication between customers and vendors. Plant modeling forms the base for
the product development as well as testing. The importance of model fidelity was
highlighted. It can be concluded that the model fidelity is application and imple-
mentation method dependent and no specific measurement technique exists. Most
of the measurement techniques are more quantitative rather than qualitative. The
model can be said to be of high fidelity if it meets the desired requirements. The
paper described the use of MBD and the role of HIL in software verification.
Employment of MBD for software verification involves high initial cost, but can
then be reused for different applications and the other advantages of early error
detection, reduced use of hardware components and model reuse that it brings, leads
to the reduction in the overall development cost and time.
References
1. Socci, Vince. “Implementing a model-based design and test workflow.” In Systems

Engineering (ISSE), 2015 IEEE International Symposium on, pp. 130–134. IEEE, 2015.
2. Preuße, Sebastian, Hans-Christian Lapp, and Hans-Michael Hanisch. “Closed-loop system
modeling, validation, and verification.” Emerging Technologies & Factory Automation
(ETFA), 2012 IEEE 17th Conference on. IEEE, 2012.
3. Zhou, Datong, Qie Hu, and Claire J. Tomlin. “Quantitative Comparison of Data-Driven and
Physics-Based Models for Commercial Building HVAC Systems.”
4. Kelemenová, Tatiana, et al. “Model based design and HIL simulations.” American Journal of
Mechanical Engineering 1.7 (2013): 276–281.
5. Kaczmarek, Michał, and Przemysław Koralewicz. “Hardware in the loop simulations of
industrial application using system on the chip architecture.” Signals and Electronic Systems
(ICSES), 2016 International Conference on. IEEE, 2016.
6. Kelemenová, Tatiana, et al. “Model based design and HIL simulations.” American Journal of
Mechanical Engineering 1.7 (2013): 276–281.
7. http://in.mathworks.com/solutions/model-based-design.
8. Köhl, Susanne, and Dirk Jegminat. “How to do hardware-in-the-loop simulation right.”
No. 2005-01-1657. SAE Technical Paper, 2005.
9. http://www.manmonthly.com.au/features/plant-modelling-a-first-step-to-early-verification-of-
control-systems/WP_Plant_Moeling.pdf.
10. Murphy, Brett, Amory Wakefield, and Jon Friedman. “Best practices for verification,
validation, and test in model-based design.” No. 2008-01-1469. SAE Technical Paper, 2008.
Silhouette-Based Human Action
Recognition by Embedding HOG and PCA
Features
A. S. Jahagirdar and M. S. Nagmode
Abstract Human action recognition has become vital aspect of video analytics.
This study explores methods for the classification of human actions by extracting
silhouette of object and then applying feature extraction. The method proposes to
integrate HOG feature and PCA feature effectively to form a feature descriptor
which is used further to train KNN classifier. HOG gives local shape-oriented
variations of the object while PCA gives global information about frequently
moving parts of human body. Experiments conducted on Weizmann and KTH
datasets show results comparable with existing methods.
Keywords Action recognition PCA HOG Silhouette image
1 Introduction
In the recent days, there has been huge swell in video data as social media has
become popular for education, information, and entertainment. Video surveillance
has increased considerably in last decade which produces huge video data.
Developing an intelligent video surveillance system in which human actions and
events can be detected automatically is a challenging task taken up by researchers.
This paper tries to find solution for automatic human action recognition. Human
action recognition is a procedure of detecting and labeling the actions. Human
action recognition has applications [1] in Intelligent Video Surveillance Systems,
Human–Computer Interaction, Content-Based Video Retrieval, Robotics, Sports
Video Analytics, etc.
A. S. Jahagirdar (&)
MIT College Engineering, Pune, India
e-mail: aditi.jahagirdar@mitcoe.edu.in
M. S. Nagmode
Government College of Engineering, Avasari (Kh), Pune, India
e-mail: manoj.nagmode@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_36
364 A. S. Jahagirdar and M. S. Nagmode
The various methods proposed in literature basically defer in type of feature

selected and classification algorithm used. The features can be broadly divided as
shape features, texture features, motion features, or combination of these. In this
paper, we have used Histogram of Oriented Gradients and principal component
Analysis as features and proposed a system where these two features are con-
catenated to form a feature vector. Classification of test data is done using Nearest
Neighbor Classifier. The algorithm is tested on two well-known datasets, viz.,
Weizmann and KTH and show satisfactory results.
Organization of remaining paper is as follows: In Sect. 2, review of earlier work
is done. In Sect. 3, proposed framework is discussed. Sections 4 and 5 give
experimental setup and results. Section 6 gives conclusion of the work.
2 Previous Work
The methods proposed for human action recognition by various authors can be
broadly classified on basis of features used for representing the videos. Bag of
Words and Spatio Temporal Interest Points (STIP) are most widely used local
representations [2]. In contrast to this, holistic approaches extract features directly
from raw videos. Local feature methods do not require steps like background
subtraction and tracking which are to be used in holistic approaches. The low level
features used in both local and holistic approaches are mostly appearance based or
motion features.
Recent trend in action recognition shows use of fusion of shape, motion, and
texture features for constructing a feature vector and using various methods like
covariance matrices, Fisher vector or GMM, etc., to form a feature descriptor [3–6].
Also various machine learning methods like neural network and convolutional
neural network [7, 8] are explored for classification of actions. A genetic pro-
gramming method is proposed in [8, 9] for extracting and fusing spatio-temporal
feature descriptor. It simultaneously extracts and fuses color and motion informa-
tion in one feature.
Another important question is what should be the size of feature vector. The
local features which are used by researchers have proposed dimensionality reduc-
tion techniques. Dimensionality reduction is achieved in [10, 11] using PCA and
Kernel PCA. In [12], Robust PCA is applied before extracting the features. PCA is
used in [13] along with HOG feature to achieve tracking and recognition simul-
taneously. Authors of [14, 15] propose tracking of individual body parts to form a
motion curve. Basic feature vector is formed by segmentation and curve fitting of
motion curve. Authors have used PCA for dimensionality reduction and classifi-
cation is done in eigenspace.
Use of Histogram of Oriented Gradients (HOG) descriptor is explored by many
researchers. HOG was first introduced by Dalal and Triggs in 2005 for Human
Detection in images [16]. For detecting humans in a video use of Histogram of
Silhouette-Based Human Action Recognition … 365
Optical Flow (HOF), a motion descriptor along with HOG was proposed in [17].
Many of the methods proposed in literature make use of HOG, HOF features along
with some modifications for detecting human action.
3 Proposed Framework
The aim of this work is to identify human actions from Weizmann dataset and KTH
dataset. The proposed method is divided in four major steps as shown in Fig. 1:
Foreground segmentation, Silhouette Extraction, HOG & PCA feature extraction
and classification. K-NN classifier is used for classifying the actions.
3.1 Foreground Segmentation and Silhouette Extraction
In first stage, video sequence is converted to frames and foreground segmentation is

done using background subtraction method to find Region of Interest. Here,
adaptive background subtraction method is used where background is modeled and
then progressively updated with every incoming frame.
In Eq. 1, It denotes current frame, It(x) denotes pixel value at location (x), and Lt
denotes binary motion label for that frame then
(
1; if jItð xÞ Btð xÞj s
Ltð xÞ ¼ ð1Þ
0; if jItð xÞ Btð xÞj s
where Bt is the background for the current frame and s is the threshold. In this
work, running Gaussian average model is used for computing background image.
The background model is updated for each incoming frame using Eq. 2 [18]:
SilhoueƩe AcƟon
• Background ExtracƟon • HOG Feature ClassificaƟon
Modeling ExtracƟon • K-NN Classifier
• Morphological
• Background OperaƟons • PCA Feature
SubtracƟon ExtracƟon
• SilhoueƩe
ExtracƟon
Foreground Feature Vector
SegmentaƟon FormaƟon
Fig. 1 Overview of proposed action recognition scheme

Fig. 2 Human detection and extraction from sample frames. (a–d) show detection of moving
human in sample frames from Weizmann and KTH dataset respectively. (e), (f), (g), and (h) show
object extracted from frame
Btð xÞ ¼ aItð xÞ þ ð1 aÞ Bt 1ð xÞ: ð2Þ
where a is updating coefficient. Value of a is between 0 and 1. Value of a is

determined empirically. The background subtraction process results in creating a
foreground blob, which is the required silhouette. Morphological operations, ero-
sion, and dilations are further applied to obtain robust silhouette of the object.
Feature extraction process is performed on this extracted silhouette of the object
which reduces computational complexity significantly. Figure 2 shows sample
frames with foreground blobs and corresponding object extracted.
3.2 HOG Feature Vector
Histogram of Gradient Orientation (HOG) counts occurrences of gradient orienta-

tions in a localized part of image. HOG describes appearance as well as shape of the
object in image in terms of distribution of intensity, giving the information about
edge directions. HOG descriptor is invariant to geometric and photometric trans-
formations. HOG was first introduced by Navneet Dalal and Bill Triggs in 2005
[16] and was used for human detection in image with very good results.
In this work, specifications defined in [19] are used to find HOG features. Taking
into consideration that there is redundancy in video data, 20 frames are selected for
generating the feature vector.
3.3 PCA Feature Vector
Principal component analysis is a second-order statistical measure to represent data

with maximum possible variance values and lower dimensions. Principal compo-
nents (PCs) give global information of extracted silhouettes [20].
Eigenvectors and eigenvalues are computed for the covariance matrix of zero
mean data. Eigenvectors having higher eigenvalues are selected for representing the
data. A feature vector is formed by converting diagonal of eigenvalues to row matrix.
A matrix of eigenvalues computed for each frame represents a video. Choice of
number of frames used to represent one video is not defined and is a tradeoff
between computational complexity and accuracy of recognition.
3.4 HOG-PCA Feature Vector
HOG feature and PCA features extracted from silhouette of object are embedded
together to form HOG-PCA feature vector. HOG gives appearance information in
form of shape and PCA gives motion information in form of principal component.
3.5 KNN Classifier
KNN is a simple classification algorithm which classifies data based on selected

similarity measure. It is basically a lazy learning algorithm which is memory based
rather than learning based. Even if it requires storing all training data and does
almost all computations at the time of classification, it gives good results when
number of classes is more in number.
4 Experimental Setup
In this element, we explain the data sets used for evaluating the algorithm. The
algorithms are tested on two well-known publically available and widely used
action dataset, viz., Weizmann and KTH.
4.1 Weizmann Dataset
Weizmann Dataset [21] used for testing the algorithm is specifically recorded for
action recognition. Database consists of 90 low-resolution video sequences.
9 different have performed 10 natural actions. Some actions are periodic and others
are nonperiodic. Videos are recorded in controlled environment using static camera.
Only one moving object is present in frame. Background is uniform and
uncluttered.
4.2 KTH Dataset
KTH Dataset [22–25] contains six natural human actions like walking, jogging,
running, etc. Actions are performed by 25 different people in 4 different situations.
All videos are recorded with uniform backgrounds with 25 fps frame rate. Videos
are downsampled to the spatial resolution of 160 120 pixels. Total 600 videos
are available in this dataset.
5.1 Results with Weizmann Dataset
Figure 3 show results obtained with Weizmann dataset in form of confusion matrix.
Threefold cross validation was used for evaluating the performance.
It is observed that average recognition accuracy achieved with PCA-HOG
embedded feature is much higher than any one of these features used individually.
In Weizmann dataset, action of waving hand is confused with jumping jack
where hands are moved along with jumping. Similarly, running action is confused
with walking action because of similar characteristics of frames. Also, it is seen that
recognition rate is higher for actions which are performed multiple times in a video
(i.e., jack, jump) than actions performed once in full video (i.e., bend). Table 1
shows comparison of results obtained with our method.
Fig. 3 Confusion Matrix on the Weizmann Dataset. a HOG Feature, b PCA Feature, c HOG
+PCA Feature
Table 1 Comparison of Method used Recognition accuracy

results obtained with
Weizmann (%) KTH (%)
Weizmann and KTH Dataset
PCA+KNN 76.66 81.66
HOG+KNN 85.83 90.83
HOG+PCA+KNN 94 91.83
Fig. 4 Confusion Matrix on the KTH Dataset. a HOG Feature, b PCA Feature, c HOG+PCA
Feature
5.2 Results with KTH Dataset
Figure 4 show results obtained with KTH dataset in the form of confusion matrix.
Threefold cross validation was used for evaluating the performance.
It is observed that recognition accuracy achieved with PCA-HOG embedded
feature is much higher than any one of these features used individually for KTH
dataset. It is seen that action of boxing is confused with action of clapping as in
many frames the human pose (shape) is similar. Also jogging action is confused
with running and walking action for same reason. Table 1 shows comparison of
results obtained with our method.
6 Conclusion
In this paper, we propose a method to use HOG and PCA as features for action
recognition from video data. The essential thought behind using these features
together was that the local object appearance and shape of an object within an
image can be described using HOG and motion parameter can be represented by
PCA. Since HOG operates on local cells, it is invariant to geometric and photo-
metric transformations. As feature extraction is applied on extract silhouette image,
computational cost is reduced considerably.
The classifier used here is simple NN classifier with Euclidian distance measure.
It simply computes the nearest neighbor of a test feature and sets its neighbor’s
label as a result. Number of frames to be used for detecting any action depends on
type of action. Selecting the frames from which features are to be extracted plays an
important role.
In future work, nonlinear Kernel PCA (KPCA) can be explored as a feature for
representing an action.
References
1. Ahmad, Mohiuddin, Irine Parvin, and Seong-Whan Lee: Silhouette History and Energy Image
Information for Human Movement Recognition, Journal of Multimedia, 2010.
2. Zhen X, Shao L: Action recognition via spatio-temporal local features: A comprehensive
study, Image and Vision Computing. 2016 Jun 30, 50:1–3.
3. Cheng, Shilei, et al.: Action Recognition Based on Spatio-temporal Log-Euclidean
Covariance Matrix, International Journal of Signal Processing, Image Processing and
Pattern Recognition 9.2 (2016): 95–106.
4. Carvajal, Johanna, et al.: Comparative Evaluation of Action Recognition Methods via
Riemannian Manifolds, Fisher Vectors and GMMs: Ideal and Challenging Conditions, arXiv
preprint arXiv:1602.01599 (2016).
5. Patel C, Garg S, Zaveri T, Banerjee A, Patel R: Human action recognition using fusion of
features for unconstrained video sequences, Computers & Electrical Engineering, 2016
Jun 18.
6. Zhen, Xiantong et al.: Embedding motion and structure features for action recognition, IEEE
transaction on Circuits and Systems for Video Technology 23.7 (2013): 1182–1190.
7. Feichtenhofer C, Pinz A, Zisserman A: Convolutional two stream network fusion for video
action recognition, in IEEE conference on Computer Vision and Pattern Recognition 2016
(pp. 1933–1941).
8. Ijjina EP, Chalavadi KM: Human action recognition using genetic algorithms and
convolutional neural networks, Pattern Recognition, 2016 Nov 30: 59: 199–212.
9. Liu L, Shao L, Li X: Learning spatio-temporal representations for action recognition: A
genetic programming approach, IEEE Transaction on cybernetics, 2016 Jan;46(1):158–70.
10. Kuehne H, Gall J, Serre T: An end to end generative framework for video segmentation and
recognition, In Applications of Computer Vision (WACV), 2016 IEEE Winter Conference on
2016 Mar 7 (pp. 1–8).
11. Oruganti VR, Goecke R: On the Dimensionality reduction of Fisher vectors for human action
recognition, IET Computer Vision. 2016 Feb 26;10(5):392–7.
12. Huang S, Ye J, Wang T, Jiang L, Wu X, Li Y: Extracting refined low rank features of robust
PCA for human action recognition, Arabian Journal for Science and Engineering. 2015 May
1;40(5):1427–41.
13. Lu WL, Little JJ: Simultaneous tracking and action recognition using the PCA-HOG
descriptor, in Computer and Robot Vision, 2006, The 3rd Canadian Conference on 2006 Jun
7 (pp. 6–6) IEEE.
14. Chivers, Daniel Stephen: Human Action Recognition by Principal Component Analysis of
Motion Curves, Diss. Wright State University, 2012.
15. Vrigkas, Michalis et al.: Matching Mixtures of curves for human action recognition,
Computer Vision and Image Understanding 119(2014):27–40.
16. Dalal Nvneet, Bill Triggs: Histogram of oriented gradients for human detection, 2005 IEEE
Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05).
Vol. 1 IEEE, 2005.
17. Dalal Navneet, Bill Triggs and Cordelia Schmid: Human detection using oriented histograms
of flow and appearance, European conference on computer vision, Springer Berlin,
Heidelberg, 2006.
18. Bouwmans, Thierry: Traditional and recent approaches in background modeling for
foreground detection: An overview, Computer Science Review 11 (2014): 31–66.
19. Chris McCormick: HOG Person Detector Tutorial 09 May 2013, http://mccormickml.com/
2013/05/09/hog-person-detector-tutorial.
20. Kim, Tae-Seong and Zia Uddin: Silhouette-based Human Activity Recognition Using
Independent Component Analysis, Linear Discriminant Analysis and Hidden Markov Model,
New Developments in Biomedical Engineering. InTech, 2010.
21. Blank, Moshe, et al.: Actions as space-time shapes, Computer Vision, 2005. ICCV 2005.
Tenth IEEE International Conference on. Vol. 2. IEEE, 2005.
22. Schuldt, Christian, Ivan Laptev, and Barbara Caputo: Recognizing human actions: a local
SVM approach, Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International
Conference on. Vol. 3. IEEE, 2004.
23. Siddiqi, Muhammad Hameed, et al.: Video based Human activity recognition using multilevel
wavelet decomposition and stepwise linear discriminant analysis, Sensors 14.4 (2014):6370–
6392.
24. Junejo, Imran N., Khurrum Nazir Junejo, Zaher Al Aghbari: Silhouette-based human action
recognition using SAX-Shapes, The Visual Computer 30.3 (2014):259–269.
25. Chaaraoui, Alexandros Andre, Pau Climenr_Perez, Francisco Florez-Revuelta:
Silhouette-based human action recognition using sequences of key poses, Pattern
Recognition Letters 34.15 (2013): 1799–1807.
Unified Algorithm for Melodic Music
Similarity and Retrieval in Query
by Humming
Velankar Makarand and Kulkarni Parag
Abstract Query by humming (QBH) is an active research area since a decade with
limited commercial success. Challenges include partial imperfect queries from
users, query representation and matching, fast, and accurate generation of results.
Our work focus is on query presentation and matching algorithms to reduce the
effective computational time and improve accuracy. We have proposed a unified
algorithm for measuring melodic music similarity in QBH. It involves two different
approaches for similarity measurement. They are novel mode normalized frequency
algorithm using edit distance and n-gram precomputed inverted index method. This
proposed algorithm is based on the study of melody representation in the form of
note string and user query variations. Queries from four non-singers with no formal
training of singing are used for initial testing. The preliminary results with 60
queries for 50 songs database are encouraging for the further research.
Keywords QBH Music similarity Pattern matching Information retrieval
1 Introduction
Query by humming (QBH) is one of the most natural ways of expressing query and
search a song from the musical database. Researchers in this domain of computa-
tional musicology are working on QBH since last decade or so. Content-based
retrieval with the melody of the song as input with QBH is a challenge and has very
limited success in commercial applications so far. The interfaces for song search
used and available today are mainly based on keywords on metadata associated
with the song.
V. Makarand (&)
Cummins College of Engineering, Pune, India
e-mail: makarand.velankar@cumminscollege.in
K. Parag
Iknowlation Research Labs Pvt. Ltd, Pune, India
e-mail: paragindia@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_37
374 V. Makarand and K. Parag
A typical query by humming system can be represented with block diagram as

shown in Fig. 1. Users submit the query in the form of hummed query. At data
cleaning stage, different signal processing techniques are applied, such as back-
ground noise removal, melody segmentation, etc. to make query ready for further
processing. The query is processed to identify and represent an underlying pattern
which is matched with song patterns already stored in the database to find best
matches. Depending on the best possible matches results are displayed with
ranking.
One of the major challenges in QBH is different users have different tonic or
fundamental frequency and identifying tonic becomes an initial challenge. The
query can be presented by the same user using different tonic at different instances
and at a different tempo. The dynamic time warping algorithm is used for nor-
malizing queries on timescale. Query transposes are used for same query repre-
sentation considering possible tonic information. This leads to many possible
representations of the same query and increases the query processing time. Since
queries are likely to be run for millions of songs in the commercial applications,
acceptable time complexity, and results are necessary for any commercial
application.
A novel mode normalized frequency (MNF) algorithm for edit distance measure
is proposed which eliminates the need of query transposes. N-gram inverted index
method uses precomputed n-grams of songs. Inverted index of each n-gram
includes a list of songs having these patterns. It is useful for the efficient matching
of the identified n-grams from the QBH. The performed experiment uses the raw
query data and presented results for different algorithms and unified algorithm. The
novel unified algorithm narrow downs the possible results using n-gram approach
and then apply MNF algorithm for edit distance on potential shortlisted songs.
The paper is organized in the following manner. Section 2 gives necessary
musical background and considerations used in the computation. Section 3 refers to
Fig. 1 Query by humming

Query by Data Ready
system
Humming Cleaning Data
Processing
Query
Pattern
Song Matching
Patterns
Results with Ranking

Unified Algorithm for Melodic Music Similarity … 375
related work, methodologies, and different approaches used by the researchers.

Section 4 detailed basic framework and approximate pattern matching algorithms
used for the study. Section 5 explains computational approach using MNF algo-
rithm. Section 6 covers details of results obtained using different algorithms. The
contribution of the paper is summarized with future directions in Sect. 7.
2 Music Background and Contemplations
Musical notations are represented on the logarithmic scale. The sample octave is
represented with 12 notes having frequencies F1 to F12. The equal tempered scale is
one of the most widely used musical scales. General formula for the calculation of
frequencies Fk from base reference frequency F1 for an octave is
Fk ¼ F1 2k=12 where k ¼ 2 to 12:

ð2:1Þ
F12 ¼ F1 2 12=12
¼ F1 2 ¼ 2 F1
The 12th note as F12 has a frequency double of F1 and it marks the end of the
current octave and beginning of next octave. For processing of the audio query, 100
samples are considered per second. In order to perceive any note, general minimum
duration of note to be sung is about 30–40 ms. Observations of the query by
untrained singers and general song patterns revealed 100 ms as the minimum
duration of any note as a rough estimate., i.e., 10 consecutive samples of the same
frequency are required to decode a specific note. These estimate found to be
appropriate for QBH.
Notes sung during query by users are not likely to be perfect as singers. They are
likely to deviate from the original note frequency. For such imperfect queries to
accommodate possible errors in the query, the algorithm design has considered a
range of frequency keeping actual frequency as a center to represent the specific
note. It has covered entire frequency spectrum using the range such that each
frequency will be associated with some or the other note. Another important
observation is users can have different pitch range and it reflects in the queries with
different frequencies for same song rendition by different users. The same person
can also submit the query in different scales and octaves. This leads to the concept
of frequency transpose considerations. As note sequence, F1 F3 F2 will be perceived
similar to all its aliases as F2 F4 F3 or F9 F11 F10, etc. These sequences represent
same relative difference as +2, −1 in subsequent notes as compare to previous note.
Relative differences are used for precalculation of n-grams for songs which elim-
inates the need of any transposes according to the tonic.
3 Related Work
Research on QBH has been prevailing since last two decades with many researchers
attempting to get better results. These results are influenced by two major factors as
quality of query itself and the algorithms used for search and comparison.
Algorithms not only need to be efficient in terms of accuracy but also needs to have
better time complexity considering millions of songs to be searched for the query in
the application. Queries are represented in the form of a string of notes in the
majority of the work with some exceptions as additional information as rhythm or
duration of notes with a pitch. Results are evaluated as top 10 hit ratios or mean
reciprocal rank. Top 10 hit ratios give a percentage of queries for which intended
song appears in top 10. Mean reciprocal rank (MRR) is calculated using the fol-
lowing formula. In case of four queries for the algorithm, the intended song appear
at say position 1, 5, 2, and 10, then MRR will be (1/1 + 1/5 + 1/2 + 1/10)/4 = 0.45.
MRR value is having a range between 0 and 1 with more value indicating better
results and lower value represents poor results.
Some standard datasets are available for western music but as per our knowl-
edge, no standard dataset is available for the queries and Hindi songs to compare
algorithms and test accuracies. Tansen [1] QBH for Hindi songs tested queries
using different syllables, such as “Ta”, “La”, “Da”, and evaluated the impact of
different syllable on the performance of the system. Tonal representation of melody
with semitone and octave abstraction along with Q max algorithm is used to
compute similarity [2]. Authors emphasized the need of query self-tuning to
improve results considering the results for the subjects with good tuning the top 10
hit ratio was 64%. Existing QBH systems have slow searching speeds and lack
practical applications and proposed GPU-accelerated algorithm with parallel pro-
cessing to improve search efficiency by 10 times [3]. Considering wide use of
mobile for QBH, the query needs to be small in duration and testing of commercial
applications revealed the time required is a crucial factor [4].
Importance of tracking fundamental frequency and its impact on the perfor-
mance of the system is studied and proposed further research for deeper insight into
different methods [5]. Different approaches for melodic similarity for Indian art
music are evaluated and it is observed that results are sensitive to distance measures
used by the algorithms [6]. The results are not comparable, as in the majority of the
cases datasets used are different by the different researchers.
Recently, new approaches have been applied to improve the accuracy of QBH
system by researchers. In addition to the conventional method, the results are
ranked according to user preferences or history to identify user taste and the songs
most likely queried at top positions [7]. Use of combining lyrics information
available with the user in melody achieved 51.19% error reduction for the top 10
results [8]. Segment-based melody retrieval approach can solve the problem of
noise and tempo variation in the query better than global linear scaling [9]. Despite
extensive research on the QBH, present systems for song search mainly relies on
keyword-based search with limited use in the commercial applications. This is our
main motivation for the research work in QBH domain of computational musi-
cology. It has revealed many further challenges and opportunities for the research
with various directions.
4 Basic Framework and Pattern Matching Algorithms
The humming query is processed using open-source tool PRAAT for experiments.
The default settings of PRAAT are used for query data processing [10]. Framework
for content-based music retrieval is proposed using human perception [11] and this
work fits as a module in it. The sampling rate for pitch samples used is 100 samples
per second. No data cleaning signal processing methods used for our experimen-
tation, however, removal of noise can improve accuracies further.
Pitch listing information with samples is further converted to note string with the
novel approach Mode Frequency Normalization. This algorithm is the outcome of
study and experimentation with queries and song patterns. Approximate pattern
matching algorithms are necessary to match the humming query patterns as exact
matches are very unlikely for the non-singers queries. Observations about hummed
queries revealed that some queries are time stretched, whereas some of them are
compressed on the timescale. In some cases, few notes are prolonged as compared
to others. It prompted us to consider compressed queries on the timeline with
removing duplicate notes in succession.
For query pattern matching, edit distance and n-gram methods are used. Use of
Euclidian distance and n-grams for pattern matching on compressed and non-
compressed queries generated four possible alternatives as (1) Simple Euclidian
distance. (2) Compressed Euclidian distance. (3) N-gram method. (4) Compressed
n-gram method. For compressed queries, the original song patterns are also com-
pressed for calculating distance measure. The cost of calculation for compressed
queries is less but it eliminates the comparison of pattern with same successive
notes.
Edit distance method uses Euclidian distance measure with distance as a mini-
mum number of insertions, deletions, or replacements required to match two
strings. The less the edit distance more similar the song will be for query. Dynamic
programming approach is used to find edit distance. For n-gram approach, the
calculations are restricted till 4-grams as it was observed unnecessary to calculate
further grams as occasionally matches are found for higher grams. In case of 2-gram
approach, all note patterns with two successive notes with relative distance are
identified from the query and compared with song patterns to an identified number
of occurrences of 2-grams. 2-gram approach is like substring matching algorithm.
The more number of n-gram matches indicate more likely the pattern matching for
the given query.
The inverted index structure is used with pre-calculated relative pitch 2-grams
(RP2G), 3-grams (RP3G) and 4-grams (RP4G) from the song patterns. This
approach eliminates the need for substring matching of the query with each song
pattern. To find matching n-grams of the query, pre-calculated values of n-grams

are used. For query example string with relative note distances as +1 +1 +4, the
2-grams are +1 and +4. An inverted index for string +1 will contain a list of songs
containing string +1 with in sorted order of a number of occurrences of +1. For each
song, 2-gram measure will be the sum of all matching 2-grams from the query. The
more the value of 2-grams, the more the song is likely to similar to the query.
Similarly 3-grams (+1 +1 and +1 +3 in the above example) and 4-grams (+1 +1 +4
in the above example) are also pre-calculated for song list. Results are evaluated for
different pattern matching algorithms using n-grams and edit distance methods.
The unified algorithm uses four individual algorithms in combination as per the
need. It advances to the next algorithm if desired results are not obtained in the
previous algorithm. At first, it uses RP4G then RP3G and RP2G followed by MNF
algorithm. It uses minimum cutoff value for n-grams. Present cutoff values are
decided considering our observations. Cutoff is minimum grams matched of the
query with the song patterns. Cutoff values considered are 2, 5, and 8, respectively,
for RP4G, RP3G, and RP2G at present, however, further fine tuning is possible.
The implementation of individual algorithms and unified algorithms is done using
C programming language and the results are tested for queries from different users.
The proposal is to submit the tested code to PRAAT open-source tool at the later
stage and make it available to all.
5 Mode Normalized Frequency (MNF) Algorithm
The note patterns for about 50 Hindi Popular songs studied which are selected
randomly from a different era. It was observed that few notes are repeated in the
majority of the note sequence pattern. These notes are different in different song
patterns, however, the trend prevails. Further introspection revealed that one note is
appearing more number of times compared to others. These observations lead us to
consider mode concept in statistics which is maximum occurrence symbol or value
in the set. Assuming this note as a reference, the song sequences are normalized and
converted them to mode normalized frequency form.
For normalized notes pattern, notes are represented using A–Z alphabetic
sequence with N as middle alphabet considered as a reference. All notes are con-
sidered relative to it, as a note with distance +2 as P or −2 as L with respect to N
and so on. The original song notes pattern is converted to the normalized pattern by
keeping a relative distance of notes same.
MNF algorithm for query processing:
1. Read pitch listing for the Query
2. Find maximum occurring frequency in the pitch listing (N)
3. Find other notes and their frequencies using the formula 2.1
4. Find range of each note to cover entire frequency range
5. Convert pitch listing of Query to notes representation

6. Read converted QBH for notes
7. If note appears 10 times in a consecutive manner add to output sequence
8. Continue reading the sequence till end
9. Output mode normalized sequence.
MNF algorithm eliminates the need for computation for transposes in Euclidian
distance measure; however, some exceptions for songs or queries do need transpose
calculations considering the pattern with non-repeating note for better results.
Following sample for song and query with appropriate representations gives an idea
about the conversion and representations used. * followed by note represents next
octave in the notations. Sa* appearing having more occurrences in compressed note
pattern is replaced by N in the normalized query pattern as shown in the example
below. Pattern matching is done using edit distance method in MNF algorithm.
Compressed song note sequence pattern: g Sa* ni Sa* re* ni Dh ni Sa* Dh dh Dh ni
Relative note sequence pattern: +8 −1 +1 +2 −3 −2 +2 +1 −3 −1 +1 +2
Normalized Query pattern for MNF edit distance algorithm:
FNMNPMKMNKJKM
6 Analysis of Results
Our experiments include humming samples recorded using ordinary recording mike
of laptop and mobile devices. High-quality recording system was not used pur-
posely as it would not be available with common listeners to generate queries. For
the humming queries, the samples are generated from people with no formal
training of singing. This has helped us to test our algorithms for imperfect samples
from the unprofessional users. In the experiments, 60 humming queries are used
with a rendition by four amateur youngsters. The queries are supposed to retrieve
one of the songs from the list of 50 sample songs as a part of the top songs of the
ranked list.
Analysis of results is done for compressed queries and compressed song patterns
using RP4G, RP3G, RP2G, MNF algorithm, and Unified approach. Present
experiments show unified algorithm is promising as it is time efficient for the
majority of queries to reduce the average computational time and overall accuracy
is also improved. It was observed in the run stage of a unified algorithm that out of
60 queries, for 24 queries only 4-grams got results, for 21 queries 3-grams got
results, for 8 queries 2G used and for 7 queries need to be occurred to use MNF
algorithm for top 10 results.
Individual user query and overall results for top 3 hit ratios are shown in Table 1
for all algorithms. Considering the mobile query interface, top 3 results carry more
importance than top 10, thus top 3 results are shown. It can be observed from the
table that for individual algorithms, RP3G is better compare to rest followed by
Table 1 Top 3 hit ratios in Queries RP4G RP3G RP2G MNF Unified
percentage
User 1 54.5 81.8 43.4 54.5 81.8
User 2 36.4 63.6 30.0 60.0 81.8
User 3 0 53.3 20.1 46.1 46.6
User 4 30.8 79.9 36.1 46.1 76.9
Overall 30.4 68.9 33.9 51.7 71.9
Table 2 MRR for different Algorithms RP4G RP3G RP2G MNF Unified
algorithms
MRR 0.2473 0.5076 0.232 0.45 0.59
MNF algorithm. The proposed unified algorithm has the best accuracy out of all
algorithms. Another observation is that results vary according to the user’s query
quality. It can be seen from the table that user 3 generated the poor-quality queries
among all others.
The mean reciprocal ratio or MRR gives an accuracy of methods. More the
MRR value, better the algorithm. Table 2 shows MRRs for different algorithms.
The unified algorithm has obtained better accuracy as compared to others for the
test data used. These preliminary results are encouraging for further
experimentations.
7 Summary and Future Directions
The unified algorithm a novel approach proposed in this paper is promising as it

reduces the cost of computation time with precomputed the inverted index for fast
look up. For 75% of the queries, the results obtained using RP4G and RP3G in the
unified algorithm and for the rest of the queries, RP2G and MNF algorithm were
used. The initial results show that the unified algorithm is time efficient considering
the average time and the results are also improved with MRR as 0.59 as compared
to 0.5076 of the best individual algorithm of RP3G and 0.45 of the developed MNF
algorithm.
Since users normally refer to the first page of results, if the intended song
appears quite below in the list it is not likely to be found. Thus, top 3/5/10 hit ratios
become more important than MRR from user’s point of view, however, MRR is a
better major for the accuracy of the algorithm. Further experimentation with more
queries from different users will help to fine-tune the unified algorithm for better
accuracy. Query and song segmentation approach needs to be evaluated to see
further precision. This will help us to fine-tune the unified algorithm and build a
robust system.
Acknowledgements The authors gratefully acknowledge the support by MKSSS’s Cummins

College of Engineering for providing experimental setup and the efforts by our UG students Aditi
Pawle, Snehal Jain, Sonal Gawande, and Sonal Avhad for the active help in preparing query
samples and experiments. Volunteer singer’s contribution for generating queries is highly
appreciable. We would like to thank Dr. Sahasrabuddhe H. V. for his valuable suggestions and
inputs related to musical knowledge and experiments.
References
1. Kumar P, Joshi M, Hariharan S, Rao P: Sung note segmentation for a query-by-humming

system. Intl Joint Conferences on Artificial Intelligence IJCAI (2007).
2. Salamon J, Serra J, Gómez E: Tonal representations for music retrieval: from version
identification to query-by-humming. International Journal of Multimedia Information
Retrieval. 2(1) pp 45–58, (2013).
3. Ruan L, Wang L, Xiao L, Zhu M, Wu Y: A Query-by-Humming System based on Marsyas
Framework and GPU Acceleration Algorithms. Appl. Math. pp 261–72, Feb (2013).
4. Chandrasekhar V, Sharifi M, Ross DA: Survey and Evaluation of Audio Fingerprinting
Schemes for Mobile Query-by-Example Applications. ISMIR Vol. 20, pp. 801–806 (2011).
5. Molina E, Tardón LJ, Barbancho I, Barbancho AM: The Importance of F0 Tracking in
Query-by-singing-humming. In ISMIR pp. 277–282, Nov (2014).
6. Gulati S, Serra J, Serra X.: An evaluation of methodologies for melodic similarity in audio
recordings of indian art music. In Acoustics, Speech and Signal Processing (ICASSP), IEEE
International Conference pp. 678–682 Apr, (2015).
7. Liu NH: Effective Results Ranking for Mobile Query by Singing/Humming Using a Hybrid
Recommendation Mechanism. IEEE Transactions on Multimedia. pp 1407–20, (2014).
8. Wang CC, Jang JS: Improving query-by-singing/humming by combining melody and lyric
information. IEEE/ACM Transactions on Audio, Speech, and Language Processing. pp 798–
806 (2015).
9. Liu, N. H: Effective Results Ranking for Mobile Query by Singing/Humming Using a Hybrid
Recommendation Mechanism. IEEE Transactions on Multimedia, 1407–1420(2014).
10. Ramesh V: Exploring Data Analysis in music using tool praat. ICETET. IEEE International
Conference, pp. 508–509, (2008).
11. Makarand, Velankar, and H. V. Sahasrabuddhe: Novel Approach for Music Search Using
Music Contents and Human Perception, IEEE International Conference on Electronic
Systems, Signal Processing and Computing Technologies ICESC, (2014).
Predict Stock Market Behavior: Role
of Machine Learning Algorithms
Uma Gurav and Nandini Sidnal
Abstract The prediction of a dynamic, volatile, and unpredictable stock market

has been a challenging issue for the researchers over the past few years. This paper
discusses stock market-related technical indicators, mathematical models, most
preferred algorithms used in data science industries, analysis of various types of
machine learning algorithms, and an overall summary of solutions. This paper is an
attempt to perform the analysis of various issues pertaining to dynamic stock
market prediction, based on the fact that minimization of stock market investment
risk is strongly correlated to minimization of forecasting errors.

Keywords Machine learning algorithms Stock market prediction

Efficient market hypothesis Ensemble machine learning
1 About Stock Markets
Stock exchanges are considered as a backbone of financial sectors of any country.

Most of the stock brokers, as shown in Fig. 1, dealing with sellers, buyers and
trading in the stock market (SM) use technical, fundamental, or time series analysis
for stock market (SM) prediction purpose. These methods do not ensure correct SM
prediction results, as these are based on the past stock market trends and not on
stock prices [1, 2]. It is, therefore, necessary to explore some improved methods,
such as ensemble machine learning (EML) algorithms of prediction [3]. Basically,
ML is the ability of a program to learn, adapt, and take intuitive decisions in the
face of different real-life applications which processes nonlinear, dynamic big
U. Gurav (&)
K.I.T’s College of Engineering, Kolhapur, India
e-mail: umabgurav@gmail.com
N. Sidnal
K.L.E’s College of Engineering, Belgaum, India
e-mail: sidnal.nandini@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_38
384 U. Gurav and N. Sidnal
Buyers
Prices obtained
Stock Exchange Brokers

by Share brokers
Index created to
inform people about
Market Performance Sellers
Stock Index
Fig. 1 How stock market works
datasets. EML works on the principal, various algorithms may overfit the data in
some area, but the correct combination of their predictions will have better pre-
dictive power.
The Efficient Market Hypothesis (EMH): EMH is a popular SM hypothesis
[2] that states the future stock price is completely unpredictable knowing the past
trading history of the stock. In EMH, any information acquired from examining the
history of the stock is immediately reflected in the price of the stock. The funda-
mental problem of this “hypothesis”, is that any market strategy will be unable to
survive large swings in the market.
The Random Walk Hypothesis: This hypothesis [2] states that stock prices do
not depend on the past stock prices, so patterns/trends cannot be explicitly deter-
mined. Further, it adds, with the advancement of more powerful computing
infrastructure, efficient ML algorithms and underlying pricing patterns analytics
optimal solution can be found [2]. Hence, ML Techniques can seriously challenge
traditional EMH like hypothesis [4].
1.1 Stock Market Prediction Methods
1. Technical Analysis: This method determines the stock prices based on the past
performance value of the stock (using time series analysis) [2]. Technical
Indicators used for market analysis are Fibonacci retracements and extensions,
Moving Averages, and Momentum indicators, like RSI (Relative Strength
Index), volume analysis, Mean reversion and co-integration, Bollinger bands,
MACD, EMA (Exponential Moving Average), Stochastic Oscillators, etc. The
main problem with these technical indicators is that these are applicable for a
specific time frame. There are different models for HFT (High-frequency trad-
ing), either intra-day or long-term trading. As per the historical observations,
long-term trading seems more beneficial in terms of profit making [5]. In some
cases, reverse engineering strategy works well, once graph reverses, needs to
take another path or follow another strategy. For the statistical models, time
Predict Stock Market Behavior: Role of Machine … 385
series analysis, autoregression, logistic regression strategies work well [6, 7]. To
solve the probability distribution problem of datasets, outliers, or extreme values
are considered and mean reversion is calculated. All above-mentioned technical
indicators are based on the price, which always reflects what has already hap-
pened in the market. Thus, the conclusion is that technical analysis is reactive
not truly predictive of what will happen [1] and stock prices are a result of
multiple variables.
2. Fundamental Analysis: This method deals with the past performance of the
company rather than the actual stock. For more complex indicators to capture
the price in a more precise manner, indicators used are Regression analysis,
Autocorrelation function, GARCH Models, Regime shifting models like
ANOVA, Markov Chains, or any other Stochastic model [1, 2]. Hence, the stock
price is a function of multiple factors which are volatile in nature.
3. Discounted Cash Flow analysis: known to evaluate the intrinsic value of a
stock but it is not applicable for cyclical. From Price to Sales judgments can be
done only whether the stock is cheap on a certain sector or on the overall market
[8, 9]. The problem with this is that experts in SM are calculating stock values
manually.
4. Sentiment Analysis (SA): eToro platform allows the user to capture the activity
of traders on the social network that are successful in their trades [1] based on
the SA. Silicon Market is a combination of advanced AI with ML algorithms
that learns from past trades, to inform future trading decisions and utilizes
adaptive and flexible ML to create an automated trading strategy [1]. Social
media can predict stock up to 80% revenue correlations with Google trends.
Reference [10] uses data from Google trends to predict stock market move-
ments, using a GARCH model to replicate. StockTwits which uses Twitter
mentions/searches to discover emotions of people. Twitter trends are better than
Google trends because twitter news is quicker, and accessible. Gaining profit
from Market Mood or Market Sentiment, works on the principle be greedy when
people are fearful and be fearful when people are greedy [10]. Ergo,
DeepStreetEdge are some more predictive methods.
1.2 Models Used for Stock Market Prediction
1. Six sigma (SS): SS method has its strength in the elimination of variations,
fluctuations from repeatable processes. The main issues in trading are when to
buy and sell shares? If these issues are addressed then the final outcomes are
better. SS can help refine, enhance the performance, improve a system that
predicts the direction of a stock, for example, managing a hedge fund [11].
2. Time series-based mathematical models: In this model shown in Fig. 2, the
correlation between datasets can be used with the help of statistical tools. Time
series models are built to study how a time series moves over time and
Time Series Analysis Based Models
ARIMA models 3. Exponential Smoothing

(3 basic Methods)
1. Auto Regression
3. Moving Averages
2. Differentiation
Fig. 2 Time series analysis-based models
application of volatility forecasting models to predict changes in volatility.

High-frequency data is prone to all sorts of biases and there has to be correction
mechanisms. Hence for the perfect dataset, even a simple linear model or
logistic regressions can predict something as complex as stock market prices [7].
Two types of time series analysis are ARIMA and Exponential Smoothing [12].
ARIMA (Autoregressive integrated moving average): The concept of moving
averages is incorporated in ARIMA models. Integrated moving average detects
movement, speed, and acceleration. The drawback of standard ARIMA mod-
eling is that it only considers proximal entries in the time series to forecast
subsequent entries [12]. A more advanced hybrid version of ARIMA models
with the combination of EML algorithms can consider cyclicality problems of
stock markets. Further performance enhancements can be done by including
data as external regressions which are outliers error value between actual and
forecast values.
Exponential smoothing: A second popular technique in which previous time
series entries contribute as a weighted moving average to forecast the next term.
Mean reversion and random walk theory, earning momentum are powerful
long-term forecasting tools in finance [12].
2 Methodologies: Machine Learning (ML) Algorithms
Improving on trading strategy iteratively with the help of different types of ML

algorithms as shown in Fig. 3 is the solution to this problem [13–15]. This is a
continuous process, capturing exact point in time snapshot of the fluctuating stock
value and then using it for further iterations is the challenge [16].
Clasiification of ML Algorithms
1. Highly Automated
2. Self modifying
Reinforcement ML algorithms
Supervised
Unsupervised 1. External corrective feedback is
provided on intermediate results.
1. Trained models 1.No labels associated with
2. This makes good decision better
2. Searches for patterns data points
3. Learning activity is based on combination
as per the value labels 2. Hence models are not trained
of stastical modelling to genrate predictions
assigned to intermediate
with some form of rule based algorithms
results
4. To interpret the significance of error rate fluctuations
Fig. 3 Classification and types of ML algorithms
2.1 Analysis of Various Machine Learning Algorithms Used

by Data Scientists
See Table 1.
Table 1 Analysis of ML algorithms

Analysis
ML algorithms When and How to use for Advantages/disadvantages/applications
stock Market prediction [14]
[14]
1. Linear 1. LR used for continuous Advantages: It is one of the most
Regression (LR) values (Time series interpretable ML algorithms, as it
analysis) and requires minimal tuning, it runs very
multidimensional problems fast. Generally, used in prediction
like SM for predicting problems. In LR mathematical forms,
exact point in time value SM estimate Et for the next iteration is
and not a category calculated based on previous iteration’s
forecasting values. For given instance x, at time t,
for k number of iterations, considering
error e, LR equation, is formulated as
follows:
Et ¼ ao þ ax1 ; t1 1 þ . . .axk ; tk 1 þ e
Applications: Estimating Sales, Risk
Assessment, Regression Trees
2. Logistic 1. LogR are used with Advantages: 1. It is very easy to debug
Regression manual discretization and tune
(LogR) which is very important 2. It extracts as much signal as possible
considering dynamicity to and leaves out noise. The fundamental
introduce nonlinearity problem is to have the data that
2. LogR applies the logistic maximizes the signal. Stock market
function to linear firms are in HFT. In low to medium
(continued)
Table 1 (continued)
Analysis
[14]
combination of features for frequency space, this technique gives
prediction of a dependent accurate results
variable, which analyze Disadvantages: Over fitting
individual components of Applications: Credit Scoring,
the markets such as the predicting the revenues of a certain
dependency of certain product, earthquake prediction
stocks on the basis of news
[14]
3. LR and LogR are the
most utilized technique in
High Frequency Trading
(HFT)
3. Decision Trees DT are of 2 types Advantages: DTs are self explanatory.
(DT) depending on outcome DT efficiently works with both
type. categorical and numerical variables.
1. Continuous Variables Even in case of missing data and
2. Binary Variables DT outliers it works very well, which in
creates a model of optimal turn reduces data preparation time
decisions and their possible Disadvantages: Accuracy of expected
consequences, including outcome is directly proportional to
chance event outcomes. number of decisions in DT. Sometimes,
DT are used in unrealistic DT leads to bad decision
circumstances where making. Not efficient for continuous
parameters are nonlinearly variables, creating large DT with more
related, in data exploration number of branches leads to higher time
and for feature selection in complexity which is a major drawback
predictive analytics Applications: In finance for option
pricing, in remote sensing applications
for extraction of similar patterns [14]
4. LDA/QDA LDA/QDA Gaussian-based Applications: For complex prediction
Gaussian-based classification are used for problems
classification samples within the same
distribution
5. K-Means 1. For Unsupervised Advantages: In case of global clusters,
Clustering Learning/Classification K-means results in tighter clusters than
2. Routinely used in hierarchical clustering algorithms. Time
applications like Google complexity is less, hence faster than
news to cluster similar hierarchical clustering for large number
stories, can be applied to of variables [14]
group the similar web Applications: Google, yahoo uses for
pages. clustering web pages, estimates the
3. Unsupervised learning predictability of the direction of
for outlier data in SM movement of stock prices correctly
prediction
(continued)
Table 1 (continued)
Analysis
[14]
6. Support Vector SVM classified into two Advantages: Gives best classification
Machine (SVM) categories: Linear and performance in terms of time
Nonlinear complexity and accurate results, on the
1. In nonlinear SVMs it is training data, overfitting does not occur
very difficult to classify Applications: SM predictions used by
training data using a hyper finance institutions. Enhancements can
plane because of be done using a stock-screener
nonlinearity [14] (classification), rather than trying to
2. If scalability is main predict exact stock price
criteria then SVM works (nondeterministic regression) problem,
well SVMs with reinforcement learning
3. In SVM each iteration achieves better results to compare
causes the “error/residue” relative performance of stocks, which
to become 1/2 of the helps in SM investment related
previous iteration, leading decisions
to very rapid convergence
resulting in faster
computations
4. Dynamic/nonlinear
classification problems
5. High-dimensional data
like text [17]
7. Neural 1. Used for prediction and Advantages: NN is a “black box”
Networks (NN) classification, where method, complex part of NN is
manipulating/weighting Gaussian mixture model or deep
features can give more learning algorithm which is more
precise results than simple advanced
classification Disadvantages: The nature of the
2. Can be used for stock predictions is unclear, reusability of the
market predictions, prices same NN model is questionable. Out of
fed into the network, with sample overfitting is the measure issue,
several technical indicators NN training is tiresome and output is
as input features, performs unstable, hence NN is poor for
adaptive filtering of the raw investment purpose problems like SM
price data, feed filtered data
into the Recurrent NN
(RNN) model, train it, and
predict next week’s move,
learning nonlinear models
[18]
8. Deep Learning 1. DL Convolutional or Advantage and Application: SM
(DL) Recurrent Networks are prediction, other complex problems 2.
specifically used for Drawback of NN is autoencoder
real-life dynamic problems (adjusting of weights) which requires
2. Learning nonlinear frequent setup, for all intermediate
models layers but advantage of DL is, to
(continued)
Table 1 (continued)
Analysis
[14]
3. Drawbacks of NN are memorize single node which captures
overcome in DL the information of almost all of the
intermediate nodes. Hence, it’s always
good to innovate on existing NN
models rather than using them as it is
[19]
9. Random Forest 1. When data is limited for Advantages: Unlike DT, no overfitting
(RF) classification/decision issue, hence pruning RF is not always
problems necessary. Time complexity of
2. RF handles different Hybrid RF is less. Algorithms are fast
types of features like but with some exceptions in case of
numerical, binary, or noise. Parallel RF processing makes it
categorical without any run faster on distributed databases as
transformation. It focuses accuracy is much more compared to DT
on the important variables since no over fitting of data, works well
estimation for classification in case of missing data, resistant to
purpose by doing implicit outliers, data scientists save data
feature selection preparation time, input data preparation
3. Unlike Random walk, not needed
RF for regression tasks Disadvantage: Analysis of RF becomes
does not predict beyond the difficult if large number of DT are
given range of target values included which takes lot of time to
in training dataset [14]. process the data, performance decreases
Mathematical Random in real-time predictions. If the data
Walk equation is as shown: consists of multilevel categorical
R(t + 1) = R(t) + e variables, then, there are chances of
4. Hybrid RF with algorithm bias in favor of those
clustering and for feature multilevel attributes which in turn
selection works well for increases time complexity compared to
stock predictions. It is fast, DT.
requires nearly no Applications: All sorts of predictions in
parameter tweaking to get banking sectors, the automobile
some results, very powerful industry, health care industry used for
technique that performs regression tasks like prediction of
well in hybrid models average number of social media shares
combined with other ML and likes. Pattern matching in speech
techniques recognition, classification of images and
texts [14]
10. Feature 1. Variation of FS is Advantages: Features are added/
Selection (FS) (chi-squared feature removed recursively
selection) Disadvantages: It does not give
2. Helps in model size accuracy benefits in discriminative
reduction models
11. Ensemble 1. EML (Tree Ensembles Advantages: EML are accurate than
Machine learning including RF and Gradient single ML algorithms. EML are
Boosted DT) is nothing but combination of different predictive
(continued)
Table 1 (continued)
Analysis
[14]
Techniques correct combinations of all models which reduces error variance
(EML) predictions blended with hence accuracy of automated system
statistics gives good results improves [21]
[3, 20] Applications: EML are used in variety
2. For time series of ML problems, such as feature
prediction, algorithm like a selection, confidence estimation,
NN, SVM, or DT are used extraction of missing features,
in ensemble ML [20] incremental reinforcement learning,
3. Unlike traditional handling imbalanced data [3]
regression approach,
provides with a way to take
multiple ML algorithms,
models, and combine their
predictions. Classifiers that
make them up are complex,
nonlinear, multivariate
problems like stock
predictions [3]
12. Naive Bayes Based on the Bayes Advantages: NB classifier converges
(NB) theorem faster [17]
1. NB works effectively Applications: To classify a news
with categorical input article, sentiment analysis, document
variables categorization, used for classifying
2. Requires relatively little news articles, Spam email filtering
training data than other
discriminative models like
logistic regression,
provided NB conditional
independence assumption
holds
3. Good for multi-class
predictions
13. Conditional 1. CRF are of 2 Types: Advantages: Extremely helpful when
Random Fields Markov random fields, dealing with graphical models. Because
(CRF) Maximum Entropy Markov of conditional nature of CRF,
Models independent assumptions by hidden
2. CRF are used in models markov model (HMM) are not required
where we have to predict to be done
multiple unknowns and Disadvantage: CRF avoids label bias
there exists some special problem which is a drawback of
structure of dependencies Maximum entropy Morkov models
between the unknowns (MEMMs)
Applications: Bio-informatics,
computational linguistics and speech
recognition
(continued)
Table 1 (continued)
Analysis
[14]
14. Regression RT Are of two types: Advantages: Adaptive, complexity can
trees(RT) Multivariate RT, be increased
Multivariate Adaptive Applications: Regression is most
Regression Splines in widely used in stock market prediction
which multiple random
variables can be used as
parameters
2. RT are used to predict a
score rather than a category
15. A/B testing Removes some of the Applications: 95% of the stuff done in
and data analytics issues with confounding the industry with big data is A/B testing
variables and allows and analytics
separation of correlation
16. Apriori ML A priori ML makes use of Advantages: Easy to implement, can be
Algorithm large item set properties processed in parallel
Applications: Detecting adverse drug
reactions, used for association analysis,
Market basket analysis, Google
autocomplete forms
17. Reservoir useful in a number of Advantages: A variation of this
sampling applications that require us algorithm, a weighted distributed
(Weighted to analyze very large data reservoir sample, where every item in
Distributed sets the set has an associated weight, with
Reservoir sample data such that the probability of
sampling) item selection is proportional to it’s
weight. Hence, number of passes are
reduced over the i/p data
Applications: CloudEraML [22] has a
sample command that can be used to
create stratified samples for text files
and Hive tables (the HCatalog interface
to the Hive Metastore)
2.2 The Challenge
The problem here is that prediction of price movement is not enough for successful
application in trading. ML prediction model tells us the move will be +1% or −1%
tomorrow. Because this 1% might mean −10% with some probability or +20% with
some other much smaller probability. SPX is always going up almost all the time
for long-term investments, but every 1% of precision improvement in prediction is
very difficult to predict. Hence, instead of focusing on the stock prices, focusing on
a probability distribution, is a better solution.
3 Conclusion
ML algorithms form the basis for big data analysis related complex problems like
SM prediction. In this paper, various technical indicators, challenges in SM pre-
diction, classification types, analysis of ML algorithms as shown in Table 1 are
discussed. The analysis part can be used to design EML which is a correct com-
bination of all predictive models in order to make a good decision better. This forms
the basis for subsequent research. Further, research stages would be the design of
the functional prototype, the design of EML algorithms, comparative analysis of
EML algorithms on the common benchmark platform based on the certain technical
indicators and performance evaluation, which can be done in near future.
References
1. www.StocksNeural.net.
2. www.vatsals.com.
3. Robi Polikar, “Ensemble Learning, Methods and applications”, Date: 19 January 2012,
Springer Verlag.
4. Machine learning algorithms, http://www-formal.stanford.edu/jmc/whatisai/node2.html.
5. Robert Nau, http:blog.optiontradingpedia.com, besttradingplatformfordaytraders.blogspot.com.
6. Rupinder kaur, Ms. Vidhu Kiran, “Time Series based Accuracy Stock Market Forecasting
using Artificial Neural Network”, IJARCCE. doi:https://doi.org/10.17148/IJARCCE, 2015.
7. AI Stock Market Forum: http://www.ai-stockmarketforum.com/.
8. How to Value Stocks using DCF and the Dangers of Doing, Do Fundamentals Really Drive
The stock market? www.mckinsey.com, A Mckinsey report.
9. Ben McClure, Discounted Cash Flow Analysis, Investopedia, Investopedia.com, 2010 http://
www.investopedia.com/university/dcf/.
10. https://www.udacity.com/course/machine-learning-for-trading-ud501.
11. http://www.angoss.com, Key Performance Indicators Six Sigma and Data Mining.pdf, 2011.
12. Forecasting home page (Introduction to ARIMA models), price Prediction Using the ARIMA
Model-IEEE Xplore, UKSim-AMSS, 16th ICCMS, 2014.
13. www.kdnuggets.com.
14. www.dezyre.com.
15. Paliouras, Georgios, Karkaletsis, Vangelis, Spyropoulos, Machine Learning and Its
Applications: advanced lectures, Springer-Verlag New York, Inc. New York, NY, USA
2001, table of contents ISBN-978-3-540-44673-6.
16. Alexandra L’Heureux; Katarina Grolinger, Hany F. El Yamany; Miriam Capretz, Machine
Learning with Big Data: Challenges and Approaches, IEEE Access, 2017, doi:https://doi.org/
10.1109/ACCESS.2017.2696365.
17. J. Huang, J. Lu and C. X. Ling, “Comparing naive Bayes, decision trees, and SVM with AUC
and accuracy,” Third IEEE International Conference on Data Mining, 2003, pp. 553–556. doi:
https://doi.org/10.1109/ICDM.2003.1250975.
18. Neelima Budhani, Dr. C.K. Jha, Sandeep K. Budhani, Application of Neural Network in
Analysis of Stock Market Prediction, www.ijcset.com, IJCSET, 2012.
19. Maryam M. Najafabadi, Flavio Villanustre, Taghi M. Khoshgoftaar, Naeem Seliya, Randall
Wald and Edin Muharemagic, Deep learning applications and challenges in big data analytics,
Journal of Big Data, SpringerOpen Journal, 2015.
20. L. K. Hansen and P. Salamon, Neural network ensembles, IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 12, no. 10, pp. 9931001, 1990.
21. W. N. Street and Y. Kim, A streaming ensemble algorithm (SEA) for large-scale
classification, Seventh ACM SIGKDD International Conference on Knowledge Discovery
Data Mining (KDD-01), pp. 377382, 2001.
22. blog.cloudera.com.
Stability of Local Information-Based
Centrality Measurements Under Degree
Preserving Randomizations
Chandni Saxena, M. N. Doja and Tanvir Ahmad
Abstract Node centrality is one of the integral measures in network analysis with
wide range of applications from socioeconomic to personalized recommendation.
We argue that an effective centrality measure should undertake stability even
under information loss or noise introduced in the network. With six local
information-based centrality metric, we investigate the effect of varying assorta-
tivity while keeping degree distribution unchanged, using networks with scale free
and exponential degree distribution. This model provides a novel scope to analyze
the stability of centrality metric which can further find many applications in social
science, biology, information science, community detection and so on.
Keywords Centrality Local information Stability Assortativity

Degree distribution
1 Introduction
The relationship among nodes in the network has varied meaning and analogy.
These relations are observed in different domains and fields, such as; the relationship
among neurons in the neural network, among web-pages in web graph, among
online users in social relationship graph, and among authors in scientific
C. Saxena (&) M. N. Doja T. Ahmad

Department of Computer Engineering, Jamia Millia Islamia, New Delhi, India
e-mail: cmooncs@gmail.com
M. N. Doja
e-mail: ndoja@yahoo.com
T. Ahmad
e-mail: tahmad2@jmi.ac.in

https://doi.org/10.1007/978-981-10-7245-1_39
396 C. Saxena et al.
collaborations network. Complex networks have been explored to study these

relations at the microscopic level of elements like nodes and their relations as links.
Structural centralities are such measures which uncover the explicit roles played by
the nodes and the links. The concept of centrality is a characterization of a node,
subject to its importance according to the structural information of the network.
Centrality has become an important measurement with vast implications in theo-
retical research such as physical science [1] or biological science [2], and practical
significance in applications, such as e-commerce [3] and social networks [4].
However, the node centrality concept is vast and can be furnished in various ways,
putting forward multiple coexisting centrality measures. Among various aspects,
centrality based on the local information is of greater importance due to its low
computation complexity and simplicity. Also, in the evident scenario of real com-
plex networks having incomplete information or loss of information and dynami-
cally changing the topological behavior of the network, local information-based
centralities provide best solutions. The important issue here is to have the ability of
this metric to be stable and robust to noise in the network or to any randomization,
such created under real situations. It has been proven that different network
topologies, such as scale free, exponential, and heavy-tailed degree distribution [5]
can affect network-based operations, such as robustness of network [6] and epidemic
spreading [7]. So far the issue of the topological structure has not been investigated
for different centralities aiming its stability issue. This has motivated us to explore
the stability of local information-based centrality with different topologies of
the networks. We studied these centralities under scale-free and exponential distri-
bution of networks and examined the effect of network randomization (perturbation)
when the degree of the nodes was kept preserved. To examine the stability problem
in local information-based centrality, we have analyzed six centrality measurements
namely, H-index [8] (h), leverage centrality [9] (lc), local structural entropy [10]
(lse), local clustering coefficient [11] (lcc), topological coefficient [12] (tc), and local
average connectivity [13] (lac). We provide a framework to categorize these metrics
in a novel dimension. In our experiments, we vary assortativity of the network
leaving the crucial property of degree distribution unchanged according to noise
model 1. We evaluate these metrics on the basis of average rank difference and
standard deviation of average rank difference for original and randomized networks.
We also compare overlaps of sets of top-ranked nodes, varying range of perturbation
with the original network as explained in noise model 2 defined by Yang [21]. We
evaluate this model empirically using real and synthetic networks under scale free
and exponential degree distribution.
The rest of the paper is organized as follows. Section 2 presents a framework of
the stability issues covering theoretical details and related works in this direction. In
Sect. 3 results and experiments evaluate the different stability performance metric
on different datasets, and finally, the paper is concluded in Sect. 4.
Stability of Local Information-Based Centrality Measurements … 397
2 Background
2.1 Local Information-Based Centrality Measure
There have been dozens of centralities based on the global and local topological
information of a network. A local network around each node in terms of its con-
nectedness, topology, and type form a network which can formulate its influence.
We have examined six local information-based measures, which require only a few
pieces of information around the node, defined in this section. H-index of a node vi
is the largest value h, such that vi has at least h neighbors of degree no less than
h. Leverage centrality finds nodes in the network which are connected to more
nodes than their neighbors and determines central nodes. The local structural
entropy of a node is based on Shannon entropy of local structure which depends on
the degree of its neighboring nodes in the local network of the target node. The
local clustering coefficient of a node is the ratio of a number of links among its
neighbors to possible numbers of links that could formulate. The topological
coefficient of a node is the sum ratio of a number of neighbors common to a pair of
nodes among directly linked neighbors to the number neighboring nodes of target
node. Local average connectivity of a node is the sum of local average connectivity
of its neighbors in the subnetwork induced by the target node.
2.2 Degree Preserving Randomization
Rewiring network randomly while keeping its degree distribution, and hence degree
of each node constant is to enact real-time perturbation of the network that mapped
incompleteness or state of loss of information. As local information-based centrality
measurements are degree dependent, therefore, preserving degree of each node
while perturbation is an essential step. Primarily, this leads to change in mixing
pattern of a network, and hence changing the assortativity of a node. A pair of edges
are picked randomly and rewired as illustrated in Fig. 1, ensuring degree preserving
randomization, e.g., edges AC and BD are picked as candidates, now the con-
nections between end points are exchanged to form new edges AD and BC. To
study the effect of randomization on centrality stability of the network nodes, we
use noise model 1 and noise model 2 on six different datasets in this paper. These
datasets differ both in size and subject to type as the scale free and the random
exponential networks as shown in Table 1. The facebook dataset consist of nodes as
people with edges as friendship ties extracted from social network facebook. The
brain-network is medulla neuron network of fly-drosophilla. The foodweb is the
food web network data of Little Rock Lake, Wisconsin USA. The c. elegans-
metabolic is a metabolic network data. Synthetic scale free and synthetic expo-
nential distribution datasets are also considered for the validation of real data. The
basic topological features of the networks are listed in Table 1.
Fig. 1 Network perturbation

preserving degree of nodes
Table 1 Topological statistics of datasets: number of nodes (n), number of edges (m), maximum
degree (kmax), best estimate of the degree exponent of the degree distribution (c), assortativity of
unperturbed network (r), and assortativity of perturbed network (rp)
Network n m kmax c (r) (rp)
Scale-free
Synthetic 2000 5742 1378 2.1 −0.4157 −0.4179
Facebook 6621 249,959 840 2.3 0.1225 0.1149
Brain 1781 33,641 16,224 1.9 −0.3235 −0.3252
Exponential
Synthetic 1000 9945 495 – −0.1252 −0.1174
C. elegans 453 4596 644 – −0.0625 −0.2276
FoodWeb 183 2494 108 – −0.2374 −0.1585
2.3 Related Work
There are a few numbers of studies on finding stability of pagerank algorithm under
randomization [14–17]. Goshal and Barabasi [18] have proposed stability criteria for
scale free and exponential network when degree preserving random perturbation and
incompleteness in network are observed. Authors have studied the ranking stability
of top nodes and manifested the initial success of pagerank algorithm to scale-free
nature of the underlying WWW topology. Lempel and Moran [14] have examined
the stability of three ranking algorithms—pagerank, HITS, and SALSA—to
perturbation, while Senanayake et al. [16] have studied the page rank algorithm
performance with underlying network topology. Andrew et al. [17] have analyzed
the stability of HITS and PageRank to small perturbations. Sarkar et al. [19]
investigated community scoring and centrality metric in terms of different noise
levels. So far no work has been observed carrying stability study of local
information-based centralities.
2.4 Stability
To examine the stability of centrality measurements we furnish the data with degree
preserving randomization according to two noise models as explained. Noise
Model 1: We implement number of perturbations in the network of size more than
2000 nodes equal to the size of network and for other networks it is 10 times the size of
the network. The reason lies on the fact to ensure the range of perturbation depending
upon the size of network and to realize meaningful change in assortativity of the
network. Where assortativity is defined as a measure that evaluate the tendency of a
node to be connected to similar nodes in the network. Figure 2 reports the distribution
of ranking scores for perturbed and unperturbed networks for three centrality mea-
sures—h-index, leverage centrality, and local structure entropy—on synthetic
scale-free and exponential network data, real scale-free facebook, and exponential
c. elegans network data. The h-index confronts the most concentrated distribution as
compared with other two measures, which are less concentrated. This indicates that
when data is exposed to degree preserving perturbations, the node ranking may also
change and stability problem exists to diffscore of various centrality indices,erent
ranges. To characterize the stability of centrality measure, we define three metrics:
Degree preserving randomized network vs True Network
(a) (d) (g) (j)

(H-index)
(b) (e) (h) (k)

( LC )
( ENTROPY)
(c) (f) (i) (l)
Fig. 2 Effect of degree preserving randomizations on h-index, leverage centrality and local
structure entropy. Synthetic Scale free and exponential distribution also real facebook and
c. elegans networks are examined. Figure shows the scatter plot of ranks in both original network
and degree preserved randomized network for the same. Synthetic scale-free network (a), (b), (c),
Facebook (d), (e), (f), C. elegans (g), (h), (i), and synthetic exponential (j), (k), (l)
Mean Bias, µ calculates the average of rank difference of same node from
perturbed and unperturbed network for some centrality measurement and is defined
as:
P
xy dxy
l¼ ; ð1Þ
n ð n 1Þ
where n is the size of the network and dxy is the difference between true rank of a
node calculated from centrality measure on original network (x) and randomized
network (y) as shown in Fig. 2. Lower the values of µ, stable the centrality of nodes
would be.
Standard Deviation of Mean Bias, r measures the susceptibility of ranks
against the change in data occurred due to randomization. It is defined as:
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sP 2ffi
xy d xy l
r¼ ð2Þ
n ð n 1Þ
High value of r indicates that ranks of a few nodes are quite unstable and low
value indicates the similar unstable level of ranks for all nodes due to randomization
or unstable centrality.
Jaccard Similarity Index [20], measures the overlap between two rank vectors.
In order to check how top ranking nodes change under different proportion of
perturbation, we compare top 25 nodes according to rank given by each centrality
measures for different networks. Highest value (1) indicates that the set of these
nodes is not changed and lowest value (0) indicates that it has changed totally due
to randomization offered to the network when compared with original vector of
rank for top nodes.
3 Experiments and Results
For experimentation, we have two base topological configurations for network;

scale free and exponential and generated synthetic data for the same also considered
two real networks for each configuration as mentioned in the Table 1. We perform
degree preserving randomization on empirical datasets. For each mentioned cen-
trality measure, we calculate the ranking vectors for perturbed and unperturbed
networks. The results of mean bias µ, standard deviation r of mean bias are shown
in Fig. 3. It is well noticeable from the histograms for inverse mean bias and inverse
standard deviation of bias score of various centrality indices, that overall h-index
outperforms other centrality measurements for each topological configuration.
However, all other centrality measures show up higher stability for the scale-free
topology of underlying network. Apparently higher values for centrality stability
score of scale-free configuration suggests the role of network characteristic in
determining its performance under mentioned stability metrics. This fact is also
supported by the work of Ghoshal and Barabasi [18], that the accomplishment
achieved by page rank algorithm in ranking the web contents is equally credited to
the scale-free characteristic of www. For investigating configuration of top ranked
nodes in randomized networks according to different centrality metrics we use noise
model 2 as introduced by Yang [21]. Jaccard overlap of top rank nodes for all
centrality measures on different datasets are evaluated with varying assortativity of
the network as shown in Fig. 4. In majority of the cases, h-index performs better
than other metrics retaining top ranked node under varied level of randomization. It
can be realized from the experimental evaluations that the scale-free networks are
more stable for local information-based centrality stability performance under
perturbation of network and h-index is the most stable metric when compared with
the present set of centrality metrics considered for evaluation.
Fig. 3 Inverse of mean bias (1/µ) and inverse of standard deviation of mean bias (1/r) for
different centrality measures under different topology networks
Fig. 4 The Jaccard similarity index between the top vertices of the original and the perturbed
networks for varying levels of assortativity changes with increasing number (degree) of
perturbations (noise model 2)
4 Conclusion
In this work, we investigate how changing assortativity affect different centrality

metrics under given topologies of the network. We find that h-index centrality
outperforms other benchmark centrality metrics based on the local information in
the network, when network is perturbed keeping its degree distribution constant
according to noise model 1. To further explore stability notion, we use Jaccard
similarity index for different centrality rank vectors of top nodes when network is
randomized according to noise model 2 with original top nodes vector. We find that
the top rank nodes also called stable nodes are more prevalent to networks with
scale-free degree distribution and h-index shows high value for Jaccard overlap on
majority of datasets. Our method introduces contemporary measures to analyze and
evaluate stability of different centrality metrics which can be further explored to
investigate other important networks parameters. This work can be enhanced by
studying behavior of network with other properties like transitivity, also different
noise models to effect networks could be studied in this regard.
References
1. Guimera, R., Mossa, S., Turtschi, A., & Amaral, L. N.: The worldwide air transportation
network: Anomalous centrality, community structure, and cities’ global roles. Proceedings of
the National Academy of Sciences, 102(22), 7794–7799 (2005).
2. Jeong, H., Mason, S. P., Barabási, A. L., & Oltvai, Z. N.: Lethality and centrality in protein
networks. Nature, 411(6833), 41–42 (2001).
3. Lamberti, F., Sanna, A., & Demartini, C.: A relation-based page rank algorithm for semantic
web search engines. IEEE Transactions on Knowledge and Data Engineering, 21(1), 123–
136 (2009).
4. Malliaros, F. D., Rossi, M. E. G., & Vazirgiannis, M.: Locating influential nodes in complex
networks. Scientific reports, 6 (2016).
5. Strogatz, S. H.: Exploring complex networks. Nature, 410(6825), 268–276 (2001).
6. Borge-Holthoefer, J., Rivero, A., García, I., Cauhé, E., Ferrer, A., Ferrer, D., & Sanz, F.:
Structural and dynamical patterns on online social networks: the spanish may 15th movement
as a case study. PloS one, 6(8), e23883 (2011).
7. Barthelemy, M., Barrat, A., & Vespignani, A.: Dynamical processes on complex networks
(2008).
8. Hirsch, J. E.: An index to quantify an individual’s scientific research output. Proceedings of
the National academy of Sciences of the United States of America, 16569–16572 (2005).
9. Joyce, K. E., Laurienti, P. J., Burdette, J. H., & Hayasaka, S.: A new measure of centrality for
brain networks. PLoS One, 5(8), e12200 (2010).
10. Q. Zhang, M. Li, Y. Du, Y. Deng, “Local structure entropy of complex networks”, arXiv
preprint arXiv:1412.3910. 2014 December.
11. Zhang, Q., Li, M., Du, Y., & Deng, Y. Local structure entropy of complex networks. arXiv
preprint arXiv:1412.3910 (2014).
12. Xu, J., & Li, Y.: Discovering disease-genes by topological features in human protein–protein
interaction network. Bioinformatics, 22(22), 2800–2805(2006).
13. Li, M., Wang, J., Chen, X., Wang, H., & Pan, Y.: A local average connectivity-based method
for identifying essential proteins from the network level. Computational Biology and
Chemistry, 35(3), 143–150 (2011).
14. Lempel, R., & Moran, S.: Rank-stability and rank-similarity of link-based web ranking
algorithms in authority-connected graphs. Information Retrieval, 8(2), 245–264 (2005).
15. de Kerchove, C., Ninove, L., & Van Dooren, P.: Maximizing PageRank via outlinks. Linear
Algebra and its Applications, 429(5–6), 1254–1276 (2008).
16. Senanayake, U., Szot, P., Piraveenan, M., & Kasthurirathna, D.: The performance of page
rank algorithm under degree preserving perturbations. In Foundations of Computational
Intelligence (FOCI), 2014 IEEE Symposium on (pp. 24–29). IEEE (2014, December).
17. Ng, A. Y., Zheng, A. X., & Jordan, M. I.: Link analysis, eigenvectors and stability.
In International Joint Conference on Artificial Intelligence (Vol. 17, No. 1, pp. 903–910).
LAWRENCE ERLBAUM ASSOCIATES LTD (2001, August).
18. Ghoshal, G., & Barabási, A. L.: Ranking stability and super-stable nodes in complex
networks. Nature Communications, 2, 394 (2011).
19. Sarkar, S., Kumar, S., Bhowmick, S., & Mukherjee, A.: Sensitivity and reliability in
incomplete networks: Centrality metrics to community scoring functions. In Advances in
Social Networks Analysis and Mining (ASONAM), 2016 IEEE/ACM International Conference
on (pp. 69–72). IEEE (2016, August).
20. Gower, J. C.: Similarity, dissimilarity, and distance measure. Encyclopedia of Biostatistics
(2005).
21. Yang, R.: Adjusting assortativity in complex networks. In Proceedings of the 2014 ACM
Southeast Regional Conference (p. 2). ACM (2014, March).
Hybrid Solution for E-Toll Payment
Ajinkya R. Algonda, Rewati R. Sonar and Saranga N. Bhutada
Abstract E-toll system has been huge improvement in decreasing the over traffic
jams that have become a big problem in metro cities nowadays. It is the best method
to handle the huge traffic. The traveler passing through the traditional type of
transport have to pay toll bill by waiting in the long queue, so it results in loss of
petrol, loss of time, pollution, and tension of carrying cash with us. So by using our
system of E-Toll payment the traveler revokes the tension of waiting in the queue to
make the payment, which decreases the fuel consumption and taking cash with
them can be avoided.
Keywords RFID FASTag OCR OBU IU AVC AVI

ETC CCH Gantry
1 Introduction
Nowadays the amount of traffic has been increasing due to increase in vehicles.
Everybody prefers their own vehicles as compare to public transport. It increases
traffic and the solution over this is wide roads, hence tooling system is used to collect
tolls. The traditional toll collection system is like vehicles come at toll gate and owner
of that corresponding vehicle has to pay money according to the toll charged onto his
vehicle. This process of toll payment is so time-consuming because manual pro-
cessing requires more time often leads to delay and customer has to wait. So one-step
toward making India a digital India is making traditional toll payment system as a
E-toll payment system, in which a user has to buy FASTag just at Rs. 250 only from
A. R. Algonda (&) R. R. Sonar S. N. Bhutada

Department of Information Technology, MIT College of Engineering, Pune, India
e-mail: ajinkyaalgonda449@gmail.com
R. R. Sonar
e-mail: revasonar9@gmail.com
S. N. Bhutada
e-mail: saranga.bhadade@mitcoe.edu.in

https://doi.org/10.1007/978-981-10-7245-1_40
406 A. R. Algonda et al.
Axis or from ICICI bank and he has to recharge it for payment for toll. The FASTag
will work as a prepaid card. The amount of toll will be debited from the account of
owner of the respective vehicle. If any vehicle is not having FASTag then he can be
identified by its registered vehicle number which can be sensed by OCR and for users
registered nothing, bill will be delivered to his postal address with validity of
80 days. These approach increases speed of toll collection process.
2 Related Work
1. South Korea. Motorway tolls can be paid using cash, major credit cards, or a
Hi-Pass card [1]. The Hi-Pass system permits owner to pay tolls without any
interference at the tollbooth. To use it, owner must have both device name as
OBU which is installed on the windscreen of the vehicle and a Hi-Pass card
which is inserted into the OBU. The OBU, sometimes marketed as a “Hi-Pass
device” or ETC unit (Electronic Toll Collection), is produced by a variety of
manufacturers. It can be purchased from any store which sells automobile
electronics. A Hi-Pass card can be topped up in advance (Hi-Pass Plus card) or
used in conjunction with credit cards issued specifically for the toll system. The
card can often be found in stores that sell the OBU.
2. Italy. Telepass is the name of the system used to collect the toll on a vehicle in
Italy. The Telepass is used for all the vehicles which are traveling on Italian
roadways. Telepass comprise a device named as OBU which is placed on the
windscreen. The OBU is operated on power supply like battery [2]. The com-
munication between OBUs and the E-Toll booths can be done through dedicated
short-range communication. Telepass can be used in the opened and the closed
system on the roadways in both system, i.e., the open or closed system the toll
which will be charged on the vehicles depends on the type of vehicles (car, bus,
auto, etc.). The speed of the vehicle of Telepass users should not exceed 30 km/h
Fig. 1 Telepass in Italy [2]

Hybrid Solution for E-Toll Payment 407
during it is in Telepass lane when OBU has been checked then the OBU gives a
high beep and barricade lifts. When Telepass users leave the toll root, the OBU
gives another a high beep. Then a number plate of the vehicle is captured and the
vehicle is allowed to pass through the tollbooth. OCR detects vehicle’s number
plate and the user of the respective vehicle gets a bill statement (Fig. 1).
3. Canada. The Electronic toll Collection (ETC) is used in Canada is known as
Canada 407 Express toll route [3]. There are no toll gate hence the name Express
toll route (ETR). This system uses Transponder and Cameras to toll collection
automatically. It is primitive example of Highway that exclusively uses Open
road toll collection. A radio antenna recognized when vehicle having transponder
has arrived and leaved the road and calculates total fare. For vehicle without
having transponder, OCR is used for number plate reorganization and the toll fare
sent to user. A small electronic “transponder” is placed on the windscreen of the
vehicle. The system tests transactions automatically at arrival and end of the toll
booths for supporting a distance-based tolling system. The sensors placed on
upper side of gantry log the 407 ETR arrival and exit point. On successful
transaction, a system beeps at four times and indicates green light. The optical
character recognition (OCR) uses for capture a no. plate of vehicle without having
transponder. The statement will be send to enrolled user’s address (Fig. 2).
4. Singapore. The ERP (Electronic Road Pricing) is a E-Toll collection system to
manage traffic. Singapore is the first city in all over the world who have
implemented electronic tolling system to address road blockage [4]. The cars in
Singapore city are having a special gadget placed inside the car which is called
as the IU (In Vehicle unit) gadget. This gadget is generally placed at the
windscreen of the vehicle. A card is inserted into the IU called as cash card or
value card. Each and every IU has its own unique ID for every vehicle (Fig. 3).
Electronic gantry. The gantry recognizes the vehicle with its unique IU. When
a vehicle having the unique IU passes through the gantry then the gantry
Fig. 2 Canada 407 ETR for ETC [2]

Fig. 3 IU gadget
Fig. 4 Electronic gantry in Singapore
recognizes the vehicle and its IU and debit amount of the appropriate toll for the
associated vehicle. Sensors placed over the gantry communicate with IU
through the short-range communication system and the deducted amount is
shown on the LCD screen on the IU. The gantry is electrically controlled and it
will generate the amount of the toll as per time of day thus effectively adjusting
and moderating the traffic (Fig. 4).
5. United States. E-ZPass is an E-Toll collection system used on most of the toll
roads, bridges, tunnels in the Midwestern and the northwestern US [5]. The
E-ZPass communicates with the reader equipments built into the lane of toll-
ways by transmitting a unique radio signals. The E-ZPass is mounted on the
vehicle’s windscreen of the vehicle. Some of the vehicles has the windscreen by
which the radio signal gets blocked; for such type of vehicles or for the his-
torical type of vehicles, an external tag is provided, which is designed to attach
Fig. 5 E-ZPass in US
the vehicle’s front license plate. The E-ZPass lanes having a low-speed limit of
8–24 km\h so that it would be easy to sense the E-ZPass on the vehicle (Fig. 5).
Each E-ZPass tag is programmed for a specific type of the vehicle, so that the toll
charges for each vehicle is also different from others. It is possible to charge
wrong toll amount by any vehicle by interchanging the tag over it; means if the
tags are programmed vehicle’s class does not match the vehicle. This will be
considered as a violation of E-Toll payment rules and fine according to the toll can
be charged to the user having a tag on their vehicle. To eliminate this, the com-
mercial vehicle uses E-ZPass, which are blue in color and the white tag is given
for standard passenger’s vehicles. These E-ZPass are also used by government
servant for their vehicles. In New York, the E-ZPass tags which are orange in
color are given to emergency vehicles or to the employee of the metropolitan
Transportation authority, an employee of Port authority, employees of New Jersey
and New York State Thruway Authority. All these agencies are interconnected to
each other for exchanging the tag data and for processing toll transactions.
3 Proposed Idea
All above-mentioned countries have been adopted same technology for E-toll
collection. But in India, NHAI ties up with Axis and ICICI bank for E-toll col-
lection. They use FASTag which is easily available in these banks at very low cost.
The FASTag is available just at Rs. 200 only. But the thing is that the Axis and
ICICI bank are the private sector banks, so each and every person cannot afford
taxes charged by these kinds of banks. So, we can propose our idea with the help of
all other nationalized as well as private sector banks. So that each and every person
can be able to get the benefit of the E-toll payment system.
4 How Will ETC Work in India
Traffic jam is a headache at the tollways. Sometimes, it occurs due to lack of

change, sometimes due to heavy vehicles and sometimes while collecting infor-
mation about the vehicles. As a solution over it, our system will provide separate
lanes for the E-Toll payment users, so that the traffic can be managed very easily.
The FASTag E-toll payment system has been adopted by India over national
highways only, but our proposed idea will make it for all the tollways in India.
This idea is very suitable for the Indian system. There are 350 toll plazas in all
over the India. NHAI is making cashless payment mechanism (FASTag) working at
123 toll plazas. NHAI will give 10% cash back on toll payments for FASTag users.
The cash back amount for a particular month will be credited into the FASTag
account at the start of the next month [6].
Ministry of Road Transport and Highways has decided to make all toll plazas on
national highways across India ETC enabled. As we know, generally people keep
their vehicle on while waiting in toll payment queue. This results in more fuel
consumption. If E-toll payment comes into the picture, then this will save fuel
around Rs. 86,000 crore [7].
5 Strategy Used for Govt. Vehicles and Ambulance
In India, some of the vehicles do not have any kind of toll on all tollways. Like
Govt. vehicles and the ambulance. So we can do one thing for avoiding the
ambiguity regarding this is, we will register these vehicles as noncharged vehicles
in the database. Means whenever the Govt. vehicle or the ambulance will arrive at
toll gate the system will not charge any kind of toll to the corresponding vehicle and
these type of vehicles are allowed to go without waiting in a queue.
6 How the Payment Can Be Made
(1) For users registered with FASTag

• The FASTag can be recharged from Rs. 100–100,000.
• So when the FASTag will be sensed by the sensors present at the tollways
the amount charged by the particular vehicle will be deducted from the bank
account linked to that particular FASTag user.
• We can promote the BHIM app for payment purpose.
(2) For users registered with number plate
• For non-FASTag users, the OCRs will be used.
• These OCRs will sense the number plate of the vehicle and through the
number plate information of the particular owner will be taken and the toll
will be deducted from the registered user of the given number plate.
(3) For users registered with nothing
• For these types of users, the bill will be delivered to his postal address with
the validity of 80 days from his journey.
• If the particular user will not be able to pay his toll till due date, he will be
charged some fine according to the rules.
7 Violation Handling
Vehicles which are Stolen, vehicles with a class mismatch, vehicles with insuffi-
cient balance and those that run through without payment are blacklisted. List of
blacklisted vehicles is shared with all toll plazas. These vehicles are stopped when
they cross the next toll plaza and appropriate actions are taken.
8 Implementation Procedure of the Idea
1. Start
2. User with registered FASTag arrives at the toll gate, FASTag will be sensed and
the toll will be calculated. Go to step 5.
3. User with registered number plate arrives at the toll gate, the info of the owner
will be fetched and the toll will be calculated. Go to step 5.
4. User arrives at the toll gate who registered nothing, the toll will be calculated
and send to his postal address.
5. Toll paid
6. End.
Yes
N
9 Documents Needed to Apply for FASTag
• Customer will need to carry original as well as copy of your KYC document.
• Customer needs to submit the following documents along with the application
for FASTag:
• Registration Certificate (RC) of the vehicle.
• Passport size photograph of the vehicle owner.
• KYC documents as per the category of the vehicle owner.
10 Procedure of Getting the FASTag
• Customer can avail ICICI Bank FASTag just by sending an SMS and their
executive will call customer within 2 working days.
• SMS—ITOLL(Space)Pincode(Space)Name to 5676766.
• E.g., ITOLL 400051 Roshan Kumar.
• Alternatively, customer can call them at their helpline number 1860-267-0104.
11 Advantages
• By adopting this system, we can take a step toward digital India.

• The corruption can be avoided.
• Using FASTag will manage the traffic over the tollways.
• The problem of change would not make any trouble.
• Time of waiting in a queue at tollway can be saved.
• Saves the fuel wastage at the toll gate.
• Man power reduced at the toll gate.
• Highly secured and reliable.
12 Conclusion
In this system, we have constructed the E-toll payment system with the help of
FASTag to offer an efficient way to pay the tolls to enhance the user service quality.
It enables the system to classify vehicles for toll payment according to the type of
payment method.
The proposed system will clear the problem of traffic jams, lack of change at the
money counters at tollways, etc. E-toll users can thus appreciate the high quality of
service provided which in turns highly promotes Digital India.
References
1. Evaluations and improvements of the Korean highway electronic toll system M.-S. Chang,
K.-W. Kang, Y.-T. Oh, H.-W. Jung.
2. Study of Different Electronic toll Collection Systems and Proposed toll Snapping and
Processing System. Apurva Hemant Kulkarni. 2014.
3. KHADIJAH KAMARULAZIZI, DR. WIDID ISMAIL electronic toll collection system using
RFID technologies Journal of Theoretical and Applied information technology © 2005–2010
JATIT & LLS.
4. Electronic Road Pricing: Experience & Lessons from Singapore. Prof. Gopinath Menon
Dr. Sarath Guttikunda 2010.
5. Electronic toll collection syetm using passive RFID technology, Khadijah Kamarulazizi,
Dr. Widad ismail.
6. Cashless Payment, http://economictimes.indiatimes.com/news/economy/infrastructure/e-toll-
starts-from-monday-now-make-cashless-payments-at-275-toll-plazas/articleshow/51909252.cms.
7. How Indian System work in India, https://www.quora.com/How-will-electronic-toll-collection-
work-in-India.
Enhancing Distributed Three Hop Routing
Protocol in Hybrid Wireless Network
Through Data Weight-Based Scheme
Neha S. Rathod and Poonam Gupta
Abstract The reason for improvising of our concept is to involve multiple nodes in
data transmission that diminishes delay in a routing of a network. In hybrid wireless
networking, most of the existing routing performs in a linear way. In such possi-
bility sometimes a maximum number of nodes are in idle state. So for that, there is a
concept of distributed three hop routing that depreciates the burden of a load on the
network. Dividing data and load on a node is an emphatic method of transmission
and pronouncement the shortest path using a fuzzy classifier that makes the faster
routing approach. But one of the biggest flaws of the existing scheme is that it does
not perform as per exceptions due to which it has higher overhead and fails to
achieve better congestion control. This paper promoted weight-based assignment
technique by using distributed three hop routing protocol. By considering imme-
diate one hop that evolved in shortest path and analyzing a response delay on two
nodes on that basis data is fragmented by using weight-based assignment technique.
To intensify the efficiency of a routing protocol in the hybrid wireless networks a
weight-based data assignment technique is used for data allocation in distributed
routing protocol using the artistry of least delay detection to maintain less data
congestion in the network.

Keywords Response delay Data fragmentation Data assignment

Shortest path identification Three hop routing
N. S. Rathod (&) P. Gupta

Computer Engineering Department, G. H. Raisoni College of Engineering and Management,
Wagholi, Pune, India
e-mail: neharathod73@gmail.com
P. Gupta
e-mail: Poonam77gupta@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_41
416 N. S. Rathod and P. Gupta
1 Introduction
Hybrid wireless network assimilates infrastructure network and ad hoc network. It

has been proven the best substitute for the future generation wireless network will
add infrastructure network to ad hoc network. By the hybrid wireless network,
deficiency of ad hoc network and infrastructure network can be overcome, such as
weak connection, poor communication. In hybrid wireless networks, the routing
protocol is a conspicuous part of the stable designing. One of the biggest drawbacks
face by the network is, when a maximum number of nodes are in an idle state that
aims to delay in routing process as well as it increases the load on the network due
to which concept of DTR with weight-based assignment technique arises. DTR
represent as distributed three hop routing protocol in which data is fragmented at
the source node and transmitted parallelly in a distributed manner. To increase
network amplitude for high-performance applications, the system has to work on
hybrid networks [1, 2].
In a wireless network, shortest path problem is one of the basic and valuable for
network optimization obstacle and it is used for many real applications. The motive
behind the shortest path is to find out the shortest way between the two nodes in a
graph or network. The shortest path is been calculated based on the fuzzy classifier.
In the fuzzy classifier, crisp inputs are taken, which are evolved due to the result of
response delay extraction. After fuzzification, the inference engine refers to the rule
base containing fuzzy IF-THEN rules. If-then rules are been set to check the
possibilities of the shortest path with defined rules which are normalized in between
0 and 1. The final path score and the essential paths will be extracted according to
the rules of features criteria. Now based on the if-then rules, the paths which are
criticized under the lowest ranks are considered for the data transmission and
ultimately they are treated in the route, so they form the shortest path.
In networking, data sends from source to destination, the immediate nodes are
selected by the source node. For that purpose sender sending a “Hello” message to
all the nodes in the network and on the basis of route reply of the immediate node, it
Fig. 1 Analyze response

delay
Enhancing Distributed Three Hop Routing Protocol … 417
classifies the response delay. Sorting all the delays and select the merest delay
which has to be given as a first priority among all the delays (Fig. 1).
Once analyzing response delay of each node in a network, all routing delays are
been gathered then based on the weight of the routing delays ratios are being
evaluated for the data fragmentation process. Based on the response delay ratios of
the nodes, splinter ratios are being decided and then data is divided in the form of
bit heap for the communication process. By using some formulas data is segmented
these are as follows:
Total response time ðRsÞ ¼ Rd of N1 þ Rd of N2 þ Rd of N3 ð1Þ
ðRd of node NnÞ

Rd % Per Node ¼ ð2Þ
ðTotal response timeÞ
Here, Rs is a total response time, Rd is response delay and N1, N2, N3…Nn are
nodes.
Hop is one portion of the path between source and destination. Based on the
response delay ratios of the three nodes data is segmented into three parts, the three
fragments are been delivered at the receiver end. In the two hop routing protocol,
the node transmission occurs within a single cell. It takes only a single path
transmission and produces high overhead and low reliability [3]. By the distributed
three hop routing protocol can improve efficiency and reduce the Overhead. As
network consists of three hops in DTR, where first two hops are in the ad hoc
transmission mode and one hop was in the cellular transmission mode (Fig. 2).
Most of the methodologies in wireless paradigm are to participate the number of
nodes in the network. Most of those are assigning this task to the routing nodes
which can create the more havoc at the node end to decrease the routing perfor-
mance of the nodes. By keeping this issue as of priority proposed system fixed to
data weight-based assignment scheme.
Fig. 2 Three hop routing

This whole paper is segmented into five sections. Section 2 describes past work
as related work. Section 3 reveals the detailing of proposing methodology. The
evolution of our methodology is performed in Sect. 4. Finally, Sect. 5 concludes
the paper with some scope for future extension.
2 Related Work
Various numbers of methods were discovered to fasten the routing process and
involving a number of nodes in the network.
2.1 Shortest Path Identification
K. Sasikala, V. Rajamani et al. reveal about the concept of neuro-fuzzy logic which
helps network to reduce delay in the network by performing high-level data to reach
the destination in less time and increase time efficiency, network throughput by
choosing shortest path [4].
Dimitris Sacharidis, Panagiotis Bouros et al. narrates [5] that fastest route is not
considered and examine the simplest route with longer distance. By this, they solve
the address issue by the simplest route which will be shortest among all analysis.
Sujata V. Mallapur and Siddarama R Patil et al. introduce the FLSBMRP in which
by the help of candidate node network can find the most stable path between nodes.
In this candidate node used a fuzzy logic technique for nodes collect performance
following parameters are considered like enduring bandwidth, enduring power, link
quality, node mobility, and reputation index. Other multiple paths can be entren-
ched through candidate nodes between source and destination [6].
2.2 Analyze Response Delay
Qin Zhu, Li Yu, and Sangsha Fang et al. propose a stochastic network calculus
(SNC) scheme which helps to analyze end-to-end delay of a data flow through the
stochastic end. For the characterize of data flow it is done on the basis of Linear
Fractional Stable Noise (LFSN)-based traffic model [7]. Xianghui Cao, Lu Liu,
Wenlong Shen and Yu Cheng et al. propose a distributed delay aware multipath
routing method that helps to minimize end-to-end delay of all the commodity nodes of
a network [8]. Ze Li and Haiying Shen et al. narrate for the transmission delay
represent, another design QoS guaranteed neighbor selection algorithm. For traffic
transmission, a short delay is the crucial real-time QoS requirement, the primitive
deadline first scheduling algorithm is consolidates by QOD which is a deadline-driven
scheduling algorithm for data traffic scheduling in intermediate nodes [9].
2.3 Weight-Based Data Assignment Technique
D. Sumathi, T. Velmurugan, S. Nandakumar, and S. Renugadevi et al. presented a

strong load balancing algorithm M-OPTF that work on the basis of dynamic weight
assignment scheme, which is for fresh calls and handoff calls. In this, the stray
nodes are distributed as per the previously loaded on access points of the base
stations. Suppose load is more, then height factor will be small [10].
Haiying Shen, Ze Li, and Chenxi Qiu et al. give the idea how to choose
neighbors for data forwarding, for that node require capacity information such as
bandwidth and queue size of nodes. In this, an elected neighbors consist of suffi-
cient storage space for each segment. All times current capacity and storage
information was updated by exchanging its data to each neighbor [11]. Weiyi
Zhang, Benjamin Bengfort, Xiaojiang Du et al. proposed a various decisive
bandwidth allotment techniques for the hybrid network like auction based, top
down, bottom up allotment techniques. The top- down is lightweight and minimum
computation radical among the other techniques [12].
2.4 Hop Routing
Y. Wei and D. Gitlin et al. depict in terms of an omission of route maintenance and
a limited number of hops in data routing, it behaves like the two hop transmission.
In this, when the bandwidth of node is greater than the neighbor, it sends a message
to base station directly. If it not happens it will select a neighbor which has a higher
channel and then send it [13].
Nithin Michael and Ao TangHALO et al. ensure to minimize the cost of carrying
traffic through packet-switched networks, first link state routing protocol was there
in packet forwarding of hop to hop [14]. Jaeweon Choand Zygmunt J. Haas et al.
suggested how throughput gain can be achieved in two and three hops relaying
scheme. For this technique of multihop relaying technology were used to mitigating
unfairness of QOS that totally depends on location wise signal quality [15].
In this section, it is discussed how the system works regarding provisions of the
perfect network in which data are transmitted by using the weight-based assignment
technique with distributed three hop routing protocol. By using this artistry, the
maximum number of nodes is involved in a network that diminishes delay in
routing. To understand this following are the diagram which demonstrates our
proposed methodology (Fig. 3).
Following are the steps which were used in the weight-based assignment with
distributed three hop routing method for the wireless communication by that our
system is more effective for its working.
3.1 Create Base Station
This is the initial step of the proposed, where every node in a network is registered
with a base station on their activation. The base station provides the shortest path to
source node on its request for 3 hop routing process. And instance routing path is
also provided by the base station on request of instance source node in the process
of 3 hop routing protocol.
3.2 Shortest Path Identification Using Fuzzy Logic
In this step, the model user selects the data that need to be delivered to the receiver.
Once the receiver and data is selected for routing then system evaluates the shortest
path using fuzzy logic theory. According to this theory, all the nodes will assign
random integers and these random values are forming a matrix of node weight. This
node weight matrix subjected to the evaluation of the shortest path, where five
fuzzy crisp values are being created like low, very low, medium, high, and very
high based on the node weight. These fuzzy crisp values are forming fuzzy rules,
which have eventually given rise to the shortest path based on the IF-THEN rules.
Whatever shortest path obtained that considered as immediate one hop from the
source node. This can be summarized along with the following algorithm.
Algorithm 1 For Shortest Path Calculation using Fuzzy Logic
//Input: Node Weight as Nwi
//Output: Shortest path
Step 0: Start
Step 1: Node Weight Matrix as MAT
Step 2: FOR i=0 TO row
Step 3: FOR j=0 TO column
Step 4: Fuzzy Crisp values Fc
Step 5: Fuzzy Rules as Fr
Step 6: Fuzzy IF-THEN rules
Step 7: Shortest path Sp = R Fc ! Fr
Step 8: END Inner FOR
Step 9: END Outer FOR
Step 10: Return Sp
Step 11: Stop
3.3 Analyze Response Delay
Once the shortest path is calculated then system identifies all other nodes that are
not involved in the shortest path. And then system automatically sends a hand-
shaking message to all these nodes and receives the reply from all of them, thereby
records the time delay of all nodes. Now based on the best top two time delay nodes
are considered along with the next immediate node from the source node which is
evolved in shortest path. So these three nodes are finalized for data fragmentation
process and for three HOP process.
3.4 Data Fragmentation Using Weight-Based Assignment

Scheme
After analyzing response delay, least delays of two nodes along with the immediate
one hop which is ripen through shortest path selects for a data fragmentation
purpose. The data are being fragmented based on the ratio of the time delay got
from the three nodes. The node which is having highest time delay receives the
chunk with lowest size and vice versa. The process of data partition and assignment
can be clearly shown below with the algorithm.
Algorithm 2 For Weight-based Data Allocation based on the Response Delay
//Input: User data as D
//Output: Assignment of D1, D2, D3,…Dn to N1, N2, N3…Nn
Where,
Dn Represents Data Segment
Nn Represents Node
Step 0: Start
Step 1: Ping String S to N1, N2, N3
Step 2: Response
P Delay Collection set as Rd={Rd1, Rd2, Rd3}
Step 3: Rs ¼ Rd
Step 4: Dw=R Rdi/Rs (Data Weight Ratio)
Step 5: R Dw!Di
Step 6: Di!Ni
Step 7: Stop
3.5 Data Routing
This step involves the mechanism of data routing to the receiver. Where parallel
data routing process triggers with the three data chunks along with the shortest path.
In the process, all the data chunks are labeled and handled according to the label.
Once all the chunks reaches to the receiver then based on the labels and data is been
merged at the receiver end to form the original data with the desired extensions.
An explicit experiment is conducted for the performance evaluation of the system is

estimated using average end-to-end delay of the transmission of data from the
sender node to receiver node. So to ensure that proposed model method is of 3 HOP
routing protocol average end-to-end delay is better than other systems. Our model
compares with the method mentioned in [16] that deals with the technique of iETT
(Improved Expected Transmission Time). iETT is an improved version of the ETT
(Expected Transmission Time), a routing sequence selection method which selects
the routing path based on the availability of the wireless links and other link quality
conditions of the network nodes. Whereas proposed system applies and analyzes
the availability of the all other nodes in the network based on the DTR. The
observed end-to-end delay is recorded in Table 1.
Table 1 Average latency (in Packet size (Bytes) ETT metric iETT metric DTR
seconds)
500 1.25 0.8 0.65
1000 1.5 1.2 1
1500 1.7 1.4 1.1
2000 1.8 1.5 1.18
Fig. 4 Average time for different data size
On observing and analyzing above plot some facts are revealed DTR protocols
having the excellent hold on transmission mode to select the strong link wireless
nodes. Due to this, the end-to-end delay can be diminished reasonably as shown in
Fig. 4.
As we are appraised of the fact that in almost all the routing protocols in hybrid
wireless network routing used to happen based on the routing sequence provided by
shortest path algorithms. This technique always utilizes the wireless nodes that are
part of the shortest path sequence for data routing, this makes all other nodes to stay
idle. Due to this a considerable delay in data transmission is always existed. So the
proposed system uses the distributed three hop routing protocol technique, where it
involves maximum possible nodes from the network for routing process to decrease
the end-to-end transmission delay. This process is powered with the theme of
weight-based data assignment technique based on the neighbor node response
delay. This empowers the system to perform better to select the best shortest and
fastest path of routing.
The proposed system can be enhanced in the future by considering multiple hop
routing protocol to get the best end-to-end transmission delay on increasing the
number of nodes in the network.
References
1. H Luo, R. Ramjee, P. Sinha, L. Li, and S. Lu, “Ucan: A unified cell and ad-hoc network
architecture”, In Proc. of MOBICOM, Volume 1, September 2003.
2. P. T. Oliver, Dousse, and M. Hasler, “Connectivity in ad hoc and hybrid networks”, IEEE In
Proc. of INFOCOM, 2002.
3. Ms. MALARVIZHI, Mrs. P.RAJESWARI, “Improving the Capacity of Hybrid Wireless
Networks by Using Distributed Three-Hop Routing Protocol”, International Conference on
Engineering Trends and Science & Humanities (ICETSH), 2015.
4. K. Sasikala, V. Rajamani, “A Neuro Fuzzy based conditional shortest path routing protocol
for wireless mesh network”, International Journal of Enhanced Research in Management &
Computer Applications, Vol. 2 Issue 5, May-2013.
5. Dimitris Sacharidis, Panagiotis Bouros, “Routing Directions: Keeping it Fast and Simple”,
ACM, November 05–08-2013.
6. Sujata V. Mallapur, Siddarama R Patil, “Fuzzy Logic Based Trusted Candidate Selection for
Stable Multipath Routing”, I.J. Information Technology and Computer Science, 2015.
7. Qin Zhu, Li Yu and Sangsha Fang, “Per-Flow End-to-End Delay Bounds in Hybrid Wireless
Network”, IEEE transaction on Wireless communication and Networking, 2013.
8. Xianghui Cao, Lu Liu, Wenlong Shen, Yu Cheng, “Distributed Scheduling and Delay-Aware
Routing in Multi-Hop MR-MC Wireless Networks”, IEEE transaction on Vehicular
technology, 2015.
9. Ze Li, Haiying Shen, “A QoS-Oriented Distributed Routing Protocol for Hybrid Wireless
Networks”, IEEE transaction on Mobile Computing, 2014.
10. D. Sumathi, T. Velmurugan, S. Nandakumar, S. Renugadevi, “Dynamic Weight Assignment
based Vertical Handoff Algorithm for Load Optimization”, Indian Journal of Science and
Technology, Vol 9, 2016.
11. Haiying Shen, Senior Member, Ze Li and Chenx Qiu, “A Distributed Three-hop Routing
Protocol to Increase the Capacity of Hybrid Wireless Networks,” IEEE Transactions on
Mobile Computing, 2015.
12. Benjamin Bengfort, Weiyi Zhang, Xiaojiang Du, “Efficient Resource Allocation in Hybrid
Wireless Networks”, IEEE transaction on WCNC, 2011.
13. Y. Wei and D. Gitlin, “Two-hop-relay architecture for next generation WWAN/WLAN
integration”, IEEE Wireless Communication, April 2004.
14. Nithin Michael, Ao Tang, “HALO: Hop-by-Hop Adaptive Link-State Optimal Routing”,
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 23, NO. 6, DECEMBER 2015.
15. Jaeweon Cho, Zygmunt J. Haas, “On the Throughput Enhancement of the Downstream
Channel in Cellular Radio Networks Through Multihop Relaying”, IEEE JOURNAL ON
SELECTED AREAS IN COMMUNICATIONS, VOL. 22, NO. 7, SEPTEMBER 2004.
16. Saad Biaz, Bing Qi, Yiming Ji, “Improving Expected Transmission Time Metric in Multi-rate
Multi-hop Networks”, IEEE CCNC, 2008.
VLSI-Based Data Hiding with Transform
Domain Module Using FPGA
Latika R. Desai and Suresh N. Mali
Abstract In this rapidly growing internet era, researchers are giving more and
more attention toward robust, secure, and fast communication channels while
hiding sensitive data. The concealment steps can be done through a spatial domain
or the transform domain. This paper proposes a data hiding system with an adaptive
Very Large-Scale Integration (VLSI) module to enhance the security and robustness
of embedded data. The Field Programmable Gate Arrays (FPGA) implementation
approach of data hiding technique provides not only pipelined and parallel opera-
tions, but also gives the perfect guard against malicious attacks. The proposed
algorithm is implemented on a Xilinx Virtex 5 FPGA board. Further, the transform
domain technique also optimizes memory space and reduces the execution time
through pipelining. The performance of the implemented system is measured using
different parameters like resource utilization, Mean Squared Error (MSE), and Peak
Signal-to-Noise Ratio (PSNR).
Keywords Security Data hiding Field programmable gate arrays

DCT PSNR
1 Introduction
The demand for robust, secure data hiding techniques in today’s real-time applica-
tions are increasing. Spatial domain and frequency domain approaches were widely
used previously. In the spatial domain, different types of least significant bit
(LSB) techniques are used to manipulate pixel values to hide a data. These are simple
methods and can be easily detected by human eye, as the converted image gets
L. R. Desai (&)
Dr. D.Y. Patil Institute of Technology, Pimpri, Pune, Maharashtra, India
e-mail: latikadesai@gmail.com1
S. N. Mali
Sinhgad Institute of Technology and Science, Narhe, Pune, Maharashtra, India
e-mail: snmali@rediffmail.com2

https://doi.org/10.1007/978-981-10-7245-1_42
426 L. R. Desai and S. N. Mali
significantly and noticeably deteriorated, even though by taking care of intensity,

hiding in the edge area, or corner or dark area of the image. Further, as you increase
the data density in the cover image, the security gets compromised accordingly.
Embedding and extracting time becomes challenging. To overcome these traditional
challenges, frequency domain can be a perfect alternative over spatial domain.
Discrete Cosine Transforms (DCT) and Discrete Wavelet Transform
(DWT) methods are more preferable in the frequency domain. The authors have
implemented the data hiding technique using transform domain through DCT, where
the adaptive hardwired mechanism for embedding the sensitive data is proposed.
Authors are working on a generic data which is today’s demand for the file transfer.
Section 2 covers related work in the field. Section 3 throws light on proposed
research work by the authors. Section 4 discusses results obtained through experi-
mentation. Section 5 concludes the paper with future scope and directions.
2 Related Work
The data hiding is basically divided into spatial domain and frequency domain.
Bassam et al. research work in [1] proposed a novel algorithm to embed and extract
the entire secret data in the Haar wavelet-based transform. Special clipping
mechanism enhances the PSNR of the stego-image. Authors in [2] introduced high
speed, flexible linear feedback shift register (LFSR)-based approach with the
reconfigurable hardware implementation. Research work in [3] demonstrated an
image Steganography technique to hide the audio signal in the image using
transform domain. Authors in [4] give a hardware realization of new algorithm with
two stego-keys helping random selection of pixels. Researchers in [5] introduced a
novel and fully reversible data embedding algorithm for digital images. Simplicity
and efficiency of reversible data embedding algorithms enhance their payload
capacity. Research work in [6] implemented a mapping-based hiding technique
which involves the different value of X box for encoding 4:2 cover with a message
which is difficult to decode. Authors in [7] introduced embedding process in data
hiding techniques are classified into spatial domain and frequency domain.
Hardware Security modules provide a way for shielding the information from harm
in the memory of the tamper proof security service which is the key feature of
hardware technique. Research work in [8] compared spatial domain techniques
versus transform domain and they demonstrated that spatial domain is simple, but
can embed high-quality image while transform were robust in nature. Researchers
in [9] proposed improved genetic algorithm getting PSNR upto 38.47 while scope
suggested other algorithms such as cuckoo search algorithm and bacteria foraging
optimization algorithm is implemented to obtain accurate modeling of
Steganography. Authors in [10] demonstrated a fast and real-time, secure and safe
communications over networks with a multipoint algorithm for text Steganography
hardware implementation. Scope of future work is parallel processing design to
optimize the system encryption speed and power consumption. Research work in
VLSI-Based Data Hiding with Transform … 427
[11] concluded a novel methodology behaving like invisible in real-time data

transfer. This can be done through the trusted key exchange; it is an extension of
existing informed embedding technology. Research work in [12] introduced mul-
tilevel embedding technique; they worked on different types of multiplexing and
modulation to hide multiple bits. Research work in [13] specified encryption which
is one of the solutions for information security, but encrypted messages once
intercepted, can easily provide a clue to the adversary or attacker that some message
of importance is being communicated. Data hiding, on the other hand, takes the
opposite approach and attempts to hide all evidence that communication is taking
place. Authors in [14–20] presented information hiding through different techniques
like LSB, DCT, DWT, simulation, and implementation using FPGA and microar-
chitecture and its comparison also information hiding without loss of file format-
ting, a combination of cryptography and Steganography. Research work in [21]
gives an analysis of different method under LSB while authors in [22] used MSB
approach in the spatial domain. Researchers in [23] proposed spatial domain
technique using LSB through the pipeline. As seen from related work, the authors
have observed that the existing most of the available algorithms are available in
spatial domain so to improve security, authors need to follow the frequency domain
and try to keep tradeoff between capacity, security, and robustness.
3 Proposed Research Work
Data hiding System consists of two main parts, i.e.,

1. Embedder
2. Extractor
To design Embedder in the transform domain (i.e., using DCT) following
challenges were introduced while embedding the secret message:
• The random number generation can greatly have a high bandwidth requirement
due to the huge number of iterations in embedding process (16 kb 4 kb
READ operations from memory) is getting reduced to only read/write where the
message is embedded.
• The 64-bit external memory bus width adds bottleneck for embedding process.
It is resolved by LFSR used for address translation.
• By comparing fractions and fixed-point numbers for proposed work, efficient
representation for fraction number mechanism is selected because the floating
point method has an extremely slow operation. A number of iterations required
in floating point is a lengthy while for fixed point gives us low or moderate
accuracy.
• The computation overheads for DCT and IDCT are more, but it gets solved by
using symmetry for DCT and IDCT computations design structure. That is why
authors are using core reusability.
3.1 Embedder
Embedder which is used to form stego is as shown in Fig. 1. We have considered

generic data which are in the form of single dimension for cover as well as the
message. The host PC provides us a cover data and message data. By dividing
cover data into 1 8 windows, calculated DCT through Multiply and Accumulate
(MAC) till the end of the cover. In the DCT form cover message bits get embedded
through the adaptive embedding technique with LFSR than calculating IDCT it
forms the Stego.
3.2 Extractor
In the proposed system, part two of data hiding module is used as a message
extractor from stego which is shown in Fig. 2. The host PC provides us stego data
by calculating DCT through MAC till the end of the cover. In the DCT form, stego
gets extracted through adaptive embedding techniques using LFSR and then the
message data gets extracted from the cover data. As we mentioned earlier, we are
operating generic data which are in the form of single dimension for cover as well
as the message. A Message is sent to the host PC. Here, the process of data hiding is
completed successfully.
3.3 Embedding Algorithm

• Data is considered for cover as well as information in the form of a single
dimension.
Fig. 1 Proposed system embedder (part1)

Fig. 2 Proposed system extractor (part2)
• Obtain the number of bits of the message that need to be embedded into the
cover.
• Read 8 values of cover (Time/Spatial domain), i.e., 1 8 at one time.
• Perform DCT transform of each of them one by one.
• Store the DCT coefficients of the cover into storage buffer/memory.
• Assign instantaneous LFSR value for n-bit embedding.
• For every 1 8 coefficient, apply the embedding of message bit in DCT
coefficient of cover.
• LSFR is used in order to point a cover location for data hiding, i.e., An adaptive
approach of stego.
• Perform the IDCT after the stage, where Message is embedded into DCT
coefficients of cover. This gives the stego output.
• Repeat the process till the last bit of the message that is to be strategically
embedded into the cover.
• Store the result as stego data.
• Stego data are transferred to host personal computer.
• There may be the possibility of an attack at this end.
3.4 Extractor Algorithm
• From host personal computer, communicate the Stego to extraction at receiving

the end of the hardware.
• However, first, the Stego undergoes the DCT transform in order to transform
from time domain/spatial domain into the frequency domain.
• The message bits are extracted from the same location of DCT coefficient of
Stego which is pointed by LFSR.
• Hence, in the pointed DCT coefficient of stego, the bit position is assigned in
order to obtain message bit is same as that selected at the time of embedding.
• By taking consideration of applicability we have extracted message bit from the
cover.
• Any noise that is possible along the channel might corrupt this bit and it is
possible that this bit may not match to bit from message expected.
3.5 Hardware-Experimental Setup
This section demonstrates the FPGA-based hardware implementation of the pro-

posed DCT-based data hiding system. Prior to implementation, a system model was
developed to validate the algorithm functional correctness and examine its per-
formance. Next, the resistor-transistor logic (RTL) was developed, simulated, and
verified using VHDL language and Modelsim software tool 10.4 through synthesis
and analysis summarized in Table 1.
Figure 3 depicts the proposed system, where the cover data (C) and data to be
embedded (M) are transferred to the FPGA RAM and there after it is processed
block-wise as explained above as per the logic downloaded from the PC to
Configurable FPGA board. Then, the bits of M eventually get embedded into
various blocks. All the blocks are then reconfigured to reconstruct the stego data
(C’). The process of embedding can be validated for various attacks by transferring
C’ back to PC. The data hiding system is proposed to be implemented initially
using Virtex 5, XC5VLX50T, FFG1136C Board.
Figure 4 shows hardware implementation of the data hiding system. On board,
FPGA uses the memory of 256 MB 64 bit. Authors are handling the cover of
16 kb 8 data of window size for 8 values from the cover and data 4096
(512 8 bit). So to map data line of memory with cover size, we need to have
memory mapping. 2 kb 64 memory segmentation of the cover and data is
modified like 64 64 to match the address range for input and output. The location
Table 1 Design parameters

Parameter’s Value
Target FPGA device XilinxVirtex 5 XC5VLX50T
Frequency 100 MHz
Input/output signal timing constraint 15 ns
Signal activity source Simulation file
Fig. 3 Proposed system
Fig. 4 Hardware implementation of the data hiding system
which is used for DCT computes the cover and after embedding cover with the
message is stored which is the same. Here, we optimized the use of memory space.
The performance of DCT data hiding FPGA designs helps us to get good PSNR as
compared with other methodologies mentioned in the paper. Furthermore, it is
desirable to estimate the resource utilization and performance in the data hiding
system. We are trying to achieve our goal by using efficient use of FPGA for
pipelining processing, optimizing memory utilization, and time.
4 Results
4.1 Experimental Results
The results regarding attack (noise) on the stego are provided in Table 2. Authors
demonstrated the stego in terms of PSNR value for various levels of noise as shown
in Table 2. The system without noise has PSNR of 60 dB with MSE of 0.06433.
And with the noise of 10%, the PSNR value reached 39 dB with MSE of
8.0817260. As we know better PSNR gives robust system. Table 2 and Fig. 5
shown below give our experimental results in the form of PSNR and MSE with
noise variation.
4.2 Performance Evaluation
Figure 6 shown below gives a performance evaluation by considering the parameter

PSNR of our result with previous researcher’s results referred in this paper. Where
different methodologies of data hiding are considered like LSB, DWT, MSB, and
DCT. It shows that using proposed system gives us better PSNR using DCT.
Table 2 Proposed results for Sr. No. NOISE PSNR MSE

different parameter with noise
variation 1 0% 60.04659689 0.064331055
2 5% 42.12002246 3.990966797
3 10% 39.05576235 8.081726074
Fig. 5 Proposed PSNR and MSE value with noise variation

Fig. 6 PSNR comparative

evaluation with different
methodologies
5 Conclusions and Future Scope
By considering the limitation of the spatial domain, authors have implemented data
hiding technique in DCT transform domain. Data hiding using DCT gives us better
PSNR as compared with previous other methodologies used by the researchers. The
algorithm is implemented in Xilinx Virtex 5, XC5VLX50T, FFG1136C
FPGA-based hardware and its performance measures are examined including PSNR
and MSE. After adding noise how PSNR gets affected is shown by the results. This
research work can be extended to analyze additional parameters like area, power
dissipation, and time complexity in the future.
References
1. Bassam Jamil Mohd, Thaier Hayajneh, et al “Wavelet-transform steganography: Algorithm

and hardware implementation”, International journal of electronic security and digital
forensics January 2013.
2. R. Sundararaman and Har Narayan Upadhyay, “Stego System on Chip with LFSR based
Information Hiding Approach,” International Journal of Computer Application, Vol. 18,
No.2, pp. 24–31, 2011.
3. Hemalatha Sa,1, U. Dinesh Acharyaa, “Wavelet transform based steganography technique to
hide audio signals in image. Procedia Computer Science 47 (2015) 272–281. © 2015 The
Authors. Published by Elsevier B.V.
4. Saeed Mahmoudpour, Sattar Mirzakuchaki, “Hardware Architecture for a Message Hiding
Algorithm with Novel Randomizers”, International Journal of Computer Applications, Vol
37, No7, Jan 2012.
5. S. Al-Fahoum and M. Yaser Reversible Data Hiding Using Contrast Enhancement Approach
International Journal of Image Processing (IJIP), Volume 7: Issue 3: 2013.
6. Mr. Jagadeesha, Mrs. Manjula, “FPGA Implementation X-Box Mapping for an Image
Steganography Technique”, International Journal of Advanced Research in Electrical,
Electronics, Instrumentation Engg. Vol.2 Issue 6, June 2013.
7. Hao Chen, Yu Chen, “A Survey on the Application of FPGA’s for Network Infrastructure
Security”, IEEE Communications Surveys & Tutorials 2010.
8. Satwinder Singh and Varinder Kaur Attri, “State-of-the-art Review on Steganographic
Techniques”, International Journal of Signal Processing, Image Processing and Pattern
Recognition Vol.8, No.7 (2015), pp. 161–170.
9. Rinita Roy, SumitLaha, “Optimization Stego Image rataining secret information using
Genetic Algorith with 8-connected PSNR”, Published by Elsevier B. V. 2015.
10. Ammar Odeh, Khaled Elleithy, Miad Faezipour, “A Reliable and Fast Real-Time Hardware
Engine for Text Steganography” from https://www.researchgate.net/publication/262013703.
11. Dr. S Sivasubramanian, Dr, Janardhana Raju, “Advanced Embedding of Information by
Secure Key Exchange via Trusted Third Party using Steganography”, International Journal of
Latest Research in Science and Technology Vol., 2 Issue 1 Page No. 536–540, 2013.
12. XzMin Wu, et al. “Data Hiding in Image and Video: Part I—Fundamental Issues and
Solutions”, IEEE Transactions on Image Processing, vol. 12, no. 6, June 2003.
13. Y. K. Lin, “High capacity reversible data hiding scheme based upon discrete cosine
transformation,” Journal of Systems and Software, vol. 85, no. 10, pp. 2395–2404, 2012.
14. Shabir A. Parah, Javaid A. Sheikh and G.M. Bhat, “Data Hiding in Intermediate Significant
Bit Planes, A High Capacity Blind Steganographic Technique,” IEEE International
Conference on Emerging Trends in Science, Engineering and Technology, pp. 192–197,
2012.
15. Edgar Gomez-Hernandez, Claudia Feregrino-Uribe, Rene Cumplido, “FPGA Hardware
Architecture of the SteganographicConText Technique”, 18th ICECC,0–7695-3120-2/
2008IEEE.
16. Hala A. Faroul, MagdiSaeb, “An Improved FPGA Implementation of the MHHEA for Data
Communication Security”, proceeding of the Design, 1530–1591/05, IEEE.
17. Po-Yueh Chen, Hung-Ju Lin, “A DWT based Approach for Image Steganography”,
International Journal of Applied Science and Engineering 2006, Vol. 4 No-3, Page No 275–
290, 2006.
18. H.B. Kekre, A. Athawale, “Performance Comparison of DCT and Walsh Transform for
Steganography”, International Conf& Workshop on Emerging Trends in Technology, 2010.
19. Maninder Singh Rana, Bhupender Singh Sangwan, “Art of Hiding: An Introduction to
Steganography”, International Journal Of Engineering And Computer Science Vol. 1, Issue 1,
Page No. 11–12, Oct 2012.
20. Dr. S Sivasubramanian, Dr, JanardhanaRaju, “Advanced Embedding of Information by
Secure Key Exchange via Trusted Third Party using Steganography”, International Journal of
Latest Research in Science and Technology Vol., 2 Issue 1 Page No. 536–540, 2013.
21. BassamJamilMohd, Sa’ed Abed, et. at., “Analysis and Modeling of FPGA Implementations
of Spatial Steganography Methods” article in journal of circuits system and computers.
February 2014.
22. Multiplexed Stego path on reconfigurable hardware: A novel random approach,
BalakrishnanRamalingam, Rengarajan Amitharajan, John Bosco Balaguru Rayappan,
Elsevier (science direct) 2016.
23. S. Raveendra Reddy and S. M. Sakthivel, “A FPGA Implementation of Dual Images based
Reversible Data Hiding Technique using LSB Matching with Pipelining”, Indian Journal of
Science and Technology, Vol 8(25) IJST/2015/v8i25/80980, October 2015.
A Novel Approach of Frequent Itemset
Mining Using HDFS Framework
Prajakta G. Kulkarni and S. R. Khonde
Abstract Frequent itemset extraction is a very important task in data mining

applications. This is useful in applications like Association rule mining and
co-relations. They are using some algorithms to extract the frequent itemsets, like
Apriori and FP-Growth. The algorithms used by these applications are inefficient to
support balancing, distributing the load, and automatic parallelization with good
speed. Data partitioning and fault tolerance is also not possible because of excessive
data. Hence, there is a need to develop algorithms which will remove these issues.
Here, a novel approach is used to work on the extracting the frequent itemsets using
MapReduce. This system is based on the Modified Apriori, called as Frequent
Itemset Mining using Modified Apriori(FIMMA). To automate the data paral-
lelization, well balance the load and to reduce the execution time FIMMA works
concurrently and independently using three mappers. It uses decomposing strategy
to work concurrently.

Keywords Association rules Frequent itemsets Data partitioning

Load balancing MapReduce Hadoop FIMMA
1 Introduction
Frequent itemset extraction is the basic problem in data mining applications, such as
association rule, correlations, sequences, and many more data mining tasks. Hence,
this becomes an important research topic to extract the frequently used itemsets.
These frequent patterns are useful to take decisions in product marketing, sales, etc.
[1]. Association rule mining is popular in data mining [2]. The main goal of
Association rule is to find all the rules that fulfill a user-defined threshold. The first
P. G. Kulkarni (&) S. R. Khonde

Modern Education Society’s College of Engineering, Pune 411001, India
e-mail: prajakta.r999@gmail.com
S. R. Khonde
e-mail: shraddha.khonde@mescoepune.org

https://doi.org/10.1007/978-981-10-7245-1_43
436 P. G. Kulkarni and S. R. Khonde
phase of association rule is to identify frequent itemsets whose support is greater

than the threshold and the second phase is to form conditional implication rules,
among the frequent itemsets. Frequent itemsets generation defines the two similar
itemsets. The first itemset is similar to another. Now a day, there are enormous data
generated from different areas such as IT companies and web applications. Existing
data mining applications are unable to handle vast data and are only suited for a
typical database. Thus, to extract the frequent itemsets from the excessive database
is a very critical task [3]. For better utilization of frequent itemsets using large size
database, speed is very important. Speeding up the process of FIM is very complex
because it consumes most of the time to calculate the input/output intensity. In this
modern era, datasets are excessively large and sequential FIM algorithms are unable
to compute large database. They, however, failed to analyze data accurately and
they suffer from performance degradation. To solve these problems, MapReduce is
used to calculate frequent itemsets. Using this approach, the data will not only be
distributed in an efficient way but also balanced in the cluster. Hence, the perfor-
mance of finding frequent itemsets will be optimized [3].
This MapReduce is using the FIM which is based on the Modified Apriori,
called FIMMA. In this strategy, we are focusing the data partitioning method, load
balancing of data with a parallel approach. FIMMA consumes less time compared
with the traditional Apriori. The working of mappers and reducers is done con-
currently to optimize the speed, well balancing the load across various clusters [3].
The rest of this paper is partitioned as follows. Section 2 gives the review of the
literature. Section 3 defines the problem statement. Section 4 gives the present
system architecture. Section 5 explains the algorithms and methodology for the
system and discussed the expected results in Section 6. Section 7 concludes this
paper.
2 Related Work
The authors of “Association Rule mining extracting frequent itemsets from the large
database” have presented a problem of finding the frequent items from the excessive
database. The authors have developed the rules that have minimum transactional
support and minimum confidence. For this, an algorithm is used that carefully
estimates the itemsets for one pass. It adjusts the data between the number of passes
and itemsets that are measured in a pass. This process uses pruning system for
avoiding certain itemsets. Hence, this gives exact frequent itemsets from excessive
databases [4, 5]. A number of parallelization procedures is used to increase the
performance of Apriori-like algorithms to find frequent itemsets. MapReduce has
not only created but also exceeds in the mining of datasets of gigabyte scale or
greater in either homogeneous or heterogeneous groups. The authors have imple-
mented three algorithms, DPC, FPC, and SPC [6]. SPC has straightforward func-
tions and the FPC has static passes merged checking capacities. DPC consolidates
the dataset of various lengths by utilizing dynamic strategy and it gives good
A Novel Approach of Frequent Itemset Mining Using HDFS Framework 437
performance over the other two calculations. Accordingly, these three calculations
will scale up to the expanded dataset [6].
When dataset gets larger the mining algorithms becomes inefficient to deal with
such excessive databases. The authors have presented a balanced parallel
FP-Growth algorithm BPFP [7], a revised version of PFP algorithm [4]. FP-growth
algorithm is utilized with the MapReduce approach named as Parallel FP-growth
algorithm. BPFP balances the load in PFP, which boosts parallelization and auto-
matically enhances execution. BPFP gives a good performance by utilizing PFP’s
grouping system [7].
FIUT suggests a new technique for mining frequent itemsets called as Frequent
Itemset Ultrametric Tree(FIUT) [8]. It is a sequential algorithm. It consist two main
stages to scan the database. First-stage calculates the support count for all itemsets
in a large database. The second stage uses pruning method and gives only frequent
itemsets. While calculating frequent one itemsets, stage two will construct small
ultrametric trees. These results will be shown by constructing small ultrametric trees
[8]. Dist-Eclat, BigFIM are two FIM algorithms used with MapReduce Framework.
Dist-Eclat focuses on speed by load balancing procedure using k-FIS. BigFIM
concentrates on hybrid approach for mining excessive data [9]. Apriori algorithm is
additionally used to create kth FIS itemsets. The Kth FIS is used to search frequent
itemsets based on the Eclat system. These three algorithms are used with
round-robin technique which achieves a better data distribution [9].
PARMA uses parallel mining approach with the benefits of Randomization for
extracting frequent itemsets from a vast number of databases. This divides the
functionality into two parts, gathering the arbitrary data samples and secondly it
uses parallel computing method that is utilized to increase the mining speed. This
method avoids the replication that is very expensive. A mining algorithm applies to
every segment individually with parallel approach [10]. K-Nearest Neighbor Joins
utilizes MapReduce and distributes the excessive information on the number of
machines. This is done by the mappers and the reducers give the results in terms of
the KNN join. KNN Join is the key component to search the kth-nearest neighbor.
MapReduce is utilized for effective computing the data to obtain the best perfor-
mance result [11, 12]. To diagnose the Heterogeneous Hadoop Cluster and to
search primary faults, this paper is used Hadoop schedulers to produce efficient
Hadoop clusters even if they are in heterogeneous clusters [13]. It proposes the
DHP algorithm (direct hashing and pruning) which is used for minimized candidate
set generation for large itemsets. It solves performance degradation problem of
large dataset mining. It minimizes candidate itemsets.
FIUT is used with MapReduce to find frequent itemsets. MapReduce is a
popular programming approach used for computing massive datasets [14]. It
divides into three MapReduce phases. The database is divided into number of input
files and given to each mapper. The first MapReduce phase finds out frequent-1
itemset and Second MapReduce phase scans the frequent one itemsets and gener-
ates k-frequent itemsets. Third MapReduce phase uses FIUT algorithm and it will
create ultrametric tree [3]. FiDoop-DP Data Partitioning uses the Map Reduce
programming and gives the effective data partitioning technique for frequent
itemset mining. This increases the performance by using the data partitioning
technique, which is based on the Voronoi diagram. It extracts the correlations
between the transactions. By consolidating the similarity and the Locality-Sensitive
Hashing strategy, FiDoop-DP puts most similar records in data partition to increase
locality and this is done without repeating records [5, 15]. To differentiate and
extract frequent and infrequent itemsets from the massive database two-phase
scanning will be done here. In the first scan, it accepts input and distributes it into
mappers and finds out infrequent itemsets using minimum support. The reducer
combines the result and sends it to the second phase. In this phase, it scans first
phase output and gives the final result [16].
3 Problem Statement
Problem statement concentrates on the investigation of Frequent Itemset

Ultrametric Tree (FIUT) and to find the efficient way for its execution in HDFS
framework Implementation on FIMMA. To show that the proposed algorithm on
the cluster is sensitive to data distribution and dimensions, as itemsets with different
lengths have different decomposition and construction costs. Improving energy
efficiency of FIMMA running on the Hadoop clusters. To improve the performance,
a workload balance metric to measure across the clusters computing nodes is
developed.
4 System Architecture
FIMMA suggests parallel frequent itemset mining algorithm which uses

MapReduce programming technique for development. This removes the issues of
existing system and applies automatic parallelization, balancing the load of the
excessive database, and well distribution of given data. FIMMA is based on
Modified Apriori algorithm to overcome the issues of FIUT algorithm with reduced
time. It uses hash-based technique [17].
4.1 Objectives of Proposed System
• Better performance and improved accuracy using automatic parallel processing.

It performs with less time execution to scan the database.
• Keeping the cost constraint as it is and dealing with load balancing with
automatic parallelization.
• Hashing technique is used to differentiate the traditional Apriori.
The proposed system consists three MapReduce phases. The user is responsible
to give the input. This input will be accepted by the job manager as shown in Fig. 1.
Job manager splits the input into a number of blocks, processes it and gives the
output to the reducer. This output is in the form of frequent one itemset.
The output of the first MapReduce is applied to the mapper of the second phase.
It scans all the data to give frequent-k itemsets. It uses pruning method to find out
frequent and infrequent itemsets. The result is obtained from reducer, in the form of
k-frequent itemsets. The system architecture is as shown in Fig. 1. This is dis-
tributed into two main parts.
1. HDFS framework
2. MapReduce Approach.
HDFS is a Hadoop distributed file system and used to store the log files. HDFS
framework accepts the data from the user. The user gives the input using SQL
queries or in the form structured data and uploads it. This uploaded dataset accessed
by the Job manager. Job manager is responsible to distribute the dataset to the
available mappers of each data node.
Here, the transactions from input dataset are distributed to the mappers. Each
mapper access the input scans it and generates the results in terms of key and
values. Key is the item-name and value is the item-count. Using the minimum
support the mappers gives the result to the reducers in the form of key and values.
Each mapper calculates frequent-1 itemset. Reducer combines the result of each
Fig. 1 System architecture

mapper, sorts it and generate a final list. This is frequent-1 itemset list. This output
gives to the next MapReduce. It applies another round of scan. In this stage, it
accepts the frequent-1 items and compares it with the minimum support and
removes the infrequent items. This is called pruning system. Depending upon the
users given threshold value of k this MapReduce generates k-frequent itemsets.
(where k-itemsets < n number of dataset) It makes the possible combination of
frequent itemset from each mapper and gives to the reducer. Second MapReduce
updates the list and gives k-frequent itemset. This frequent-k-itemsets applies to the
third MapReduce. In this MapReduce, it accepts all k- itemsets and gives the result
in terms of top most k-frequent itemset.
FIMMA uses hash-based technique. There is a hash table used after each result
of MapReduce. These hash tables are used to store the result generated from
frequent-1 itemset to the k-itemset. These hash tables give a unique value to the
stored frequent items. Unique value is obtained by calculating mod hash formula.
Whenever there comes new input then hash table compares that items with stored
one. When new input from first mappers matches with stored one then hash table
sets a bit to 1 otherwise 0. This same procedure applies to the result of second
MapReduce (i.e., k-frequent itemsets) and on the third MapReduce. In last
MapReduce, it checks all set bit and stores into a new list, update it and gives the
final result. FIMMA helps to reduce the candidate generation of items by using hash
tables, therefore automatically it reduces the time. It controls the huge generation of
candidates with minimum support. It avoids the transaction record which does not
have any frequent items by comparing hash tables.
5 System Analysis
5.1 Algorithm for Frequent 1 Itemsets [3]
Input: minimum-support, Database D

Output: Frequent-1 itemset
Mapper Algorithm
Step 1: Mapper function is used with-MAPPPER (key offset, values Database D)
Step 2: //TR shows the transaction in Database D
Step 3: for loop is used for all Transactions TR in Database D do
Step 4: Candidate-items < – Splited each transaction TR.
Step 5: Use for loop for all candidate-items in all items, do
Step 6: output(candidate-item, 1)
Step 7: Here ends second for loop
Step 8: Here ends first for loop
Step 9: Mapper Function ends with each items count.
Step 10: The reducer takes input from mappers as input = (candidate-item, 1)
Reducer Algorithm
Step 11: REDUCE function starts with key, value (key candidate-item, values 1)
Step 12: take a variable total to store output, i.e., total = 0;
Step 13: Use for loop to calculate all candidate-item do
Step 14: Add new frequent item in total +=1 // Here ends for loop
Step 15: Output (frequent1-itemset, total).
5.2 Algorithm for Frequent-K-Itemsets
Input: minimum-support, Database D

output: frequent-k-itemsets.
Mapper Algorithm:
Step 1: In step 1 mapper function is used i.e. MAPPER (key offset, values
Database)
Step 2: //TR shows the transaction in Database D
Step 3: for loop is used for all Transactions TR in Database D do
Step 4: Candidate-items ← Splited each transaction TR. //Here input database D is
frequent-1 itemset
Step 5: Use for loop for all candidate-items in all item, do
Step 6: Step 6 applies pruning system using if condition for infrequent items
if Candidate-item = infrequent item then
Step 7: Remove the Candidate-item which is infrequent from the Transaction TR;
Step 8: If conditions ends here
Step 9: variable Frequent-k-itemset is used to store all k-frequent items and shown
by-Frequent-k-itemset ← (frequent-k, fr-set)
//After applying pruning system fr-set is the result with frequent-k items
Step 10: output (Frequent– itemset, 1);
Step 11: Here ends second for loop
Step 12: Here ends first for loop
Step 13: Mapper function is ends here by giving all mappers output.
Reducer Algorithm
Step 14: Reducer starts from this step using function REDUCER (key k-itemset,
values 1)
Step 15: Total = 0;
Step 16: used for a loop to count all items from mapper as-for all (k-itemset): do
Step 17: Total += 1;//For loop ends here
Step 18: output = (frequent-k items + total)
Step 19: Reducer function ends here with final k output.
5.3 Algorithm for FIMMA
Step 1: Consider C be the variable for selection of one cluster at a time

Step 2: Here, the database will be scanned using minimum support and it will
generate frequent items. It will combine all the possible combinations of
frequent itemsets
Step 3: Function Fre1 stores the frequent itemsets ! Fre1 = find-freq-1itemset(T)
Step 4: for k = 2 to fk−1 6¼ Ø; generate Ø from fk−1 items
Step 5: Consider H1 is the hash table of size 8. B1 is buckets in the hash table and
A1 is Unique value to the frequent itemsets. V1 is the bit vector
Step 6: Calculate the items I up to user threshold value w from ck with min support;
ft(1 w k); end for
Step 7: Calculate frequent items with minimum support
Step 8: y variable to store the result of minimum support, i.e., y = min support
(ck, ft)
Step 9: get transaction id in variable ! target = get-trans id (y)
Step 10: compare the target values with the hash tables
Step 11: for each transaction id in target increment count of all ck
Step 12: if b1 >=min support; then bit vector v1 = 1; otherwise v1 = 0
Step 13: prune the 1 = 0 itemsets and modify the list
Step 14: fk = items in ck (min support)\\ end for
Let S be the system which do analysis and read documents; such that:
S ¼ fS1; S2; S3; S4; S5g where—S1 represents a query requesting by the user;
S2 represents authentication; S3 represents MapReduce module; S4 denote the sql
injection techniques; S5 gives the graphical presentation.
S1 ¼ fU1; U2; U3. . .. . .::Ung; Where, S1 contains SQL query–If S1 is valid

then proceed, Else discard.
S2 define user is authenticated or not; Where, Ui ¼ fUI1; UI2:UI3. . .. . .UIng
• Ui is the master node which having different storage nodes as clusters.
S3 = Functionality for three MapReduce phases:
• Input: database DI,min sup; Output: Frequent itemsets
• Let DI ¼ fDI1; DI2; ::. . .Ding; where, DI is Input Database
Applying algorithms to FindP 1DIand k-frequent itemsets, Hash-based Algorithm

F1(freq 1-itemset) = scan P 0
; Fk (freq k-itemset) = scan (F1)
FI ¼ fFI1; FI2; :. . .FIKg;Where FI is the final frequent itemsets

S4 ¼ fpatten 1; pattern 2. . .. . .:pattern ng
Each pattern checks the behavior of query created by end user’s S1 (query
module). S1 ¼ avg þ min þ max þ round þ floor þ todate == possible queries
S5 ¼ graphical representation for time comparison graph
The experimental results evaluated with the minimum support (sometimes larger
data sizes). This upgradation comes at no execution cost, as they prove the way that
this implementation, achieves the good performance, compared to other techniques
with reduced time. By examining with this work, it demonstrates that, the execution
of FIUT is slow. It results that, whenever increased the minimum support and
dataset, it gives the balanced output. FIMMA gives an improved performance by
using hash table concept. When the threshold value decreases, other methods
occupy more memory as well as consume more time.
Table 1 shows the time required to extract frequent itemsets. As shown in table
I, first column shows the Size of Dataset. Other columns show the methodologies to
find frequent itemset. Proposed system requires less time compared to existing
systems. It shows the time in seconds. The performance of this system against the
Frequent Itemset Ultrametric Tree (FIUT) method is as shown in Fig. 2. Modified
Apriori algorithm is a good algorithm to give the correct results as compared to
existing systems. Also, this algorithm shows the faster execution even for a large
database. As synthetic dataset is used here, the user can make their own dataset to
run the tests.
Table 1 Time required for Size of dataset FIUT Modified apriori (FIMMA)
finding the frequent itemset
from dataset 1000 139 113
2000 165 144
3000 218 196
5000 277 234
10,000 317 293
Figure 2 shows the graphical representation of the methodologies and datasets.

These are the results of two methodologies when the dataset is increased. X-axis
shows the methodologies and y-axis shows the time in seconds. Figures 3 and 4
shows the Hadoop implementation results. Figure 3 shows the FIUT implementa-
tion with required time in ms. In this, it takes 43,780 ms (shown in red color
rectangle) to execute a job. Figure 4 shows the FIMMA implementation. Here it
shows the time required to execute the same job. FIMMA takes 21,090 ms to
execute a job. Hence, it shows that FIUT takes more time than the FIMMA method.
FIMMA gives better performance than the FIUT with reduced time.
Fig. 2 Comparison graph
Fig. 3 Hadoop implementation using FIUT

Fig. 4 Hadoop implementation using FIMMA
7 Conclusion
FIMMA technique is used to defeat the issues which are available in existing
methods like parallel mining and load adjusting algorithms. In this approach,
FIMMA algorithm (Hash-based) is proposed using MapReduce. The Comparison
between existing method and proposed methods shows that there is up to 60%
reduction in time. Hash-based Apriori is most efficient for generating the frequent
itemset than existing methods. Data partitioning and data distribution is done by
using this proposed system. FIMMA stores the previous results in hash tables and
therefore it reduces time. At each result stage, the new input will be compared with
the hash table and if matches with hash table’s stored result then vector bit is set to
1 and named as a frequent item. All the set vector’s list will be combined by the
reducer, modify it and gives the final result. The proposed system works efficiently
on MapReduce stages and gives the final result, rather than using the FIUT algo-
rithm and increases the performance.
References
1. Bechini, Alessio, Francesco Marcelloni, and Armando Segatori. “A MapReduce solution for
associative classification of big data”, Information Sciences, 2016.
2. X Zhou, Y Huang - Fuzzy Systems and Knowledge Discovery. An Improved Parallel
Association Rules Algorithm Based on MapReduce Framework for Big Data”, pp. 284–288,
2014 11th International Conference on Fuzzy Systems and Knowledge Discovery.
3. Yaling Xun, Jifu Zhang, and Xiao Qin, FiDoop: Parallel Mining of Frequent Itemsets Using
MapReduce” IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS:
SYSTEMS, VOL. 46, NO. 3, pp. 313–325, MARCH 2016.
4. R. Agrawal, T. Imieli nski, and A. Swami, “Mining association rules between sets of items in
large databases,” ACM SIGMOD Rec., vol.22, no. 2, pp. 207–216, 1993.
5. S Deshpande, H Pawar, A Chandras, A Langhe, Data Partitioning in Frequent Itemset Mining
on Hadoop Clusters”- 2016 – irjet.net, https://irjet.net/archives/V3/i11/IRJET-V3I11229.pdf.
6. M.-Y. Lin, P.-Y. Lee, and S.-C. Hsueh, “Apriori-based frequent itemset mining algorithms on
MapReduce,” in Proc. 6th Int. Conf. Ubiquit. Inf. Manage. Common. (ICUIMC), Danang,
Vietnam, 2012, pp. 76:1–76:8.
7. L. Zhou et al., “Balanced parallel FP-growth with MapReduce,” in Proc. IEEE Youth Conf.
Inf. Compute. Telecommun. (YC-ICT), Beijing, China, 2010, pp. 243–246.
8. Y.-J. Tsay, T.-J. Hsu, and J.-R. Yu, “FIUT: A new method for mining frequent itemsets,” Inf.
Sci., vol. 179, no. 11, pp. 1724–1737, 2009.
9. Kiran Chavan, Priyanka Kulkarni, Pooja Ghodekar, S. N. Patil, Frequent itemset mining for
Big data”, IEEE, Green Computing and Internet of Things (ICGCIoT), pp. 1365–1368, 2015.
10. M. Riondato, J. A. DeBrabant, R. Fonseca, and E. Upfal, “PARMA: A parallel randomized
algorithm for approximate association rules mining in MapReduce,” in Proc. 21st ACM Int.
Conf. Inf. Knowl. Manage.,Maui, HI, USA, pp. 85–94, 2012.
11. Wei Lu, Yanyan Shen, Su Chen, Beng Chin Ooi, “Efficient Processing of kNearest Neighbor
Joins using MapReduce” 2012 VLDB Endowment 2150-8097/12/06, Vol. 5, No. 10,
pp. 1016–1027.
12. [Online]. Available:www.vldb.org.
13. Shekhar Gupta, Christian Fritz, Johan de Kleer, and Cees Witteveen, “Diagnosing
Heterogeneous Hadoop Clusters”, 2012, 23 rd International Workshop on Principles of
Diagnosis.
14. J. Dean and S. Ghemawat, “MapReduce: A flexible data processing tool,” Commun. ACM,
vol. 53, no. 1, pp. 72–77, Jan. 2010.
15. Yaling Xun, Jifu Zhang, Xiao Qin and Xujun Zhao, “FiDoop-Dp Data Partitioning in
Frequent Itemset Mining on Hadoop clusters”, VOL. 28, NO. 1, pp. 101–113, 2017.
16. Ramakrishnudu, T, and R B V Subramanyam. Mining Interesting Infrequent Itemsets from
Very Large Data based on MapReduce Framework”, International Journal of Intelligent
Systems and Applications, Vol. 7, No. 7, pp. 44–49, 2015.
17. Jong So Park, Ming Syan Chen, Philip S, “An Effective Hash based Algorithm for mining
Association rule”, ‘95 Proceedings of the 1995 ACM SIGMOD international conference on
Management of data, held on May 22–25, 1995, Pages 175–186.
Issues of Cryptographic Performance
in Resource-Constrained Devices:
An Experimental Study
Balaso Jagdale and Jagdish Bakal
Abstract Many gazettes, toys, sensors, instruments, etc., devices are mushrooming
in the electronics market these days which are running with different operating
systems and application frameworks. In Mobile phones also many platforms are
available with its own stack, right from hardware level to application level. So it is
difficult to all the mobile phones or devices with same characteristics, performance,
and features. The conclusive success of a platform entirely depends on its security
to the user data ultimately, it constructs the global market. We present here a
comparative study of four Mobile Operating Systems architecture and security
perspectives of each. Moreover, experiments are carried out with Windows and
android smartphone to study the cryptographic performance in these devices which
plays as decisive parameters for future applications and system developers to
implement secured devices and applications. We show that although algorithms
characteristics are fixed, due to different architectures and dependencies in different
devices with different operating systems, same efficiency, and performance is not
guaranteed by cryptographic as well as other data structure algorithms.
Keywords Security Cryptography Mobile platform Performance
1 Introduction
Users prefer storing data on the mobile device so that they have access to data as
and when they need it. The data stored is private and sensitive in nature and hence it
is desirable to have an operating system on these devices that secure the data from
malicious applications. Here, we look at how the leading mobile operating systems
B. Jagdale (&)
G. H. Raisoni College of Engineering, Nagpur, India
e-mail: bjagdale@gmail.com
B. Jagdale J. Bakal
MIT College of Engineering, Pune, India
e-mail: bakaljw@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_44
448 B. Jagdale and J. Bakal
implement the security. The operating systems under consideration are Android,
iPhone, Symbian (being absolute), and Windows mobile.
Naseer Ahmed and other [1], have compared different operating system used in
mobile devices and shown how to utilize resources efficiently. Ahmed and others
demonstrated the security algorithms performance in RFID resource-constrained
devices [2], where authors have discussed energy and performance challenge.
Authors N. R. Potlapally [3] and other have worked with SSL and energy and
performance efficient cryptography for mobile platforms. Daniel [4] has worked
with Texas and light microcontrollers and studied the performance of
ultra-lightweight cryptography. Payal patel [5] and other have worked with ECC
implementation for resource-constrained devices. Mathias and other have discussed
flexibility and strength of a Mobile OS. Viewpoints can be the manufacturer and
end user [6]. In another study [7], the author suggests an application-based selection
of operating system and based on the technical parameters and features, the user
should select mobile devices.
In Fig. 1, mobile operating system layers are shown which are generic in nature.
At the bottom, hardware boards with processors are controlled by the operating
system and the user has not control. Middle library API can be accessed by user
programs. But application framework controls permission during installations.
Moreover, communication services are also staked in the form of HTTP and
similar.
There are various places to apply the security controls, such as architectures,
hardware, etc., cryptography is a prime security control as compared other security
controls Many authors have done applied research of cryptographic algorithms in
resource-constrained devices. The reason is the feasibility of performance, power
consumption, and strength. The research work of S. Prasanna Ganesan and others
[8] shows RSA-based asymmetric cryptography is still a challenge in mobile
devices due to required computing power, Infrastructure support, and memory
capacity. Creighton T. R. Hager and others [9], studied the performance and energy
requirements for various cryptographic algorithms in personal digital assistants and
tablets.
Fig. 1 Stack of mobile

operating system
Issues of Cryptographic Performance in Resource-Constrained … 449
2 Architectures of Mobile Device Operating System
In this section, we are discussing the architecture of the mobile operating systems
Android, iPhone, and Windows mobile.
Figure 2 shows android layers. Open Handset Alliance launched android device
platform in 2007. The goal was to keep it open, free and diverse for the Internet of
things. OS Layers are shown in Fig. 2. Many mobile manufacturers like Motorola,
Ericson, HTC, and Samsung have adopted the platform. It usually is loaded on the
ARM-Snapdragon processor designed by Qualcomm. It has Good energy moni-
toring and memory management. Now that android is used in many
electronics-enabled appliances, it has become important to study performance and
other utilization of resources.
Android implementation is based on Linux kernel [10, 11]. Native libraries are
written in C/C ++ to achieve the stable performance of various elements. The main
component of Android Runtime, Dalvik Virtual Machine. The architecture of
iPhone is shown in Fig. 3. ARM series of processors are used by iPhone. SDKs and
drivers are mainly developed in Objective C runtime. Figure 4 shows, windows
mobile layers [12], which uses Microsoft Win32 API. Design and applications are
similar to the desktop flavor of windows. Windows CE.NET-based platform is used
Fig. 2 Architecture of
Android
iPhone OS
Windows mobile OS
[12]. Performance is dependent on OEM adaptation layer (OAL) which is placed

between the Microsoft Windows CE.NET kernel and target hardware mobile
device.
3 Security Aspects and Analysis in Mobile Devices
Users increasingly use their smartphones for Internet access, accessing their email,
and store more and more sensitive information on their mobile phones. This, on the
one hand, increases the ways in which mobile phones can be infected with malware,
and on the other hand, makes mobile devices increasingly attractive as an attack
target. The operating systems, [11] Symbian, Windows Mobile, iPhone OS, and
Android use different concepts to protect the integrity of Software and Firmware on
the mobile device and prevent users from installing malware-infected programs.
Mobile device’s security can be categorized as follows: Platform security,
Application configuration security, Network services security, Data storage secu-
rity, Hardware security, Transport layer security, and Application layer security.
Platform security: Platform security provides protection to the users’ information
such as secondary data storage, phone memory data storage, and location infor-
mation protection, protection to different hardware elements, generating, or storing
information, present on the device. Following are the important areas, where
security plays a key role. Application configuration security: whereby application to
be fixed on the device is verified by the software developer but is tested for safety
by the platform. Moreover, in application configuration one can restrict the appli-
cations access to the mobile resources. Network services security: the network
access through the device must be secure enough to keep the login data, i.e., user
information safe. The data transfer that would take place in the sessions must not be
vulnerable. Proper encryption-decryption methods must be used. Proper policies
must be employed to ensure the integrity, secrecy of the data. Data storage security:
applications may require the data stored on the device. Proper authentication and
authorization techniques must be used to restrict the open access to the data on the
device. Hardware security: security provided for different hardware modules can be
categorized here. Locking keypad, passwords set against usage of phone memory,
sim card. Transport layer security: can be implemented in the form of secured
socket layer which is implemented just above transport layer which provides API to
applications. This is implemented during the development of client–server mobile
applications with the help of programming libraries. Browser and OS network
applications support SSL to access secured sites which is an inherent security
provided in the mobile device platform. Application layer security: applications
may be installed explicitly or may be installed as a part of other application.
Application layer security may be implemented in mobile applications with the help
of cryptography libraries provided with the programming SDKs. The authenticity of
the application must to be checked. These applications may access data from other
applications which may harm the data or functioning of other applications.
(A) Android: Android supports security mode to cater mobile situation. Android
has system features similar to shared memory, files access control, user
Identifiers and Preemptive multitasking. The Dalvik virtual machine imple-
ments main security of the android operating system. The main design aspect of
Android is that it has control over applications due to which applications cannot
adversely affect system [11]. Android reduces abuse of access of device data
from applications by employing permission features statically and dynamically
[13]. Rooted phone concept is also implemented in Android, where only
advanced system developers can root the device and perform core changes in
the system as per the requirements. Through Google play, applications are
going through security check before they are deployed by users in the phone.
(B) iPhone: The iPhone OS runs applications with root access. Thus, it makes the
operating system vulnerable to security attacks. iPhone has security features
which help to protect sensitive data collected by applications. iPhone OS
deploys BSD (Berkeley Software Distribution) and CDSA (Common Data
Security Architecture) that supports protection of services that are implemented
by using CDSA. APIs support send, receive, and store data in the encrypted
form [14]. File access permissions are similar to UNIX BSD kernel, as a part of
security features.
(C) Symbian: Symbian has several security features for protecting integrity and
privacy. Symbian security architecture concentrates mainly on threats of dis-
tribution of malicious applications. Applications are signed by users or
developers whenever installed on the device. After that access to resources is
limited. But there is a risk of using nonsigned applications running on the
device. Control of API’s is limited for the programs running in the device, thus
restricting access of file system for trivial reasons. Each application as public
and private file system as a part of security. Moreover, API is provided to
implement security features for cryptography, hashing, and random number
generation.
(D) Windows mobile: Programmers are allowed to use lower APIs. So programmer
has the freedom to control hardware. The build has restricted security features
in the operating system. Programs do not have their own boundaries and dif-
ficult to protect from each other. Even though, programs cannot access directly,
but with DLL one program can access maliciously other user’s data. Small
process bugs can be tested but this is not enough to protect from the malicious
activity of applications. Security policy features are available so that trusted
API’s can refer before granting permissions to access low-level resources
including hardware [12].
4 Cryptographic Experiments
We carried cryptographic experiments in Windows Mobile and Android mobile.

Experiments are run for digest algorithm, Symmetric algorithm, Public key algo-
rithm. In digest category, we used the popular SHA1 algorithm. In the same key
category, we chose AES (advanced encryption standard). Finally, for Public key
cryptography, we experimented RSA algorithm. The experimental setup diagram is
as follows.
As shown in the Fig. 5, the experimental setup consists of a personal computer
loaded with the three different or possibly a single universal SDK, which could be
used to write, compile, and test the application on the emulator. The deployment
packages would then be created. These packages will be deployed on the mobile
device through a USB cable and will be executed to study their behavior.
The time required to execute a task is measured with timers available in pro-
gramming languages. Time can be measured in nano or milli seconds resolutions in
modern SDKs. Same experimental setup is also applicable for Win mobile instead
of android SDK, it is Visual studio 2008, Windows mobile SDK. Cryptographic
algorithms were run in both the platforms with 20, 40, 60, 100 K text file samples.
Cryptographic task ()//Time measurement in mobile phone

Start
Note the start time ()
Open the data file (file1)
Fig. 5 Device Development

System
Output = Encrypt (file1, algorithm, key)

StoreTheCipherOutput (Output, file2)
Note the end time ()
Time required i.e. end time – start time;
End
We used HTC touch diamond2 phone to conduct the experiments. This phone is
dual OS phone. Android java is used to run these programs in android device.
C Sharp is used to win mobile programming. Multiple readings were taken in which
minimum time reading is selected as proper cryptographic time. Also, mobile is
kept in idle mode while this program is run by killing unwanted processes. Also,
SIM card is not installed in mobile to avoid communication events and in the effort
to keep the mobile device in silent mode.
This section shows the results from different experiments in graphical form and
security analysis is done based on the performance results and other parameters. As
shown in the graph, Fig. 1 shows the performance for the SHA1 algorithm in which
time required less means good performance. In SHA1 Android is scoring well as
compared to Windows mobile, but in graph 2, we can infer that windows mobile is
doing well as compared to the android platform for Advanced encryption standard.
In graph 3, again RSA is doing well which is a well-accepted Public key algorithm.
Here are some Experimental Development Observations (Fig. 6).
We found that Android has good programming model and many libraries as
compared with windows mobile system. The security model of android is better that
windows mobile. In windows mobile, live device interface and programming is
excellent (plug and play) because of Active Sync utility from Microsoft (Figs. 7, 8).
Moreover, windows have upper hand for GUI design and ease of programming.
It is difficult to carry out same experiments of same logic on different platforms and
study benchmarking. The main reason is that iPhones, Symbian’s, Win mobile,
Fig. 6 SHA1 Algorithm

Android WinMobile SHA1
performance
700
Time Required ms
600
500
400
300
200
100
0
20K 40K 60K 100K
Data File Size
Fig. 7 AES 128 (Rijndel)

Android WinMobile AES
Algorithm performance
800
Time Required ms
700
600
500
400
300
200
100
0
20K 40K 60K 100K
Data File Size
Fig. 8 RSA 1024 Algorithm Android WinMobile

performance RSA
7000
Time Required ms
6000
5000
4000
3000
2000
1000
0
20K 40K 60K 100K
Data File Size
Android platforms are ported on numerous hardware platforms with variable size
form factors including Processors, Clock frequencies, RAM, ROM, buses, displays,
sensors, etc.
6 Conclusion
Cryptographic results are important to implement security for mobile data while
data is being transmitted or received over the network as well as data residing in the
mobile phone. As can be seen from experiments, android performs better in public
key and hash algorithms while windows perform better in symmetric key algo-
rithms. One important thing to observe is that algorithms are same but OS and
lower layers have an impact on performance. There are other ways for security in
term of signing the applications or during installations, OS provides security for
verification of applications. One can also carry out this kind of applied research in
the field of hybrid cryptographic protocols.
The android operating system runs applications in a closeted virtual machine

environment. Therefore, the android security architecture is strong. Symbian has a
security system in place for authenticating applications and windows mobile also
supports this, although the system is less sophisticated. iPhone implements security
services at OS level by using BSD and CDSA, leaving less freedom to the
developer. It is not possible to point out one as the best OS but we can certainly
choose the one over other based on their application development framework and
security.
References
1. Naseer Ahmad, Boota, M.W. and Masoom, A.H.: Comparative Analysis of Operating System
of Different Smart Phones, Journal of Software Engineering and Applications, 8, 114–126,
2015.
2. Ahmed Khattab, Zahra Jeddi, Esmaeil Amini, Magdy Bayoumi: RBS Performance
Evaluation, Springer International Publishing, 10.1007/9783319475455_6, 2016.
3. N. R. Potlapally, S. Ravi, A. Raghunathan, N. K. Jha: A study of the energy consumption
characteristics of cryptographic algorithms and security protocols, IEEE Transactions on
Mobile Computing (Volume: 5, Issue: 2, Feb. 2006).
4. Daniel Engels, Xinxin Fan, Guang Gong, Honggang Hu, Eric M. Smith: Hummingbird:
UltraLightweight Cryptography for Resource Constrained Devices, Springer Lecture Notes in
Computer Science, ISBN 9783642149917, 2010.
5. Payal Patel, Rajan Patel, Nimisha Patel: Integrated ECC and Blowfish for Smartphone
Security, Elsevier Procedia Computer Science, International Conference on Information
Security & Privacy (ICISP2015).
6. C. Mathias: Why mobile operating systems could fade away, Computerworld, 2007.
7. Hielko van der Hoorn: A Survey of Mobile Platforms for Pervasive Computing, University of
Groningen Computing Science Master of Science Thesis, Groningen, 2010.
8. S. Prasanna Ganesan, “An Asymmetric Authentication Protocol for Mobile Devices Using
Elliptic Curve Cryptography”, 978-1-4244-5848-6/10/ IEEE 2010.
9. Creighton T. R. Hager, Scott F. Midkiff, Jung-Min Park, Thomas L. Martin, “Performance
and Energy Efficiency of Block Ciphers in Personal Digital Assistants”, Proceedings of the
3rd IEEE Int’l Conf. on Pervasive Computing 2005.
10. Black hat report on android, http://www.blackhat.com/presentations/AndroidSurgery-PAPER.
pdf, 2009.
11. Google android architecture, http://developer.android.com/guide/basics/ android.html.
12. Microsoft mobile report, http://msdn.microsoft.com/en-us/library/ms920098.aspx.
13. Google Android security report, http://code.google.com/intl/nl/android/devel/security.html.
14. Denmark IT mobile security report, http://itsec.rwth-aachen.de/research/comparative-study-
of-protection-echanisms-for-mobile-operating-systems.
Assessment of Object Detection Using
Deep Convolutional Neural Networks
Ajeet Ram Pathak, Manjusha Pandey, Siddharth Rautaray

and Karishma Pawar
Abstract Detecting the objects from images and videos has always been the point
of active research area for the applications of computer vision and artificial intel-
ligence namely robotics, self-driving cars, automated video surveillance, crowd
management, home automation and manufacturing industries, activity recognition
systems, medical imaging, and biometrics. The recent years witnessed the boom of
deep learning technology for its effective performance on image classification and
detection challenges in visual recognition competitions like PASCAL VOC,
Microsoft COCO, and ImageNet. Deep convolutional neural networks have pro-
vided promising results for object detection by alleviating the need for human
expertise for manually handcrafting the features for extraction. It allows the model
to learn automatically by letting the neural network to be trained on large-scale
image data using powerful and robust GPUs in a parallel way, thus, reducing
training time. This paper aims to highlight the state-of-the-art approaches based on
the deep convolutional neural networks especially designed for object detection
from images.

Keywords Computer vision Deep convolutional neural networks

Deep learning Object detection
A. R. Pathak (&) M. Pandey S. Rautaray

Data Science Center of Excellence, School of Computer Engineering,
Kalinga Institute of Industrial Technology (KIIT) University, Bhubaneswar, India
e-mail: ajeet.pathak44@gmail.com
M. Pandey
e-mail: manjushapandey82@gmail.com
S. Rautaray
e-mail: sr.rgpv@gmail.com
K. Pawar
Department of Computer Engineering & IT, College of Engineering Pune (COEP),
Pune, India
e-mail: kvppawar@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_45
458 A. R. Pathak et al.
1 Introduction
Computer vision technology has been extensively used in different segments like
industry, automation, consumer markets, medical organizations, entertainment sec-
tors, defense, and surveillance, to mention a few. The ubiquitous and wide applica-
tions like scene understanding, video surveillance, robotics and self-driving cars
triggered vast research in the domain of computer vision during the most recent
decade. Visual recognition systems encompassing image classification, localization,
and detection have achieved great research momentum due to significant development
in neural networks especially deep learning, and attained remarkable performance [1].
The last 4 years witnessed a great improvement in performance of computer vision
tasks especially using deep convolution neural networks (DCNNs) [2].
Several factors are responsible for proliferation for DCNNs viz. (i) Availability
of large training datasets and fully annotated datasets (ii) Robust GPU to train
large-scale neural network models in a parallel way (iii) State-of-the-art training
strategies and regularization methods. Object detection is one of the crucial chal-
lenges in computer vision and it is efficiently handled by DCNN [3], Restricted
Boltzmaan Machine (RBM) [4], autoencoders [5], and sparse coding representation
[6]. This paper aims to highlight state-of-the-art approaches for object detection
based on the DCNNs.
The contents of the paper are portrayed as follows. Section 2 introduces object
detection. The fundamental building blocks of DCNNs from the perspective of
object detection are enunciated in Sect. 3. The state-of-the-art DCNN-based
approaches for object detection are discussed in Sect. 4. The paper is concluded in
Sect. 5.
2 Object Detection
An image or video contains single or more than one classes of real-world objects
and abstract things like human, faces, building, scene, etc. The aim of object
detection is to determine whether the given instance of the class is present in the
image, estimate the location of the instance/instances of the all the classes by
outputting the bounding box overlapping the object instance along with obtained
accuracy of detection irrespective of partial occlusions, pose, scale, lightening
conditions, location, and camera position. It is generally carried out using feature
extraction and learning algorithms. Object detection is a preliminary step for var-
ious computer vision tasks like object recognition, scene understanding from
images and activity recognition, anomalous behavior detection from videos.
Detecting instance or instances of the single class of object from image or video is
termed as single object class detection. Multi-class object detection deals with
detecting instances of more than one class of objects from the image or video.
Following challenges need to be handled while detecting objects from the images.
Assessment of Object Detection Using Deep Convolutional … 459
• Image-based challenges
Many computer vision applications require multiple objects to be detected from
the image. Object occlusions (partial/full occlusion), noise, and illumination
changes make detection challenging task. Camouflage is a challenge in which
object of interest is somewhat similar to the background scene. This challenge
needs to be handled in surveillance applications. It is also necessary to detect
objects under conditions of multiple views (lateral, front), poses, and resolutions.
The object detection should be invariant to scale, lighting conditions, color,
viewpoint, and occlusions.
• Processing challenges
Detecting objects at large scale without losing accuracy is a primary requirement
of object detection tasks. Some applications require robust and efficient detection
approaches, whereas others require real-time object detection. Thanks to specialized
hardware like GPUs and deep learning techniques which allow to train multiple
neural networks in parallel and distributed way helping to detect objects at
real-time.
3 Building Blocks of Convolutional Neural Network
DCNNs was first used for image classification. After achieving state-of-the-art
performance in image classification, DCNN has been used for more complex tasks
like object detection from images and videos.
3.1 CNN Architecture
Convolutional neural network (CNN) is a kind of feedforward neural network in

which the neurons are connected in the same way as the neurons present in the brain
of animal’s visual cortex area. Figure 1 shows the architecture of CNN. Being
hierarchical in nature, CNN encompasses convolution layer with activation function
like Rectified Linear Unit (ReLU), followed by pooling layer and eventually fully
connected layers. The pattern of CONV-ReLU-POOL is repeated in such a manner
that image reduces spatially. The neurons are arranged in the form of three
dimensions—width, height, and depth. Depth corresponds to color channels in the
image. The image to be recognized is fed as input in terms of [width height
depth]. For the desired number of filters, the image is convolved with the filter
function in order to get the specific feature. This process is repeated for the desired
number of filters and accordingly feature map is created. This is done by applying
the dot product of the weight assigned to the neuron and the specified region in the
image. In this way, the output of a neuron is computed by convolution layer.
Fig. 1 Architecture of convolutional neural network
Rectified Linear Unit (ReLU) applies activation function at elemental level making
it nonlinear. Generally, convolution layer is followed by pooling layer to reduce the
complexity of the network and the number of parameters in learning by
down-sampling the feature map along spatial dimensions. The last layer in the CNN
is a fully connected layer which gives the output of the image recognition task in
the form of scores representing the object classes. Highest score represents the
presence of a corresponding class of the object in the image.
3.2 Pooling Layers
Addition of pooling layer amidst the consecutive layers of convolutional layer

reduces the number of parameters and complexity of the network, and thus, control
overfitting. Pooling layer is translation-invariant and it takes activation maps as
input and operates on every patch of the selected map. There are various kinds of
pooling layers.
• Max pooling: In this pooling, each depth slice of input is operated using
pooling. Figure 2 shows working of max pooling where a filter of size 2 2 is
applied over a patch of an activation map with a stride of 2. The max value
among each entry of 4 numbers is chosen and stored into the matrix, getting a
spatially resized map. Another pooling approach is average pooling which takes
an average of neighborhood pixels. Max pooling gives better results compared
to average pooling [7].
• Deformation constrained pooling (Def-pooling): In order to apply deformation
of object parts along with geometric constraints and associated penalty,
def-pooling is applied [8]. It has the ability to learn deformable properties of
object parts and shares visual patterns at any level of information abstraction and
composition.
Fig. 2 Max pooling over the slice of single depth
• Fully connected layers: These layers perform high-level reasoning in CNN.

They exhibit a full connection to all the decision functions in the previous layer
and convert 2D features into one-dimensional feature vector. Fully connected
layers possess a large amount of parameters and so require powerful compu-
tational resources.
3.3 Regularization
“Regularization is defined as any modification we make to a learning algorithm

that is intended to reduce its generalization error but not its training error [9].”
Due to a large amount of parameters for training and ability of architectures to learn
more abstractions using deep learning, there are chances of obtaining negative
performance on test data from the model, i.e., model learns too well such that it
poorly generalizes in case of new data. This is known as overfitting. Regularization
strategies are required in order to stop overfitting.
4 State-of-the-Art Object Detection Approaches Using

DCNN
Table 1 compares the state-of-the-art discussion of DCNN-based approaches for

object detection. DCNNs have been extensively used for image classification and
achieved state-of-the-art results [1].
For the very first time, DCNNs have been used for object detection by Szegedy
et al. [10]. The authors have formulated object detection as a regression problem for
object bounding box masks and defined object detection as estimation and local-
ization of class and object from the image, respectively. Their approach is known as
DetectorNet in which the last layer of AlexNet [1] architecture is replaced with
regression layer in order to localize the objects using DNN-based object mask
regression. To precisely detect the multiple instances of the same object,
DetectorNet applies multi-scale box inference with refinement procedure. But, this
approach lacks multiple classes of objects for detection since it uses only single
mask regression.
To demystify the working of features extracted in CNN model and diagnose the
errors associated with the model, Zeiler and Fergus (ZFNet) [3] put forth a novel
visualization technique based on multilayered deconvolution network (deconvnet).
This model used deconvnet in order to project features back into pixel space of the
image.
Deformable DCNN (DeepID-Net) encompasses feature representation learning,
part deformation learning, context modeling, model averaging, and bounding box
location refinement [11] and uses cascaded CNN for object detection. It works on
deformable part objects. Regions with CNN features (R-CNN) [12] take an input
image and evaluate ‘n’ number of bottom-up region proposals using segmentation.
Once region proposals are obtained, it classifies proposals using class-relevant
SVMs to get classified regions. This method acts as a baseline model for a large
number of approaches put forth for object detection. Fast R-CNN [13] is the
extended version of R-CNN to improve the speed of training and testing phase and
improve the detection accuracy. Fast R-CNN suffers from the drawback of calcu-
lating the proposals for each region in the image, thus, incurring the large cost of
computation, this drawback has been removed in Faster R-CNN by Ren et al. [14].
The authors put forth region proposal network (RPN) which is a fully convolution
network in which input image is shared with the detector network, this network
simultaneously calculates object bounds and object features at each point, thus
freeing cost of region proposals.
This method merges RPN with fast R-CNN and creates a unified network. It
works on the principle of “attention mechanism” in which RPN guides network
where to search for the object bounds. In the paper, by Markus et al. [15], multiple
CNN models are used to detect objects at multiple scales. The papers by Lee et al.
[16] and Cheng et al. [17] are based on the region-based proposals. Lee et al. [16]
handled the issue of intra-class and interclass variability among objects using
multi-scale templates of CNN and non-maximum suppression method. On the other
Table 1 Comparative study of State-of-the-approaches for object detection
Paper Approach Issues Features CNN configuration Pooling Classifier Regularization Deep network training
technique
ImageNet CNN Reducing Image 5 CNN layers + 3 Max Softmax DropOut Supervised learning
[1] detection errors classification Fully conn. Layers pooling mode, Local response
for normalization and
high-resolution stochastic gradient
images descent
ZFNet [3] CNN and Demystifying Visualization of Fully supervised Max Softmax DropOut, ReLU Back-propagating
deconv. the working of intermediate CNN models pooling as activation derivative of cross
Neural intermediate features in the function entropy function for
network CNN layers and network training and SGD
classifier
DetectorNet CNN Object detection Multi-object 5 CNN layers + 2 Max Softmax Use of negative Adaptive gradient
[10] and localization detection, Fully conn. Layers pooling classes to algorithm based on
DNN-based regularize the Stochastic
regression for network optimization
localization
DeepID-Net Cascaded Handling of Generic object 13 CNN layers + 3 Def-pooling SVM Pre-training Multi-stage deep
[11] CNN deformable part detection Fully conn. Layers scheme & jointly training model
objects (O-Net and T-Net) learning feature
Assessment of Object Detection Using Deep Convolutional …
representation
Faster Region-based Computation Cost-free VGG-Net (13 CNN RoI pooling R-CNN acts as “Image-centric” “Attention”
R-CNN [14] proposal cost of region & 3 fully conn. layer a classifier sampling strategy mechanism of Neural
network region-Based proposals, layers) + ZF Net based on network,
proposal Real-time back-propagation back-propagation
handling object and SGD
translation
(continued)
463
Table 1 (continued)
464
Paper Approach Issues Features CNN configuration Pooling Classifier Regularization Deep network training
technique
Markus et al. Multi-scale Visual Multi-scale 3 stages of CNNs Max Non-maximum DropOut SGD with mini-batch
[15] model Pedestrian person processing image pooling suppression training
detection with detection patches
high accuracy
RIFD-CNN Region-based Object rotation, Rotation AlexNet + VGG-Net Max Softmax, Discrimination CNN-based
[17] proposal intra-class and invariant, pooling Linear Support regularization pre-training and
(Fisher interclass interclass & Vector Machine fine-tuning using SGD
discriminative) similarity of intra- class (SVM)
objects object detection
MSS-CNN CNN working Handling of Multi-scale, Image Max structured Training based on Multi-scale training,
[18] on image scale variation context-aware pyramid-based CNN pooling SVM, the data non-max suppression
pyramids object detection bounding box augmentation to resolve multiple
and localization regression detection issue
A. R. Pathak et al.
hand, in RIFD-CNN by Cheng et al. [17], the issue of object rotation and intra-class
and interclass variability is handled by introducing rotation-based layer and Fisher
discriminative layer in the network, respectively. For detecting small objects and
localizing them, contextual information based on the multi-scale model of CNN is
used in [18], handling the issue of variation in scaling of objects. The aforemen-
tioned approaches mainly focus on the specific challenge in object detection viz.
multi-scale model, fast and real-time detection, detection accuracy, and localization,
interclass and intra-class variation of objects.
It is worth important to amalgamate challenges and address them using unified
object detection framework applicable to detect objects in different complex sce-
narios and thereby enhance the usability of such object detection systems.
5 Conclusion
This paper compares some of the noteworthy approaches to object detection based
on DCNNs. DCNN-based approaches are found to be suitable for images and can
also be applicable to detect moving objects from the video [19]. The need of the
hour is to develop object detection model which can be generalized to work in
different application scenarios like face recognition, emotion detection, abandoned
object detection (Suspicious object detection), etc. The role of “transfer learning”
method for training deep networks would help to cope with the issue [20].
The efficacy of object detection frameworks mainly depends on the learning
mode, method of processing the images (parallel programming) and also the plat-
form (CPU, GPU). Continuous change in the scene implies a change in the behavior
of objects to be detected, therefore, it is mandatory for such systems to continuously
learn the multitude features of objects and detect them despite of a change in their
orientation, views, and forms. In addition to this, real-time detection of objects [21]
helps to take proactive measures or acts as alarming conditions for effectively
monitoring and controlling the public and private places requiring utmost security.
Object detection is a very promising area which can be applied in computer
vision and robotics systems, surveillance based on the drone cameras, etc. It is
extremely useful in places like deep mines, expeditions to exploring deep ocean
floor where human presence is not feasible.
References
1. Alex Krizhevsky, Ilya Sutskever, and Geoff Hinton. Imagenet classification with deep
convolutional neural networks. In: Advances in Neural Information Processing Systems.
(2012).
2. He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep
convolutional networks for visual recognition. In: European Conference on Computer Vision,
pp. 346–361. Springer International Publishing. (2014).
3. M.D. Zeiler, R. Fergus.: Visualizing and understanding convolutional neural networks. In:
ECCV. (2014).
4. R. Salakhutdinov, G.E. Hinton.: Deep boltzmann machines. In: AISTATS, (2009).
5. S. Rifai, P. Vincent, X. Muller, et al.: Contractive auto-encoders: explicit invariance during
feature extraction. In: ICML (2011).
6. Yang, Jianchao, Kai Yu, Yihong Gong, and Thomas Huang.: Linear spatial pyramid matching
using sparse coding for image classification. In: Computer Vision and Pattern Recognition,
CVPR. pp. 1794–1801. IEEE. (2009).
7. D. Scherer, A. Müller, S. Behnke.: Evaluation of pooling operations in convolutional
architectures for object recognition. In: ICANN. (2010).
8. W. Ouyang, P. Luo, X. Zeng, et al.: Deepid-net: Deformable deep convolutional neural
networks for object detection. In: Computer Vision and Pattern Recognition, pp. 2403–2412.
IEEE. (2015).
9. I. Goodfellow, Y. Bengio, and A. Courville.: Deep Learning. MIT Press. (2016).
10. C. Szegedy, A. Toshev, D. Erhan.: Deep neural networks for object detection. In: Proceedings
of the NIPS. (2013).
11. W. Ouyang, P. Luo, X. Zeng, et al.: DeepID-Net: multi-stage and deformable deep
convolutional neural networks for object detection. In: Proceedings of the CVPR. (2015).
12. R. Girshick, J. Donahue, T. Darrell, et al.: Rich feature hierarchies for accurate object
detection and semantic segmentation. In: Proceedings of the CVPR. (2014).
13. R. Girshick.: Fast R-CNN. In: ICCV. (2015).
14. S. Ren, K. He, R. Girshick, and J. Sun.: Faster R-CNN: Towards Real-Time Object Detection
with Region Proposal Networks. In: TPAMI, pp. 91–99. IEEE. (2016).
15. Eisenbach, Markus, Daniel Seichter, Tim Wengefeld, and Horst-Michael Gross.: Cooperative
multi-scale Convolutional Neural Networks for person detection. In: International Joint
Conference on Neural Networks (IJCNN), pp. 267–276. IEEE. (2016).
16. B. Lee, E. Erdenee, S. Jin, and P. K. Rhee. Efficient object detection using convolutional
neural network-based hierarchical feature modeling. In: Signal, Image Video Process. vol. 10,
no. 8, pp. 1503–1510, (2016).
17. Cheng, Gong, Peicheng Zhou, and Junwei Han.: RIFD-CNN: Rotation-invariant and fisher
discriminative convolutional neural networks for object detection. In: Computer Vision and
Pattern Recognition, pp. 2884–2893. IEEE. (2016).
18. E. Ohn-Bar and M. M. Trivedi.: Multi-scale volumes for deep object detection and
localization. In: Pattern Recognition, vol. 61, pp. 557–572. Elsevier (2017).
19. S. H. Shaikh, K. Saeed, and N. Chaki.: Moving Object Detection Approaches, Challenges and
Object Tracking. In: Moving Object Detection Using Background Subtraction, pp. 5–14.
Springer International Publishing (2014).
20. Dauphin, G.M. Yann, X. Glorot, S. Rifai, Y Bengio, I. Goodfellow, E. Lavoie, X. Muller
et al.: Unsupervised and transfer learning challenge: a deep learning approach. In: ICML
Workshop on Unsupervised and Transfer Learning, pp. 97–110. (2012).
21. P. Viola, M. Jones.: Rapid object detection using a boosted cascade of simple features. In:
Computer Vision and Pattern Recognition. pp. I-511-I-518. IEEE. (2001).
Implementation of Credit Card Fraud
Detection System with Concept Drifts
Adaptation
Anita Jog and Anjali A. Chandavale
Abstract There is a large number of credit card payments take place that is tar-
geted by fraudulent activities. Companies which are responsible for the processing
of electronic transactions need to efficiently detect the fraudulent activity to
maintain customers’ trust and the continuity of their own business. In this paper, the
developed algorithm detects credit card fraud. Prediction of any algorithm is based
on certain attribute like customer’s buying behavior, a network of merchants that
customer usually deals with, the location of the transaction, amount of transaction,
etc. But these attribute changes over time. So, the algorithmic model needs to be
updated periodically to reduce this kind of errors. Proposed System provides two
solutions for handling concept drift. One is an Active solution and another one is
Passive. Active solution refers to triggering mechanisms by explicitly detecting a
change in statistics. Passive solution suggests updating the model continuously in
order to consider newly added records. The proposed and developed system filters
80% fraudulent transactions and acts as a support system for the society at a large.

Keywords Credit card Concept drift adaptation Credit card fraud detection
Network-based extension
1 Introduction
Credit cards and debit cards are used extensively in today’s world. People prefer
executing banking transaction online rather than going to the branch. This is due to
easy availability of modern resources like laptops, phone, and tablets. Most of the
banks switched to centralize processing to support this trend. Another reason for
credit card gaining popularity is online shopping trend. Lots of deals, promotions
A. Jog (&) A. A. Chandavale

Department of Information Technology, MIT College of Engineering, Pune, India
e-mail: Anita.visal@gmail.com
A. A. Chandavale
e-mail: Anjali.chandavale@mitcoe.edu.in

https://doi.org/10.1007/978-981-10-7245-1_46
468 A. Jog and A. A. Chandavale
attract customers to buy online. Online shopping allows customers to explore more
items in short time. Customers can also compare the prices of the same item from
different vendors. All these online activities are more prone to fraudulent cases due
to lack of proper training or awareness before using digital transaction. During
online transactions, personal data like account number, password, birth date, credit
card details, etc. is exposed over the network resulting in financial fraud. Financial
fraud causes loss of huge amount worldwide each year. There is a huge amount of
loss occurs due to online fraud. As per cybercrime report [1], total card sales
volume is $28.844 trillion while the losses occurred from fraud on cards is $16.31
billion. Comparing to past year fraud increased by 19%, while the overall volume of
card sales only grew by only 15%. Companies that are responsible for processing
electronic transactions need to efficiently detect the fraudulent activity to maintain
customers’ trust and the continuity of their business. Data mining offers a number of
techniques to analyze patterns in data, differentiating genuine transaction from
suspicious transactions. The most important aspect of fraud detection is to correctly
identify the atypical character of fraud, since the fraudulent transactions are very
few as compared to the legitimate transactions. A critical step in an efficient fraud
detection process is creating a set of significant characteristics that capture irregular
behavior. The detection of fraud is a complex computational task and there is no
ideal system that accurately predicts fraudulent transaction. These systems only
predict the probability of the transaction to be fraudulent. The paper is divided into
following sections. Section 2 presents the related work. Section 3 describes the
proposed work. In Sect. 3.3.3 expected result is outlined and Sect. 4 describes the
conclusion.
2 Related Work
Various data mining techniques exist for fraud detection [2, 3, 4, 5]. Support Vector
machine, genetic programming, Artificial Intelligence, Neural networks, Decision
tree are few of them [6, 7, 8, 9]. Using decision tree, the system is described by
Yusuf Sahin and Serol Bulkan. In this paper, cost-sensitive decision tree approach
is suggested which minimizes the sum of misclassification costs while selecting the
splitting attribute at each nonterminal node [10]. Author has compared the per-
formance of this approach with the well-known traditional classification models on
a real-world credit card data set. Andrea Dal Pozzolo, Giacomo Boracchi designed
two fraud detection systems using ensemble and sliding window approach [11].
Customer feedbacks and alerts provided by the system are used as supervised
samples. One more novel approach is described by Véronique Van Vlasselaer and
Cristián Bravo [12]. This approach is a combination of the characteristics of
incoming transactions along with the customer spending history based on the
fundamentals of RFM (Recency–Frequency–Monetary); and characteristics by
exploring the network of credit card holders and merchants. A time-dependent
suspiciousness score is calculated for each network object. As seen from the
Implementation of Credit Card Fraud Detection System … 469
literature, it is essential to have a more efficient system capable of detecting the

frauds at the very first instance. The following section describes the proposed and
developed system based on the concept drift adaption which values the customer
feedback. It performs a proactive heuristic search across all the customers.
3 Proposed Work
The proposed and developed system works on the principle of the active solution to
detect the fraud. It follows the layered architecture namely data layer, utility layer,
manager layer, and controller layer. The data layer is responsible for storing the
historical data, transactional data, the model which contains the set of attributes to
validate for each customer. Utility layer acts as a support layer. Manager layer is
responsible for executing each job separately while controller encapsulates all the
flows for every user action. The concept drift adaptation technique is implemented
by invoking the scheduler periodically to update the model. Below section covers
the overview of how credit card processing happens followed by the technical
details of the proposed system.
3.1 Overview of Credit Card Processing
1. Authorization: It is the first step in which a merchant swipes the card. Then, the
login details are then passed to an acquirer. Acquirer authorizes the transaction.
The acquirer is nothing but a bank. It is responsible for processing and settling
credit card transactions by consulting a card issuer. The acquirer then sends the
request to the card-issuing bank. Card-issuing bank either authorizes or denies
the request. Depending upon the decision the merchant is allowed to proceed
with the sale.
2. Batching: Merchant reviews all the day’s sales to make sure all of them were
authorized and endorsed by the cardholder. It then submits all the transactions in
a batch, to the acquirer to process payment.
3. Clearing: Once acquirer receives the batch, it sends it through the card network.
Card network is responsible for routing each transaction to the appropriate
issuing bank as shown in Fig. 2. The issuing bank then deducts its interchange
fees and sends the remaining amount to acquirer through the network. The
interchange fee is the amount paid by merchants to a credit card issuer and a
card network for accepting credit cards. It usually ranges between 1 and 3%.
Card network is Visa, MasterCard, or other networks that act as a mediator
between an acquirer and an issuer. They are responsible for authorizing credit
card transactions.
4. Funding: This is the last step in processing a credit card. Acquirer deducts its
discount charges and sends the remaining amount to the merchant after
receiving payment from the issuer. Thus, the merchant got paid for the trans-
action. Thus, the cardholder is charged. Merchants pay a discount fee to
acquirers to incur the processing cost of the credit cards payment.
3.2 Proposed System
Figure 1 shows the block diagram of the proposed system. Using the historical
transactional data, the model builder will the model. This model is nothing but the
set of differentiating attributes for identifying the fraudulent transaction. This model
is used by online fraud detector to decide if each incoming transaction contains any
suspicious characteristic. Based on the decision given by online fraud detector,
transaction processor either aborts the transaction or proceeds for further process-
ing. Separate scheduler which runs daily triggers the offline calibrator. The offline
calibrator is responsible for correcting the model as per customer feedback. This
model is updated weekly to consider newly added transactions. The application has
following modules starting from user login, viewing account statement, down-
loading in PDF or excel format to actual transaction execution which covers dif-
ferent validations. Some validations are performed during actual transaction while
some heuristic checks are applied across all customers, where the pattern is rec-
ognized and suspicious transactions are marked in the database table.
As shown in Fig. 2, the proposed system is broadly divided into four layers
1. Controller Layer: It controls the flow of actions. It calls appropriate manager to
execute each action. It calls alert utility depending upon the output received
from manager layer. For every user action, there is one controller class. It
decides different steps that need to be performed to accomplish given task.
2. Manager Layer: It executes below described action using database tables. It
sends notifications, alerts to the controller.
(i) Authentication Manager: First calls authentication manager to verify the
userId, password. If the authentication manager successfully returns as a
valid user, then the transaction controller passes the request to the pre-
liminary validator.
(ii) Fraud detection manager: After passing all the validation checks, the
controller calls fraud detection manager to detect any unusual behavior of
transaction. It performs few fraud detection checks which include
amount, location, valid merchant, last transaction time, etc.
(iii) Transaction Manager: Finally, the controller calls Transaction manager to
execute the transaction dealing with debit/credit, transfer, etc.
3. Utility Layer: It takes care of reporting. Notifications, Alerts triggered by
manager layer are converted into appropriate Actions.
Fig. 1 Block diagram
4. Database Layer: It consists of transactional data, historical data, rules model

data. Transactions flagged by the algorithm kept in the separate table. Rules
Model table is populated based on historical data. And, it gets updated peri-
odically using self-learning algorithm. The following figure shows a quick
overview of the database layer.
3.3 Implementation
The system is implemented in JAVA, JSP. The model is determined by WEKA

output. SQL server database is used for backend data storage. For download
functionality in PDF format, PDFBOX library is used and APACHE POI library is
used to support Excel format. After testing the system, it is found that using the
fraud detection system, 80% fraudulent transactions are filtered. Heuristics searches
also help to detect fraud at the initial instant.
3.3.1 DB Design
The main tables used for fraud detection are:
1. Customer_Vendor_tbl is used to store the RFM attributes, i.e., Recency, fre-

quency, and monetary value for each customer–vendor pair. For each incoming
transaction, the frequency is checked. If it is monthly, then current transaction

date and last transaction date for the corresponding vendor is compared. If the
difference is less than a month, then after considering certain threshold, the
transaction is put on HOLD. Customer care will reach out to the customer to
confirm if the transaction is genuine or not.
2. Account_validation_tbl is storing all the different validations for each account.
Attribute name in this table represents attribute like location, amount, etc.
attribute value represents set of allowed values separated by a comma. For
example, it is comma separated list of cities, where the customer can use the
card for the transaction. If new city is encountered then the transaction is put on
HOLD.
3. Fraud_log_tbl is used to log every transaction that is blocked. The design allows
having different attributes for different accounts.
3.3.2 Algorithm
When a customer performs a transaction, then the request comes to credit card fraud
detection system and the following algorithm is executed.
Step 1: Perform preliminary checks given below on the incoming transaction.
• Balance Check: The credit card balanced is checked to confirm if the there is
sufficient balance is the credit card.
• Validate customer: Customer is validated to check if he has paid previous bills.
• Validate card status: Card status is verified to check if it is in active state.
Step 2: Perform online fraudulent checks given below:
Min/Max amount check: Using the historical data, threshold values for the
account are determined and stored in the database. For the incoming transaction, the
transaction amount is compared to check if it falls within the account threshold. If
transaction amount falls out of threshold, it is marked as suspicious.
• Customer–Vendor validation: Using the historical data, the relationship between
customer and vendor is determined with respect to frequency and monetary
values. For each customer–vendor pair, the minimum amount transacted, max
amount transacted, and frequency of the transaction is calculated using historical
data and stored in the database. The incoming transaction is verified against
corresponding customer–vendor pair for threshold values and frequency. For
example, customer ABC transacts with vendor PQR every month between the
amounts 1000 to 3000. If the incoming transaction has customer ABC and
vendor PQR with amount 7000, the system marks it as suspicious.
• Location check: Using historical data, different locations are identified for each
customer. These locations are stored as comma separated list in the database. If
the incoming transaction has a location which was never found before, it is
marked as suspicious.
Step 3: If anyone of above check fails, then transaction is suspicious and is put on
HOLD
All transactions that are marked as suspicious then confirmed with the customer.
Below algorithm is executed when confirmation is received from the customer.
Step 1: Notify the customer about fraud transaction. Show the validations status to
the customer. If the customer confirms the transaction to be fraudulent, abort the
transaction. Mark transaction status as “CANCELLED” and no further action is
taken
Step 2: If a customer claims that the transactions are genuine, then proceed with
transaction successfully and mark as COMPLETED.
Step 3: After the transaction is executed successfully, update the model as given
below.
• Recalculate the account threshold values.
• Recalibrate customer–vendor frequency, amount threshold.
• If there is new location found in the current transaction, then update the list of
locations.
With this updated model, the similar transaction will be executed successfully in
future.
Step 4: Recalibrate amount threshold, and location list at every end of the day to
consider newly added records.
3.3.3 Project Results
Below are the screenshots shown in Figs. 3 and 4 for transaction status, fraud
summary, and action taken on suspicious transactions after receiving feedback from
the customer.
3.4 Fraud Detection System Evaluation
The data is downloaded from kaggle.com. The datasets contain transactions made
by credit cards in September 2013 by European cardholders. Along with this,
synthetic data is also added to the database to test different scenarios. The resulted
dataset has 152,706 credit card transactions of 120 cardholders. 151,768 transac-
tions legitimate are and 1026 transactions are fraudulent. Criteria used to evaluate
the system is alarm rate, sensitivity, false positive rate, precision, negative pre-
dictive value, and accuracy as shown in Table 1 [13–15].
Fig. 3 Transaction status of fraud check
Fig. 4 Fraud summary page

Table 1 Evaluation results Sensitivity Precision Accuracy

43.24% 11.20% 80.12%
4 Conclusion
Credit card fraud detection algorithms have emerged remarkably over last few
years. The algorithm used for credit card detection needs to be self-learning and
periodically updated. The approach suggested in this paper is most suitable for the
changing environment since it changes dynamically. The work is a hybrid system
based on Network-based extension and concept drift adaptation. Using past history,
the model is determined. The model consists of different conditional attributes. The
set of attributes can be different for different customers. Hence, the suggested model
is flexible and dynamic with 80% fraudulent transactions filtering rate. Thus, the
proposed and developed system is providing the essential backbone to promote the
plastic money.
References
1. http://www.pymnts.com/news/2015/global-card-fraud-damages-reach-16b/.
2. Emanuel MinedaCarneiro, “Cluster Analysis and Artificial Neural Networks: A Case Study in
Credit Card Fraud Detection,” in 2015 IEEE International Conference.
3. Dhiya Al-Jumeily, “Methods and Techniques to Support the Development of Fraud Detection
System”, IEEE 2015.
4. Mukesh Kumar Mishra, “A Comparative Study of Chebyshev Functional Link Artificial
Neural Network, Multi-layer Perceptron and Decision Tree for Credit Card Fraud Detection”,
2014 13th International Conference on Information Technology.
5. Andrea Dal Pozzolo, Olivier Caelen,” Learned lessons in credit card fraud detection from a
practitioner perspective”, Expert Systems with Applications 41, 2014.
6. V. Mareeswari, “Prevention of Credit Card Fraud Detection based on HSVM”. 2016 IEEE
International Conference On Information Communication And Em- bedded System.
7. Carlos A. S. Assis, “A Genetic Programming Approach for Fraud Detection in Electronic
Transactions” in Advances in Computing and Communication Engineering (ICACCE), 2015
Second International Conference.
8. Dustin Y. Harvey, “Automated Feature Design for Numeric Sequence Classification by
Genetic Programming”, IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION,
VOL. 19, NO. 4, AUGUST 2015.
9. Kang Fu, Dawei Cheng, Yi Tu, and Liqing Zhang, “Credit Card Fraud Detection Using
Convolutional Neural Networks”, Neural Information Processing, Springer.
10. Sahin Yusuf, BulkanSerol, DumanEkrem,” A Cost-Sensitive Decision Tree Approach for
Fraud Detection”, Expert Systems with Applications,vol.40, pp. 5916–5923, 2013.
11. Andrea Dal Pozzolo, “Credit Card Fraud Detection and Concept-Drift Adaptation with
Delayed Supervised Information”.
12. Véronique Van Vlasselaer, “APATE: A novel approach for automated credit card transaction
fraud detection using network-based extensions” published in Decision Support Systems
2015.
13. Yiğit Kültür, “A Novel Cardholder Behavior Model for Detecting Credit Card Fraud”, IEEE
international conference on commuting and communication engineering, 2015.
14. https://www.creditcards.com/credit-card-news/assets/HowACreditCardIsProcessed.pdf.
15. https://www.kaggle.com/dalpozz/creditcardfraud.
Intelligent Traffic Control by Multi-agent
Cooperative Q Learning (MCQL)
Deepak A. Vidhate and Parag Kulkarni
Abstract Traffic crisis frequently happens because of traffic demands by the large
number vehicles on the path. Increasing transportation move and decreasing the
average waiting time of each vehicle are the objectives of cooperative intelligent
traffic control system. Each signal wishes to catch better travel move. During the
course, signals form a strategy of cooperation in addition to restriction for neigh-
boring signals to exploit their individual benefit. A superior traffic signal scheduling
strategy is useful to resolve the difficulty. The several parameters may influence the
traffic control model. So it is hard to learn the best possible result. The lack of
expertise of traffic light controllers to study from previous practice results makes
them to be incapable of incorporating uncertain modifications of traffic flow.
Defining instantaneous features of the real traffic scenario, reinforcement learning
algorithm based traffic control model can be used to obtain fine timing rules. The
projected real-time traffic control optimization model is able to continue with the
traffic signal scheduling rules successfully. The model expands traffic value of the
vehicle, which consists of delay time, the number of vehicles stopped at the signal,
and the newly arriving vehicles to learn and establish the optimal actions. The
experimentation outcome illustrates a major enhancement in traffic control,
demonstrating the projected model is competent of making possible real-time
dynamic traffic control.
Keywords Cooperative learning Multi-agent learning Q learning
D. A. Vidhate (&)
Department of Computer Engineering, College of Engineering, Pune, Maharashtra, India
e-mail: dvidhate@yahoo.com
P. Kulkarni
iKnowlation Research Lab. Pvt. Ltd, Pune, Maharashtra, India
e-mail: parag.india@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_47
480 D. A. Vidhate and P. Kulkarni
1 Introduction
Thousands of vehicles distribute in a large and board urban area. It is a difficult and
complicated work to effectively take care of such a large scale, dynamic, and
distributed system with a high degree of uncertainty [1]. Though the number of
vehicles is getting more and more in major cities, most of the current traffic control
methods have not taken benefit of an intelligent control of traffic light [2]. It is
observed that sensible traffic control and enhancing the deployment effectiveness of
roads is an efficient and cost-effective technique to resolve the urban traffic crisis in
majority urban areas [3]. Major vital part of the intelligent transportation system is
traffic signal lights control strategy becomes necessary [4]. There are so various
parameters that have an effect on the traffic lights control. The static control method
is not feasible for rapid and irregular traffic flow. The paper suggests a dynamic
traffic control framework which is based on reinforcement learning [5]. The rein-
forcement learning can present a very crucial move to resolve the above cited
problems. It is effectively deployed in resolving various problems [6]. The
framework defines different traffic signal control types as action selections; the
number of vehicles arriving and density of vehicle at a junction is observed as the
context of environment and common signal control indicators, including delay time,
the number of stopped vehicles and the total vehicle density are described as
received rewards. The paper is divided as Sect. 2 gives the insights about the related
work done in the area of traffic signal control. Section 3 describes Multi-agent
Cooperative Q learning algorithm (MCQL). Section 4 explains about the system
model. Experimental results are given in Sect. 5. The conclusion is presented in the
Sect. 6.
2 Related Work
The traffic control systems can be categorized into offline traffic control systems and
online traffic control systems. Offline methods make use of theoretical move toward
optimizing the controls. Online methods regulate traffic regulator period dynami-
cally as per instantaneous traffic conditions. Many achievements in collaborative
traffic flow guidance and control strategy have been made. The approach of [7] the
transportation industry. By means of F-B method approach, the traffic jam difficulty
was partially resolved [7]. After that, several enhanced methods based on the F-B
approach had developed [8]. Driving reimbursement coefficient and delay time was
used to estimate the effectiveness of time distribution system given in [9]. The
approach reduces the delay of waiting time, making the method appear to be sharp
and sensible. There is a need to discover a new proper technique as this method
could hardly solve the heavy traffic problem. Traffic congestion situation has been
addressed using intelligent traffic control in [10] but congestion problems among
neighboring junctions required better technique. The local synchronization
Intelligent Traffic Control by Multi-agent Cooperative … 481
demonstrated a fine result to this problem discussed in [11]. Because of compli-

cation and unpredictability, it is of limited opportunity to construct a precise
mathematical model for traffic system in advance [12]. It has turned out to be a style
to resolve traffic problems by taking benefit of computing expertise and machine
learning [13]. Among many machine intelligence methods, reinforcement learning
is feasible for the finest control of the transport system [14]. The study using the
learning algorithm [15] achieved online traffic control. The approach was able to
choose the optimal coordination model under different traffic conditions. Some
applications [16] that utilize learning algorithm have received much significant
effect. A paper implemented an online traffic control through learning algorithm,
yielding good effort in the normal state of traffic congestion [17].
2.1 Traffic Estimation Parameters
Signal lights control has a very crucial responsibility in traffic management.

Normally applied traffic estimation parameters [18] comprises of delay time, the
number of automobiles stopped at a signal, and a number of newly arriving cars.
Delay Time. The delay between the real time and theoretically calculated time
for a vehicle to leave a signal is defined as the delay time. In practice, we can get
total delay time during a certain period of time and the average delay time of a cross
to evaluate the time difference. The more delay time indicates the slower average
speed of a vehicle to leave a signal.
Number of Vehicles Stopped. How many vehicles are waiting behind stop line to
leave the road signal gives the number of vehicles stopped. The indicator [18] is
used to measure the smooth degree of the road as well as the road traffic flow. It is
defined as
stop ¼ stopG þ stopR, ð1Þ
where stopG is the number of automobiles stopped before the green light and stopR
is the number of vehicles stopped at the red light.
Number of Vehicles Newly Arrived. The ratio of the actual traffic flow to the
maximum available traffic flow gives the signal saturation. Newly arrived vehicle is
calculated as
traffic flow
S¼ ; ð2Þ
ðdr sf Þ
where dr is the ratio of red light duration to green light duration and sf is traffic flow
of the signal.
Traffic Flow Capacity. The highest number of vehicles crossing through the
signal is shown by traffic flow capacity. The result of the signal control strategy is
given by the indicator. Traffic signal duration is associated with traffic flow capacity.
2.2 Reinforcement Learning
Reinforcement learning describes about maximize the numerical reward and

mapping the state into actions through different way [18, 19]. Signal agents identify
situation and responses from traffic scenarios, learn information depend on learning
algorithms. Then it makes action choice with respect to its own accumulated
information. Increase the traffic flow and decrease the average delay time is the
purpose of traffic light control system. In this traffic arrangement, signals at one
intersection coordinate with the signals at other intersection for better transport
flow. Throughout the process, signals at each intersection develop a strategy of
cooperation to maximize their individual benefit. Cooperation between agents is
accomplished by distributing partial information of the states with the adjacent
agents. In view of changing scenarios of the real traffic situation, multi-agent
cooperative Q learning algorithm (MCQL) is developed for intelligent traffic con-
trol approach [19–21]. The Q update equation is given as:
Qðs, aÞ Qðs; aÞ þ aðr þ cQðs'; a'Þ Qðs; aÞÞ; ð3Þ
where the state, action, immediate reward, and cumulative reward at time t corre-
spondingly stands for st, at, rt, and Qt. Qt(s, a) is called policy function. a 2 [0, 1]
refers to the learning rate and c 2 [0, 1] indicates the discount rate.
3 Multi-agent Cooperative Q Learning (MCQL)
Synchronization in multi-agent reinforcement generates a complex set of presen-

tations achieved from the different agents’ actions. A portion of good performing
agent group (i.e., a general form) is shared among the different agents via a specific
form(Qi) [22]. Such specific forms embrace the limited details about the environ-
ment. Such strategies are incorporated to improve the sum of the partial rewards
received using satisfactory cooperation prototype. The action plans or forms are
created by the way of multi-agent Q learning algorithm by constructing the agents
to travel for the most excellent form Q* and accumulating the rewards. When forms
Q1, …, Qx are incorporated, it is possible to construct new forms that is General
Form (GF = {GF1, …, GFx}), in which GFi denotes the outstanding reinforce-
ment received by agent i all through the knowledge mode [5]. Algorithm 1
expresses get_form algorithm that splits the agents’ knowledge. The forms are
designed by the Q learning used for all prototypes. Outstanding reinforcements are
liable for GF which compiles all outstanding rewards. It will be shared by the way
of the added agents [21, 22]. Transforming incomplete rewards as GF is considered
for outstanding reinforcements to achieve the cooperation between the agents.
A status utility gives the outstanding form among the opening states and closing
state for a known form which approximates GF with the outstanding
reinforcements. The status utility is calculated by summation of steps the agent

needed to get to the destination at the closing state and the sum of the received
status in the forms among each opening and the closing state [22].
Algorithm 1 Cooperative Multi-agent Q Learning Algorithm
Algorithm get_form (I, technique)
1. Initialization Qi(s, a) and GFi(s, a)
2. Coordination of the agents i є I;
3. Agents collaborate till the target state is found; episode episode +1
4. Renewal rule which estimates the reward value;
Qðs, aÞ Qðs; aÞ þ aðr þ cQðs0 ; a0 Þ Qðs; aÞÞ
5. Fcooperate (episode, tech, s, a, i);

6. Qi GF that is Qi of agent i є I is updated by means of GFi.
Cooperation Models
Various collaboration methods for cooperative reinforcement learning are pro-
posed here:
(i) Group model—reinforcements are distributed in a sequence of steps.
(ii) Dynamic model—reinforcements are distributed in each action.
(iii) Goal-oriented model—distributing the sum of reinforcements when the agent
reaches the goal-state (Sgoal).
Algorithm 2 Cooperation model

Fcooperate (episode, tech,s,a,i)/*cooperation between agents as four cases*/
q: count of sequence
1. Switch between cases
2. In case of Group method
if episode mod q = 0 then
get_Policy(Qi, Q*,GFi);
3. In case of Dynamic method
Px
r j¼1 Qjðs; aÞ;
Qi(s, a) r;
get_Policy(Qi, Q*, GFi);
4. In case of Goal-oriented method
if S =PSgoal then
x
r j¼1 Qjðs; aÞ;
Qi(s,a) r;
get_Policy(Qi,Q*,GFi);
Algorithm 3 get_Policy
Function get_Policy(Qi, Q*, GFi) /*find out universal agent policy */
1. for loop for each agent i є I
2. for loop for each state s є S
3. if value(Qi, s) value(Q*, s) then
GFi(s,a) Qi(s,a);
4. end for loop
Group Model: During the learning process each agents receives reinforcements
for their actions. At the last part of the series (step q), each agent gives out the cost
of Qj to GF. If reward value is suitable, that is it improves the usefulness of another
agents for given state the agents will afterward donate to these rewards [21–23].
4 Model Design
In a practical environment, traffic flows of four signals with eight flow directions are
considered for the development. As shown in Fig. 1, there are altogether four
junctions at each signal agent, i.e., agent 1, agent 2, agent 3, and agent 4 for Ja, Jb,
Jc, and Jd, respectively.
The control coordination between the intersections can be viewed as a Markov
process, denoted by ⟨S, R⟩, where represents the state of the intersection, stands for
the action for traffic control, and indicates the return attained by the control agent.
Definition of State: Agent receives instantaneous traffic state and then returns
traffic control decision by the present state of the road. Essential data such as a
number of vehicles newly arriving and number of vehicles currently stopped at
signal are used to reflect the state of road traffic.
Fig. 1 Traffic flow and

control of four intersections Xin
with eight flow directions Signal Signal
Agent 1 Agent 2
Signal Signal
Agent 4 Agent 3
Number of vehicles newly arriving = Xmax = x1, x2, x3, x4 = 10

Number of vehicles currently stopped at junction J = Imax = i1, i2, i3, i4 = 20.
State for agent 1 become (x1, i1), e.g., (5,0) that means 5 new vehicles are
arriving to agent 1 with 0 vehicles are stopped at junction 1. State for agent 2
become (x2, i2), State for agent 3 become (x3, i3) and State for agent 4 become (x4,
i4). State of the system become Input as (xi, ii). Here, it can get together 200
possible states by combining maximum 10 arriving vehicle and maximum 20
vehicles stopped at signal (10 20 = 200).
Definition of Action: In reinforcement learning framework, policy denotes the
learning agent activities at a given time. Traffic lights control actions can be cat-
egorized to three types: no change in signal duration, increasing signal duration,
reducing signal duration.
Value Action
1 No change in signal duration
2 Increase in signal duration
3 Reduce the signal duration
Action set for signal agent 1 is A1 = {1, 2, 3}, action set for signal agent 2 is
A2 = {1, 2, 3} and action set for signal agent 3 is A3 = {1, 2, 3}.
Each of them is for one of the following actual traffic scenarios. The strategy of
no change in signal duration is used in the case of the normal traffic flow when the
lights control rules do not change. The strategy increasing the signal duration is
mostly used in the case that in one route traffic flow is stopped and the other route is
regular. Increasing the signal duration extends the traffic flow while signal lights are
still timing. The strategy decreasing signal duration is mostly used in the case that
in one route of traffic flow is little while that of the other route is big. Decreasing
signal light duration reduces the waiting time of the other route and lets vehicles of
that route pass the junction faster, while signal lights keep timing [23, 24].
Definitions of Reward and Return: Agent makes signal control decisions under
diverse traffic circumstances and returns an action sequence. We use traffic value
display to estimate the traffic flows as [26].
Reward is calculated in the system as given below:
Assume current state i = (xi, ii) and next state j = (xj, ij).
current state i ! next state j
Case 1: [xi, ii]![xi, ii−1]
½Xmax ¼ 10; Imax ¼ 20 ! ½Xmax ¼ 10; Imax ¼ 19
That means: one vehicle from currently stopped vehicle is passing the junction
Case 2: [xi, ii] ! [xi+1, ii−1]
That means: one newly arrived vehicle at junction and one vehicle is passing
Case 3: [xi, ii] ! [xi, ii−3]
That means: More than one stopped vehicles are passing the junction
Case 4: [xi, 0] ! [xi+1, 0]
That means: new one new vehicle is arriving and no stopped vehicle at the
junction. Depending on above state transitions from current state to next state,
reward is calculated as [24]
Reward is rp ði, p, jÞ ¼ 1 if x01 ¼ x1 þ 1. . .. . .. . .:Case 4

¼ 2 if i01 ¼ i1 1. . .. . .. . .::Case 1
¼ 3 if i01 ¼ i1 3. . .. . .. . .. . .Case 2 & 3
¼ 0 otherwise:
The study learns a controller with learning rate = 0.5, discount rate = 0.9, and
k = 0.6. During the learning process, the cost was updated 1000 with 6000
episodes.
Figure 2 shows that delay time versus a number of state given by simple Q
learning (without cooperation) and group method (with cooperation). Delay time
obtained by cooperative methods, i.e., group method is much less than that of
without cooperation method, i.e., simple Q learning for agent 1 in the multi-agent
scenario.
Figure 3 shows that delay time versus a number of state given by simple Q
learning (without cooperation) and group method (with cooperation). Delay time
Fig. 2 States versus Delay time for Agent 1 by Q learning and Group Method
Fig. 3 States versus Delay Time for Agent 2 by Q learning and Group Method
without cooperation method, i.e., simple Q learning for agent 2 in the multi-agent
scenario.
Figure 4 shows that delay time vs number of state given by simple Q learning
(without cooperation) and group method (with cooperation). Delay time duration
without cooperation method, i.e., simple Q learning for agent 3 in multi-agent
scenario.
Fig. 4 States versus Delay Time for Agent 3 by Q learning and Group Method
6 Conclusion
Because traffic control system is so complicated and variable that a Q learning

model (without cooperation) with defined strategy can rarely manage with the traffic
jam and sudden traffic accidents which actually may occur at any time, the demand
for combining timely and intelligent traffic control policy with real-time road traffic
is getting more and more urgent. Reinforcement learning gathers tests and infor-
mation by keeping communication with the situation. Although it usually needs a
long duration to complete learning, it has good learning ability to a complex sys-
tem, enabling it to handle unknown complex states well. The application of rein-
forcement learning in traffic management area is gradually receiving more and more
concerns. The paper proposed a cooperative multi-agent reinforcement
learning-based models (CMRLM) for traffic control optimization. The actual con-
tinuous traffic states are discretized for the purpose of simplification. We design
actions for traffic control and define reward and return by mean of traffic cost which
combines with multiple traffic capacity indicators.
References
1. F. Zhu, J. Ning, Y. Ren, and J. Peng, “Optimization of image processing in video-based traffic
monitoring,” ElektronikairElektrotechnika, vol.18, no.8, pp. 91–96, 2012.
2. B. de Schutter, “Optimal traffic light control for a single intersection,” in Proceedings of the
American Control Conference (ACC ’99), vol. 3, pp. 2195–2199, June 1999.
3. N. Findler and J. Stapp,“A distributed approach to optimized control of street traffic signals,”
Journal of Transportation Engineering, vol.118, no.1, pp. 99–110, 1992.
4. L. D. Baskar and H. Hellendoorn, “Traffic management for automated highway systems using
model-based control,”IEEE Transactions on Intelligent Transportation Systems, vol. 3, no. 2,
pp. 838–847, 2012.
5. R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, MIT Press,
Cambridge, Mass, USA, 1998.
6. Artificial Intelligence in Transportation Information for Application, Transportation

Research CIRCULAR, Number E–C 113, Transportation On Research Board of the
National Academies, January 2007.
7. Deepak A. Vidhate, Parag Kulkarni “New Approach for Advanced Cooperative Learning
Algorithms using RL methods (ACLA)” VisionNet’16 Proceedings of the Third International
Symposium on Computer Vision and the Internet, ACM DL pp 12–20, 2016.
8. K. Mase and H. Yamamoto, “Advanced traffic control methods for network management,”
IEEE Magazine, vol. 28, no. 10, pp. 82–88, 1990.
9. Deepak A. Vidhate, Parag Kulkarni “Innovative Approach Towards Cooperation Models for
Multi-agent Reinforcement Learning (CMMARL)” in Smart Trends in Information
Technology and Computer Communications, Springer Nature, Vol 628, pp 468–478, 2016.
10. L. D. Baskar, B. de Schutter, J. Hellendoorn, and Z. Papp, “Traffic control and intelligent
vehicle highway systems: a survey,” IET Intelligent Transport Systems, vol. 5, no. 1, pp. 38–
52, 2011.
11. M. Broucke “A theory of traffic flow in automated highway systems,” Transportation
Research C, vol. 4, no. 4, pp. 181–210, 1996.
12. D. Helbing, A. Hennecke, V. Shvetsov, and M. Treiber, “Micro and macro-simulation of
freeway traffic,” Mathematical and Computer Modelling, vol. 35, no. 5–6, pp. 517–547, 2002.
13. S. Zegeye, B. de Schutter, J. Hellendoorn, E. A. Breunesse, and A. Hegyi, “A predictive
traffic controller for sustainable mobility using parameterized control policies,” IEEE
Transactions on Intelligent Transportation Systems, vol. 13, no. 3, pp. 1420–1429, 2012.
14. Deepak A. Vidhate, Parag Kulkarni “Enhancement in Decision Making with Improved
Performance by Multiagent Learning Algorithms” IOSR Journal of Computer Engineering,
Volume 1, Issue 18, pp 18–25, 2016.
15. A. Bonarini and M. Restelli, “Reinforcement distribution in fuzzy Q-learning,” Fuzzy Sets
and Systems, vol.160, no.10, pp. 1420–1443, 2009.
16. Y. K. Chin, Y. K. Wei, and K. T. K. Teo, “Qlearning traffic signal optimization within
multiple intersections traffic network,” in Proceedings of the 6th UKSim/AMSS European
Symposium on Computer Modeling and Simulation (EMS ’12), pp. 343–348, Nov 2012.
17. L.A. Prashanth and S. Bhatnagar, “Reinforcement learning with function approximation for
traffic signal control,” IEEE Transactions on Intelligent Transportation Systems, vol. 12, no.
2, pp. 412–421, 2011.
18. Deepak A. Vidhate, Parag Kulkarni “Multilevel Relationship Algorithm for Association Rule
Mining used for Cooperative Learning” in International Journal of Computer Applications
(IJCA), Volume 86 Number 4- 2014 pp. 20–27.
19. Y. K. Chin, L. K. Lee, N. Bolong, S. S. Yang, and K. T. K. Teo, “Exploring Q-learning
optimization in traffic signal timing plan management,” in Proceedings of the 3rd
International Conference on Computational Intelligence, Communication Systems and
Networks (CICSyN ’11), pp. 269–274, July 2011.
20. Deepak A. Vidhate, Parag Kulkarni “Multi-agent Cooperation Methods by Reinforcement
Learning (MCMRL)”, Elsevier International Conference on Advanced Material Technologies
(ICAMT)-2016}No. SS-LTMLBDA-06-05, 2016.
21. S. Russell and P. Norvi, Artificial Intelligence: A Modern Approach, PHI, 2009.
22. Deepak A. Vidhate, Parag Kulkarni “Performance enhancement of cooperative learning
algorithms by improved decision making for context based application”, International
Conference on Automatic Control and Dynamic Optimization Techniques (ICACDOT) IEEE
Xplorer, pp 246–252, 2016.
23. Deepak A. Vidhate, Parag Kulkarni “Improvement In Association Rule Mining By Multilevel
Relationship algorithm” in International Journal of Research in Advent Technology (IJRAT),
Volume 2 Number 1- 2014 pp. 366–373.
24. Young-Cheol Choi, Student Member, Hyo-Sung Ahn “A Survey on Multi-Agent
Reinforcement Learning: Coordination Problems”, IEEE/ASME International Conference
on Mechatronics Embedded Systems and Applications, pp. 81–86, 2010.
Digital Tokens: A Scheme for Enabling
Trust Between Customers and Electronic
Marketplaces
Balaji Rajendran, Mohammed Misbahuddin, S. Kaviraj

and B. S. Bindhumadhava
Abstract In electronic marketplaces, when the supply of a particular product is

limited, and when there is a huge demand for the same, the questions of trans-
parency and integrity prop up. We propose Digital Tokens—defined using proven
cryptographic techniques—as a mechanism to assure trust for customers, and issued
by a reliable, transparent and third-party intermediary, called digital token service
provider (DTSP). The digital tokens are issued to a customer on behalf of a vendor
and could be authenticated by both Vendor and the DTSP. This paper details the
architecture involving the DTSP, protocols for communication, implementation
details, the potential uses and benefits of the system and performance evaluation of
such a system.

Keywords Digital token Timestamp DTSP Digital token service provider

Certifying authorities Electronic market place
1 Introduction
The electronic marketplaces [1] have been witnessing huge demand for select
brands of products such as mobile phones from customers. These electronic mar-
ketplaces have limited stock of the products, primarily owing to the constraints of
B. Rajendran (&) M. Misbahuddin S. Kaviraj B. S. Bindhumadhava

Computer Networks and Internet Engineering (CNIE) Division,
Centre for Development of Advanced Computing (C-DAC),
Electronics City, Bangalore 560100, India
e-mail: balaji@cdac.in
M. Misbahuddin
e-mail: misbah@cdac.in
S. Kaviraj
e-mail: skaviraj@cdac.in
B. S. Bindhumadhava
e-mail: bindhu@cdac.in

https://doi.org/10.1007/978-981-10-7245-1_48
492 B. Rajendran et al.
the supplier, and therefore release the product in batches, which leads to huge
competition among potential buyers as they easily outnumber the available quan-
tity, and many buyers return disappointed, as the entire stock goes off in a flash.
We propose a mechanism of Digital Token, wherein a Digital Token is issued to
the customer by a trusted intermediary, on behalf of the vendor. The token is then
presented by the customer to the vendor, which can be verified for its authenticity
and validity by both the intermediary and the vendor. The trusted token can also be
validated by the customer if required, by communicating with the intermediary.
This mechanism establishes the much-needed trust between the customer and the
vendor.
Digital Tokens are useful, whenever the demand is more and supply is less, and
when the buyers need a reliable and trustable intermediary. In general, the concept
could be applied to any real-life token system, and the potential applications are
many as detailed in the later sections of this paper. The benefits of the digital
token-based system include: the introduction of a trustable third party that could be
relied by the buyer and because of which the vendor relies on it—as it brings in
transparency and the vendor can offload the complexities involved in generating the
tokens, authenticating and validating them and also will have lesser loads on its
servers.
This paper explains the elements of such a Digital Token, the process of
authentication and validation of a Digital Token that can be carried out by Vendor
and the Intermediary, the protocols for communication between the three parties—
Customer, Vendor (eMarketplace) and the Intermediary (referred as the DTSP—
Digital Token Service Provider), and finally, the implementation detailing the
experimental setup and study of performance issues.
2 Related Work
The concept of Digital Token as a cryptographically encoded message with Digital

Time Stamps has not been mentioned elsewhere, although the reference of tokens
has been widely used in the literature.
However, the concepts of Digital timestamping [2] as a time validation
methodology has been described and used in several PKI-based applications,
including e-Tendering, e-Auction, etc. Digital timestamping is comprehensively
described in the IETF RFC 3161 standard [3]. The format of timestamping request
and the response are detailed. Digital timestamping as a service has been offered by
several vendors all over the world, with most of the vendors offering the core CA
(Certifying Authority) services.
Digital Watermarking is a technique used for copy-protection of media content,
by tracking any unauthorized distribution of the digitally watermarked content.
A buyer–seller watermarking protocol [4] was proposed, in such a way that neither
Digital Tokens: A Scheme for Enabling Trust … 493
the buyer will be able to distribute the watermarked contents received from the
seller, nor the seller would be able to create unauthorized copies of the buyer
watermarked contents. The proposed technique uses digital signatures and
encryption for its protocol. Holmquist et al. [5] proposed the concept of tokens as a
physical object required to access digital information. A Credit card company
aimed to replace the Credit Cards with a Digital Token [6]. The concept of Token
based Secure Electronic Transaction [7] was proposed by Rajdeep et al., which
focused on the customers can be sure about the trustworthiness of the Seller before
indulging any transaction, they also mentioned that faulty transactions never take
place by implementing token-based SET mechanism in electronic commerce sites.
Matsuura et al. [8] proposed digital timestamp mechanism for dispute settle-
ments in electronic commerce by providing long-lived authenticity and archiving
by the server. Shazia et al. [9] explained the importance of E-commerce security,
implementing PKI, digital signature and certificate-based cryptographic techniques
in E-commerce security.
3 Proposed System: Architecture and Processes
3.1 Entities
There are three entities involved in this System: Customer—who is interested in

buying a product; Vendor—selling a product through an electronic marketplace;
DTSP—Digital Token Service Provider—a trusted intermediary for both the
Customer and the Vendor, who issues Digital Tokens, and who can verify the
authenticity and time validity of a token.
3.2 Digital Token
A Digital Token is a cryptographic entity derived from customer’s identity, time

value and a timestamp by the DTSP. The time value indicates the time validity of
the token, the timestamp mentions the time the token was generated that will
uniquely identify a transaction.
3.3 Architecture
Figure 1 illustrates the architecture containing the overall sequence of communi-

cations between Customer, Vendor and DTSP detailing the life cycle of a Digital
Token.
Fig. 1 Digital token life cycle
3.4 Processes
The system is comprised of two main processes: One, issuance of a token to a

customer, on behest of the Vendor; Two, authentication and validation of the digital
token—which can be either carried out by the Vendor or by the DTSP on request.
Process 1: Digital Token Issuance
Step 1: This is a base operation, wherein the customer visits the vendor website
(typically an online marketplace) and selects a particular product, and requests for
an advanced registration or booking. The vendor then records the request and
assigns a time slot—during which the customer has to appear for buying of the
product—and sends a request to DTSP to issue a Digital Token to the customer on
its behalf. During the above process, the vendor constructs the Digital Token
request—DTRQ, which consists of customer’s id (UID), start time (ST) and end
time (ET). DTRQ is then digitally signed by the vendor and sent to the DTSP.
Step 2: DTSP, upon receipt of the request—DTRQ, verifies the authenticity and
integrity of the request by verifying the digital signature sent along with the request.
If authenticated successfully, a timestamp value is obtained and sent as a response
to the request in a cryptographically hashed form [10], as follows:
H = Hash (Customers ID + Start Time + End Time + TimeStamp + Nonce);
Here, the nonce represents a random secret value generated by DTSP. Now the
DTSP proceeds to generate a digital token, by digitally signing the request and
response as
DT = <S, H> where S = Digital Sign (H)
Therefore, a Digital Token is a digital signature of the DTSP along with the gen-
erated Hash, which is sent to the customer. Figure 2 illustrates the Token issuance
process:
Fig. 2 Digital token issuance process
Process 2: Authentication and Validation of Digital Tokens

The authenticity check will tell whether the token has been really issued by the
DTSP to the particular customer in question, on behalf of the said vendor. The
validation will have two cases—one, the customer presenting it at the right time
slot, two—the customer presenting either before time or after time—the latter case
being an expired token, but both invalid.
Step 1: The customer logs into the vendor site (at the specified time slot) and
uploads the Digital Token, received from the DTSP corresponding to a particular
product. The authenticity and validity of the token can now be carried out either by
the vendor or by the DTSP.
Step 2(a): Authentication and Validation by the Vendor: The vendor verifies the
Digital Signed Hash from Digital Token by using the public key of the DTSP. The
vendor then searches his database for the H value from DT using the customer’s ID
as a filter, and if a match is found then the token is authentic—i.e. the token has
been issued by the DTSP, for the particular customers. The vendor will also be able
to retrieve the time values—start time and end time and validate the token for the
time, and if found to be valid, the vendor may permit the customer to proceed for
the subsequent steps involved in buying a said product. Figure 3 illustrates the
process of authentication and validation carried out by the vendor.
Fig. 3 Authentication and validation of digital token by vender
Step 2(b): Authentication and Validation by the DTSP: The vendor can decide to
offload the authentication process to the DTSP, by simply forwarding the digital
token. In such a case, the DTSP will receive the digital token presented by the
customers, and the customer’s id from the vendor. The DTSP will then verify the
digital signature from Digital Token using its public key. Now the DTSP will
search its database of active tokens issued at behest of the particular vendor,
containing H value from DT using the Customers Id as the filter. If found, the token
is authenticated successfully, meaning the token has been issued by the DTSP only,
and then the time validation can be done by fetching the corresponding start and
end time. The result of the validation will then be communicated to the vendor by
the DTSP. Figure 4 illustrates the process of authentication and validation carried
out by the DTSP.
4 Implementation
A prototype implementation has been developed using Java EE wherein the entities
—Vendor and DTSP are modelled as distinct entities as in real situations, and
cryptographic operations are carried out with the native Java Crypto libraries.
4.1 Features and Assumptions
1. The DTSP does not store any product information that the customer is buying,
therefore helping to protect the privacy.
Fig. 4 Authentication and validation of digital token by DTSP
2. The DTSP is running a timestamp server or at least has an access to a timestamp

service that will give a reliable time.
3. There is binding between the customer and a Digital token, as DTSP uses
customer identification (CustomersID) information supplied by the vendor. The
Customers Id is assumed to be an email address as typically is the case with
most vendors.
4. The start time could be null by default, which means that the customer can buy
the product after it is open for sale at any time, but before the end time.
5. A customer can obtain a digital token only for one quantity of a product at a
time. If more quantities are required, the customer has to obtain those many
digital tokens.
4.2 Digital Token Encoding
The digital token created by the DTSP on behalf of the vendor is emailed as a file to
the customer. The digital token as described above is actually a signature of the
DTSP on a message containing a time stamp. The digital signature is encoded in the
standard Base64 [11] format, and emailed to the customer. Upon receipt of the
token, for authentication and validation, either by the vendor or by the DTSP, both
have to decode it back to the standard bytes for further processing. A sample digital
token is given below:
–Begin Signature–
ckvZu-
oSlH3+670PLu2QwoXUd2COoKt9GkutCBVi9nmyyeSII/pNDCPLlg8/2U
+WrjfG62
HJatUs-
vaAY2G4UUfvFsipVQFDmH4K4PftSaHdG+/RD+VIG43n5lZxerUy0oB/
QfCKTLAk9aHw/KrF8NxQTdnckw8gkTtzfo/t1LjOw=
–End Signature–
–Begin Hash–
bf65c0043aec52ec26f4c67014e5960d40a4755d
–End Hash—
4.3 Schema
The schema of the DTSP is given as below:

{ID, VendorID, CustomersID, StartTime, EndTime, Timestamp, Hash, Status}
As it can be observed, the following information comes from the vendor—
CustomersID, StartTime, EndTime and VendorID—each vendor is assigned a
unique ID by the DTSP. The Hash value computed by the DTSP would be used for
searching for a token, in case if the same customers had booked for multiple
quantities of the same product or booked for different products from the same
vendor having overlapping timeslots.
We use MongoDB [12] for storing the data, as it yields readily itself to JSON
format so that it could be used for communications straightaway. The status field is
used to store the status of the record/digital token—Active, Expired, etc. All expired
tokens are moved out periodically to separate databases, keeping only the active
ones, and grouped by vendor to improve the search performance of DTSP. A sample
of a fully constructed a document with the DTSP would look as follows:
DTSP Document:
{
``_id'' : ObjectId(``55e7f79d8ad5d90404000029''),
``VendorId'':64,
``CustomersId'':``xyz@abc.in'',
``StartTime'':new Date(``3-9-2015 12:30:00''),
`ÈndTime'':new Date(``3-9-2015 19:30:00''),
``TimeStamp'':new Date(``25-8-2015 16:35:00''),
``Hash'': ``bf65c0043aec52ec26f4c67014e5960d40a4755d'',
``Status'':`Àctive''
}
The schema at the Vendor for the purposes of Digital Tokens is given as below:
{ID, CustomersID, ProductID, StartTime, EndTime, Hash, Status}
As it can be observed, the vendor has details of the customers, along with the
product information that the customers are seeking to book, and also the time in
which the vendor wants to allow the customers to actually buy the product. The
vendor can set the ‘StartTime’ as NULL, if he wants to only set the end time to the
customer. The Hash value is obtained from the DTSP, which is a combination of
the above information plus the timestamp.
Vendor Document:
{
``_id'':ObjectId(``55e7f79d8ad5d90404000029''),
``CustomersId'':``xyz@abc.in'',
``ProductId'':M101256,
``startTime'':new Date(``3-9-2015 12:30:00''),
`ÈndTime'':new Date(``3-9-2015 19:30:00''),
``Hash'': ``bf65c0043aec52ec26f4c67014e5960d40a4755d'',
``Status'':`Àctive''
}
4.4 Communication Between Vendor and DTSP
The Vendor communicates with DTSP during the process of token issuance and
during the process of authentication and validation. During Issuance, the vendor
opens up a connection with the DTSP, and communicates its identity, the customer
identity information and the time slot, as a key-value pair in JSON format and
redirects the customer to the DTSP’s page. The Vendor then waits to listen from the
DTSP for the corresponding H value and closes or resets its connection with the
customer.
In the process of authentication and validation, if the vendor decides to offload
the process to the DTSP, the vendor has to communicate its and the customer’s
identity information, and wait for the result. If a successful result is announced by
the DTSP, then the vendor puts the customer through the subsequent process of
buying the product. It may be noted that the DTSP communicates the results only
with the vendor, which will be useful for the latter to detect any multiple attempts
by an attacker to hog the resources by sending fake tokens, or expired tokens.
4.5 Screenshots
A sample of the screenshots depicts the flow of the process as illustrated in Fig. 5.
Fig. 5 Sample screenshots
5 Performance Analysis
The performance is analyzed by looking at the time it takes to create a digital token,
and the time it takes to authenticate and validate a token.
5.1 Time Taken to Generate a Digital Token
Generating a Digital Token is carried out by the DTSP and involves the following
steps, and the average time for each activity is given in Table 1. As it can be seen,
approximately 470 digital tokens can be created in less than a minute, by a system
that is not fully optimized for this activity. The mailing of digital tokens to the
customer is considered as a separate process, outside of the critical sections
affecting the performance of the system. Also, it may be noted that the timestamps
are obtained from a dedicated timestamping service that implements the RFC 3161,
and hence, the time taken to obtain it is higher than obtaining the local server time
value.
5.2 Time Taken for Authentication and Validation
The authentication and validation can be carried out by either the Vendor or the
DTSP and the logic is same—verification of the presented Digital token with the
Table 1 Average time taken for generation of a digital token

S. Activity Average
No. duration (in
ms)
1 Generating the timestamp—typically done through a timestamping 122
server operated by the DTSP
2 Creating the hash of all the values {CustomersID, StartTime, 0.75
EndTime, TimeStamp, Nonce}
3 Creating the signature by encrypting the Hash created in step 2 with 4.01
its private key
4 Encoding the signature in base64 format and writing into a file 1.66
Total 128.42
Table 2 Average time taken for authentication and validation of a digital token
S. Activity Avg. Remarks
No. time (in
ms)
1 Reading and decoding the 0.55 Failure at this step can mean the token is
signature given in base64 corrupted
format
2 Verification of the signature 2.56 Failure here means, the customers may be
using public key of DTSP presenting a fake token
3 Searching for the H value 0.26 Failure here means, the customer has
(part of DT) presented an invalid token
4 Perform time validation 0.01 Failure here means, the token has been
presented before time (as expired tokens have
been moved to different database and in that
case would be eliminated in previous step)
Total 3.38
public key of DTSP and search for the record matching the Hash value, and if found
then perform the time validation. The above steps and time taken are given below in
Table 2.
In the case of authentication and validation, failure is possible at each and every
step, and failure at an earlier stage is better than at the later stage—as the latter
stages will add to the cost of time. Step 3 and 4 have been optimized using efficient
data structures and clean-up techniques.
5.3 Analysis
Following observations can be made from the above experiments:

1. The process of digital token issuance is costlier than the process of authenti-
cation and validation.
2. In the event of authentication and validation by DTSP, the overheads involved

in communication between the vendor and the DTSP is primarily determined by
the latency factor, rather than the process itself.
3. The cryptographic operations–especially the encryption and decryption are
relatively faster.
When the vendor is experiencing peak-traffic loads, it may offload the process of
authentication and validation of digital tokens to the DTSP, but in such scenarios
the vendor may have to take into account the latency delays also, as the time
required to do the authentication and validation is meagre 3.38 ms, which means an
approximate 17,000 digital tokens could be evaluated in less than a minute. The
vendor can use the above value as guidance to timeslot the customers accordingly.
6 Conclusion and Future Work
In this paper, we presented the concept and process of a Digital Token, issued by a
trusted intermediary, playing the role of a DTSP—a role that could be played by a
Certifying Authorities (CA) which could probably add to another business or
revenue line to their existing services—the Digital tokens being issued to the
customer who is attempting to buy a high-demand product or service from an
electronic marketplace. The DTSP can conduct audit trails of the digital tokens
issued, authentication and validation done by it and can publish the same.
The process could have been strengthened for the better by having a tight
binding between a digital token and the customer, but then the customer should
have a mechanism to apply their digital signature. Server-based digital signing
approaches like eSign [13] could be used for high-value transactions. This approach
can be extended to create a competitive market place by decoupling the tokens from
the market places. Therefore, in our future work, we aim to shift the balance in
binding proportions wherein we envisage a strong binding between the customer
and the token, while a light binding between the token and the vendor or
eMarketplace.
References
1. Sonja Grabner-Kraeuter, The Role of Consumers’ Trust in Online-Shopping, July 2002,

Journal of Business Ethics. http://link.springer.com/article/10.1023/A:1016323815802
2. S. Haber, WS. Stornetta, “How to Time-Stamp a Digital Document” Springer Berlin
Heidelberg, pp. 437–455, 1991
3. C Adams, P Cain, D Pinkas, R Zuccherato, Internet X.509 Public Key Infrastructure
Time-Stamp Protocol (TSP), 2001 – IETF RFC 3161 - https://www.ietf.org/rfc/rfc3161.txt
4. “A Buyer-Seller Watermarking Protocol”, Nasir Memon, Ping Wah Wong, IEEE
Transactions on Image Processing, pp: 643–649, 2001
5. L. Holmquist, J. Redström, and P. Ljungstrand, “TokenBased Access to Digital Information,”

Proceedings of HUC’99 (1999), pp. 234–245
6. Amex to Implement Digital Tokens to Replace Cards, http://www.infosecurity-magazine.
com/news/amex-to-implement-digital-tokens/
7. Rajdeep Borgohain, Chandrakant Sakhatwade, Sugata Sanyal, “TSET: Token based Secure
Electronic Transaction”, IJCA, Volume 45 – Number 5, 2012.
8. Kanta Matsuura, Hideki Imai, “Digital timestamps for dispute settlement in electronic
commerce: Generation, Verification and renewal”, CiteSeerx, 2002
9. Shazia Yasin, Khalid Haseeb, Rashid Jalal Qureshi, “Cryptography based E-Commerce
security: A Review”, Volume 9, Issue 2, IJCSI, March 2012
10. L. Damgard. “Collission-free hash functions and public-key signature schemes.” In Advances
in Cryptology – Eurocrypt’87, pp. 203–217. Springer-Verlag, LNCS, vol. 304, 1988.
11. S. Josefsson, SJD, “The Base16, Base32, and Base64 Data Encodings”, RFC 4648,Network
Working Group, October 2006.
12. MongoDB-Documentation. http://docs.mongodb.org/manual/tutorial/
13. eSign – Online Electronic Signature Service, http://www.cca.gov.in/cca/?q=eSign.html, Last
accessed May 12, 2017.
BMWA: A Novel Model for Behavior
Mapping for Wormhole Adversary Node
in MANET
S. B. Geetha and Venkanagouda C. Patil
Abstract Wormhole attack has received very less attention in the research com-
munity with respect to mobile ad hoc network (MANET). Majority of the security
techniques are toward different forms of wireless network and less in MANET.
Therefore, we introduce a model for behavior mapping for Wormhole Attacker
considering the unknown and uncertain behavior of a wormhole node. The core
idea is to find out the malicious node and statistically confirm if their communi-
cation behavior is very discrete from the normal node by formulating a novel
strategic approach to construct effective decision. Our study outcome shows
enhanced throughput and minimal overhead-latency with increasing number of
wormhole node.

Keywords Mobile ad hoc network Wormhole attack Adversary

Malicious node behavior Attack pattern
1 Introduction
Security problems in mobile ad hoc network are always represented as challenging

problems owing to it inherent characteristics, i.e., dynamic topology, intermittent
link breakage, power consumption, etc. [1, 2]. In past decade, there have been
various research work carried out toward improving security system in MANET
taking case studies of various forms of adversaries [3, 4]. Although, such studies
develop a good baselines of future research but they also suffer from the limitation
S. B. Geetha (&)
Department of Computer Engineering, MMCOE, Pune, India
e-mail: sb.geetha@gmail.com
V. C. Patil
Department of Electronics & Communication Engineering, BITM, Bellary, India
e-mail: patilvc@rediffmail.com

https://doi.org/10.1007/978-981-10-7245-1_49
506 S. B. Geetha and V. C. Patil
of their narrowed applicability [5]. Existing routing protocols with security features
are quite less and highly symptomatic [6]. There are also good numbers of studies
highlighting about node behavior of MANET [7–9]. Even these studies too does not
emphasized over the computational complexities associated with the implementa-
tions of computationally complex algorithms. Moreover, with a changing time,
design aspects of a malicious node are revolutionized for which reason it becomes
sometime very hard to differentiate malicious node and regular node. Presence of
malicious node does not mean attack implication. Even such nodes could be con-
trolled to assists in packet forwarding. Unfortunately, there were no such research
works, where the malicious node does assists in packet forwarding. We comment
here that normally in wormhole attack, the malicious node will act as a regular node
so that they can divert the incoming route through newly developed wormhole
tunnel. This characteristic has not been researched by anyone with respect to
wormhole attack and hence the proposed manuscript will brief out the simple
mechanism where the malicious node will be used for this reason. This paper
presents a novel technique toward leveraging the security features in MANET
where both security features and communication capabilities are emphasized in
MANET in presence of wormhole adversary. Section 2 discusses about the existing
research work followed by problem identification in Sect. 3. Section 4 discusses
about proposed methodology followed by an elaborated discussion of algorithm
implementation in Sect. 5. Comparative analysis of accomplished result is dis-
cussed under Sect. 6 followed by the conclusion in Sect. 7.
2 Related Work
This section discusses about the relevant work being carried out toward wormhole
attack in MANET. Our prior work has already reviewed certain techniques of
resisting wormhole attack [10]. It was seen that wormhole attack was basically
studied with respect to other forms of network, e.g., sensor network [11–14],
wireless mesh network [15], and quite a very less in MANET. Singh et al. [16] have
presented a technique to countermeasure the adverse effect of wormhole attack over
on-demand routing in MANET using case study of misbehaved nodes. The study
outcome shows enhanced communication performance. Almost similar direction of
study toward wormhole attack was also carried out by Kaur and Chaba [17] to find
the technique offer lower delay and throughput. Gupta and Pathak [18] have
introduced a technique that explores substitute route for avoiding the packets go in
wormhole tunnel. The study outcome shows reduced overhead in routing, reduced
packet drop, and increased received packets. Stoleru et al. [19] have presented a
technique that emphasizes on the discovery of the neighbor nodes for a secured
communication system in MANET. The technique also assists in determining the
wormhole node position and explores the neighbor node in much secured manner.
BMWA: A Novel Model … 507
Shastri and Joshi [20] have reviewed the existing techniques toward determining
and eradicating the wormhole adversaries in MANET. Similar forms of identifying
and localizing wormhole attacks were also carried out by Ojha and Kushwah [21].
The technique has also used threshold-based approach for improving delay and
communication performance. Ali and Nand [22] have presented a discussion where
two different on-demand routing protocols were investigated in presence of
wormhole attack on MANET. In a nutshell, it can be said that there are few
techniques to resists the adverse effect of wormhole attacks in MANET. The
existing technique has emphasized on detection of the adversaries and presented
approaches to identify the tunnel. All the above-mentioned technique has associated
advantages as well as pitfalls too. Discussion of the existing research problems is
carried out in the next section.
3 Problem Description
From the previous section, it can be seen that there are a various approach that
assists in determining the adversary in wormhole where the focus is more on
exploring the tunnel and less on finding the exact adversaries. The biggest chal-
lenging situation is that exploring exact malicious node is quite a difficult task
especially in presence of dynamic topology in MANET. A malicious node can
easily act as normal node until it gain complete trust of the network to invoke a
collateral damage which is much dangerous than wormhole tunnel. In such situa-
tion, none of the existing techniques presented in literature can assist to differentiate
malicious or regular node in presence of wormhole attack. This problem is yet to be
studied as a solution to this problem will give precise information about the
adversaries.
The proposed work is a continuation of our prior model of resisting wormhole

attack [23]. The present work adopts analytical research methodology, where the
focus is laid to find the wormhole adversary node in the simulation area. We
consider the complex hiding behavior of wormhole node and chalk out a novel
implementation formulation that uses confusing characteristics of a malicious node
in order to resist other regular nodes to fall into wormhole tunnel. The schematic
architecture of the proposed system is shown in Fig. 1.
The proposed study applies three different strategies motivated from game
theory and performs the computation of three different forms of the trust-based
coefficient, e.g., positive trust, negative trust, and ambiguity trust. Sequential
Strategy Formulation
ρ 1(Pure Strategy) ρ 2(Mixed Strategy) ρ 3(PBE Strategy)
ambiguity
α 1 (+Trust) α 2 (-Trust)
Trust α 3
Analytical
Approach
trust-based φ 1 (number of identified packet

coefficients forwarding operation)
φ 2 (number of packet dropping
actions
Fig. 1 Schematic architecture of BMWA
rationality is implemented to invoke complex behavior of wormhole adversary in

MANET. The prime idea is to ensure that malicious node should be caught at any
cost and while doing so the communication performance in MANET should not be
affected. The next section briefs about algorithm implementation of BMWA.
5 Algorithm Implementation
This section discusses about the algorithm that is responsible for mapping the
malicious behavior of a mobile node in MANET. The algorithm implements a
novel logic of strategy formulation for mapping malicious actions of the mobile
node and thereby undertakes different steps to mitigate the attack. The algorithm
takes the input of the A (simulation area), n (number of the node), m (percentage of
the malicious node), T (Threshold), which upon processing leads to the generation
of the outcome of a malicious node or regular node. The basis of the algorithm
performs both detection and prevention system. The steps of the proposed system
are as shown below:
Algorithm for BMWA

Input: A, n, m
Output: adversary
Start
1. init A, n, m
2. [Rns, Mns]=[q1, q2, q3]
3. If Th*(1-Tu)<T
4. If Th<val1
5. flag!DF
6. else
7. flag strategy!PD
8. end
9. End
10. compute a1, a2, a3
11. Pcatch=si*((temp1-Damf)/statf)
12. Flag n(max(card(a3))!adversary
End
The algorithm implements the potential logic of game theory where three dif-
ferent strategies, i.e., q1 (Pure Strategy), q2 (Mixed Strategy), q3 (PBE Strategy) is
being testified. A matrix is constructed for such strategies against regular node (Rns)
and malicious node (Mns) (Line-2). We also construct various entities in the pro-
posed model, e.g., (i) Intrusion Profit, (ii) Packet forwarding profit, (iii) Updating
Profit, (iv) Resources required for Intrusion, (v) Resources required for packet
forwarding, (vi) Resources required for constructing tunneling, (vii) Damage factor
due to negative update, and (viii) statistical factor for negative update. The pro-
posed algorithm also consider a threshold T considered for a normal mobile node to
perform the update. We also apply probability modeling for constructing this
algorithm. Our first probability factor is Th that computes the possibility of one
mobile node being malicious. Similarly, our second probability factor is also
defined for if the malicious node will really perform intrusion or choose to assist in
data packet forwarding just to hide its true identity. The proposed system also
considers another probability factor Tu to understand the possibility of the uncer-
tainty of one node being a malicious node. We perform empirical formulation to
compute the ambiguity factor a3 (Line-11) as c.u1* u2/((u1 + u2)2*(u1 + u2 + 1)),
where u1 and u2 are a number of identified packet forwarding operation and pack
dropping actions, respectively. A condition is applied for mapping malicious node
behavior with Perfect Bayesian Equilibrium (Lines-3 and 4) after computing a
decision value val1 which corresponds to (Packet forwarding profit-Resources
required for packet forwarding)/(Packet forwarding profit + Resources required for
constructing tunneling). Therefore, the condition shown in Line-3 and Line-4 rep-
resents the good state of the node in which case we flag the node as to obtain a
strategy for data forwarding (DF) (Line-5) otherwise, it is flagged as a malicious
node with packet dropping (PD) action (Line-7). However, we also infer that it is
highly possible that a malicious node can act as regular node showing DF, whereas a
regular node can also show PD action. Hence, Algorithm line-3–7 just assists in
understanding the possible strategies of a malicious or regular node but it does not
confirm the fact that the node is really a malicious or regular node. For this purpose,
we consistently monitor three trust-based coefficients (a1, a2, a3) (Line-10), The first
coefficient a1 is calculated as [u1//(u1 + u2)] * a3, the second coefficient a2 is
calculated as [u2//(u1 + u2)] * a3.
We compute the probability of detecting malicious node Pcatch as shown in
Line-11. Where si is selected strategy, and the variable temp1 is equivalent to
(Th * (1 − Tu)*Profit of update)/(1 − Th*(1 − Tu)). The other variable Damage

factor Damf and statistical factor for the negative update, i.e., statf. It will mean that
if the regular node attempts to defame another regular node by offering a false
update to its neighbor that it is a malicious node, than it is considered as false
updating system. In such case, the regular node is being penalized by offering lower
utility value. This process is called as sequential rationality, where both malicious
node and the regular node will attempt to obtain the highest utility. For better game
strategy, we maintained sequential rationality in such a way that malicious node can
get the highest utility only when they attack and in such case regular node will get
highest utility if they can catch hold of such attack using updating mechanism.
Hence, in such scenario, chances of the malicious node to be get caught is quite
high. However, if the malicious node chooses not to perform intrusion but to assists
in packet forwarding, a malicious node is discouraged by offering lower utility
value. But a closer look in such strategy suggests that by adopting the practice of
data packet forwarding by a malicious node, it does not affect any communication
performance of MANET due to any security issues as long as there is no intrusion.
To further confirm about it, we compute maximum cardinality of the third trust
factor, i.e., a3. If the ambiguity factor is found to be consistently higher that it is
assured that the node is malicious and there is no routing being performed through
this node once the false reporting of its neighbor regular node is found to be lower.
Hence, in this manner, the proposed system not only identifies the malicious node
but also discourages them to initiate any forms of attacks.
6 Results Discussion
The assessment of the proposed study is carried out considering 1000 1000 m2
simulation areas with 100 mobile nodes being randomly deployed. We consider
10% of the malicious node being present during the simulation. We consider all the
three strategies of game theory, i.e., pure, mixed, and PBE for both the types of the
nodes in order to incorporate sequential rationality. Each mobile node is considered
to have a range of 20 m during the assessment of the result. For a better assessment,
we compared the outcome of our study with respect to multicast ad hoc on-demand
vector (MAODV) routing protocol in MANET (Fig. 2).
The study outcome shows that proposed BMWA offers a better improvement
over throughput obtained by existing MAODV. It also offers higher extent of
reduction in routing overhead as well as routing latency as compared to existing
MAODV. Hence, the security techniques incorporated in BMWA does not seem to
negatively affect communication quality. A closer look at the outcome will also
show the capability to enhance communication performance increase with an
increase in malicious nodes, which means the presence of malicious node does not
create security loopholes in BMWA. Hence, the equilibrium between security and
communication performance is maintained.
(a) Throughput (b) Reduced Routing Overhead
(c) Reduced Routing Latency
Fig. 2 Outcome of proposed study
7 Conclusion
The prime factor for working on this presented paper is as following: (i) existing
research studies toward MANET security considering wormhole attack is quite less,
(ii) a wormhole attacker can perform camouflaging by spoofing itself with the
identity of the regular node. Therefore, in such scenario, there are no existing
techniques that offer comprehensive disclosure of malicious node identity. Hence,
this paper presents a probabilistic technique that assists in formulating a decision by
a regular node to determine if its neighbor or its other communication node is
wormhole adversary. Proper identification of wormhole adversary will protect the
regular node from wormhole tunnel. Our outcome shows that proposed system
offers a better balance between malicious node identification with communication
performance in MANET.
References
1. R. Rajendran, A Survey on Multicast Routing Protocol and Its Challenges in Manet, Lap
Lambert Academic Publishing, 2015
2. S. Paul, Introduction to MANET and Clustering in MANET, Anchor Academic
Publishing-Computer, 2016
3. A. Nadeem and M. P. Howarth, “A Survey of MANET Intrusion Detection & Prevention
Approaches for Network Layer Attacks,” in IEEE Communications Surveys & Tutorials, vol.
15, no. 4, pp. 2027–2045, Fourth Quarter 2013.
4. S. Das and P. Nand, “Survey of hash security on DSR routing protocol,” 2016 International
Conference on Computing, Communication and Automation (ICCCA), Noida, 2016, pp. 520–
524.
5. Rawat, Danda B., Security, Privacy, Trust, and Resource Management in Mobile and
Wireless Communications, IGI Global, 2013
6. A. L. Sandoval Orozco, J. GarcıaMatesanz, L. J. Garcıa Villalba,J. D.Márquez Dıaz, and T.-
H. Kim, “Security Issues in Mobile Adhoc Networks”, Hindawi Publishing Corporation,
International Journal of Distributed Sensor Networks, 2012
7. P. M. Nanaware, S. D. Babar and A. Sinha, “Survey on bias minimization and application
performance maximization using trust management in mobile adhoc networks,” 2016 3rd
International Conference on Computing for Sustainable Global Development (INDIACom),
New Delhi, 2016, pp. 831–834.
8. P. Ramkumar, V. Vimala and G. S. Sundari, “Homogeneous and hetrogeneous intrusion
detection system in mobile ad hoc networks,” 2016 International Conference on Computing
Technologies and Intelligent Data Engineering (ICCTIDE’16), Kovilpatti, 2016, pp. 1–5.
9. [9] A. Lupia and F. De Rango, “A probabilistic energy-efficient approach for monitoring and
detecting malicious/selfish nodes in mobile ad-hoc networks,” 2016 IEEE Wireless
Communications and Networking Conference, Doha, 2016, pp. 1–6
10. S.B. Geetha and V.C. Patil, “Evaluating the Research Trends and Techniques for Addressing
Wormhole Attack in MANET”, International Journal of Computer Applications, Vol. 110,
No. 6, 2015
11. R. Singh, J. Singh, and R. Singh, “WRHT: A Hybrid Technique for Detection of Wormhole
Attack in Wireless Sensor Networks”, Mobile Information Systems, pp. 13, 2016
12. L. Lu, M.J. Hussain, G. Luo, and Z. Han, “Pworm: passive and real-time wormhole detection
scheme for WSNs”, International Journal of Distributed Sensor Networks, pp. 16, 2015
13. S. Mukherjee, M. Chattopadhyay, S. Chattopadhyay, and P. Kar, “Wormhole Detection
Based on Ordinal MDS Using RTT in Wireless Sensor Network”, Journal of Computer
Networks and Communications, pp. 15, 2016
14. H. Chen, W. Chen, Z. Wang, Z. Wang, and Y. Li, “Mobile beacon based wormhole attackers
detection and positioning in wireless sensor networks”, International Journal of Distributed
Sensor Networks, Vol. 10, No. 3, 2014
15. R. Matam and S. Tripathy, “Secure Multicast Routing Algorithm for Wireless Mesh
Networks”, Journal of Computer Networks and Communications, pp. 11, 2016
16. Y. Singh, A. Khatkar, P. Rani, and D.D. Barak, “Wormhole Attack Avoidance Technique in
Mobile Adhoc Networks”, In Advanced Computing and Communication Technologies
(ACCT), Third International Conference, pp. 283–287, 2013.
17. G. Kaur, V. K. Jain, and Y. Chaba, “Wormhole attacks: Performance evaluation of on demand
routing protocols in Mobile Adhoc networks”, In Information and Communication
Technologies (WICT), 2011 World Congress, pp. 1155–1158, 2011.
18. C. Gupta, and P. Pathak, “Movement based or neighbor based tehnique for preventing
wormhole attack in MANET”, In Colossal Data Analysis and Networking (CDAN),
Symposium, pp. 1–5, 2016.
19. R. Stoleru, H. Wu, and H. Chenji, “Secure neighbor discovery and wormhole localization in
mobile ad hoc networks”, Ad Hoc Networks, Vol. 10, No. 7, pp. 1179–1190, 2012
20. A. Shastri and J. Joshi, “A Wormhole Attack in Mobile Ad-Hoc Network: Detection and
Prevention”, In Proceedings of the Second International Conference on Information and
Communication Technology for Competitive Strategies, pp. 31, 2016
21. M. Ojha and R.S. Kushwah, “Improving Quality of Service of trust based system against
wormhole attack by multi-path routing method”, In Soft Computing Techniques and
Implementations (ICSCTI), 2015 International Conference, pp. 33–38, 2015
22. S. Ali and P. Nand, “Comparative performance analysis of AODV and DSR routing protocols
under wormhole attack in mobile ad hoc network on different node’s speeds”, In Computing,
Communication and Automation (ICCCA), 2016 International Conference, (pp. 641–644),
2016
23. S. B. Geetha, V.C. Patil, “Elimination of Energy and Communication Tradeoff to Resist
Wormhole Attack in MANET”, International Conference on Emerging Research in
Electronics, Computer Science and Technology, pp. 143–148, 2015
AWGN Suppression Algorithm in EMG
Signals Using Ensemble Empirical Mode
Decomposition
Ashita Srivastava, Vikrant Bhateja, Deepak Kumar Tiwari

and Deeksha Anand
Abstract Surface Electromyogram (EMG) signals are often contaminated by

background interferences or noises, imposing difficulties for myoelectric control.
Among these, a major concern is the effective suppression of Additive White
Gaussian Noise (AWGN), whose spectral components coincide with the spectrum
of EMG signals; making its analysis problematic. This paper presents an algorithm
for the minimization of AWGN from the EMG signal using Ensemble Empirical
Mode Decomposition (EEMD). In this methodology, EEMD is first applied on the
corrupted EMG signals to decompose them into various Intrinsic Mode Functions
(IMFs) followed by Morphological Filtering. Herein, a square-shaped structuring
element is employed for requisite filtering of each of the IMFs. The outcomes of the
proposed methodology are found improved when compared with those of con-
ventional EMD-and EEMD-based approaches.

Keywords EMG Additive white gaussian noise (AWGN) Ensemble empirical

mode decomposition (EEMD) Morphological filtering
A. Srivastava (&) V. Bhateja D. K. Tiwari D. Anand

Department of Electronics and Communication Engineering, Shri Ramswaroop Memorial
Group of Professional Colleges (SRMGPC), Lucknow 226028, Uttar Pradesh, India
e-mail: srivastavaashita95@gmail.com
V. Bhateja
e-mail: bhateja.vikrant@gmail.com
D. K. Tiwari
e-mail: srmdeepaktiwari57@gmail.com
D. Anand
e-mail: deeksharoma2012@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_50
516 A. Srivastava et al.
1 Introduction
EMG is an electrical demonstration of a shrinking muscle. The acquisition of a pure

EMG signal becomes a major constraint in its proper analysis and diagnosis [1].
Surface EMG signals are often corrupted by Baseline Wander and AWGN [2].
AWGN gets superimposed in EMG signal as a result of thermal vibration of atoms
in conductors, which is currently a more demanding task [3]. Many filtering
approaches have been developed in recent past for the effective suppression of
AWGN from EMG signals [4]. Among these, the most simple and the cost-effective
ones’ deploy Gaussian filtering for AWGN suppression. However, processing via
Gaussian filter also attenuates useful diagnostic information from EMG signals.
Hence, effective minimization of AWGN without deteriorating the signal quality
(upon restoration) is a prominent challenge towards the development of filtering
algorithms for EMG signals. Ren et al. [5] introduced an approach of background
and gaussian noise suppression in EMG signals based on the combination of
Independent Component Analysis (ICA) and Wavelet-based filtering. But, ICA has
a disadvantage that the scale of the resulting signal is not same as the original
signal. Mello et al. [6] propounded an approach of digital Butterworth filter design
to delimitate the frequency band of surface Electromyograms (EMG) and remove
the mains noise and its harmonics. Yet, it is highly impractical to design a
Butterworth Filter for the elimination of such a high-frequency noise. Phinyomark
et al. [7] performed EMG signal denoising using adaptive wavelet shrinkage
function. However, proper selection of wavelets is a necessary task which acts as a
constraint. In addition, this approach is also computationally intensive. Djellatou
et al. [8] carried out a dual-adapted fast block least mean squares algorithm
(DA-FBLMS) to remove noise contaminations from Surface Electromyogram
(sEMG). However, processing via least mean square (LMS) algorithm for denoising
signals leads to a slower convergence. Recently, Harrach et al. [9] proposed a
technique of Canonical Correlation Analysis (CCA) followed by noise intensity
ratio thresholding for AWGN elimination. But, the technique has been performance
limited as it hampered the diagnostic quality/utility of the EMG signal. In this
paper, an improved algorithm for preprocessing (AWGN suppression) of EMG
signals based on the EEMD followed by morphological filtering is presented [10].
EEMD has a major advantage that this algorithm uses the full statistical properties
of white noise and hence works effectively. Further, combinations of morphological
operators are used to retain the shape of the original EMG signal upon restoration.
After that, the signal fidelity assessment parameters, i.e., signal-to-noise ratio
(SNR) and percentage root mean square difference (PRD %) are evaluated for the
performance validation of the proposed algorithm. The rest of the paper has been
sectioned as follows: Sect. 2 reports the proposed AWGN suppression algorithm.
Section 3 comprises of the results section which demonstrates the simulation out-
comes along with their analytical discussion; whereas Sect. 4 concludes the work.
AWGN Suppression Algorithm in EMG Signals Using Ensemble … 517
2 Proposed AWGN Suppression Algorithm
The proposed AWGN suppression algorithm combines the EEMD algorithm with
morphological filtering to preprocess the corrupted EMG signals. Initially, the
unprocessed EMG signal is fed to the EEMD module, where the signal decom-
position via EEMD takes place which breaks the EMG signal into IMFs. These
IMFs are then filtered by morphological operations (opening and closing) to get the
denoised signal. At the end, the signal fidelity assessment is performed using SNR
and PRD% as the performance metrics.
2.1 EEMD
The EMD, particularly drafted for nonlinear and nonstationary signals, is a local,
data-driven technique which dissolves the signal without the use of any precedent
premise [11]. EMD decomposes a multifrequency component signal into
non-coinciding frequency constituent named intrinsic mode functions (IMFs) and a
residual signal [12]. The deduced IMFs are mono-frequency components that fulfill
two criterions:
• The total amount of maxima (both the local maxima and minima) and zero
crossings in the time series must not differ by more than one.
• The average value of the upper envelope (interpolated through maxima) and
lower envelop (interpolated through minima) is zero in the entire time series.
EMD being such an adequate dissolving technique reaps some troublesome
crunches. The main problem with the original EMD is the continual appearance of
mode mixing, i.e., the detail related to one scale can appear in two different intrinsic
modes. To dispatch the mode mixing problem, a new noise-assisted data analysis
(NADA) technique brought by Huang et al. [13], the Ensemble EMD (EEMD), that
defines the true IMFs as the mean of an ensemble of trials, each residing of the
signal plus a white noise of finite amplitude. The procedural steps of EEMD are
shown in Algorithm 1. The resultant IMFs, namely cij (t) are averaged across trials
to get the final IMFs, ci (t) as shown in Eq. No. (1) below:
1X N
ci ðtÞ ¼ cij ðtÞ; ð1Þ
N j¼1
where, i is the IMF order, j denotes the trial index, and N is the total number of
trials.
Algorithm 1 Procedural Steps for EEMD algorithm
BEGIN
Step 1: Input: Baseline Corrupted EMG signal
Step 2: Process: Add a white noise series to the corrupted EMG signal
Step 3: Compute: All Local Maxima and Minima
Step 4: Compute: Upper and Lower Envelopes by Cubic Spline function
Step 5: Compute: Average of Upper and Lower Envelopes
Step 6: Compute: Difference by subtracting average from the noisy signal
Step 7: Process: Check if the difference is an IMF
Step 8: Compute: Residue signal by subtracting difference from the noisy signal if it is an
IMF otherwise repeat steps 1–7
Step 9: Process: Check if residue signal is a monotonic function
Step 10: Process: Repeat steps 1–9 iteratively with changed versions of noise each time
Step 11: Compute: Mean (Ensemble) of corresponding decomposed IMFs as the final true
IMFs
END
2.2 Morphological Filtering
Morphological filtering is a nonlinear transformation technique mostly used for the

sake of regionally altering the geometrical attributes of a signal. It employs the
simple applications of set theory. Morphological filters are the type of set functions
which adjust the linear representation of the signal to get the desired geometrical
shape of the signal [14]. In this technique, the whole examined signal is compared
with a small template called structuring element (SE), whose suitable selection
plays a very vital role for the effective filtering of the signal. The whole signal and
structuring element are the two sets in mathematical morphology; onto which
various operations are applied to change the structural shape of the signal. A huge
variety of structuring elements are present in the mathematical morphology which
enriches its applications. The shape and size of structuring element should be
selected literally as its inaccurate choice may distort the adjoining wave of the EMG
signal [15]. Linear, Triangular, Octagonal, and Square-shaped structuring elements
are usually applied in the morphological filtering and shown in the Fig. 1a–d.
In the proposed approach, square-shaped structuring element of width 4 is
chosen as shown in Fig. 1d. Dilation, Erosion, Opening, and Closing are basic
operations in the mathematical morphology to alter the signal’s geometrical shape.
AWGN is suppressed by the usage of morphological operations comprising high
pass as well as lowpass filter properties. Hence in the proposed algorithm, opening
and closing operations have been used for the adequate suppression of AWGN. The
respective expression of opening and closing operations is given by:
Fig. 1 Structuring Elements a Linear, b Triangular, c Octagonal, d Square
A B ¼ ðAHBÞ B ð2Þ
A B ¼ ðA BÞHB ð3Þ
where: A denotes the matrix of decomposed canonical components, B represents the

structuring element and and H denotes the dilation and erosion, respectively.
Opening, i.e., erosion succeeded by dilation and Closing, i.e., dilation succeeded by
erosion is deployed on the decomposed IMFs from the AWGN corrupted EMG sig-
nals. These operations are carried out on all the IMFs one by one. After that, an average
of the outputs obtained from the opening and closing operation performed on the IMFs
is calculated to obtain the filtered EMG signal with minimized gaussian noise.
In the proposed algorithm, the EMG signals have been acquired from The UCI
Machine Learning Repository [16] which includes 10 normal and 10 aggressive
physical actions measuring the distinct human activities. Herein, for the simulation
purpose, five aggressive actions EMG signals have been taken as the test data. The
number of iterations and the noise parameter needs to be considered for applying
EEMD Algorithm. Herein, ensemble number (NE) is taken to be 10. The noise
parameter (Nstd), defined as the ratio of the standard deviation of noise to the
standard deviation of the signal power, is selected to be 0.2.
3.1 Analysis of the Structuring Element
Adequate selection of structuring element is a very crucial job in morphological

operation so here an analysis is done to select best structuring element for effective
suppression of AWGN [17]. In the process of selecting most appropriate structuring
element, different shapes/orders (linear, triangular, octagonal and square) of the
same were deployed; out of which square shape structuring element yielded best
results. In the present simulation setup, a “square”-shaped structuring element of
width: 4 has been used. The reconstruction quality comparison of results using
different structuring elements has been given in Table 1.
3.2 Analysis of Filtered EMG Signals
In the simulation process, the EMG signals were first decomposed into IMFs using
the EEMD algorithm as discussed in [11]. After this, the decomposed IMFs were
filtered using morphological filters and the filtered IMFs were reconstructed for the
AWGN suppressed signal. Figure 2. demonstrates the results obtained using the
proposed algorithm along with the outputs of the other conventional versions
(EMD-Gaussian [18] and EEMD-Gaussian [10]).
Figure 2a shows an original EMG signal (Pushing3) in which AWGN is present.
From Fig. 2b, it is observed that only a little portion of AWGN component has been
suppressed using EMD-Gaussian algorithm although it is present in huge quantity
in the original EMG signal. Figure 2c shows the output of the EEMD-Gaussian
methodology in which a major portion of the AWGN component has been sup-
pressed. But, there is an improvement in the output of the proposed algorithm over
the other two conventional ones. In this, AWGN has been suppressed significantly
Table 1 Comparison of SNR (in dB) values with different structuring elements used in the
proposed filtering algorithm
EMG Original Triangular Linear Octagon Square
signal structuring structuring structuring structuring
Ref. no. element element element element
S1 −48.5911 −20.7906 −20.2755 −20.0123 −17.3911
S2 −53.3090 −24.0179 −22.0869 −21.4212 −20.6459
(a) Noisy EMG Signal (b) WGN Removed EMG Signal

1000 3000
500 2000
0 1000
Amplitude
Amplitude
-500 0
-1000 -1000
-1500 -2000
-2000 -3000
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
time time
(c) WGN Removed EMG Signal (d) Filtered EMG Signal

3000 1500
1000
2000
500
1000
0
Amplitude
Amplitude
0 -500
-1000
-1000
-1500
-2000
-2000
-3000 -2500
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
time time
Fig. 2 (a) Corrupted EMG Signal (Pushing3), Restored Outputs Generated with
(b) EMD-Gaussian [18], (c) EEMD-Gaussian [10], (d) Proposed Algorithm-EEMD-
Morphological Filtering
from the original EMG signal which is clearly seen in Fig. 2d. The noisy EMG
signal along with the corresponding filtered outputs using the proposed algorithm
has been shown in Fig. 3a, b, respectively.
Furthermore, SNR (dB) and PRD (%) have been used as performance metrics for
the evaluation and validation of the proposed algorithm [19]. The comparison of
SNR (dB) and PRD (%) values of the AWGN suppressed EMG signal using the
proposed algorithm with the EMD-Gaussian and EEMD-Gaussian-based approa-
ches have been shown in Tables 2 and 3, respectively.
Signal-to-noise ratio is a parameter which evaluates the improvement in the
signal quality relative to noise. The SNR value of the algorithm which has been
proposed is reported to be −17.3911 for signal (S1) which is much better as
compared to EMD-Gaussian and EEMD-Gaussian. Similar results are obtained
during simulations when tested with other signals. Also, the PRD % value for signal
(S1) of the proposed algorithm is 0.3110 which is less in comparison to the other
methodologies thus signifying minimum distortion in the original EMG signal.
(a) 5000 Noisy EMG Signal (b) x 10

4
Filtered EMG Signal
1
4000 0.8
3000 0.6
2000 0.4
Amplitude
Amplitude
1000 0.2
0 0
-1000 -0.2
-2000 -0.4
-3000 -0.6
-4000 -0.8
-5000 -1
0 1 2 3 4 5 6 7 8 94 0 1 2 3 4 5 6 7 8 9
4
time x 10 time x 10
Fig. 3 a Corrupted EMG Signal (Pushing3). b Proposed Methodology Filtered EMG Signal
Table 2 Computed values of SNR (in dB) for performance comparison

EMG signal Ref. Original EMD-Gaussian EEMD-Gaussian Proposed
no. methodology
S1 −48.5911 −20.4691 −19.0516 −17.3911
S2 −50.0350 −21.6478 −19.7652 −17.9144
S3 −53.3090 −22.9461 −22.6957 −21.7131
S4 −62.4497 −30.0952 −26.9520 −21.8559
S5 −83.7804 −36.1392 −28.3775 −22.5433
Table 3 Computed Values of PRD (%) for performance comparison

EMG signal Ref. no. EMD-Gaussian EEMD-Gaussian Proposed methodology
S1 0.3451 0.3447 0.3110
S2 0.3580 0.3576 0.3216
S3 0.3525 0.3520 0.3255
S4 0.3514 0.3511 0.3465
S5 0.3583 0.3580 0.3488
Hence these two performance metrics confirm the superior performance of the
proposed algorithm over the other two algorithms used. Therefore, the proposed
algorithm certainly outshines over the other two methodologies.
4 Conclusion
This paper introduced an improved algorithm for minimization of AWGN from

EMG signals based on EEMD algorithm which is followed by morphological
filtering [20]. EEMD algorithm effectively decomposes the EMG signal into IMFs
while morphological filtering successfully reduces the AWGN present with mini-
mum distortion of the original EMG signal. The performance of this methodology
was compared with EMD-and EEMD-based approaches. The proposed algorithm
achieved better results than the other two conventional approaches in terms of
suppression of AWGN in the EMG signal as validated by values of SNR (dB) and
PRD (%).
References
1. Luca, C. D. J., Adam, A., Wotiz, R., Gilmore, L. D., Nawab, S. H.: Decomposition of
Surface EMG Signals, Journal of Neurophysiology, vol. 96, pp. 1647–1654, (2006).
2. Ahsan, M. R., Ibrahimy, M., Khalifa, O. O.: EMG Signal Classification for Human Computer
Interaction A Review, Journal of Scientific Research, vol. 33, pp. 480–501, (2009).
3. Srivastava, A., Bhateja, V., Tiwari, H., Satapathy, S. C.: Restoration Algorithm for Gaussian
Corrupted MRI using Non-Local Average Filtering, 2nd International Conference on
Information Systems Design and Intelligent Applications (INDIA-2015), vol. 2, pp. 831–
840, (2015).
4. Canal, M. R.: Comparison of Wavelet and Short Time Fourier Transform Methods in the
Analysis of EMG Signals, Journal of Medical Systems, Springer, vol. 34, no. 1, pp. 91–94,
(2010).
5. Ren, X., Yan, Z., Wang Z., Hu, X.: Noise Reduction based on ICA Decomposition and
Wavelet Transform for the Extraction of Motor Unit Action Potentials, Journal of
Neuroscience Methods (Elsevier), vol. 158, pp. 313–322, (2006).
6. Mello, R. G. T., Oliveira, L. F., Nadal, J.: Digital Butterworth Filter for Subtracting Noise
from Low Magnitude Surface Electromyogram, Journal of Computer Methods and Programs
in Biomedicine (Elsevier), vol. 87, pp. 28–35, (2007).
7. Phinyomark, A., Phukpattaranont, P., Limsakul, C.: EMG Signal Denoising via Adaptive
Wavelet Shrinkage for Multifunction Upper-Limb Prosthesis, The 3rd International
Conference on Biomedical Engineering, pp. 35–41, (2010).
8. Djellatou, M. E. F., Nougarou, F., Massicote, D.: Enhanced FBLMS Algorithm for ECG and
Noise Removal from sEMG Signals, IEEE 18th International Conference on Digital Signal
Processing, pp. 1–6, (2013).
9. Harrach, M. A., Boudaoud, S., Hassan, M., Ayachi, F. S., Gamet, D., Grosset, J. F., Marin, F.:
Denoising of HD-sEMG Signals using Canonical Correlation Analysis, Medical & Biological
Engineering & Computing, vol. 55, no.3, pp. 375–388, (2017).
10. Zhang, X., Zhou, P.: Filtering of Surface EMG using Ensemble Empirical Mode
Decomposition, Journal of Medical Engineering and Physics, vol. 35, no. 4, pp. 537–542,
(2014).
11. Vergallo, P., Lay-Ekuakille, A., Giannoccaro, N. I., Trabacca, A., Morabito, F. C., Urooj, S.,
Bhateja, V.: Identification of Visual Evoked Potentials in EEG detection by Empirical Mode
Decomposition, (IEEE) 11th International Multi-Conference on Systems, Signals and
Devices- Conference on Sensors, Circuits & Instrumentation Systems, pp. 1–5, (2014).
12. Tiwari, D. K., Bhateja, V., Anand, D., Srivastava, A.: Combination of EEMD and
Morphological Filtering for Baseline Wander Correction in EMG Signals, International
Conference on Micro-Electronics, Electromagnetics and Telecommunications (ICMEET
2017), Springer-LNEE Vishakhapatnam (India), pp. 1–8, (2017).
13. Wu, Z., Huang, N. E.: Ensemble Empirical Mode Decomposition: A Noise Assisted Data
Analysis Method, Advances in Adaptive Data Analysis, vol. 1, no. 1, pp. 1–51, (2005).
14. Bhateja, V., Verma, R., Mehrotra, R., Urooj, S., Lay-Ekuakille, A., Verma, V. D.: A
Composite Wavelets and Morphology Approach for ECG Noise Filtering, 5th International
Conference on Pattern Recognition and Machine Intelligence (PReMI2013), vol. 8251,
pp. 361–366, (2013).
15. Verma, R., Mehrotra, R., Bhateja, V.: An Improved Algorithm for Noise Suppression and
Baseline Correction of ECG Signals, Springer, vol. 199, pp. 733–739, (2013).
16. Frank, A., Asuncion, A.: UCI Machine Learning Repository, University of California, School
of Information and Computer Science, Irvine, Calif, USA, (2010).
17. Shrivastava, A., Alankrita, Raj, A., Bhateja, V.: Combination of Wavelet Transform and
Morphological Filtering for Enhancement of Magnetic Resonance Images, International
Conference on Digital Information Processing and Communications (ICDIPC 2011), pp. 460–
474, (2011).
18. Andrade, A. O., Nasuto, S., Kyberd, P., Sweeney-Reed, C. M., Kanijn, V.: EMG Signal
Filtering based on Empirical Mode Decomposition, Journal of Biomedical Signal Processing
and Control, vol. 1, pp. 44–55, (2006).
19. Verma, R., Mehrotra, R., Bhateja, V.: An Integration of Improved Median and Morphological
Filtering Techniques for Electrocardiogram Signal Processing, 3rd IEEE International
Advance Computing Conference, Springer, pp. 1212–1217, (2013).
20. Qiang, L., Bo, L.: The Muscle Activity Detection from Surface EMG Signal using the
Morphological Filter, Applied Mechanics and Materials, vol. 195, pp. 1137–1141, (2012).
Visible-Infrared Image Fusion Method
Using Anisotropic Diffusion
Ashutosh Singhal, Vikrant Bhateja, Anil Singh

and Suresh Chandra Satapathy
Abstract In this paper, Visible and Infrared sensors are used to take comple-
mentary images of a targeted scene. Image fusion thus aims to integrate the two
images so that maximum information and fewer artifacts are introduced in the fused
image. The concept of merging two different multisensor images using the com-
bination of Anisotropic Diffusion (AD) and max–min approach is carried out in this
paper. Herein, each of the registered source images are decomposed into approxi-
mation and detailed layers using AD filter. Later, max–min fusion rules are applied
on detail and approximate layer, respectively, to preserve both spectral as well as
structural information. Image-quality assessment of the fused images is made using
structural similarity index (SSIM) , fusion factor (FF), and entropy (E) which
justifies the effectiveness of proposed method.
Keywords Multisensor Entropy Anisotropic diffusion (AD)
A. Singhal (&) V. Bhateja A. Singh

Department of Electronics and Communication Engineering,
Shri Ramswaroop Memorial Group of Professional Colleges (SRMGPC),
Lucknow 226028, Uttar Pradesh, India
e-mail: ashutoshsinghal71@gmail.com
V. Bhateja
e-mail: bhateja.vikrant@gmail.com
A. Singh
e-mail: lfsanilsinghgkp@gmail.com
S. C. Satapathy
Department of CSE, PVP Siddhartha Institute of Technology,
Vijayawada, Andhra Pradesh, India
e-mail: sureshsatapathy@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_51
526 A. Singhal et al.
1 Introduction
Image fusion is a method of synthesizing different registered source images having

complementary information; enhancing the capability of human visual perception.
The aim is to produce a new image, which carries complementary as well as
common features of individual images. Multisensor image fusion which involves
visible and infrared imaging sensors can be applied for various applications
involving guidance/detection system, navigation, surveillance, and targeting as well
as remote sensing images [1]. Visible images are organized in visible spectrum and
are formed due to scattering whereas thermal radiation creates infrared images and
are related with brightness. In unpropitious weather condition and during night use
of infrared over visible sensor is required [2]. There is a need to form a composite
image so that there will be no loss of information in order to provide a compact
representation of targeted scene with enhanced apprehension capability. Literature
in the last two decades provides various studies discussing number of image
multiresolution decomposition methodologies, such as Discrete Wavelet Transform
(DWT) [3], Laplacian Pyramid (LP) [4] and data driven methods, such as Empirical
Mode Decomposition (EMD) [5], Independent Component Analysis (ICA) [6] are
employed. But the outcome of these methods employs lack of directional infor-
mation, anisotropy and contains artifacts in the fused image. To overcome this
problem AD filter [7] is used which involve multiple iteration for the purpose of
image pixel segregation and processing into approximate and detail layer, respec-
tively. Various fusion mechanisms, such as Principal Component Analysis
(PCA) [8], pixel-based image fusion techniques which includes Intensity Hue
Saturation (IHS), maximum, minimum, averaging, and weighted averaging image
fusion have been popularly used. In addition, several other edge preserving and
smoothing schemes, such as Guided and Bilateral filter [9], are widely used for
fusion purpose. These methods [10] produce gradient reversal artifacts and halo
effects [7]. Godse et al. [11] worked on wavelet fusion technique by using maxi-
mum intensity approach but the fused image suffered from contrast reduction and
blurring. Sadhasivam et al. [12] performed fusion using PCA by selecting maxi-
mum pixel intensity but the obtained fused image has low contrast and illumination.
From the above discussion, it can be inferred that there is a need to remove the
artifacts and redundant information from the fused image so that it contain the
complementary information of both the source images. AD-based approach uses
intra-region smoothing; preserving edges and removes artifacts [6]. AD processing
of source images result in preserving both high-frequency information and contrast.
Max-and Min-based approach improves the redundant information and feature
enhancement property in the fused image. Therefore, it is assimilated that combi-
nation of AD and max–min can be better modeled to provide fusion of visible and
infrared images. In this paper, the decomposition of source images is carried out by
using AD and further min rule is applied on approximate layer and max rule is
applied on detail layer to yield final fused image. The proposed fusion results are
promising when evaluated by using parameters, such as SSIM, FF, and E.
Visible-Infrared Image Fusion Method … 527
2 Proposed Visible-Infrared Fusion Method
The scope of the problem presented in this paper is to perform Visible and Infrared
image fusion by using combination of AD and Max–Min fusion rule. Max and Min
fusion rule involves selection of maximum and minimum pixel from every corre-
sponding pixel of input images, which in turns forms resultant pixel of input
images. Thus, every pixel of fused image will be pixel having maximum and
minimum intensity of corresponding pixel in the input images. It helps to suppress
the redundant information and highlights the components having maximum infor-
mation in the source image.
2.1 Anisotropic Diffusion (AD) Filtering
Perona and Malik [13] performed nonlinear filtering on source images to produce
base layer and named it as anisotropic diffusion filter. It is an inhomogeneous
process that reduces the diffusivity at that location where there is much likelihood to
be edges. AD utilizes flux Eq. (1) in order to control the diffusion of an image
I which is governed by Eq. (1).
At ¼ pðm; n; tÞDA þ rp : rA ð1Þ
Where, p(m, n, t) = Flux function also known as diffusion rate, D,r are lapla-
cian and Gradient operator, respectively, t = iteration. Equation (1) is also known
as Heat Eq. whose solution is given as:
h i
tþ1
Im;n ¼ Im;n
t
þ k cN :rN Im;n
t
þ cS :rS Im;n
t
þ cE :rE Im;n
t
þ cW :rW Im;n
t
ð2Þ
tþ1
In Eq. (2) Im;n represent coarser image which depend on previous coarser image
scale. Value of k lies between 0 k 1=4. rN , rS , rE and rW represent nearest
neighbor distance in respective direction. They can be defined as:
rN Im;n ¼ Im1;n Im;n ð3Þ
rS Im;n ¼ Im1;n Im;n ð4Þ
rE Im;n ¼ Im1;n Im;n ð5Þ
rW Im;n ¼ Im1;n Im;n ð6Þ
Perona and Malik [13] further proposed two equations as given below as
Eqs. (7) and (8).
krAk 2
sðrAÞ ¼ eð c Þ ð7Þ
1
sðrAÞ ¼ 2 ð8Þ
krAk
1þ c
Equations (7) and (8) regulates smoothing and edge-preservation. Equation (7)
is utilized if image consists of high contrast edges over low contrast edges while
Eq. (8) is effective when wider regions dominate over smaller regions in the image
[7]. These attributes of the conventional AD filter justifies its usage for multisensor
images.
2.2 Image Fusion Algorithm Based on the Combination

of AD and Max-Min Rule
The first preprocessing step involves conversion of multisource image from RGB to
gray scale. After preprocessing, visible and infrared images (source images) are
processed using AD. When AD filter is applied on both the registered multisource
images; base layer is obtained. Further, the procedural algorithmic steps for the
same are detailed in Table 1.
Table 1 Proposed fusion algorithm using AD, Max–Min fusion rule

BEGIN
Step 1: Input source images A1 and A2
Step 2: Decompose source image by AD to obtain base layer B1 and B2 . This can be achieved via
following steps:
2. (a) Determine diffusion rate of source image using Eq. (1)
2. (b) Eqs. (7) and (8) can be utilizes to yield base layer
Step 3: Detail layer D1 and D2 can be obtained by subtracting base layer from source image as
given in Eq. (9)
Dk ðm; nÞ ¼ Ak ðm; nÞ Bk ðm; nÞ (9)
Where, Dk ðm; nÞ is detail layer, Ak ðm; nÞ is source image and Bk ðm; nÞ is base layer
Step 4: Fuse detail layer of both input source images using max fusion rule given by Eq. (10)
Dk ðm; nÞ ¼ maxðD1k ðm; nÞ; D2k ðm; nÞÞ (10)
Step 5: Fuse base layer of both input source images using min fusion rule given by Eq. (11)
Bk ðm; nÞ ¼ minðB1k ðm; nÞ; B2k ðm; nÞÞ (11)
Step 6: Base and detail layer are combined using max rule to obtained final result. This is
achieved by employing Eq. (12)
Fk ðm; nÞ ¼ maxðBk ðm; nÞ; Dk ðm; nÞÞ (12)
F(m, n) is resultant fuse image
END
After the generation of final reconstructed image using the proposed fusion
algorithm; their performance parameters, such as SSIM, FF and E [14–16] are
calculated. Higher values of these parameters justify the worthiness of the proposed
fusion algorithm.
Fusion of different Visible and Infrared images is performed by decomposing it into

base layer by applying AD. Further finding detail layer and combining each layer of
corresponding source images by utilizing Max and Min fusion rule as fusion
algorithms. The proposed fusion method is tested with four different sets of visible
and infrared images IMG-1(p1, q1, r1), similarly for IMG-2, IMG-3, IMG-4 to
confirm the effectiveness. It can be inferred that in IMG-1 due to image capture
from visible sensor house rooftop and greenery in targeted scene is clearly visible
whereas roads are not clearly visible as shown in Fig. 1 (p1) whereas in Fig. 1 (q1)
due to image captured from infrared sensor roads and person are clearly visible but
the house rooftop and greenery are not clearly visible. Figure 1 (r1) represents the
fusion of visible and infrared image from the proposed algorithm which shows
house rooftop, roads, greenery and person in the fused image. High values of SSIM
depict measure similarity in structure between the source and fused images as
shown in Table 2. This has been the methodology of utilizing the fused image to
improve the loss of information. Similarly, in IMG-2, Fig. 1 (p2) clearly shows
alphabets on the roof and street lights whereas cars and people are not clearly
visible while in Fig. 1 (q2), the people and car can be clearly visualized while the
streetlight and alphabet on the roof is not. The obtained fused images show both the
complementary features which inferred the effectiveness of proposed fusion
method. Figure 1 also depicts the other test input source images and the obtained
fusion results.
The obtained results are evaluated for fusion quality using performance
parameters, such as E, FF, and SSIM. The overall impact is that the Entropy (E) of
the fused image has been improved which shows a high amount of information
content in the fused image. It is further verified through Fusion Factor (FF) shown
in Table 2. Large values of E, FF, and SSIM depicts better performance of pro-
posed fusion approach for various source images. Hence, the proposed methodol-
ogy is validated to show its worth in conserving the spectral as well as spatial
features.
(p1) (q1) (r1)
(p2) (q2) (r2)
(p3) (q3) (r3)
(p4) (q4) (r4)
Fig. 1 An illustration of image fusion results for different sets of visible and infrared images using
proposed methodology: (p1–p4) Input source images (Visible), (q1–q4) Input source images
(Infrared), (r1–r4) Output fused images
Table 2 Image quality Images E FF SSIM

computation using various
fusion metrics IMG-1 6.5518 6.4997 0.7536
IMG-2 4.9323 5.2636 0.9583
IMG-3 7.1145 6.2284 0.9342
IMG-4 6.5427 6.8427 0.8802
4 Conclusion
An AD-based fusion method applicable to visible and infrared images employing

Max and Min fusion rule is proposed in this paper. Since, the image acquired by
using infrared radiations contains few details with high contrast, whereas visible
imaginary provides plenty of high-frequency information but have low contrast. AD
processing of source images results in preserving both high-frequency information
and contrast. This yields a better result than other multiresolution decomposition
techniques [1, 3, 14]. AD and max–min-based image fusion approach leads to
minimum loss of information and artifacts. It can be easily observed from the result
and the same can be easily validated using different fusion metrics pertaining to E,
FF, and SSIM. Further, the AD can be utilized with other fusion algorithm, such as
Karhunen–Loeve (KL) Transform [7], Non-Subsampled Contourlet Transform
(NSCT) [17], etc. to obtained better results.
References
1. Srivastava, A., Bhateja, V., Moin, A.: Combination of PCA and Contourlets for Multispectral
Image Fusion, In: International Conference on Data Engineering and Communication
Technology (ICDECT-2016), Pune, India, Vol. 2, pp. 1–8, Springer (March, 2016).
2. Bulanon, D. M., Burks, T. F., Alchanatis, V.: Image fusion of visible and thermal images for
fruit detection, In: Biosystems Engineering, Vol. 103, No.1, pp. 12–22—Elsevier (May,
2009).
3. Krishn, A., Bhateja, V., Himanshi, Sahu, A.: Medical Image Fusion Using Combination of
PCA and Wavelet Analysis, In: International Conference on Advances in Computing,
Communication and Informatics (ICACCI-2014), Gr. Noida (U.P), India, pp. 986–991, IEEE
(September, 2014).
4. Sahu, A., Bhateja, V.,Krishn, A., Himanshi: Medical Image Fusion with Laplacian Pyramid,
In: International Conference on Medical Imaging, m-Health & Emerging Communication
Systems (MEDCom-2014), Gr. Noida (U.P.), pp. 448–453, IEEE (November, 2014).
5. Hui, C. S., Hongbo, S., Renhua, S., Jing, T.: Fusing remote sensing images using a trous
wavelet transform and empirical mode decomposition, In: Pattern Recognition Letters, Vol.
29, No.3, pp. 330–342, Elsevier (1 February, 2008).
6. Mitianoudis, N., Stathaki, T.: Pixel-based and region-based image fusion schemes using ICA
bases, In: Information Fusion, Vol. 8, No. 2, pp. 131–142, Elsevier (2007).
7. Bavirisetti, D. P., Dhuli, R.: Fusion of Infrared and Visible sensor images based on
Anisotropic diffusion and Karhunen-Loeve Transform, In: IEEE Sensor Journal, Vol. 16,
No. 1, pp. 203–209, IEEE (14 September 2015).
8. He, C., Liu, Q., Li, H., Wang, H.: Multimodal Medical Image Fusion Based on IHS and PCA,
In: Symposium on Security Detection and Information Processing, Vol. 7, pp. 280–285,
Elsevier (2010).
9. He, K., Sun, J., Tang, X.: Guided Image Filtering, In: IEEE Transaction on Pattern Analysis
and Machine Intelligence, Vol. 35, No. 6, pp. 1397–1408, IEEE (June 2013).
10. Li, S., Kang, X., Hu, J.: Image Fusion with Guided Filtering, In: IEEE Transactions on Image
Processing, Vol. 22, No. 7, pp. 2864–2875, IEEE (July 2013).
11. Godse, D. A., Bormane, D. S.: Wavelet based Image Fusion using Pixel based Maximum
Selection rule, In: International Journal of Engineering Science and Technology, vol. 3, no. 7,
pp. 5572–5578, (2011).
12. Sadhasivam, S. K., Keerthivasan, M. K., Muttan, S.: Implementation of Max Principle with
PCA in Image Fusion for Surveillance and Navigation Application, In: Electronic Letters on
Computer Vision and Image Analysis, Vol. 10, No. 1, pp. 1–10, (2011).
13. Perona, P., Malik, J.: Scale-Space and Edge Detection Using Anisotropic Diffusion, In: IEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol. 12, No. 7, pp. 629–639,
(July 1990).
14. Himanshi, Bhateja, V., Krishn, A., Sahu, A.: Medical Image Fusion in Curvelet Domain
Employing PCA and Maximum Selection Rule, In: International Conference on Computers
and Communication Technologies (IC3T-2015), Hyderabad, India, vol. 1, pp. 1–9, IEEE
(July, 2015).
15. Moin, A., Bhateja, V., Srivastava, A.: Multispectral Medical Image Fusion using PCA in
Wavelet Domain, In: Proc. (ACM-ICPS) Second International Conference on Information and
Communication Technology for Competitive Strategies (ICTCS-2016), Udaipur, India, pp. 1–
6, (March, 2016).
16. Krishn, A., Bhateja, V., Himanshi, Sahu, A.: PCA based Medical Image Fusion in Ridgelet
Domain, In: International Conference on Frontiers in Intelligent Computing Theory and
Applications (FICTA-2014), Bhubaneswar, India, vol. 328, pp. 475–482, Springer
(November 2014).
17. Xiao-bo, Q., Wen, Y. J., Zhi, X. H., Zi-Qian, z.: Image Fusion Algorithm Based on Spatial
Frequency-Motivated Pulse Coupled Neural Networks in Nonsubsampled Contourlet
Transform Domain, In: Acta Automatica Sinica, vol. 34, no.12, pp. 1508–1514, Elsevier
(2008).
Fast Radial Harmonic Moments
for Invariant Image Representation
Shabana Urooj, Satya P. Singh, Shevet Kamal Maurya

and Mayank Priyadarshi
Abstract The main objective of this paper is to reduce the reconstruction error and
fast calculation of Radial Harmonic Fourier Moments (RHFM). In the proposed
work, the fast RHFM has been applied on the original gray image for reconstruc-
tion. Before applying RHFM on grayscale image, the image in portioned into radial
and angular sectors. Results are compared with traditional methods. The proposed
approach results in better reconstruction error. Also, moments can be calculated at
high speed using proposed approach.
Keywords Radial harmonic Fourier moments Reconstruction error

Computational speed
1 Introduction
In the image processing technology, the immunity of an Image is a very violent

problem and quality of image also one problem. The improvement of the quality of
an image one fast technology applied by author fast Radial Harmonic Fourier
Moments (RHFM), and give the batter results by the new fast RHFMs. The author
will apply the proposed approach to the gray image for reconstruction. This will
reduce the consume time and error of an image in the reconstruction performance.
The proposed approach of the radial harmonic Fourier moments are the regenerated
techniques of the traditional approach. The Fourier Moments, the linear combina-
tion of sine and cosine wave are the Fourier series. The application of Fourier
moments is in image processing, signal processing, and quantum mechanics. First
reconstruction of an image by Zernike moments/pseudo-Zernike moments into
S. Urooj (&) S. P. Singh S. K. Maurya M. Priyadarshi

School of Engineering, Gautam Buddha University, Gr. Noida, UP, India
e-mail: shabanaurooj@ieee.org

https://doi.org/10.1007/978-981-10-7245-1_52
534 S. Urooj et al.
image watermarking conducted by Shirani et al. [1–3]. Invariant extraction has been
executed by Tchebichef moments in Deng et al. for the reconstruction of an image
by radial exponent moments [4]. The proposed reconstruction of an image through
Radial Harmonic Fourier Moments [5, 6].
2 Proposed Work
2.1 Radial Harmonic Fourier Moments
In the polar coordinate system, RHFM of an image f ðr; hÞ, 0 r 1, and

0 h 2p, are describe on unit disk. Let /nm are the RHFMs and order and
repetition of function is n, m, n 0, m 0 then RHFMs are defined as follows:
Z Z
1
Mnm ¼ f ðr; hÞPnm ðr; hÞrdrdh ð1Þ
p
where, P nm ðr; hÞ are the conjugate of Pnm ðr; hÞ and defined as follows:
Pnm ðr; hÞ ¼ Tn ðr Þ expðjm hÞ ð2Þ

8 pffiffiffiffiffiffiffi
ffiffiffiffiffiffiffi for n ¼ 0
< p1=r
Tn ð r Þ ¼ 2=r sinðn þ 1Þpr for n ¼ odd ð3Þ
: pffiffiffiffiffiffiffi
2=r cos npr for n ¼ even
The radial basis function Pnm ðr; hÞ also satisfies orthogonal conditions as
follows,
Z2p Z1
Pnm ðr; hÞPk‘ ðr; hÞrdrdh ¼ 2pdnk dml ð4Þ
0 0
where d is Kronecker delta and 2p is the normalization factor.

The image f ðr; hÞ can be reconstructed using RHFMs as
X
^f ðr; hÞ ¼ En‘ Hn‘ ðr; hÞ ð5Þ
n‘2S
In the rectangular coordinate system [7] reconstruction of RHFM two types by

this method circular inscribed mapping and on the circle circumscribed mapping. In
which circular circumscribed mapping are not suitable for the pattern recognition
image watermarking and orthogonal moments. Then we will use only circular
inscribed mapping.
Fast Radial Harmonic Moments for Invariant Image Representation 535
The image function 0 f 0 is mapped into polar coordinated to calculate the coef-
ficient of Eq. 1 [8, 9]. We divide the whole image in radial sector Nr and angular
size of Nu into polar coordinates so that the grid density is high enough as per
sampling theorem, i.e., Nu 2‘max. In this way, we partitioned the angular and
radial components of Eq. 1 into M equal parts as:
1MX1 X
2p M 1
Nr ¼ u and Nh ¼ v ð6Þ
M u¼0 M v¼0
Therefore, we obtained M 2 subregions inside the unit disk (Fig. 1). Now, the
coordinates (Nr ; Nh ) are changed to rectangular coordinates as x ¼ ðNr
2 cos Nh Þ, and y ¼ Nr 2 sin Nh . Therefore, the polar coordinate can be
N N
written in terms of portioned data as follows;
fp ðNr ; Nh Þ ¼ f ði; jÞ ð7Þ
Let us define wn‘ as partitioned RHFMs and Eq. 1 can be rewritten as follows:
Z 2p Z 1
1
wn‘ ¼ f ðNr ; Nh Þ½Tn ðNr Þ expði‘Nh ÞNr dNr dNh ð8Þ
2p 0 0
Fig. 1 Angular radial

partitioning of an image to
N angular and M radial sectors
536 S. Urooj et al.
Substituting Eq. 6 into Eq. 8 we get:
Z2p Z1
1 1 2p
wn‘ ¼ f ðNr ; Nh Þ½Tn ðNr Þ expði‘Nh ÞNr du dv
2p M M
0 0 ð9Þ
1 X
M X
1 M 1
wn‘ ¼ f ðNr ; Nh ÞHn ðNr Þ expði‘Nh Þ
M2 u¼0 v¼0
where, the term Hn ðNr Þ is the radial kernel of the proposed method partitioned
RHFMs and Eq. 3 can be rewritten as follows:
8 pffiffiffiffiffiffiffi
ffiffiffiffiffiffiffi for n ¼ 0
< p1=r
Tn ðr Þ ¼ 2=r sinðn þ 1Þpr for n ¼ odd ð10Þ
: pffiffiffiffiffiffiffi
2=r cos npr for n ¼ even
3 Reconstruction of Image and Error in Reconstruction
The image f ðr; hÞ reconstructed approximately by order and repetition of fast

RHFMs (n N, m M). Then, the reconstruction performance of an image are
denoted by ^f ðr; hÞ. The orthogonal function will gives the independence benefici-
ation for the image reconstruction.
Then normalized reconstruction error of an Image (NIRE)
R R1 2
f ðr; hÞ ^f ðr; hÞ drdh
1R R
e ¼
2
1 2 ð11Þ
1 f ðr; hÞdrdh
4 Experiment Result
In the first experiment, the proposed method is tested and validated for recon-
struction capabilities. The results are also compared with the traditional approach.
Standard grayscaled Lena image is used to validate the proposed approach. The
standard Lena image of size 512 512 is down sampled to the size of 128 128.
For this experiment a set of moments/coefficients are calculated using Eq. 7 and
again image reconstruction process is performed from these sets of moments/
coefficients. The mean square error is calculated for measuring the reconstruction
performance using the following equation:
P
f ðx; yÞ ^f ðx; yÞ
x2 þ y2 1
MSRE ¼ P
x2 þ y2 1 f 2 ðx; yÞ
Fast Radial Harmonic Moments for Invariant Image Representation 537
Method
Original
Image
Traditional
Approach
Proposed
Approach
Order 10 20 30 40 60
Fig. 2 Reconstruction performance of image by fast RHFMs
0.8
RHFM
0.6
MSRE
Proposed
0.4
0.2
0
0 5 10 15 20 25 30
Nmax=Mmax
Fig. 3 Comparative analysis of reconstruction error using proposed method and traditional
method
The results are shown in Fig. 2 for moments order and repetition 10, 20, 30, and
40. If we take a closer look at the center of reconstructed images, we can observe
that there is information loss using the traditional method as we increases the order
of moments. On the other hand, the reconstruction error continuously decreases
using proposed method. In continuation, we also plot behavior of reconstruction
error using the proposed method and traditional method. The results are shown in
Fig. 3. It can be seen that after certain degree of moments, the reconstruction error
of the traditional method starts increasing while this error continuously decrease
using the proposed method.
538 S. Urooj et al.
5 Conclusions
The traditional computing approach of RHFM suffers from numerical integration

error which results in reconstruction error. To tackle this issue, a new approach has
been proposed by portioning the radial and angular parts into small sectors. To test
the proposed methods, two experiments have been carried out. In the first experi-
ment, the proposed method is tested for its reconstruction performance and in the
second experiment; the proposed approach is tested for computational speed. The
proposed approach is also compared with the traditional approach. It is concluded
that the proposed approach significantly reduces the reconstruction error. The
proposed approach also results in high-computational speed.
In future work, the proposed approach for RHFMs can be applied on the pattern
recognition, and image segmentation. We can also apply this approach to the color
Image.
References
1. S. Shirani, M. Farzam, A robust multimedia watermarking technique using Zernike transform,

in 2001 IEEE 4th Workshop on Multimedia Signal Processing ((MMSP 01), 2001, CANNES,
pp. 529–534).
2. Singh, Satya P., Shabana Urooj, and Aime Lay Ekuakille. “Rotational-invariant texture
analysis using radon and polar complex exponential transform.” Proceedings of the 3rd
International Conference on Frontiers of Intelligent Computing: Theory and Applications
(FICTA) 2014. Springer, Cham, 2015.
3. Singh, Satya P., and Shabana Urooj. “An improved CAD system for breast cancer diagnosis
based on generalized pseudo-Zernike moment and Ada-DEWNN classifier.” Journal of medical
systems 40.4 (2016): 1–13.
4. Y. D. Zhang, C. Shao, Orthogonal moments based on exponent functions: exponent- Fourier
moments, Pattern Recognition. 47 (2014) 2596–2606.
5. C. Singh, S. K. Ranade, A high capacity image adaptive watermarking scheme with Radial
Harmonic Fourier Moments, Digit. Signal Processing. A Rev. J. 23 (2013) 1470–1482.
6. C. Singh, R, Upneja, A Computational Model of Enhanced accuracy of radial harmonic Fourier
moments, in World Congress of Engineering, 2012, London, pp. 1189–1194.
7. H.T. Hu, Y.D. Zhang, C. Shao, Q. Ju Orthogonal moments based on exponent function:
Exponent Fourier Moments, Pattern Recognition. 47(2014) 2596–2606.
8. Singh, Satya P., and Shabana Urooj. “Combined rotation-and scale-invariant texture analysis
using radon-based polar complex exponential transform.” Arabian Journal for Science and
Engineering 40.8 (2015): 2309–2322.
9. Singh, Satya P., and Shabana Urooj. “Accurate and Fast Computation of Exponent Fourier
Moment.” Arabian Journal for Science and Engineering (2017): 1–8.
A Wide-Area Network Protection Method
Using PMUs
Namita Chandra and Shabana Urooj
Abstract This paper proposes the idea of utilizing the data of PMU’s (Phasor
Measurement Unit) to detect different types of faults in transmission lines. Data
measured from PMUs is collected at a system control center. Proposed method
equates the positive sequence voltages of different buses to detect the faulted bus.
With the help of synchronized phasor measurements, bus voltages are estimated by
using different paths, furthermore, they are matched for the detection of the faulted
line also. The proposed protection scheme is verified for balanced and unbalanced
faults. Simulation has been done with MATLAB/SIMULINK for a 230 kV IEEE
9-bus system.
Keywords Wide-area measurement (WAM) PMU Backup protection
1 Introduction
In general, the aim of backup protection is to clear the faults that are not cleared by
the primary protection. Backup protection operates more slowly than primary
protection, to ensure proper coordination, and is less selective [1]. As it has to
protect a larger part of a power system, settings become more difficult and
dependent on the operating state of the system [2]. The main drawbacks of con-
ventional backup protection are coordination among several backup protections and
delay caused due to measuring devices.
With the growth of computer applications and communication networks,
wide-area measurement system (WAMS) is one of the most significant innovative
developments in the field of existing power system [3]. WAM is able to offer a
N. Chandra (&) S. Urooj

Department of Electrical Engineering, School of Engineering, Gautam Buddha University,
Greater Noida 201310, Uttar Pradesh, India
e-mail: namitachandra.94@gmail.com
S. Urooj
e-mail: shabanabilal@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_53
540 N. Chandra and S. Urooj
real-time understanding of the dynamic performance of a power system that updates

once per cycle, through inventions in the field of synchronized measurement
technology and PMUs [4]. It is a vital resource for constructing new applications
that can benefit power system protection and control.
In a recent report [4] 70% of wide-area disturbances were due to relay malop-
eration. These reported commotions were majorly caused due to poor relay settings
or unseen failures in the protection system. Incorrect operation of the relay in
wide-area disturbances is a source of concern these days, as these disturbances have
played a major part in recent power failures [5]. Considering the shortcomings of
backup security it has become essential to develop new protection scheme using
WAMS.
Various schemes have been developed in [6–9] for faulted line identification
based on the WAMS. In [6] positive sequence voltage magnitudes are paralleled to
detect bus nearest to the fault and then positive sequences current angles at each line
are associated to identify the faulty line. In [7] PMU MATLAB simulator is created.
Its aim is to help in the understanding of the behavior of algorithms internal to the
PMU. A wide-area-based backup protection scheme for series compensated line is
created. It also works in the case of voltage and current inversions. PMUs were
assumed at all the buses [8]. PMUs can also be employed in state estimation of
energy management systems [9]. The study is focused on the improving the power
systems immunity toward tragic letdowns by using PMUs within the system.
In this paper, a new protection scheme has been proposed using the data col-
lected from time-synchronized phasor measurement units placed at selected buses,
resulting in less number of PMUs. This scheme is simple in terms of computation
and is cost effective. Simulations have been done with MATLAB/SIMULINK
environment on 230 kV, 9-bus system.
In this paper proposed protection scheme is discussed in Sect. 2. Simulation
results are detailed in Sects. 3 and 4 conclude the paper.
2 Proposed Protection Scheme
Whenever the system gets unbalanced (i.e., fault condition), its unbalanced phasors
can be transformed into three balanced phasors (positive, negative, and zero) known
as sequence components. This method matches the sequence voltages of unlike
buses to decide the faulted bus. It was found that during the instance of a fault the
voltage of the bus closest to the fault will observe maximum swerve [8]. For
detection of faulted bus, positive sequence voltages of selected buses are compared.
After determining the faulted bus, line at fault is also determined by estimating
non-PMU bus voltages using voltages and currents of buses having PMU through
different paths. A single line diagram of a 9-bus system with PMUs at selected
locations used for the proposed protection scheme is shown in Fig. 1. Locations of
PMUs are decided considering that each transmission line is connected to a PMU
bus [10–12].
A Wide-Area Network Protection Method Using PMUs 541
Fig. 1 Single line diagram of a 9-bus system
2.1 Identification of Faulted Bus
For the detection of the faulted bus, positive sequence voltage magnitudes are
collected from the PMU buses at a system control center. The basic principle behind
the detection of the faulted bus is that at the instance of a fault, voltage magnitude is
reduced to a certain extent. The condition used to detect if there is a fault in the
system is as follows:
jVFAULT j\K VR ; ð1Þ
where VFAULT is the positive sequence voltage magnitude at the time of the fault,
VR is the rated voltage magnitude, and K is a constant. K is used to calculate the
voltage thresholds for detecting a fault in a system. Its value is set at 0.7 for this
system and is decided after carrying out some simulations.
If Eq. (1) is true the bus with a minimum value of positive sequence voltage
magnitude will be assumed as a faulted bus. In this scheme, if a fault is nearer to a
non-PMU bus, it will not be identified as a faulted bus as only PMU buses are
compared for detection of the faulted bus. So, until the detection of faulted line, the
nearest PMU bus will be assumed as the faulted bus.
2.2 Identification of Faulted Line
After locating faulted bus, lines connected to the faulted bus are tested for the
identification of the faulted line. For the detection of faulted line, positive sequence
Fig. 2 Single line diagram of a subnetwork
voltages are estimated from different paths, for the non-PMU buses utilizing the
values of PMU buses. For estimating the non-PMU bus voltages, a hyperbolic form
of the long transmission line equation is utilized.
VYX ¼ VX coshðc LÞ IX ZC sinhðc LÞ; ð2Þ
where VYX is the estimated positive sequence voltage for a Y (non-PMU) bus using
the positive sequence voltages and currents of X (PMU) bus. ZC is the surge
impedance of the line of length L and c is the propagation constant.
The basic concept for detection of the faulted line is that during normal con-
ditions estimated voltages of a non-PMU bus through different paths would be
almost equal. But, if a fault occurs on the line connected to the bus whose voltage is
estimated, then the estimated values of voltages through different paths may differ.
This logic is used for the detection of the line at fault. This estimation is done only
for the buses which are directly connected to the faulted bus. When bus X is the
faulted bus as shown in Fig. 2 then DVY is calculated for all the buses connected to
the faulted bus as follows:
DVY¼ jVYX VYZ j: ð3Þ
During normal conditions, the value of DV will be approximately equal to zero.

But in the case of fault, the line with the minimum value of DV will be considered as
the faulted line.
Faulted line ¼ Min fDVg ð4Þ
2.3 Flowchart of the Proposed Protection Scheme
PMUs measure the voltage and current phasors of all the buses. These measure-
ments are transferred to a phasor data concentrator (PDC) at the system control
center. The PDC links the data by time tag to build a WAMS. The flowchart of this
protection scheme is given in Fig. 3.
Fig. 3 Flowchart of the

proposed scheme
3 Simulation Results
The proposed protection scheme is tested for a 9-bus power system as shown in
Fig. 1 with faults on line 7–8. Using MATLAB R2013a, simulations are carried out
for different faults (balanced and unbalanced), using positive sequence voltages and
currents. The sequence components are calculated using the simPowerSystem tool
box present in MATLAB/SIMULINK. The number of PMUs required for the
protection of transmission lines in this system is 3. PMUs are placed on bus 4, 7,
and 9 so that each line is connected to at least one PMU.
3.1 Results from Identification of Faulted Bus
A single phase, line-to-ground (LG) fault of 5 Ω is created on line 7–8 at 0.4 s,

30 km from bus 7. Figure 4a shows the positive sequence voltage magnitudes of
(a) 1.05
1 Bus 7
Bus 4
0.95 Bus 9
Voltage (p.u)
0.9
0.85
0.8
0.75
0.7
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (s)
(b)
1
Bus 7
Bus 4
0.8
Voltage (p.u)
Bus 9
0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (s)
Fig. 4 Positive sequence bus voltage magnitudes for (a) LG fault (b) LLL fault
buses 7, 9, and 4. The simulated output shows that all the buses have been affected
from the LG fault but bus 7 has been affected the most. Bus 7 has minimum positive
sequence voltage hence it is assumed as the faulted bus. Figure 4b shows the
positive sequence voltage magnitudes for LLL (three phase) fault.
3.2 Results from Identification of Faulted Line
To verify the result of the faulted bus identification, the faulted line is identified. If
bus 7 is identified as the faulted bus then either line 7–8 or line 7–5 will be the
Table 1 Results of faulted line identification

Type of Fault resistance Fault location from bus 7 DV8 DV5
fault (Ω) (%) (kV) (kV)
L-G 5 30 5.8 9
70 1.9 6.7
50 30 4.2 6.7
70 1.7 5.5
L-L-G 5 30 11.95 18.62
70 3.71 14.14
50 30 7.6 13.1
70 3 10.8
L-L-L 5 30 18.7 29.09
70 5.54 22.04
L-L-L-G 5 30 18.7 29.09
faulted line. Hence, positive sequence voltages of the buses next to the faulted bus 7
are estimated through different paths using Eq. (2). As bus 8 and 5 are the buses
next to the faulted bus, V87, V89 are calculated for bus 8 and V57, V54 are calculated
for bus 5. Then DV is calculated from Eq. (3) as follows:
DV8 ¼ jV87 V89 j: ð5Þ
DV5 ¼ jV57 V54 j: ð6Þ
On comparing the values of DV8 and DV5, DV8 was found the minimum. Hence,
confirming the fault on line 7–8 as bus 7 was considered the faulted bus. Table 1
shows the results of faulted line identification with different types of faults, fault
locations, and fault resistances. It is apparent from the following table that this
protection scheme is capable for the identification faulted line for balanced and
unbalanced faults in all cases. Results show the effectiveness of the proposed
protection scheme.
4 Conclusion
In this paper, a backup protection scheme was proposed using WAMS (Wide-Area
Measurement System). Positive sequence voltage magnitudes of only PMU buses
were compared for the identification of faulted bus. Utilizing the measurements of
PMUs, the faulted line was identified. Positive sequence voltage of buses next to
the faulted bus was estimated using different paths. The proposed protection scheme
was tested on IEEE 9-bus system for unbalanced and balanced faults. A major
achievement of this protection scheme is the lessening in the number of PMUs
without being intricate. The main benefit of this scheme is its simplicity,
cost-efficiency, and low computation. In future, this protection scheme can also be
tested for non-fault events like load encroachment and power swing.
References
1. Jena, M.K., Samantaray, S.R., Panigrahi, B.K.: A New Wide-Area Backup Protection
Scheme for Series-Compensated Transmission System. IEEE Systems Journal https://doi.org/
10.1109/JSYST.2015.2467218 (2015).
2. Phadke, A.G., Wall, P., Ding, L. et al.: J. Mod. Power Syst. Clean Energy 4: 319. https://doi.
org/10.1007/s40565-016-0211-x (2016).
3. Urooj, S., Sood, V.: Phasor Measurement Unit (PMU) Based Wide Area Protection System.
2nd International Conference on Computing for Sustainable Global Development
(INDIACom) (2015).
4. Horowitz, S.H., Phadke, A.G.: Third zone revisited. IEEE Trans Power Deliv 21(1):23–29
(2006).
5. Report of the Enquiry Committee on grid disturbance in northern region on 30th July 2012
and in northern, eastern & north-eastern region on 31st July 2012. The Enquiry Committee,
Ministry of Commerce and Industry, Government of India, New Delhi, India, (2012).
6. Eissa, M.M., Masoud, M.E., Elanwar, M.M.M.: A novel backup wide area protection
technique for power transmission grids using phasor measurement unit. IEEE Trans. Power
Del., vol. 25, no. 1, pp. 270–278, Jan. (2010).
7. Dotta, D., Chow, J.H., Vanfretti, L., Almas, M.S., Agostini, M.N.: A MATLAB-based PMU
Simulator. Power and Energy Society General Meeting (PES), IEEE, (2013).
8. Nayak, P.K., Pradhan, A.K., Bajpai, P.: Wide-Area Measurement-Based Backup Protection
for Power Network With Series Compensation. IEEE Trans. Power Del., Nov (2013).
9. Ree, J.D., Centeno, V., Thorp, J.S., Phadke, A.G.: Synchronized phasor measurement
applications in power systems. IEEE Trans. Smart Grid, vol. 1, no. 1, pp. 20–27, Jun. (2010).
10. Roy, B.K.S., Sharma, R., Pradhan, A.K., Sinha. A.K.: Faulty Line Identification Algorithm
for Secured Backup Protection Using PMUs. Electric Power Components and Systems, 45:5,
491–504, https://doi.org/10.1080/15325008.2016.1266417 (2017).
11. Roy, B.K.S., Sinha, A.K., Pradhan, A.K..: An optimal PMU placement technique for power
system observability. Int. J. of Elect. Power Energy Syst., Vol. 42, No. 1, pp. 71–77,
November (2012).
12. Phadke, A.G., Thorp, J.S., Adamiak, M.G..: A new measurement technique for tracking
voltage phasors, local system frequency, and rate of change of frequency. IEEE Trans Power
Appar Syst 102(5):1025–1038 (1983).
Analysis and Prediction of the Effect
of Surya Namaskar on Pulse of Different
Prakruti Using Machine Learning
Jayshree Ghorpade-Aher, Abhishek Girish Patil, Eeshan Phatak,

Sumant Gaopande and Yudhishthir Deshpande
Abstract “Surya Namaskar” is the key for Good health! Today’s social life can be
made easier and healthier using the mantra of “YOGA”. Nadi Parikshan is a
diagnostic technique which is based on the ancient Ayurvedic principles of Wrist
Pulse analysis. Nadi describes the mental and physical health of a person in great
depth. This information can be used by practitioners to prevent, detect as well as
treat any ailment. Surya Namaskar is a Yoga exercise which has multiple health
benefits and a direct impact on Pulse. Prakruti of a person is a metaphysical
characteristic and a combination of the three doshas in Ayurveda viz. Vatta, Pitta,
and Kapha which remains constant for the lifetime. Experimentation was carried
out to analyze the effect of Surya Namaskar exercise on the Pulse of different
Prakruti. The Pulse was recorded for a group of young students aged between 19
and 23 years with different Prakruti before Surya Namaskar and after Surya
Namaskar for a period of 4 days. This paper analyzes the effect of Surya Namaskar
on human Pulse and proposes a framework to predict Pulse after Surya Namaskar.
The changes that Surya Namaskar causes in the Pulse are studied and used to
predict Pulse after performing Surya Namaskar. This analysis helps understand how
Surya Namaskar benefits the health of a person. Performing Surya Namaskar in our
daily routine would improve the health of the society as a whole making the
subjects energetic and active.
J. Ghorpade-Aher (&) A. G. Patil E. Phatak S. Gaopande Y. Deshpande

Department of Computer Engineering, M.I.T. College of Engineering, Pune, India
e-mail: jayshree.aj@gmail.com
A. G. Patil
e-mail: abhilampard@gmail.com
E. Phatak
e-mail: eeshan.phatak@gmail.com
S. Gaopande
e-mail: sumantgaopande@gmail.com
Y. Deshpande
e-mail: udealwithit20@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_54
548 J. Ghorpade-Aher et al.
Keywords Nadi Parikshan Wrist pulse Pulse analysis Surya Namaskar

Yoga Machine learning
1 Introduction
Ayurveda, is a traditional effective medical science which was developed in the

Indian subcontinent thousands of years ago where “Ayur” means life and “Veda”
means science. ‘Nadi Parikshan’ (Wrist Pulse Diagnosis), can accurately diagnose
both physical and mental diseases as well as the imbalances in doshas [1]. It is an
important diagnostic method used by Ayurvedic practitioners which consists of a
various examination and physical tests [2]. Surya Namaskar is a Yoga exercise
which has been practiced since ancient times by Rushimunis.
Various evidence-based research studies have suggested that Surya Namaskar
has a positive impact on the human body and its functionalities such as balancing
metabolism function, tuning the central nervous system, supporting urogenital
system, boosting gastrointestinal system, etc. In today’s busy life, practicing Surya
Namaskar regularly revitalizes body, improves energy level, and keeps mind calm
[3]. Figure 1 (src: http://www.harekrsna.de/surya/surya/surya-names.htm.) depicts
Fig. 1 The 12 Asanas of Surya Namaskar

Analysis and Prediction of the Effect of Surya Namaskar … 549
the various steps in performing Surya Namaskar. The steps include 12 different
asanas having different benefits. The 12 asanas are as follows:
1. Pranamasana
2. Hasta Uttanasana
3. Hasta Padasana
4. Aekpadprasarnaasana
5. Adhomukhasvanasana
6. Ashtanga Namaskar
7. Bhujangasana
8. Adhomukhasvanasana
9. Ashwasanchalanasana
10. Uttanasana
11. Hasta Uttanasana
12. Pranamasana
Various research studies have proved that regular practice of Surya Namaskar
exercise improves cardiopulmonary efficiency in healthy adolescents and is a
beneficial exercise for both males and females. The regular practice of Surya
Namaskar for a long time, results in reduced Pulse rate and this is attributed to
factors like the increased vagal tone and decreased sympathetic activity [4].
2 Proposed Framework
The objective of this paper is to analyze changes in Pulse caused by Surya

Namaskar and predict Pulse after Surya Namaskar.
Figure 2 shows the block diagram for the Proposed Framework of System that
includes various steps in building the predictive model. The data recorded is split
into training data and testing data. The predictive model will be built by using
feedforward neural network and backpropogation. Testing data will be used to
improve the accuracy of the predictive model using cross-validation. This predic-
tive model will predict Pulse after Surya Namaskar.
2.1 Data Acquisition
‘Nadi Tarangini’ is a noninvasive Pulse-based diagnosis system which is an

alternative to the traditional method of “Nadi Pariksha’. It has three pressure sensors
to record the Pulse which gives 3 Pulse waveforms at the 3 points on the wrist of the
tridoshas viz. Vatta, Kapha, and Pitta, respectively. The pressure is slowly
increased on the wrist until one of the 3 Pulse waveforms start to take the form of a
typical Pulse signal. The recorded data is obtained in the form of Time Series data
Fig. 2 Block diagram for the proposed framework of system
Fig. 3 Nadi Tarangini to record pulse
and is stored in the form of Comma Separated Values (.csv file) where three
columns represent the amplitude values of respective Dosha signal in Vatta–Pitta–
Kapha sequence. The sampling frequency of the device is around 500 Hz [5, 6].
Figure 3 shows the real-time images of the Surya Namaskar practical sessions that
were conducted for data acquisition.
Table 1 Data acquisition information

Sr. No. Parameter Description
1 Number of subjects 35 (23 Males, 12 Females)
2 Age group 21 ± 2 years
3 Time 3:00 P.M.–5:00 P.M. (Indian Standard Time)
4 Recording of pulse Pulse recorded for duration of 2 min,
1. Before Surya Namaskar
2. After Surya Namaskar
The physical and mental constitution of the body; which remains constant during
the lifetime of a person is identified by Prakruti of that person. In Ayurveda, the
“Panchamahabhutas” manifests the features of Prithvi (Earth), Akash (Space),
Tejas (Fire), Vayu (Wind), and Jal (Water). The amalgamation of these elements is
responsible for the tridoshas—Vatta, Pitta, and Kapha [7]. The Prakruti of a person
is identified by using a questionnaire and based on the answers given by the person.
The questions are related to the daily lifestyle and habits of the subject which
defines the Prakruti of that subject. Table 1 shows the description of the dataset for
the data acquisition process.
Disclaimer: The Pulse was recorded using a noninvasive device. No human
being was harmed for the study. Only healthy people were considered who were
without any apparent medical condition or illness. The procedure for recording
Pulse is similar to one followed by many other authors [5, 8].
2.2 Data Preprocessing
Baseline removal is a wave preprocessing technique which arranges the peaks of a

wave in a uniform manner about an axis that divides each side into equidistant
parts.
As shown in Fig. 4a, all the peaks are seen in a haphazard manner before
baseline removal. After applying baseline removal technique the wave is brought
about the X-axis in a uniform manner as shown in Fig. 4b. This allows comparing
Fig. 4 a Before baseline removal. b After baseline removal

the change in amplitude and frequency for the Pulse waves before and after per-
forming Surya Namaskar. Denoising a wave is any technique applied to remove
noise from a signal and preserve the useful information. The denoising function has
been applied that smoothens the shape of the wave such that even intricate curves
can be studied to conclude minute changes after performing Surya Namaskar.
2.3 Feed Forward Neural Network
The prediction of the Pulse signal after performing Surya Namaskar will be non-
linear. A feedforward neural network (FFNN) is a type of neural network con-
taining many neurons which are arranged in multiple layers viz. input layer, hidden
layer, and output layer. Each neuron in a hidden layer joins another respective
neuron in the previous layer and its next layer. Each such connection has a weight
associated with it which gets adjusted as the training process continues. These are
also called as multilayered perceptron (MLP). As there is no feedback involved,
they are called as FFNN. A MLP is adopted as a prediction model [9, 10–12]. The
features given to the input layer will be:
1. Point in Pulse Before Surya Namaskar
2. P to P time interval between consecutive Pulse as shown in Fig. 5
3. D to D time interval between consecutive Pulse as shown in Fig. 5
These features will be given to each neuron in the input layer for each Dosha
(Vatta, Pitta, or Kapha).
Every input has an associated weight which determines its influence on the
change in a wave. Each neuron in the network is governed by an activation function
given by:
1
; ð1Þ
1 þ ef ðxÞ
Fig. 5 P to P and D to D
interval
where,
X
4
f ðxÞ ¼ wi xi þ Bias
i¼1
W Weight associated with corresponding

X Input Parameter
Bias Measure of the threshold of a Neuron.
The FFNN will be trained using backpropagation algorithm. This is based on
calculating the error between the predicted and actual values. The change in this
error with respect to the weights is calculated. This is called the error derivative.
Error derivative will be used to readjust the weights as training progress [13]. Once
the training is completed the neural network along with the weights forms the
prediction model. A new Pulse signal before Surya Namaskar can be given to this
model which predicts the “After Surya Namaskar” Pulse. Once, the network is
trained, on providing a new input value (Pulse signal before performing Surya
Namaskar), it will predict the corresponding output. As the dataset is increased for
training the network, this degree of error can be minimized. Using MATLAB’s
“ntstool”, nonlinear prediction of time series data is possible [14–16]. It uses a
FFNN and has two parameters:
1. Input Vector: defines the set of values before Surya Namaskar
2. Target Vector: Expected set of values after Surya Namaskar
This neural network is trained and returns the output vector and the mean error in
prediction.
3 Observations and Discussion
Analyzing and observing the Pulse signals recorded under two conditions viz.
before Surya Namaskar and after Surya Namaskar; the amplitude of the Pulse
signals increased. It was also observed that the frequency increased in the latter case
(After Surya Namaskar). Figure 6a, b, c shows one such example of the 3 Pulse
signals (Vatta, Pitta and Kapha Dosha signals) recorded for both the conditions at
Afternoon Prahar.
In Fig. 6a, b, c, the y-axis represents the Pressure amplitude and x-axis represents
the timestamps. In all the three cases after Surya Namaskar, there was an increase in
amplitude and frequency. The increase in frequency is a direct indication of the
increase in Pulse rate of the Pulse. There was a significant change in all the doshas
viz. Vatta, Pitta and Kapha signals as shown in Fig. 6a, b, c, respectively. In all the
signals, the Dichrotic Notch appears to be diminished or less prominent and the
Pulse becomes smoother after Surya Namaskar. This is because while performing
Fig. 6 a Pulse signal (Vatta). b. Pulse signal (Pitta). c. Pulse signal (Kapha)
Surya Namaskar exercise, blood flow to the muscle is tightly coupled to its
metabolic rate. This is due to the increase in the cardiac output and means arterial
blood pressure and regional redistribution of blood flow through adjustments in the
vascular resistance. The arterial compliance increases and blood flows faster and as
a result the Pulse becomes smooth [8].
4 Conclusion
Surya Namaskar has an immediate impact on Pulse and improves the overall health
of an individual. The young age group considered in this study show significant
changes which signify the usefulness of this exercise and benefits of Yoga. Surya
Namaskar has an important role in our day to day life and has a positive impact on
the society as a whole, by balancing the three doshas. The changes in Pulse that
Surya Namaskar brings about can be further analyzed in detail and can have many
applications in healthcare domain. Medical practitioners can use the system and
understand the Surya Namaskar benefits on the body of a patient. Ayurveda, despite
being an effective medical system, not much technological advancement have been
there. This research is an effort toward proving these results using machine learning.
References
1. V. Lad.: Secrets of the pulse: The ancient art of Ayurvedic pulse diagnosis. Motilal
Banarasidas. Delhi 2005.
2. Prasad, G. P., Bharati, K., & Swamy, R. K.: Some important aspects of nadipariksha from
basavarajiyam. Ancient science of life, 24(1), 27–9. 2004.
3. Amit Vaibhav, Swati Shukla and Om Prakash Singh.: Surya Namaskar (Sun Salutation): A
Path to Good Health. International Journal of Pharmacological Research IJPR Volume 6 Issue
07 (2016).
4. Pratima M. Bhutkar, Milind V. Bhutkar, Govind B. Taware, Vinayak Doijad and B.R.
Doddamani.: Effect of Surya Namaskar Practice on Cardio-respiratory Fitness Parameters: A
Pilot Study. Al Ameen J Med Sci (2008) 1(2).
5. Aniruddha Joshi, Anand Kulkarni, Sharat Chandran, V. K. Jayaraman and B. D. Kulkarni.:
Nadi Tarangini: A Pulse Based Diagnostic System. Engineering in Medicine and Biology
Society, 2007.
6. Nadi Tarangini. http://www.naditarangini.com. Accessed: 25th January 2017.
7. Dinesh Kumar Meena, Devanand Upadhyay, Rani Singh, B. K. Dwibedy.: A Critical Review
of Fundamental Principles of Ayurveda. International Ayurvedic Medical Journal (IAMJ):
Volume 3, Issue 7, July 2015.
8. L.S. Xu, K.Q. Wang, L. Wang, Naimin Li.: Pulse Contour Variability Before and After
Exercise. Proceedings of the 19th IEEE Symposium on Computer-Based Medical Systems
(CBMS’06).
9. Jayshree Ghorpade-Aher, Abhishek Patil, Yudhishthir Deshpande, Sumant Gaopande and
Eeshan Phatak.: A Proposed Framework for Prediction of Pulse, based on the Effect of Surya
Namaskar on Different Prakruti at Different Prahars of the Day. 2017 International Conference
On Big Data Analytics and Computational Intelligence (ICBDACI), Chirala, India, 2017.
10. F. Scarselli and A. C. Tsoi.: Universal approximation using feedforward neural networks: a
survey of some existing methods, and some new results. Neural Networks, vol. 11, pp. 15–37,
1998.
11. Ming Yuchi, Jun Jo.: Heart Rate Prediction Based on Physical Activity using Feedforward
Neural Network. International Conference on Convergence and Hybrid Information
Technology 2008.
12. Priddy K.L., Keller P.E.: Introduction in Artificial Neural Networks: An Introduction.
Bellingham Wash. pp. 1–12. St. Bellingham USA, 2005.
13. Gashler, Michael S., and Stephen C. Ashmore.: Training Deep Fourier Neural Networks to Fit
Time-Series Data. Intelligent Computing in Bioinformatics. Springer International Publishing,
2014.
14. Neural Network Time-Series Prediction and Modeling—MATLAB & Simulink-MathWorks,
UK https://in.mathworks.com/help/nnet/gs/neural-network-time-series-prediction-and-modeli
ng.html. Accessed: 25th Jan 2017.
15. Nitin Sharma, Neha Udainiya.: Mechanism of Chakras in Suryanamaskar and its benefits: A
Conceptual Study. International Ayurvedic Medical Journal (IAMJ): Volume 3, Issue 8,
August-2015.
16. L. Xu, D. Zhang, K. Wang, N. Li, and X. Wang.: Baseline wander correction in pulse
waveforms using wavelet-based cascaded adaptive filter. Computers in Biology and
Medicine, vol. 37, pp. 716–731, 2007.
Cognitive Depression Detection
Methodology Using EEG Signal Analysis
Sharwin P. Bobde, Shamla T. Mantri, Dipti D. Patil

and Vijay Wadhai
Abstract This paper illustrates a new method for depression detection using EEG
recordings of a subject. It is meant to be used as a computerised aid by psychiatrists
to provide objective and accurate diagnosis of a patient. First, data from the
occipital and parietal regions of the brain is extracted and different channels are
fused to form one wave. Then DFT, using FFT, is applied on the occipito-parietal
wave to perform spectral analysis and the fundamental is selected from the spec-
trum. The fundamental is the wave with the maximum amplitude in the spectrum.
Then classification of the subject is made based on the frequency of the funda-
mental using rule-based classifier. Detailed analysis of the output has been carried
out. It has been noted that lower frequency of the fundamental tends to show
hypoactivation of the lobes. Moreover, low-frequency characteristics have also
been observed in depressed subjects. In this research, 37.5% of the subjects showed
Major Depressive Disorder (MDD) and in all 80% of the subjects showed some
form of depression.

Keywords Depression Electroencephalography (EEG) Discrete Fourier

Transform (DFT) Fast Fourier Transform (FFT) Alpha band

Delta band Rule-based classifier
S. P. Bobde (&) S. T. Mantri

MIT College of Engineering, Pune, India
e-mail: bsharwin@gmail.com
D. D. Patil
MKSSS’s Cummins College of Engineering for Women, Pune, India
e-mail: diptivt@gmail.com
V. Wadhai
KJEI’s Trinity Academy of Engineering, Pune, India

https://doi.org/10.1007/978-981-10-7245-1_55
558 S. P. Bobde et al.
1 Introduction
According to the WHO report issued in 2016, depression affects more than 350
million people of all ages worldwide, and this makes it the most prevalent global
disease. It is also recorded that more women are affected by depression as compared
to men. More than 8,00,000 people commit suicide due to depression each year [1].
The global burden of depression is increasing and there is an immediate need for a
coordinated response from all health and social sectors at the country level [2].
Depression is characterized by persistent low-mood state and loss of interest and
passion. In Major Depressive Disorder (MDD), difficulty with daily functioning in
personal, family, social, educational, occupational areas is observed. Many people
also show symptoms of anxiety. Depression is managed through psychological
intervention and drug prescriptions [3].
There is a need for an objective diagnostic tool, which is technologically
advanced and accurate. In this chapter, one such system has been proposed. The
input to the system is the patient’s Electroencephalography (EEG) recording. The
system extracts readings from the occipital and parietal lobes of the brain, which are
of main interest in this research. The readings are then fused to create one
occipito-parietal wave. Then, the spectral analysis of the wave is carried out using
Discrete Fourier Transform (DFT). Here, FFT has been used to perform DFT on the
wave. The classification is dependent o the fundamental in the spectrum. After the
fundamental has been selected, its frequency is recorded and the subject is classified
accordingly.
2 Previous Work
Previous studies have studied different parts of the brain separately along with
different frequency bands. There are four main frequency bands in EEG on which
there has been extensive research done; they are delta (0.5–4 Hz), theta (4–8 Hz),
alpha (8–13 Hz) and beta (13–30 Hz). It has been observed that the alpha band is
most informative about a subject’s state of depression [4]. Alpha activity mainly
corresponds to the occipital and parietal lobes [5]. Alpha band has also been used to
distinguish between depressed and normal subjects in the past. Increase in slow
alpha activity in depressed subjects in occipital and parietal regions has also been
noted in a 2010 study [6]. Relatively decreased alpha activity in patients with
(MDD) has also been noted. In fact, decreased parieto-occipital alpha power is also
stated as a biomarker of depression and excess alpha is also considered as a
favourable response to antidepressive treatments [7]. A 2005 study found abnormal
low-frequency activity in the occipital lobe of untreated patients with depression.
They also found increased occipital lobe delta dipole density as a risk factor for
depression [8]. Another paper stated that alpha and to a lesser extent delta oscil-
lations can distinguish depressed and healthy individuals [9, 10]. Taking these facts
Cognitive Depression Detection Methodology Using EEG Signal … 559
into consideration our method focuses on the occipital and parietal regions of the
brain. These facts also realise a relation between depression and the
occipito-parietal wave. To detect and classify depression, it is checked if the power
of the alpha band is low and if it shows low-frequency characteristics. The char-
acteristic frequency is the frequency of the fundamental found in the spectral
analysis of the wave. To extract these properties, DFT is performed on the
occipito-parietal wave.
3 Data Collection
EEG readings are non-stationary and non-Gaussian signals that arise from the brain.
They have high temporal resolution and low spacial resolution [11, 12]. In this
research, EEG data from 80 different subjects has been used. Each dataset contains
EEG data of 1 min. It has been recorded on a 64-channel EEG headset, which gives
high precision in differentiating between different parts of the brain. The dataset has
been obtained from PhysioNet and PhysioBank, which are research resources for
complex psychologic signals [13]. The particular dataset was recorded for motor/
imagery tasks [14].
Disclaimer
The EEG recordings used for research were collected from a reputable source. The
data has been used in past research experiments. The resulting product is usually
used for forensic experiments. No human was harmed during collection or utili-
sation of the data.
4 Methodology
In the first stage, the occipital and parietal regions of the brain are selected from the
EEG data. Then the data from the channels is fused to produce the occipito-parietal
wave. For feature extraction of DFT, using FFT, is performed on the
occipito-parietal wave. Then an apt range is defined from which the fundamental of
the wave is selected. This fundamental is the required feature for classification.
4.1 Occipito-Parietal Wave
As the occipital and parietal lobes of the brain are of main interest, data from these
parts is fused to form one wave. In a 64-channel EEG, there are 17 channels which
point to these two parts of the brain. These channels are P7, P5, P3, P1, Pz, P2, P4,
P6, P8, PO7, PO3, POz, PO4, PO8, O1, Oz and O2. The fusing of data is done by
taking the mean of the channels’ readings at every time instant. This results in a
new wave with the same number of readings. Let this new occipito-parietal wave be
W. The index position of each record in the dataset is i. Therefore, the resulting
voltage value at a recorded instant i will be Wi . Let SF be the sampling frequency of
the recording headset. Therefore, the index positions are given by Eq. 1 as follows:
fi 2 Z þ j1 i ð60 SFÞg ð1Þ
Let the set of required channel numbers be Ch as in Eq. 2, where P7; P5 O2

represent the selected channel numbers.
Ch ¼ fP7; P5; P3; Oz; O2g ð2Þ

P
EEGi;Ch
Wi ¼ Ch ð3Þ
jChj
The voltage reading of W at index i will be given by Eq. 3. Here, EEG is the data
frame containing readings of all 64 channels at all i time instances. Figure 1a shows
the EEG readings of selected 17 EEG channels of subject 30. All 17 graphs show
similar readings, but actually they all differ in voltage readings. Figure 1b shows
the resulting occipito-parietal wave of subject 30.
Fast Fourier Transform is a computational algorithm which facilitates DFT of an

input signal [15]. It helps in performing fast spectral analysis. It is a more efficient
way of performing DFT of a signal. FFT transforms the input signal into a sequence
of complex numbers which represent pure sinusoidal waves. The real part of the
number represents the amplitude and the imaginary part represents the phase of the
sinusoidal wave. For this system the amplitude of the waves is important.
Therefore, the absolute value of each complex number is taken. The feature for
Fig. 1 a Plot of selected 17 channels b Resulting occipito-parietal wave

classification here is the fundamental obtained after spectral analysis of the

occipito-parietal wave W. The FFT algorithm outputs a list of complex numbers.
The list will be of length jij, which was used before. First, the amplitude values
obtained after FFT are saved. Let this vector of amplitudes be F. Therefore, let each
element of the vector, using DFT equation, be given by Eq. 4 as follows:

XN 1
kn
Fk ¼ Wn EN ; 0kN 1 ð4Þ
n¼0
where EN ¼ e j2p=N; ki ¼ i 1; N ¼ jij; N ¼ 60 SF.

The FFT algorithm gives the output in a logarithmic scaling. Frequency scaling
is done to get the correct output. Let the new vector of correct amplitudes be F 0 . The
relation between F 0 and F is given by Eq. 5. The resulting set F 0 contains correct
values of amplitude in lV for each sinusoidal wave with a frequency equal to the
index position at which it occurs. Two such plots of subjects’ number 1 and 20 have
been shown in Fig. 2 as examples.

1 Fi
Fi0 ¼ ð5Þ
2 SF
To select the fundamental, a frequency range needs to be defined from which the
fundamental’s frequency can be selected. This range is very important as it forms
the basis of the diagnosis. It should be broad enough to cover maximum real-world
instances. It should cover only the important. Thus range from 1 Hz to the Nyquist
frequency will be apt. The Nyquist frequency in this study’s samples is 80 Hz. This
range contains all 4 major frequency bands. The Nyquist frequency is given by
Eq. 6.
SF
Nyq ¼ ð6Þ
2
Fig. 2 FFT spectrum of a Subject 1 b Subject 20

After the range is defined, the fundamental can be selected. The fundamental is
that sinusoidal wave which has the maximum amplitude in the defined range.
Because of the narrowing of range, the algorithm does not consider high and
comparatively lower frequency noise which is present because of the hardware and
can affect the output.
5 Analysis
5.1 Square Mean Voltage
To compare the fundamental’s frequency with the activation in the occipital and
parietal regions, it needs to be related to power. As power is proportional to
intensity, which is the square of amplitude (here voltage), it can be related to the
occipito-parietal wave. The previous range of frequency is used to extract a part of
the spectrum. Then, the mean amplitude in that range is found and its square is
computed. The equation for mean voltage is given by Eq. 7.
P 0
F
MV ¼ n n ; 1 n Nyq; n 2 i; jnj\jij ð7Þ
jnj
SMV ¼ MV2 ð8Þ
Square mean voltage is derived from Eq. 8. Then, the frequency of the funda-
mental was graphically compared to Square Mean Voltage (SMV). As power is
proportional to SMV the comparison gives an understanding of the relation between
the fundamental’s frequency and activation of the occipital and parietal lobes of the
brain. It is shown in Fig. 3. A regression line was also plot along with it. The line
yielded a positive slope.
This means that as the frequency of the fundamental decreases the activation of
these lobes tends to decrease. This hypoactivation of the parietal and occipital lobes
of the brain is an indication of depression [7].
5.2 Low-Frequency Activity
After analysing all 80 graphs of FFT spectral analysis manually, certain reoccurring
characteristics were realised. The subjects with low SMV had fundamental’s fre-
quency ranging from 1 to 4 Hz (delta band). The power of this band was also
greater to those in other bands. Examples of subjects 5 and 26 are shown below in
Fig. 4. This observation lines up with the study conducted in 2005. The power of
the occipito-parietal wave is less, as well as the delta band power is high [8]. This
further validates the classification of depression in the subjects.
Fig. 3 Plot of square mean voltage versus fundamental’s frequency
Fig. 4 FFT spectrum of subjects to be classified as depressed
5.3 Classification
The classification bins were defined after performing analysis and inspecting the
occipito-parietal waves and their spectrums together. The occipito-parietal waves
have been analysed for their intensity, pattern, spectrum patterns, and fundamental’s
properties. Similar wave patterns have also been seen in other studies which took
pre-classified subjects for analysis [16]. The classification bins have been decided as
severe, moderate, mild and normal. The classification is done using a rule-based
classifier, which selects bins based on the fundamental’s frequency. The funda-
mental’s frequency range from 1 to 4 Hz shows server depression of patients. Patients
lying in this category showed all signs of major depressive disorder (MDD) con-
sidered in this analysis.
Fig. 5 Plot of fundamental’s frequency versus subject number, also showing classification labels
and percentage of subjects present
6 Results
EEG data of 80 subjects has been processed and out of these 76 subjects were
classified. In the experimental analysis, four subjects were classified as outliers
based on their fundamental’s frequency. Figure 5 shows the results graphically for
clearer understanding. In this research, 37.5% of the considered subjects showed
signs of MDD. In particular, these subjects showed clear distinction from other
subjects in all statistical parameters. 80% of total subjects showed some level of
depression. The remaining 20% of the subjects exhibited no statistical parameters in
EEG which indicated depression.
7 Conclusion
Analysis of 80 subjects has been performed and 78 have been classified according
to various levels of depression. 61.25% of the subjects indicated moderate or severe
depression, which means they need professional help. Out of these 37.5% indicated
severe depression. 2.5% of the total subjects were outliers and could not be clas-
sified, mainly due to noise in the EEG signal. The main emphasis was given on the
occipital and parietal lobes of the brain, which have been studied to be good sources
of indicators for depression. The data of these lobes is fused to achieve good data.
Analysis has also been done on low-frequency characteristics of brainwave which is
also an indicator of depression. Classification bins were defined based on analysis
of the occipito-parietal wave and its frequency spectrum. The final classification has
been made using a rule-based classifier.
The system can diagnose subjects before their condition can become severe. It is
based on an objective understanding of the human brain. This can benefit victims of
depression and the society in many ways. Early diagnosis means the patient can live
a happy life. It is free from the subjective bias of psychologists and psychiatrists.
A better, objective help can improve the communal standard of living and decrease
the suicide statistics.
References
1. World Health Organization, revised on November 2016, available at http://www.who.int/

topics/depression/en/.
2. WHO: World Health Assembly (WHA) Resolution 65th Report by Secretariat, 2012, pp 1–4.
http://apps.who.int/gb/ebwha/pdf_files/WHA65/A65_R4-en.pdf?ua=1 (accessed November
28, 2014).
3. World Health Organization, mhGAP Intervention Guide - Version 2.0 for mental,
neurological and substance use disorders in non-specialized health settings, 2016.
4. Behshad Hosseinfard, Mohammad Hassan Moradi, Reza Rostami, “Classifying depression
patients and normal subjects using machine learning techniques and nonlinear features from
EEG signal,” Elsevier, Computer methods and programs in Biomedicine, 2012.
5. David M. Groppe, Stephan Bickel, Corey J. Keller, Sanjay K. Jain, Sean T. Hwang, Cynthia
Harden, and Ashesh D. Mehta, “Dominant frequencies of resting human brain activity as
measured by the electrocorticogram,” Neuroimage; 79: 223–233, 1st Oct 2013.
6. Vera A. Grin-Yatsenko, Inke Bass, Valery A. Ponomarev, Juri D. Kropotov, “Independent
Component approach to the analysis of EEG recordings at early stages of depressive
disorders,” Elsevier, Clinical Neurophysiology 121, 281–289, 2010.
7. Sebastian Olbrich & Martijn Arns, “EEG biomarkers in major depressive disorder:
Discriminative power and prediction of treatment response,” International Review of
Psychiatry; 25(5): 604–618, October 2013.
8. Alberto Fernández, Alfonso Rodriguez-Palancas, MarÍa López-Ibor, Pilar Zuluaga, AgustÍn
Turrero, Fernando Maestú, Carlos Amo, Juan José López-Ibor, Jr, and Tomás Ortiz,
“Increased occipital delta dipole density in major depressive disorder determined by
magnetoencephalography,” J Psychiatry Neurosci.; 30(1): 17–23, Jan 2005.
9. Mahdi Mohammadi, Fadwa Al-Azab, Bijan Raahemi, Gregory Richards, Natalia Jaworska,
Dylan Smith, Sara de la Salle, Pierre Blier and Verner Knott, “Data mining EEG signals in
depression for their diagnostic value,” BMC Medical Informatics and Decision Making,
15:108, 2015.
10. SUBHA D. PUTHANKATTIL and PAUL K. JOSEPH, “Classification Of Eeg Signals In
Normal And Depression Conditions By Ann Using Rwe And Signal Entropy,” Journal of
Mechanics in Medicine and Biology Vol. 12, No. 4 1240019 (13 pages), 2012.
11. D. Puthankattil Subha & Paul K. Joseph & Rajendra Acharya U & Choo Min Lim, “EEG
Signal Analysis: A Survey,” J Med Syst (2010) 34:195–212, 2010.
12. M. Rajya Lakshmi, Dr. T. V. Prasad, Dr. V. Chandra Prakash, “Survey on EEG Signal
Processing Methods,” International Journal of Advanced Research in Computer Science and
Software Engineering, Volume 4, Issue 1, January 2014.
13. Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh, Mark RG, Mietus JE,
Moody GB, Peng CK, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet: Components
of a New Research Resource for Complex Physiologic Signals. Circulation 101(23):e215–
e220 [Circulation Electronic Pages; http://circ.ahajournals.org/content/101/23/e215.full];
2000 (June 13).
14. Schalk G., McFarland D.J., Hinterberger T., Birbaumer N., Wolpaw J.R. “A General-Purpose
Brain-Computer Interface (BCI) System,” IEEE Transactions on Biomedical Engineering 51
(6):1034–1043, 2004.
15. W.T. Cochran, J.W. Cooley, D.L. Favin, H.D. Helms, R.A. Kaenel, W.W. Lang, G.C.
Maling, D.E. Nelson, C.M. Rader, P.D. Welch, “What is the fast Fourier transform?,”
Proceedings of the IEEE, Volume: 55, Issue: 10, Oct. 1967.
16. Rajendra Acharya, Vidya K. Sudarshan, Hojjat Adeli, Jayasree Santhosh, Joel E.W. Koh,
Amir Adeli, “Computer-Aided Diagnosis of Depression Using EEG Signals,” European
Neurology 2015;73:329–336, 2015.
Biogas Monitoring System Using DS18B20
Temperature Sensor and MQTT Protocol
Suruchi Dedgaonkar, Aakankssha Kaalay, Nitesh Biyani

and Madhuri Mohite
Abstract Nonrenewable energy resources such as coal, petroleum, and natural gas
are becoming extinct and thus there is a need for renewable energy resources in the
long run. One of the most important renewable energy resources is biogas. Biogas
is the gas produced after anaerobic digestion of organic matter by micro-organisms.
Biogas mainly contains methane (about 60%). There are a number of factors
affecting the production of biogas and one of them is temperature. The temperature
of the biogas plant should be held constant with variation <1 °C and within the
range 30–55 °C. Thus, there is a need of a proper monitoring system for the biogas
plant. In this study, we try to develop a monitoring system for the biogas plant using
temperature sensor DS18B20, MQTT protocol, Mosquitto broker, and Raspberry
Pi. A web-based system, where the temperature sensor values will be uploaded
periodically and an end user will monitor the temperature of biogas plant remotely,
is proposed.
Keywords Biogas monitoring system DS18B20 Temperature sensor

MQTT protocol Mosquitto Raspberry Pi
S. Dedgaonkar (&) A. Kaalay N. Biyani M. Mohite

Information Technology Department, VIIT, Savitribai Phule Pune University, Kondhwa
(BK), Pune, Maharashtra, India
e-mail: suruchigd@gmail.com
A. Kaalay
e-mail: aakanksshak21@gmail.com
N. Biyani
e-mail: niteshbi294@gmail.com
M. Mohite
e-mail: madhurimohite123@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_56
568 S. Dedgaonkar et al.
1 Introduction
The biogas is in use since a long time in India, but the technology improvement is
necessary in order to improve energy outputs. Biogas is one of the most trusted and
reliable energy resources on earth. There is a need for an eco-friendly substitute for
energy resources that are depleting continuously and will soon be exhausted
completely. Biogas serves this. Biogas is the gas produced by the complex
breakdown of organic matter in the absence of oxygen, it does not have any
geographical limitations and also biogas is very simple to use and apply. Kitchen
waste consists of major part of the feed used for the biogas plant as kitchen waste is
very high in calorific value and also it is highly accessible and cheap, kitchen waste
also helps in increasing the methane content of the gas produced. There is a need for
a monitoring system of biogas process, since a monitoring system can help to
increase the efficiency, enhance performance, and also get better results at minimal
cost. Biogas is the most important and efficient resource of energy at a cheaper price
and thus it needs to be monitored for maximum output. There is no proper moni-
toring system for the biogas plants as such that will govern the working and output
of the plant. There is no efficient system to measure the parameters of the gas
produced in the plant and take necessary actions according to them Fig. 1 shows
homemade biogas plant used for simulation purpose [1].
Fig. 1 Homemade biogas

plant used for monitoring
Biogas Monitoring System Using DS18B20 Temperature … 569
2 Message Queuing Telemetry Transport
MQTT is a protocol, which sits on the top of the TCP/IP protocol that is designed for
machine-to-machine communication and for the internet of things connectivity.
MQTT is a lightweight protocol, unlike HTTP which is very bulky in comparison.
MQTT is a protocol which is designed for publish/subscribe protocol which consists
of a broker as a medium. MQTT runs on low bandwidth and is still extremely
powerful. It can be used for high-cost connections because it is readily available [2–4,
8–10]. In this study, temperature sensor sends values continuously to MQTT. The
broker examines the values as to whom it has to be sent and sends accordingly. On the
other hand, the subscribers which are previously subscribed to that topic keep
receiving the messages from that particular topic until they cancel the subscription
[15, 16].
A. MQTT client:
All the components of the IOT environment can act as a client, which sends or
receives data in this protocol. The components can be a microcontroller, sensor,
or the end user devices. In our study, we will be using a temperature sensor,
which will send the values to the broker, which will be uploaded to a web page
and also will be stored in a database and will be sent to the user application.
A client can be a publisher or a subscriber, it depends upon the role assigned to
the device. In both the cases, once the connection is established the device must
announce itself as a subscriber or a publisher.
A publisher can publish the temperature values. A subscriber needs to subscribe
to the same topic to which the publisher is sending the values to receive the
temperature values. For a device to behave as an MQTT client, an MQTT
library must be installed on the client and the broker and the client should be on
the same network.
B. MQTT Publisher:
MQTT publisher is a client, which publishes the data to a particular topic to
which the subscriber is subscribed and sends the values to that topic. The data is
received by the subscriber through this topic.
C. MQTT Subscriber:
MQTT subscriber is a client, which is subscribed to the topic and receives
values from that topic, which are published by the publisher. The subscriber can
access data on that topic only if he is subscribed to that topic.
D. MQTT Broker:
An MQTT broker is a bridge between the publisher and the subscriber. The job
of the broker is to establish a communication between a publisher and a sub-
scriber. Broker sends values from a publisher to a subscriber. A broker can
handle up to hundreds of clients simultaneously. When the data is incoming,
the broker must identify all the clients that are subscribed to the corresponding
topic and make sure that each and every one of them receives the data. In this
study, the broker receives the values from the temperature sensor connected to
Raspberry Pi and sends the values to the subscribed user. Broker also performs
the task of authentication and authorization of the clients for security. The
clients send the username and password along with the connect message which
is verified by the broker. Broker also makes sure that the messages are received
by the subscribed users only and no other device can access the data.
There are a number of brokers which can be used for implementation of MQTT
protocol.
Mosquitto—It is an open-source MQTT broker. In this study, we will be using
Mosquitto broker [11].
Installation and testing of Mosquitto broker on Raspberry Pi:
The commands for installation of mosquito are given below. The following
commands are to be entered into the terminal of Raspberry Pi
- > sudo apt-get install mosquitto

- > sudo apt-get install mosquito –clients
- > sudo apt-get install mosquito mosquito-clients python-mosquitto
Next, we need to edit the configuration file of Mosquitto which is mosquito.conf.

We need to stop the Mosquitto service and then edit the file
- > sudo /etc./init.d/mosquito stop

- > sudo nano/etc./mosquito/mosquito.conf
Enter the following lines into the file
Pid_file /var/run/mosquito.pid
persistence true
persistence_location
/var/lib/mosquito
log_dest topic
log_type error
log_type warning
log_type notice
log_type information
connection_messages true
log_timestamp true
include_dir /etc./mosquito/conf.d
Fig. 2 Publisher window-publishing the text messages
Save the file by pressing CTRL+X and then pressing Y and ENTER
Now, start the Mosquitto service again
- > sudo /etc./init.d/mosquito start
Now, open two terminals, one for the publisher and the second one for the
subscriber.
Enter the following command into subscriber window to subscribe to the topic to
receive the messages published by the publisher
- > mosquito_sub –d –t hello/world
Enter the following command into publisher window (Fig. 2,3)
- > mosquito_pub –d –t hello/world –m “this is first test”.

Fig. 3 Subscriber window-receiving the text messages published by the publisher
3 Sensor
DS18B20 (Temperature sensor):

DS18B20 is a 1-wire digital temperature sensor. The DS18B20 digital ther-
mometer provides 9-bit to 12-bit Celsius temperature measurements and has an
alarm function with nonvolatile user-programmable upper and lower trigger points.
The DS18B20 communicates over a 1-wire bus that by definition requires only one
data line (and ground) for communication with a central microprocessor. In addi-
tion, the DS18B20 can derive power directly from the data line (“parasite power”),
eliminating the need for an external power supply. Each DS18B20 has a unique
64-bit serial code, which allows multiple DS18B20s to function on the same 1-wire
bus. Thus, it is simple to use one microprocessor to control many DS18B20s
distributed over a large area. Applications that can benefit from this feature include
HVAC environmental controls, temperature-monitoring systems inside buildings,
equipment, or machinery, and process monitoring and control systems (Fig. 4).
The connections of DS18B20 with Raspberry Pi are as follows:
Connect GND of DS18B20 to pin 6 of Raspberry Pi
Connect DATA of DS18B20 to pin 7 of Raspberry Pi
DS18B20 to pin 1 of Raspberry Pi (3.3 v supply)
Fig. 4 DS18B20
(Temperature sensor)
4 Raspberry Pi
Raspberry Pi is a small credit card-sized computer developed for basic program-

ming and learning purposes [5–7, 17]. It has 1 GB of RAM with 1.2 GHZ 64-bit
quad-core ARM Cortex-A53 processor. It has 4 USB ports and an HDMI port, also
it has 40 GPIO pins and we can install an operating system on it. In this study, we
will be installing our MQTT broker on Raspberry Pi, and also the temperature
sensor will be connected to Raspberry Pi for taking the values of the biogas plant.
To access the Raspberry Pi terminal, we will be using SSH configuration through
putty by using the IP address of the Raspberry Pi.
5 Web Interface
In this study, we will be uploading all the values coming from the temperature
sensor of the biogas plant through MQTT protocol to a web page. The values will
be uploaded continuously on the web page in real time for monitoring. The web
page is designed using the functionality of HTML along with the JavaScript and
PHP to design a dynamic page [12]. The values received from the sensor are stored
in SQL database from where the data is sent to the web page. As soon as the values
are received by the sensor, the values are sent to the web page which will be sent
the values to the users with the topic concerning the temperature values if the
temperature values are below or above the range and also if there is a continuous
change in the temperature of the biogas plant. The user can monitor the temperature
values of the biogas plant from the web page in real time. The web page consists of
graphs and charts of the incoming temperature values for a graphical view of the
current temperature of the biogas plant.
6 Android Application
This study consists of an end user android application, which is designed to monitor
and analyze the values according to time and date. The values received from the
temperature sensor are stored in a database in SQL and all the values are sorted by
date, time, and range. The application consists of an interface where the user can
select to view the data by time and date. The user can view the temperature of the
biogas plant by selecting the date and time. The user has an option of checking what
the value of the plant was on a particular date and at a particular time. This will help
in understanding the production and output of the biogas plant better. The user can
analyze the data over time and can take helpful decisions according to the data.
7 Design and Implementation
In this proposed study, we have created a biogas monitoring system with temper-
ature sensor DS18B20 and MQTT protocol for monitoring the temperature of the
biogas plant. The temperature sensor will sense the temperature values of the biogas
plant and will send these values to the MQTT protocol. The job of MQTT protocol
is to send the alerts to the registered users if the temperature goes below or above
the ideal temperature for the biogas plant. Simultaneously all the values are
uploaded to a web page for a web-based monitoring. The protocol is installed on
Raspberry Pi which acts as a client in this architecture for the protocol and all the
installations and working takes place on Raspberry Pi (Fig. 5) [13, 14].
Fig. 5 Overall flow of data

The connection of Raspberry Pi and DS18B20 temperature sensor is shown in

(Fig. 6).
The values can be received and observed on the terminal of the Raspberry
(Fig. 7).
The system also displays the temperature of the biogas plant on 2X16-bit LCD
screen in degrees as well as in Fahrenheit (Fig. 8).
The data can be sent to MQTT broker once the connection is established
between the Raspberry Pi and the temperature sensor. For sending the values from
sensor to the broker, we use Python language.
The following code helps in sending the data to the MQTT broker
Fig. 6 Connection of Raspberry Pi and DS18B20
Fig. 7 Raspberry Pi terminal

showing sensor values
Fig. 8 Temperature values on LCD screen in degrees and Fahrenheit
import paho.mqtt.client as paho

import time client = paho.Client()
client.on_publish = on_publish
client.connect(“192.168.225.105”, 1883)
client.loop_start()
while True:
temperature = read_temperature()
(rc, mid) = client.publish
(“test/temp”, str(temperature), qos = 1)
time.sleep(30)
This study proposes a proper monitoring system for analyzing real-time temperature
values of the biogas plant which helps in maintaining the temperature of the plant
according to our needs and getting desired and ideal output from the plant. The
system consists of a web interface, which helps users in monitoring the system
remotely over the Internet and taking decisions accordingly. The system consists of
an end user android application, which allows user to monitor the temperature, and
also offers users with options for monitoring the temperature of the biogas plant
according to time and date which helps user in building a study or a pattern of the
plant’s temperature over the time.
The system also facilitates with future advancements in the proposed system to
control the temperature of the biogas plant externally using a number of different
ways.
The ways by which the temperature can be controlled are as follows:

(1) Energy from cooling water and also exhaust gas can be used to control the
temperature of the biogas plant according to our needs.
(2) Insulations, heating elements, heat exchanges, steam injections and water baths,
etc., can be used to control the temperature of the biogas plant according to our
requirements.
References
1. Mohammad Shariful Islam, Asif Islam, Md. Zakirul Islam “Variation of Biogas Production
with Different Factors in Poultry Farms of Bangladesh”.
2. Aleksandar Antonić, Martina Marjanović, Pavle Skočir and Ivana Podnar Žarko “Comparison
of the CUPUS Middleware and MQTT Protocol for Smart City Services”.
3. Mohsen Hallaj Asghar, Nasibeh Mohammadzadeh “Design and Simulation of Energy
Efficiency in Node Based on MQTT Protocol in Internet of Things”.
4. Krešimir Grgić, Ivan Špeh, Ivan Heđi “A Web-Based IOT Solution for Monitoring Data
Using MQTT Protocol”.
5. F. Xia, L.T. Yang, L. Wang, and A. Vinel, “Internet of things,” International Journal of
Communication Systems, vol. 25, pp. 11011102, 2012.
6. S. Charmonman, P. Mongkhonvanit, V. Dieu, and N. Linden “Applications of internet of
things in e-learning,” International Journal of the Computer, the Internet and Management,
vol. 23, no. 3, pp. 1–4, September December 2015.
7. D. Evans, “The internet of things – how the next evolution of the internet is changing
everything,” white paper, April 2011.
8. L. Dürkop, B. Czybik, and J. Jasperneite, “Performance evaluation of M2 M protocols over
cellular networks in a lab environment,” Intelligence in Next Generation Networks (ICIN),
February 2015, pp. 70–75.
9. HiveMQ Enterprise MQTT Broker. (2016, Feb. 8). [Online]. Available: http://www.hivemq.
com/.
10. Scalagent, “JoramMQ, a distributed MQTT broker for the internet of things,” white paper,
September 2014.
11. Eclipse Mosquitto. (2016, Feb. 12). [Online]. Available: http://mosquitto.org/.
12. Eclipse Paho JavaScript Client. (2016, Feb. 12). [Online]. Available: https://eclipse.org/paho/
clients/js/.
13. Maxim Integrated. (2016, Feb. 14). [Online]. Available: www.maximintegrated.com.
14. Smart home implementation based on internet and Wi-fi technology, “YAN Wenbo, WANG
Quanyu, GAO Zhenwei Control Conference (CCC), 2015 34th Chinese.
15. MQTT for Sensor Networks (MQTT-SN) Protocol Specification, Andy Stanford-Clark and
Hong Linh Truong, Andy Stanford-Clark and Hong Linh Truong (andysc@uk.ibm.com,
hlt@zurich.ibm.com) November 14, 2013.
16. MQTT-S A publish/subscribe protocol for Wireless Sensor Networks, Hunkeler, U.; IBM
Zurich Res. Lab., Zurich; Hong Linh Truong; Stanford-Clark, A, Communication Systems
Software and Middleware and Workshops, 2008. COMSWARE 2008. 3rd International
Conference.
17. Raspberry Pi as a Wireless Sensor Node: Performances and Constraints, Vladimir Vujovic
and Mirjana Maksimovic, Faculty of Electrical Engineering, East Sarajevo, Bosnia and
Herzegovina, IEEE 2014.
Time-Efficient and Attack-Resistant
Authentication Schemes in VANET
Sachin Godse and Parikshit Mahalle
Abstract VANET (Vehicular Ad hoc Network) brings evolution in the trans-

portation system. It is vulnerable to different kind of attacks. Authentication is the
first line of defense against security problems. The previous researcher comes with
some cryptographic, trust-based, ID-based, and signature-based authentication
schemes. If we consider the performance of previous schemes, process time and
efficiency need to be improved. Faster authentication can help to establish com-
munication in a short time and RSU can serve more number of vehicles in a shorter
time span. We presented AECC (Adaptive Elliptic Curve Cryptography) and EECC
(Enhanced Elliptic Curve Cryptography) schemes to improve the speed and security
of authentication. In AECC key size is adaptive, i.e., different sizes of keys are
generated during the key generation phase. Three ranges are specified for key sizes
small, large, and medium. In EECC, we added an extra parameter during trans-
mission of information from the vehicle to RSU (Road Side Unit) for key gener-
ation. This additional parameter gives the information about vehicle ID and location
of the vehicle to RSU and other vehicles.
Keywords VANET AECC EECC Authentication ITS OBU
1 Introduction
Intelligent transportation system (ITS) is an important application of the vehicular

ad hoc network. ITS helps to automate traffic monitoring and control system.
In VANET, nodes are vehicles with each vehicle having OBU (On Board Unit)
deployed on it. RSUs are deployed at fixed distances on the roadside. Figure 1
S. Godse (&)
Department of Computer Engineering, SAOE, SPPU, Pune, India
e-mail: sachin.gds@gmail.com
P. Mahalle
Department of Computer Engineering, SKNCOE, SPPU, Pune, India
e-mail: parikshitmahalle@yahoo.com

https://doi.org/10.1007/978-981-10-7245-1_57
580 S. Godse and P. Mahalle
shows the different components of VANET and communications in it. In VANET,

nodes can communicate in three different ways: V2V (Vehicle to Vehicle), I2V
(Infrastructure to Vehicle), and V2I (Vehicle to Infrastructure) communication. Due
to open medium of VANET, security is a major issue with it. Malicious vehicle can
enter into the network and hamper the performance of the network. Authentication
is the need of vehicular network to avoid unauthorized vehicles and unauthorized
messages. Previous research gives some solutions like ID-based, trust-based,
cryptography-based, and signature-based authentication mechanisms, but overhead
is a major issue in these solutions [1, 2]. The time required for authentication
process is another important issue in VANET as nodes are moving fast compared to
other ad hoc networks. The time span between nearest RSU and vehicles is very
low. The presented schemes, improving the speed of authentication and improve the
efficiency of RSU to serve more number of vehicles. These schemes are designed
based on ECC. In Sect. 2, a survey of existing authentication schemes is discussed.
Mathematical analysis of the time required and vehicle served by RSU with
respective key sizes is explained in Sect. 3. Section 4 presents our authentication
schemes AECC and EECC.
Fig. 1 Different Components in VANET and their communication

Time-Efficient and Attack-Resistant Authentication Schemes … 581
2 Existing Authentication Schemes in VANET
1. PKI (Public-Key Infrastructure): It is used to verify the user. PKI is based on

asymmetric key cryptography. It uses two keys: public and private key. CA
issues certificates to the vehicles to decide the validity of the users. A vehicle
manufacturer or government authority can act as a CA. Certificate revocation
takes place at predefined time intervals, through this certificate revocation a list
is generated.
2. TESLA (Timed-Efficient Stream Loss-Tolerant Authentication): TESLA
provides authentication for multicast and broadcast network communications. It
uses symmetric key cryptography. The main flaw of TESLA is that it susceptible
to attacks that happen due to memory-based denial of service.
3. TESLA ++: It is an advanced version of TESLA, which overcomes the
drawbacks of TESLA.
4. ECDSA (Elliptic Curve Digital Signature Algorithm): It includes three
phases: key generation, signature generation, and signature verification.
In ECDSA, a scalar multiplication of a given random point is used in signature
generation and verification.
Table 1 summarizes existing authentication schemes and their merits and
demerits.
3 Experimental Analysis of Authentication in VANET
There are two types of authentication in VANET: vehicle authentication and

message authentication. The vehicle authentication verifies the identity of the
vehicle. Vehicle authentication is further classified into 5 types: Node-level
authentication, Group-level authentication, Unicast authentication, Multicast
authentication, and Broadcast authentication [3]. Here, we studied and analyzed
RSA- and ECC-based authentication scheme on the simulator.
3.1 Mathematical Analysis of Time Required

for Authentication and Vehicles Serves by the RSU
We have given an analysis of authentication time required for RSA (Rivest–

Shamir–Adleman) and ECC (Elliptic Curve Cryptography) algorithm. Consider the
communication range of RSU is 10 km, i.e., 10,000 m. RSU requires 100 ms for
authenticating one vehicle. The radius of RSU is 10 km. Figure 2 shows RSU
communication range and vehicle “A” traveling from point S to D.
Table 1 Existing authentication scheme survey

References A: Authentication Scheme; M: Merit; D: Demerit
[4] A A Hierarchical Privacy Preserving Pseudonymous Authentication Protocol
for VANET
M 1. No need of storage for storing a large pool of pseudonyms
2. Not need of Certificate Revocation List (CRL)
3. Valuable information is secured as compared to server
D 1. If CA, RA, or RSU compromise scheme can fail
[5] A Security Enhancement in Group-Based Authentication for VANET
M 1. A group-based V2 V communication framework is designed to secure
VANET and preserve privacy
2. Eliminates the need of sign message in V2 V communication which leads
to faster authentication
D 1. Digital signature generation and verification of message in V2 V
communication requires more time, which degrades performance of the
network
[6] A Vehicular Authentication Security Scheme (VASS)
M 1. The computational effort is much lower than other methods in the hash
function
2. VASS provide security features such as privacy, authentication, and Sybil
attack detection
D 1. Vehicle to infrastructure communication not considered
[7] A Secure and distributed certification system architecture for safety message
authentication in VANET
M 1. False public-key certification is avoided
2. Provide secure and distributed certification system
D 1. The storage required more for each vehicle to maintains a private key, a
public key, an implicit certificate, a short-term key pair, and a public-key
certificate delivered by the RCA
2. The high transmission range required to transmit various safety messages
[8] A A secure and efficient V2 V authentication method in heavy traffic
environment
M 1. Improve the speed of message processing by sending a low data volume in
areas of heavy traffic
2. Detect replay attacks by checking timestamps of messages
D 1. Vehicle to infrastructure communication not considered
[9] A PKI (Public-Key Infrastructure)
M 1. The user ensures the integrity of the message by signing the encoded
message using digital sign
2. Certificate authorities (CA) digitally sign the data and binds the public
keys with private keys
D 1. Cryptographic keys get compromised
2. Slower because of complexity
[10] A TESLA (Timed-Efficient Stream Loss-Tolerant Authentication)
M 1. Lightweight broadcast authentication protocol
2. It uses symmetric key cryptography so it is faster
D 1. TESLA is vulnerable to memory-based denial-of-service attack
(continued)
Table 1 (continued)
References A: Authentication Scheme; M: Merit; D: Demerit
[11] A TESLA ++
M 1. Protect against memory-based denial-of-service (DoS) attacks
2. Reduces the memory requirements at the receiver’s end
3. Protect against computation-based denial-of-service (DoS) attacks with
equal priority
4. Offers a more secure user authentication mechanism than TESLA
D 1. In the lossy networks, performance of TESLA ++ gets degraded
2. TESLA ++ does not offer non-repudiation
[12] A ECDSA (Elliptic Curve Digital Signature Algorithm)
M 1. It is mathematically derived form of Digital Signature Algorithm (DSA)
2. As an elliptic curve is used for key generations, strength per key bit is
significantly greater
D 1. Attacks on Elliptic Curve Discrete Logarithmic Problem (ECDLP)
2. Hash function can be hacked through attack
A literature survey of the existing authentication scheme shows that speed of authentication and
security of authentication need to improve. Our research work is focused on how to improve speed
of authentication and how to maintain security while authentication
Fig. 2 RSU Communication

range
20 KM
A A
• Vehicle A required 20 s for traveling from S to D.

• Number of vehicles authenticated by RSU are 20,000 ms/100 ms = 200
vehicles.
• 200 vehicles serves by RSU in 20 s.
3.2 Analysis of Time Required for Authentication

and Vehicles Serves by RSU with Respective Key Size
(RSA Algorithm)
Table 2 shows the time required for authentication with respective key sizes.
A number of vehicles served by RSU is calculated as follows:
• RSA with key size 512 bits required 461 ms for authentication.
• Number of vehicles authenticated by RSU are 20,000 ms/461 ms = 43 vehicles.
• 43 vehicles serves by RSU in 20 s.
Table 2 Vehicle served by RSU in 20 s for the RSA algorithm

Sr. no. Key size Authentication time (ms) Vehicles serve by RSU in 20 s
1 256 64 312
2 512 128 156
3 1024 256 78
4 768 192 104
5 1536 384 52
6 2048 512 39
3.3 Analysis of Time Required for Authentication

and Vehicles Serves by RSU with Respective Key Size
(ECC Algorithm)
Table 3 shows the time required for authentication with respective key sizes.
A number of vehicles served by RSU is calculated as follows:
• ECC with key size 112 bits require 301 ms for authentication.
• A number of vehicles authenticate by RSU are 20,000 ms/301 ms = 66
vehicles.
• 66 vehicles are served by RSU in 20 s.
Figure 3 summarizes time required for authentication with respective key sizes
in RSA and ECC. We can see that the time required for authentication is less using
ECC. Because of this, we took ECC for our authentication schemes.
4 Presented Authentication Schemes for VANET
The time required for authentication is an important factor for authentication. There
is a need for efficient authentication protocol for VANET, which not only provides
security but also take less time. Here, we proposed time-efficient authentication
system based on ECC. Along with authentication, it provides the node location
verification, replica attack, and Sybil attack detection. Figure 4 shows the
Table 3 Vehicle served by RSU in 20 s for ECC algorithm

Sr. no. Key size Authentication time (ms) Vehicles serve by RSU in 20 s
1 40 12 1666
2 80 24 833
3 160 48 416
4 120 36 555
5 240 72 277
6 320 96 208
Fig. 3 Time required for authentication with respective key sizes
Simulation
Load Simulator Load Scenario Load Map Simulation Parameter
AECC EECC
Random Key Authentication
ECC Authentication Find Malicious
ECC Key Pool Alert System
RSU & Server

Authentication Key Generator Malicious Node Detection
Fig. 4 Architecture for time-efficient attack-resistant authentication protocol
architecture of the proposed authentication protocol. The architecture shows three

layers, simulation (front end), algorithms (processing part), and Server/RSU-side
activities.
4.1 Simulation
We are using VSIM (VANET Simulator) for our protocol implementation and
testing. VSIM is a Java-based simulator, which provides different classes to create
an environment for VANET. We can implement our protocols and ideas in the
simulator using Java. Different maps are available to test protocols, we can upload a
map. Different road scenarios are available for each map; we can upload those
scenarios in the simulator after uploading the map. The simulator has different input
parameters like a number of vehicles, time stamp of each vehicle, road traffic
density, event priority, time slot, etc., that we can provide to the simulator.
4.2 Algorithms/Processing
Middle layer in architecture shows processing part, i.e., algorithms which give
authentication. Two types of algorithms are presented for authentication, AECC
algorithm and EECC algorithm. Table 4 shows notations used in the algorithms.
4.2.1 AECC (Adaptive Elliptic Curve Cryptography)-Based

Authentication
This algorithm uses random key size where no attacker can guess the key size at the
current time and tries to break it. This system uses a cooperative system to decide
the key size after every defined timeslot. When an attacker tries to guess the key to
break the system as ECC is strong enough this not happens easily. But when an
Table 4 Terms/Notations Used in Algorithm

Terms/notations Meaning
P Key pool
Ts Time slot. (Re-generate keys after every Ts seconds)
G Key Generator
m, a, b Unique parameters
K Keys
Pu Public key
Pr Private key
Vc Current vehicles
NR Neighbor RSU
Ks Key size
Kx New Key
Us Public-key server
Re Verify—Sybil attack, replica attack
attacker succeeds to do so, because of our adaptive key size (AKS) algorithm that
the key is no longer relevant to that attacker.
Algorithm/Pseudocode for AECC-based authentication:
Input: G, {Ts}, {Ks, P}, {V}
Output: Random_Keys, Access Granted/Rejected
Algorithm:
1. Sync {V,RSU,S} ! Ts-Time Slot
2. Ser-Generated TimeSlots {Ts} and KeySizePool {Ks,P}
3. Generate ECC initial parameters G,PW
4. SessionKeyDistribution {Rc, Rs}
a. Generate Random variable rA
b. Compute Ra and Wa
c. Get Ks ! {Ks,P}
d. Generate K ! Ks size Client Side
e. Generate K ! Ks at Server Side
5. Session Key verify – H{K}
a. Generate Hash{P}
b. Verify
6. Session Granted/Rejected
End
4.2.2 EECC (Enhanced Elliptic Curve Cryptography)-Based

Authentication
In enhanced ECC algorithm, we added an extra parameter during transmission of

information from the vehicle to the RSU for key generation. This additional
parameter gives the information about vehicle ID and location of the vehicle to
RSU and other vehicles. This additional parameter also used in key generation. This
algorithm provides replica and Sybil attack detection along with authentication.
Algorithm/Pseudo code for EECC-Based Authentication:
Input: G, {V}, {Ts}, {Ks, P}
Output: Detect Attack, Access Granted/Rejected
Algorithm:
1. Generate ECC initial parameters G,PW
2. SessionKeyDistribution {Rc, Rs}
a. Generate Random variable rA
b. Compute Ra and Wa
c. Get Ks ! {Ks,P}
d. {ID, K,L,TS} ! RSU
e. Verify V by RSU
f. If Verifed
i. Generate K ! Ks size Client Side
ii. Generate K ! Ks at Server Side
g. End IF
h. Else
i. Start Re_verify
1. Vehicle shares new {id, TS, L}
2. Verify by RSU and Server
ii. End
3. Session Key verify – H{K}
a. Generate Hash{P}
b. Verify
4. Session Granted/Rejected
End
4.3 RSU and Server
Key generation and authentication take place at server side and malicious node
detection is take place at RSU.
5 Conclusion
In VANET, nodes are highly mobile and there is need to improve the speed of
authentication mechanism. A cryptographic solution provided by the previous
scheme need to enhance as per VANET requirement. Considering this we proposed
a new effective protocol for VANET authentication. An ECC provides faster
authentication than RSA, Diffie–Helman and other traditional cryptographic solu-
tions with same key sizes. We proposed a modification in existing ECC algorithm
to improve performances of VANET. The adaptive key size algorithm makes the
system more secure against password cracking attack. Enhanced ECC algorithm
provides security against Sybil attack and replica attack.
References
1. Saira Gillani, Farrukh Shahzad, Amir Qayyum et. al.: A Survey on Security in Vehicular Ad
HocNetworks. In: Springer, LNCS 7865, pp. 59–74, 2013.
2. Jonathan Petit, Zoubir Mammeri.: Authentication and consensus overhead in vehicular ad hoc
networks. In: Springer Science, Telecommun Syst (2013).
3. Khalid Haseeb, Dr. Muhammad Arshad, et. al.: A Survey of VANET’s Authentication. In: PG
NET 2010. Liverpool. John Moores University UK, June 2010.
4. Ubaidullah Rajput, Fizza Abbas, Heekuck Oh.: A Hierarchical Privacy Preserving
Pseudonymous Authentication Protocol for VANET. In: IEEE Access, October 25, 2016.
5. Rajkumar Waghmode, Rupali Gonsalves et.al.: Security Enhancement in Group Based
Authentication for VANET. In: IEEE International Conference on Recent Trends in
Electronics Information Communication Technology, May 20–21, 2016, India.
6. Yongchan Kim, Jongkun Lee.: A secure analysis of vehicular authentication security scheme
of RSUs in VANET. In: Springer-Verlag France 2016.22.
7. Tiziri Oulhaci, Mawloud Omar, Fatiha Harzine, et. al.: Secure and distributed certification
system architecture for safety message authentication in VANET. In: Springer
Science + Business Media New York 2016.
8. Myoung-Seok Han, Sang Jun Lee,Woo-Sik Bae.: A Secure and Efficient V2 V Authentication
Method in Heavy Traffic Environment. In: Springer Science + Business Media New York
2016.
9. A. Rao, A. Sangwan, A. Kherani, A. Varghese, B. Bellur, and R. Shorey.: Secure V2 V
Communication With Certificate Revocations. In: proceedings of the IEEE Infocom 2007.
10. Yih-Chun Hu and Kenneth P. Laberteaux.: Strong VANET security on a budget. In:
Proceedings of the 4th Annual Conference on Embedded Security in Cars (ESCAR 2006).
11. Arzoo Dahiya, Mr. Vaibhav Sharma,: A survey on securing user authentication in vehicular
adhoc networks. In: 6th EmbeddedSecurity in Cars Conference.
12. Don Johnson, Alfred Menezes and Scott Vanstone.: The Elliptic Curve Digital Signature
Algorithm (ECDSA). In: Published in IJIS, Vol. 1 (2001) pp. 36–63.
Inferring User Emotions from Keyboard
and Mouse
Taranpreet Singh Saini and Mangesh Bedekar
Abstract This chapter emphasizes on retrieving user emotions from keyboard and
mouse using different parameters. These parameters can be user keyboard typing
style, mouse movements, and some physiological sensors are used. This field of
retrieving emotions from machines comes under the field of affective computing.
Keywords Affective computing Emotions Keyboard Mouse

Physiological sensors
1 Introduction
Emotions are the form of human feelings and sentiments that results in physical and
psychological changes that affect human behaviors. The physiology of emotion is
closely linked to arousal of the human nervous system with various emotional states
and power of arousal relating, apparently, to particular emotions. Emotion is also
linked to the behavioral tendency which changes over time. Emotions involve
different components, such as subjective experience, cognitive processes, expres-
sive behavior, psychophysiological changes, and instrumental behavior. The ability
to recognize, interpret, and express emotions plays a key role in human commu-
nication and increasingly in HCI. Recent research could demonstrate that humans
have an inherent tendency to interact with computers in a natural and humane way,
mirroring interactions between humans in social situations. Computer systems
currently are not able to recognize or respond to these complexities of a natural
social interaction. Emotion recognition has become an important subject when it
comes to human–machine interaction. Various methods have been used in the past
to detect and evaluate human emotions. The most commonly used techniques
T. S. Saini (&) M. Bedekar

Department of Computer Engineering, Maharashtra Institute of Technology, Pune, India
e-mail: saini_13sf@yahoo.co.in
M. Bedekar
e-mail: mangesh.bedekar@mitpune.edu.in

https://doi.org/10.1007/978-981-10-7245-1_58
592 T. S. Saini and M. Bedekar
include the use of textual information, facial expressions, speech, body gestures,
and physiological signals. These signals are obtained from a skin temperature
sensor, a heart rate sensor, and a force sensor.
The study of emotions from machines has developed into a new field known as
affective computing. Affective computing is the process to learn, observe, and
develop systems and applications that can sense, interpret, procedure, and pretend
human-like effects. It is a field of work spanning computer science, psychology, and
cognitive science. While the origins of the topic may be traced to far back in history
as with humans early philosophical reasoning into emotion, the more quantifiable
recent work based on computer science got shape with the path-breaking study of
R. W. Picard’s 1995 paper on affective computing.
2 Literature Survey
2.1 Affective Haptics
Affective haptics is the newest emerging area of study, which emphases on the
study and plan of devices and structures that can incite, improve, or affect the
emotional state of a human by means of intelligence of touch. The study area is
highlighted in the research paper by Dzmitry Tsetserukou et al. Focused by the
inspiration to enhance social communication and expressively immense under-
standing of users’ real-time messaging, virtual, augmented realities, the idea of
strengthening (intensifying) own moods, and replicating (simulating) the emotions
felt by the communicating partner was scheduled [1].
Four simple haptic (tactile) networks leading our emotions can be demonstrated,
(1) Functional changes
(2) Bodily stimulation
(3) Social touch
(4) Emotional haptic design.
2.1.1 Applications of Affective Haptics
• Treating nervousness
• Monitoring and adjusting moods on the basis of physiological signals,
• Affective and combined games,
• Testing psychology,
• Communication systems for kids with autism (Fig. 1).
Inferring User Emotions from Keyboard and Mouse 593
Fig. 1 Areas of affective computing
2.2 Detecting and Recognizing Emotional Information
Sensing strong feelings begins with information gathering from action-less sensors,
which take in facts about the user’s physical state or behavior. The facts gathered
are of the form of human cues which are used to perceive feelings. For example, a
camera might click photographs of the human face; recognize body position thus
illuminating human feeling, while a microphone may transform human sound to
electric signals based on verbal words. Many other explicit sensors can be used to
determine specific feelings by directly calculating physiological facts, such as the
human’s skin, body temperature, etc. [1].
This information is treated to extract decisive information from the gathered
facts. This is finalized using machine-learning methods that process unalike
modalities, such as usage of words, natural language processing, or of the face
words discovery, and create either tickets giving name-value pairs or orders in a
valence-arousal space [1].
2.3 Emotions in Machines
The new area in computing feelings refers to the design of computational systems
which display strong feelings or a part of the feelings. A more valuable move of
linking computers is based on the current technology based on powers, which is the
simulation of feelings in communicating agents to improve interactivity between
human-like and machine-like communications. While human-like moods are often
associated with strong waves in hormones and other neuropeptides, emotional state
in machines might be connected with outline states connected with forward growth
(or exist without of forward development) in self-ruled learning structures. In this
viewpoint, feelings are solid states which are time derivatives (perturbations) in the
learning turn of a computer-based learning system based on instructions [1].
3 Mathematical Modeling
The user can interact with the computer. We can capture user attributes like typing
speed, key press—key release time, mouse movements, mouse scroll, mouse hover,
mouse click, physiological signals—heartbeat sensor, force sensor, and temperature
sensor.
Then, we will map all these attributes with emotions to capture user’s current
emotional state.
The system is defined by the tuple, “S”, such that
• S = {I, Ic, O, Oc, F}
• I = Input to the system (Web pages = W and User Activities = U),
• Ic = Input Conditions (Mouse behaviors = MB and Keyboard behaviors = KB,
time spent on web page, sensors),
• O = Output of the system (User Profile = I + Ic + F)
• Oc = Output Conditions (as behavior changes, user profile also changes),
• F = Functions I to O, through Ic to Oc.
(Table 1).
There can be four conditions,
• First condition where mouse, keyboard, and sensors work together.
• Second condition, only mouse will work.
• Third condition will be for keyboard only.
• Last condition where readings can be taken from sensors only.
4 Proposed Work
System flow describing the proposed work is as shown in Fig. 2.

Table 1 Mathematical modeling

VM Vertical mouse movement HSM Mouse horizontal scroll
H0VM Mouse hover BK Keyboard bookmark
HIM Mouse highlight SK Keyboard save
HM Horizontal mouse movement HSM Mouse horizontal scroll
BM Mouse bookmark BK Keyboard bookmark
SM Mouse save SLR Shift left/right
F/B Forward and backward CA Control all
SE Send email CK Keyboard copy
PM Mouse print PAK Keyboard paste
STM Mouse select text PK Keyboard print
CM Mouse copy VK Keyboard vertical
PAM Mouse paste HK Keyboard horizontal
VSM Mouse vertical scroll TK Keyboard trace
S Sensors Hs Heartbeat sensor
Fs Force sensor Ts Temperature sensor
Fig. 2 System flow

5 Implementation
5.1 Detecting User Web Activities
• Time Spent on Web Page—User Engagement—Time was considered as one of

the interest indicators. It consists of passive intervals and active intervals.
• User Click Behavior—Users’ clicks on links is a factor for implicitly learning
the user’s interests.
• Copying Text from Web Page—The text copied indicates relevancy and that too
in which parts of the web page the user is interested in.
• Mouse Movements on Web Page—Mouse movements can also determine the
interest of the user.
• Amount and Speed of Scrolling on Web Page—several implicit measures
related to the interests of the user and scrolling was one of the factors.
• Number of Time a Web Page is visited—greatest implicit indicators of interest
were the time spent on the web page, number of link clicks, and arrow presses.
• User Behavior Based on Age and Gender—For understanding user’s emotions,
it is important to understand user’s gender and age. It can be done by the process
of gender identification and age detection.
• Most Number of Files Viewed—Types of files viewed (image, text) by the user
also helps to understand the emotional state of the user.
5.2 Material Used for Physiological Signals
• ATMEGA 328—The ATMEGA 328 is a CMOS 8-bit microcontroller with low

power based on the AVR gave greater value to RISC buildings and structure
design. The ATMEGA 328 operates at 1 MIPS per MHz which allows the system
designed to make the most out of the power consumed for processing [2].
• Bread Board—This component is used because it will be the base of our device
with all the connections being connected with its help.
• Heart Beat Sensor TCRT1000—The use of this sensor simplifies the build
process of the sensor part of the project as both the infrared light emitter diode
and the detector are arranged side by side in a leaded package, thus blocking the
surrounding ambient light, which could otherwise affect the sensor performance
[3]. This sensor continuously monitors the pulse rate of the user, which is
directly connected with the emotions.
• Temperature Sensor LM35—This component is selected because this is the only
sensor, which is used to measure the body temperature accurately [4].
• Force Sensor MF01-N-221-A01—This sensor is mainly used on the keys, it can

be a computer keyboard, mouse buttons, musical instruments like piano, har-
monium, etc., or also in vehicles to detect force or the pressure [5].
• Bluetooth Module HC05—This Bluetooth module is used for transmitting data
to the smartphone and to the computer also. This module directly sends data
from the sensors and it is easily connected to any device [6]. We are using
ArduDroid application on the android device, which directly gets data from this
Bluetooth module.
(Fig. 3; Tables 2 and 3).
In Fig. 4, it is shown that human physiological signals from the heartbeat sensor,
force sensor, and temperature sensor are the direct input into the microcontroller
where all the processing takes place. The microcontroller used is the ATMEGA
328, where all the sensors are attached and LCD shows all the results. Heartbeat
sensor continuously checks the pulse rate of the user. Force sensor senses the user
pressure on the mouse button and keyboard keys. This will give result in the form
of percentage value out of 100%. The temperature sensor senses the user temper-
ature in the form of °C from users hand while working with a mouse. The
microcontroller transmits data to Android and computer device through Bluetooth
module which is wireless. Then, the data can be analyzed using different algo-
rithms, techniques, and tools.
Fig. 3 Schematic diagram

Table 2 Heart rate analysis

Normal Systolic rate is lesser than 120 mmHg and diastolic rate is lesser than
80 mmHg
At risk Systolic rate is higher than 120–139 mmHg and diastolic rate is higher
(pre-hypertension) than 80–89 mmHg
High Systolic rate is 140 mmHg or higher diastolic rate is 90 mmHg or
higher
Table 3 Temperature analysis

Normal The average normal temperature is 98.6 °F (37 °C). But “normal” varies from
person to person
Abnormal Oral, temporal artery temperature: Fever: 100.4 °F (38 °C)–103.9 °F (39.9 °C),
High fever: 104 °F (40 °C) and higher
Armpit (auxiliary) temperature: Fever: 99.4 °F (37.4 °C)–102.9 °F (39.4 °C),
High fever: 103 °F (39.5 °C) and higher
A rectal or ear temperature of less than 97 °F (36.1 °C) means a low body
Fig. 4 Working of sensors
6 Results
Some of the results are captured from users using software and hardware. These
results show the user data with respect to the website and their usage. The more the
usage the more will be the engagement with the website. The heat maps show the
user movements on the web page. This can help in capturing the user pattern
(Figs. 5, 6, 7 and 8).
Fig. 5 User overview
Fig. 6 User engagement data
Fig. 7 Gender data

Fig. 8 Heat maps
Fig. 9 Android results
Figure 9, shows the Android window where all results are shown in the get data
column. These results are captured by the ArduDroid Android application in get
data column, where T shows the temperature readings, F1 AND F2 are the readings
for the force sensor.
7 Emotion Detection
Emotions play a very important role in problem solving, decision-making, com-

municating, and adapting to the unpredictable environment. These emotion detec-
tion methods have been used for improving the quality of human life. This model of
emotion detection and recognition will help the users in early detection of any
problem. The major agenda of this model is to recognize user emotions while
working with keyboard and mouse. For the purpose of emotion detection, we will
capture all the data from the sensors through Bluetooth module and map the values
with different negative and positive emotions using the circumplex model.
8 Conclusion
The advantage of this system is that we are using keyboard and mouse, which are
the natural type of biometrics, which does not require any special hardware. The
sensors are small and are not very costly, so they can easily fit in the mouse and the
keyboard for the continuous reading. These readings will help the computer users in
evaluating and recognizing their emotions. This model can be used at any place
where computers are used like universities, schools, offices, banks, hospitals, etc.
This system can also be used in the home for the old people, who are using
different devices like television remote or computers, for continuously monitoring
their health.
References
1. Taranpreet Singh Saini, Mangesh Bedekar, Saniya Zahoor. “Analysing human feelings by
Affective Computing—survey”, 2016 International Conference on Computing
Communication Control and automation (ICCUBEA), 2016.
2. Russell, J. (1980). A circumplex model of affect. Journal of Personality and Social
Psychology. http://doi.org/10.1037/h0077714.
3. Performance, L. Power, A. V. R. Microcontroller, A. R. Architecture, I. Programming, O.
B. Program, S. P. W. M. Channels, O. A. Comparator, S. Sleep, M. Idle, and A. D. C.
N. Reduction, “Features—32 8 General Purpose Working Registers—Optional Boot Code
Section with Independent Lock Bits True Read-While-Write Operation—Programming Lock
for Software Security—Internal Calibrated Oscillator—External and Internal Interrupt
Sources and Extended Standby I/ O and Packages Operating Voltage: Temperature Range:
Speed Grade: Microcontroller Bytes In-System Programmable Flash ATmega88PA
ATmega168PA ATmega328P.”.
4. “Heartbeat sensor information TCRT 1000.pdf”.
5. Description, “Precision Centigrade Temperature Sensors LM35 LM35A LM35C LM35CA
LM35D Precision Centigrade Temperature Sensors,” no. December 1994, 1995.
6. Information, “l A7800 Series Positive-Voltage Regulators Internal Thermal-Overload
Protection l A7800 SERIES,” no. May 1976, 2003.
An Overview of Automatic Speaker
Verification System
Ravika Naika
Abstract Biometrics is used as a form of identification in many access control

systems. Some of them are fingerprint, iris, face, speech, and retina. Speech bio-
metrics is used for speaker verification. Speech is the most convenient way to
communicate with person and machine, so it plays a vital role in signal processing.
Automatic speaker verification is the authentication of individuals by doing analysis
on speech utterances. Speaker verification falls into pattern matching problem.
Many technologies are used for processing and storing voice prints. Some of them
are Frequency Estimation, Hidden Markov Models, Gaussian Mixture Models,
Neural Networks, Vector Quantization, and Decision Trees. Mainly speaker veri-
fication depends upon speaker modeling and this paper represents a brief overview
of the speaker verification system with feature extraction and speaker modeling.
Bob spear toolkit is used for evaluation and experiment for the result and analysis.
Bob spear is an open-source toolkit for speech processing. For evaluation purpose,
three algorithms are proposed which are GMM, ISV, and JFA with the same
preprocessing and feature extraction techniques.
Keywords ASV MFCC LPCC LLR PLDA GMM UBM

SVM
1 Introduction
Various methods and technologies are used for biometric authentication which
include face, fingerprint, iris, palm print, retina, hand geometry, signature, and
voice. Face recognition works with live capturing or digital image data. Fingerprint
recognition works with the printed pattern of particular person’s fingerprint.
Likewise, speaker verification works on the basis of voice. As our primary need for
R. Naika (&)
Institute of Technology, Computer Science and Engineering, Nirma University, S.G.
Highway, Ahmedabad 382481, India
e-mail: ravika.naika94@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_59
604 R. Naika
Fig. 1 Classification of speaker recognition [2]
communication is speech, we can say it is an appealing modality which includes

individual differences in both physiological and behavioral manner [1]. For
example, we can take vocal tract shape and intonation for use in automatic speaker
verification.
The automatic speaker verification system is built on the bases of speech science,
speaker characterization, and pattern recognition algorithm. It checks whether the
person is who he/she claims to be. Speaker verification is the subpart of speaker
recognition [2].
There are two parts of speaker recognition (1) speaker identification (2) speaker
verification as described in Fig. 1.
Speaker verification is related to the physical shape of individual’s vocal track
with airways and soft tissues from where the sound is originated. All these com-
ponents work in combination which includes physical movement of the jaw, larynx,
and nasal passages. Speech acoustic patterns come from airways’ physical char-
acteristics and the motion of mouth and pronunciations. They are used as behavioral
components of biometrics [3].
Several commercial applications are using speaker verification as a biometric for
controlling or accessing important information or services and accounts. Speaker
verification offers the replacement of PINs and passwords with something that
cannot be stolen or forgotten.
2 Automatic Speaker Verification
Speaker verification is the task of determining whether a person is who he/she

claims to be. Automatic speaker verification works on speech signals and the
information which it conveys to the listener on different levels. First level conveys
the message and the other levels convey the emotion, gender, and identity. This
verification task is based on closed set and open set. In closed set, the input is
matched with the stored samples and the closest to the input is returned, there is no
rejection in closed set [4].
An Overview of Automatic Speaker Verification System 605
ASV can be classified in text-dependent and text-independent. In text-

dependent, the speaker has to speak the predefined text while in text-independent
there is no prior knowledge of any text that is why the text-independent is more
difficult yet more flexible. Text-dependent mostly used with authentication systems
while text-independent is more suitable for surveillance applications. Text-
dependent technology is nowadays the most commercially viable and useful
technology. Although the research has been done on both the task. The basic model
for automatic speaker verification system is given here. As mentioned before, the
result is given based on the accept or reject decision and its similarity is compared
with a threshold. Here, the threshold is taken as likelihood ratio [5].
Figure 2 describes how automatic speaker verification works. First, input speech
is given and then input speech is compared with the reference model and then it
calculates the similarity score. Based on the threshold value, the decision of
acceptance or rejection is taken. This is the basic flow of speaker verification
system.
Mostly speaker verification is based on feature extraction and speaker modeling.
Speaker verification is the pattern matching problem so different technologies like
GMM (Gaussian Mixture Model), HMM (Hidden Markov Model), SVM (Support
Vector Machine), VQ (Vector Quantization), ANN (Artificial Neural Network),
JFA (Joint Factor Analysis), ISV (Inter-session Variability), etc., are used for
performing the speaker verification. Different features like MFCC (Mel-Frequency
Cepstral Coefficients), LPCC (Linear Predictive Cepstral Coefficients), PLP
(Perceptual Linear Predictive), etc., and their fusions are used for better perfor-
mance of the system.
Research and development techniques in this area have been undertaken for
almost 3–4 decades, so it has been an active area for research. Different tech-
nologies and methods had been used for speaker verification. Its year wise repre-
sentation is shown in the Fig. 3 [6–11]. From starting with pattern recognition
technology to using GMM, HMM, and SVM.
Fig. 2 Model of automatic speaker verification system [4]

606 R. Naika
Fig. 3 Evaluation in speaker verification technologies [6–11]
Speech signal contains timber of voice, prosody, and content of language.

Classification of speaker individuality can be done in three parts short-term spectral,
prosodic, and high-level idiolectal features. MFCC, LPCC, and PLP are all spectral
features.
Prosodic features are extracted from longer duration of time. Pitch, rhythm,
energy, and duration can be included in prosodic features. However, these features
are sparse in nature so it takes a large amount of training data to extract the features,
also they are unreliable in a noisy environment.
The third one is high-level idiolectal features and they are extracted from lexi-
cons, which represents speaker behavior or lexical cues [12]. They are less sensitive
to the noise and channel effects. They require more complex front ends to extract
the features.
2.2 Speaker Modeling and Classifiers
As mentioned above speaker verification is characterized in text-dependent and

text-independent and both are using different datasets. Text-dependent uses rela-
tively small size of data in comparison of text-independent verification because in
text-independent the speech is not predefined so the data becomes large. Many
classifiers are used for text-dependent and text-independent both. But in many
classifier theoretical frameworks uses log-likelihood ratio [13].
pðXjH0 Þ
l ¼ log ð1Þ
pðXjH1 Þ
Its comparison is done with pre-obtained threshold to check the hypothesis of

same speaker H0 or different speaker H1 where p(X|H0) and p(X|H1) are the like-
lihood for each. We can take X = x1, x2 …, and p(X|H1) is used to know the
common effects which are not connected with speaker identity. In most approaches,
X are considered as MFCCs and with that different classifier like GMM, PLDA,
etc., is used.
2.3 System Fusion
It is a recent trend in ASV community. One of the examples of this fusion system
we discussed in text-independent classifiers which is the combination of GMM,
SVM, and UBM (Universal Background Model). It gives better result in compar-
ison of single classifiers. Because it captures different aspects of speech together,
this is not possible with the single classifier. Fusion system provides a platform for
research on large-scale collaboration. Using the fusion of different classifier dif-
ferent feature extraction can be done together. But it uses the weighted summation
of base classifiers and these weights are optimized.
3 Experiment Information
The experiment is done on bob spear toolkit. It is an open-source tool for signal
processing and the machine learning algorithms. In the recent past few years,
researchers are using this tool, so it has been a very popular tool in signal pro-
cessing. It implements the whole chain from feature extraction to evaluation. It
includes several techniques like GMM, SVM, JFA, i-vector with different databases
like MOBIO, NISTSRE, VoxForge [14], etc. It also provides the system fusion for
more accuracy in result and analysis. This tool provides the flexibility of using any
data sets.
Speaker verification includes three things which are speech activity detection,
feature extraction, and background data modeling with score computation. In order
to do this, datasets are divided into background training set, DEV (Development)
set, EVAL (Evaluation) set so the computation can be done easily [15]. Bob spear
uses different techniques at each stage which are explained below.
Preprocessing
Mostly speech activity detection can be done in two ways. The first one is unsu-
pervised and it computes energy values frame level wise. The other one computes
the values based on the combination of energy and modulation 4 Hz. In both the
techniques, a class with higher mean are considered as speech class.
Feature Extraction
This tool provides an efficient way to implement spectral and cepstral features.
According to that, it computes its MFCC values. Cepstral mean and normalization
of variance is also integrated. This tool provides different features for different
datasets. For example, MFCC, LPCC, PLP are used with different number of
Gaussians like MFCC60, MFCC20.
608 R. Naika
Modeling, Enrollment, and Score Computation

GMM, JFA, and ISV are some art modeling techniques which are integrated into
the toolbox. GMM includes the background model of UBM, which uses maximum
likelihood ratio. Enrollment is using Maximum-A Posteriori adaption. Matching is
done using log-likelihood ratio. JFA works on the training set and used to estimate
eigenvoice and eigenchannel and matrix which helps in modeling the residual
noise. Matching is done based on linear score approximation. ISV works like JFA
up to certain extents.
Generally, UBM model represents general person-independent feature charac-
teristics in biometric verification system. So that they can be compared with the
model of person-specific feature characteristic at the decision time. Here, UBM
model is trained with speech samples of the specific enrolled speaker and then
likelihood ratio test is done for a specific model with the UBM. It can also be used
as a prior model, while training the speaker-specific model. That can be done with
Maximum-A Posteriori adaption. Here, GMM-UBM works as the priority model
for JFA and ISV model.
3.1 Information of Database
VoxForge is a collection of transcribed speech samples. It is free and can be easily

used with speech recognition and verification engines. This database supports
Windows, Linux, and Mac environment and it can be downloaded from www.
voxforge.org. VoxForge collects these speech samples from volunteers and they
make it available to the open-source speech recognition engine. Here, it is used with
Bob spear toolkit. All these samples come from volunteers who visit this website
and record their voice while reading written prompts provided by the website.
These voice samples are recorded by volunteer’s own microphone so the quality
varies significantly. This corpus includes 6561 files and all are containing English
audio files [16]. Audio files of 30 speakers are taken and all the speakers are
selected randomly. This database is distributed in three parts: Training (10 speak-
ers), Development (10 speakers), and Test (10 speakers). This database is of 1 GB
size.
The experiment is performed on VoxForge database with different algorithms and

they are UBM-GMM, JFA, and ISV. According to different algorithms, it gives
different values of EER for development and evaluation set. It also gives Cllr and
Min Cllr values for both the development and evaluation set [17].
Table 1 Results of baseline Measure UBM-GMM ISV JFA

ASV system for development
set Cllr 0.72080 0.68779 0.73322
Min Cllr 0.04816 0.024024 0.10287
EER 1.981% 1.037% 3.15%
Table 2 Results of baseline Measure UBM-GMM ISV JFA

ASV system for evaluation set
Cllr 0.0633 0.03870 0.30423
Min Cllr 0.81157 0.076971 0.84141
HTER 2.074% 2.370% 8.944%
Cllr
It is the measure of log-likelihood ratio, which can properly evaluate the discrim-
ination of all the log-likelihood ratio scores. It also evaluates the quality of the
calibration.
Min Cllr
Min Cllr represents the minimum possible value of Cllr and that can be achieved by
the optimally calibrated system. It ranges from zero to infinity but mostly the value
is equal to one or below one. The smaller the Min Cllr better the indication of the
classifier [18].
EER
It is an equal error rate, and it represents an overall measure of system performance.
It corresponds to the threshold value where false acceptance rate and the false
rejection rate is same.
With the algorithm GMM-UBM, it uses 25 training iterations and 256 Gaussian
for the evaluation. Here, we have used the MFCC60 extractor and energy-2 Gauss
pre-processor. The database has total 30 wave files so after completion of every
iteration, it gives scores for both development and evaluation dataset and according
to that it calculates the EER and HTER.
All the values of development and evaluation set is given in the below Tables 1
and 2.
5 Conclusion
In this paper, a brief summary is given about automatic speaker verification system.
Also, we have discussed various techniques which were used in different years.
Speaker modeling, feature extraction, and score computation are discussed in detail.
Bob spear toolkit is a useful kit for speech and signal processing and it is helpful for
implementing different algorithms. Results of evaluation and development dataset
are given in the table of the algorithm UBM-GMM, JFA, and ISV. From the results,
we can conclude that UBM-GMM algorithm gives the best result with the
VoxForge database.
610 R. Naika
References
1. Campbell, Joseph P. “Speaker recognition: a tutorial.” Proceedings of the IEEE 85.9 (1997):
1437–1462.
2. Tranter, Sue E., and Douglas A. Reynolds. “An overview of automatic speaker diarization
systems.” IEEE Transactions on Audio, Speech, and Language Processing 14.5 (2006): 1557–
1565.
3. Voice Acoustics: an introduction, http://newt.phys.unsw.edu.au/jw/voice.html.
4. Rosenberg, A. E. “Evaluation of an Automatic Speaker Verification System Over Telephone
Lines.” Bell System Technical Journal 55.6 (1976): 723–744.
5. Reynolds, Douglas. “An overview of automatic speaker recognition.” Proceedings of the
International Conference on Acoustics, Speech and Signal Processing (ICASSP) (S. 4072–
4075). 2002.
6. Sturim, Douglas E., et al. “Speaker indexing in large audio databases using anchor models.”
Acoustics, Speech, and Signal Processing, 2001. Proceedings. (ICASSP’01). 2001 IEEE
International Conference on. Vol. 1. IEEE, 2001.
7. Sant’Ana, Ricardo, Rosngela Coelho, and Abraham Alcaim. “Text-independent speaker
recognition based on the Hurst parameter and the multidimensional fractional Brownian
motion model.” IEEE Transactions on Audio, Speech, and Language Processing 14.3 (2006):
931–940.
8. Campbell, William M., et al. “Speaker verification using support vector machines and
high-level features.” IEEE Transactions on Audio, Speech, and Language Processing 15.7
(2007): 2085–2094.
9. Markov, Konstantin P., and Seiichi Nakagawa. “Text-independent speaker recognition using
non-linear frame likelihood transformation.” Speech Communication 24.3 (1998): 193–209.
10. Murthy, Hema A., et al. “Robust text-independent speaker identi cation over tele-phone
channels.” IEEE Transactions on Speech and Audio Processing 7.5 (1999): 554–568.
11. Moattar, Mohammad Hossein, and Mohammad Mehdi Homayounpour. “Text-independent
speaker verification using variational Gaussian mixture model.” ETRI Journal 33.6 (2011):
914–923.
12. Larcher, Anthony, et al. “Text-dependent speaker verification: Classifiers, databases and
RSR2015.” Speech Communication 60 (2014): 56–77.
13. Bimbot, Frdric, et al. “A tutorial on text-independent speaker verification.” EURASIP journal
on applied signal processing 2004 (2004): 430–451.
14. Khoury, Elie, Laurent El Shafey, and Sbastien Marcel. “Spear: An open source toolbox for
speaker recognition based on Bob.” 2014 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP). IEEE, 2014.
15. https://pythonhosted.org/bob.bio.base/installation.html.
16. VoxForge, Free Speech Recognition http://www.voxforge.org/.
17. https://pythonhosted.org/bob.bio.spear/baselines.html.
18. N. Brummer and D. Van Leeuwen, On calibration of language recognition scores, in IEEE
Odyssey Speaker and Language Recognition Workshop, 2006, pp. 18.
Topic Modeling on Online News
Extraction
Aashka Sahni and Sushila Palwe
Abstract News media includes print media, broadcast news, and Internet (online
newspapers, news blogs, etc.). The proposed system intends to collect news data
from such diverse sources, capture the varied perceptions, summarize, and present
the news. It involves identifying topic from real-time news extractions, then per-
form clustering of the news documents based on the topics. Previous approaches,
like LDA, identify topics efficiently for long news texts, however, fail to do so in
case of short news texts. In short news texts, the issues of acute sparsity and
irregularity are prevalent. In this paper, we present a solution for topic modeling,
i.e, a word co-occurrence network-based model named WNTM, which works for
both long and short news by overcoming its shortcomings. It effectively works
without wasting much time and space complexity. Further, we intend to create a
news recommendation system, which would recommend news to the user according
to user preference.

Keywords Data mining Topic modeling Document clustering

Online news Recommendation
1 Introduction
Recently, a generative probabilistic model of textual corpora has been considered,

to segregate representations of the news(information). It decreases the depiction
length and discloses inter- and intra-document factual structure. Such models
ordinarily will be classified into one of two arrangements that produces every word
on the argument of some number of preceding words or word classes and those that
A. Sahni (&) S. Palwe

Department of Computer Engineering, MITCOE, Pune, India
e-mail: aas.sah91@gmail.com
S. Palwe
e-mail: sushila.aghav@mitcoe.edu.in

https://doi.org/10.1007/978-981-10-7245-1_60
612 A. Sahni and S. Palwe
produce words based on latest variables induce from word correlations independent
of the request in which the words show up.
Topic models are mainly applied to text in spite of a willful need of alertness of
the hidden semantic structures that exist in natural language. In a topic model, the
words of each document are accepted to be exchangeable; their probability is
invariant to stage. This improvement is proven valuable for inferring proficient
derivation procedures and quickly examining huge corpora.
However, exchangeable word models have limitations. For the classification or
data recovery, these model are used, where a coarse statistical footprint of the topics
of a document is adequate for achievement, exchangeable word models are weakly
prepared for issues relying on all the more “fine-grained characteristics of language”
[5]. For example, though a topic model can recommend documents important for a
query, it cannot find particularly important phrases for question answering.
Therefore, while a topic model could find a pattern, for example “eat” happening
with “cheesecake”, it does not have the illustration to represent selection preferences
the procedure where certain words restrict the decision of the words that follow.
Due to the rapid growth of Internet and the development of different kinds of
web applications, online news have become one of the most leading topic of
Internet. Internet news covers topics ranging from politics, weather to sports,
entertainment, and different important events, and, in this way, news captures, in a
major way, the cultural, social, and political picture of society. Apart from the
media networks, self-governing news sources also play a role in reporting the
events which may get ignored in the midst of the major stories. Hence, accurately
mining topics behind these online news texts are essential for a wide range of tasks,
including content, query suggestion, document classification, and text clustering.
Due to the increasing volume of short text data, it is essential to gather the latent
topics from them, which would be useful for an extensive variety of substance
examination applications, for example, content analysis, user interest formulation,
and emerging topic detecting. The previously used topic models like LDA and its
variations are successful in efficiently detecting topics from long texts (news),
however, failing to do so in the case of short texts (news).
Short texts are common on the web, regardless of traditional websites, for
example, titles of webpage, text endorsements, and image subtitles, or in social
networking sites, e.g., status updates, tweets, and questions in Q and A websites.
Dissimilar from traditional normal texts (for example, articles, scholarly paper
publications). Short texts, as the name indicates, include a few words. As there is an
increasing volume of short text datasets, gathering latest topics from them is
essential for an extensive variety of substance examination applications, e.g.,
content analysis, user interest formulation, etc.
The content sparsity in short text data leads to some new difficulties in topic
detection. “PLSA and LDA” [5] are an example of previously used topic models
which explain that a document could contain a number of topics, where a topic
conveys subject of the document by a set of related or connected words. To study the
topic components (that is topic-word distributions) as well as mixture coefficients
(that is topic proportions) of every document, statistical procedures are then used.
Topic Modeling on Online News Extraction 613
Traditional topic models detect topics within a text corpus by first capturing the
document-level word co-occurrence patterns implicitly, and then, directly applying
these models on short texts to tackle the acute data issues. In individual short text, the
frequency of words assume less differentiate role than in case of long text, by making
it difficult to collect the words which are more associated in every document.
Document clustering is grouping or categorizing documents or text. Document
Clustering is an unsupervised learning technique; it is very useful for automatic
topic detection and faster information retrieval. For example, as the volume of
online data are increasing rapidly, users and information retrieval system needed to
classify the desired document against a specific query. There are two types of
clustering approaches used, top-down and bottom-up. In this paper, focus is on the
performance of K-means clustering algorithm, which is a top-down clustering
algorithm; every document is assigned to the cluster whose center also called
centroid is nearest. Here, documents are presented as document vector and the
center is the average of all the documents in the cluster, in vector space model.
In this paper, a study about the related work done in Sect. 2, the proposed
approach modules description, mathematical modeling, algorithm, and experi-
mental setup in Sect. 3 and finally, we provide a conclusion in Sect. 4.
2 Literature Survey
Cheng et al. a novel topic model, “biterm topic model (BTM)” [1], has been
proposed for short text topic modeling. Are sparse and experience data ambiguity
and imbalance due to lack of contextual information, which are not solved by
traditional topic modeling approaches like LDA and its variants. In order to address
these problems, in this paper, BTM has been proposed, which works on the
corpus-level information (rather than document level) and is based on the word
co-occurrence patterns. Also, online algorithms have been presented to manage
large amount of incoming short content information.
Sahami et al. a “similarity kernel” [2] function has been described and its prop-
erties have been mathematically analyzed. Conventional document similarity mea-
sures (e.g., cosine) are not able to determine the similarity between two short text
portions efficiently. In order to solve this problem, a new technique, a similarity
kernel function, to measure the similarity between short text portions has been
introduced in this paper. The author has also performed experiments using this kernel
function in an expansive-level framework by suggesting associated queries to web.
In Phan et al. [3], a generic structure for building classifiers has been presented,
which deals with short and sparse text and web portions by discovering hidden
points found from large-scale information collections. Numerous examples of
working with short portions of web, such as, forum messages, blog discussions,
news headlines, image captions, etc., do not aim at gaining high precision because
of the information sparsity. The author, hence, came up with an idea of gaining
external learning, by collecting an external data compilation called “universal data
set”, to make the information more related and additionally extend the scope of
classifiers to handle future information better.
Jin et al. [4] presents an unconventional approach to cluster short text messages
via transfer learning from auxiliary long text data. To bridge the possible incon-
sistency between source and target data, we propose a unconventional topic model
—Dual Latent Dirichlet Allocation (DLDA) model, which jointly learns two sets of
topics on short and long texts and couples the topic parameters to cope with the
potential inconsistency between data sets.
Ramage et al. present an adaptable execution of a partially “supervised learning
model (Labeled LDA)” [5] that maps the substance of the Twitter channel into
dimensions. These dimensions compare generally to substance, style, status, and
social attributes of posts. The author characterizes users, tweets utilizing this model,
and presents results on two data consumption oriented tasks.
Zhao et al. presents the comparison of the Twitter content and a customary news
medium, New York Times, utilizing unsupervised topic modeling. The author
utilizes a “Twitter-LDA model” [6] to find topics from a representative sample of
the entire Twitter. The author then utilizes content mining strategies to compare
these Twitter topics with themes from New York Times, taking into consideration
topic classes and types. The author has also studied the relation among the extents
of opinionated tweets and retweets and topic classifications.
Wang et al. a recent topic model, “Temporal-LDA” [7] or TM-LDA has been
proposed, for proficiently modeling streams of social content such as a Twitter
stream for a creator, by detecting the topic transitions that arise in these data.
Existing models such as Latent Dirichlet Allocation (LDA) were developed for
static corpora of relatively large documents. However, much of the textual content
on web today, is temporarily sequenced and comes in short fragments, for example,
microblog posts on sites such as Twitter, Weibo, etc.
Pramod Bide et al. an “improved document clustering calculation” [8] is taken it
creates number of clusters for any content reports and to put similar documents in
appropriate groups it uses cosine similarity measures.
Raihana Ferdous et al. proposed a “modified k-means algorithm” [9] which
chooses distinct k documents as centroids for k clusters using a distance similarity
measure. In experiments, it is proved that proposed K-means algorithm with Jaccard
similarity distance measure to compute the centroid gives better results than tra-
ditional K-means algorithm.
3 System Architecture/System Overview
3.1 Problem Statement
The previous algorithms (LDA and its variants) work for long texts (news articles),
however, fail to give efficient results for short texts (news articles) due to data
sparsity problem. Our aim is to develop a novel way for topic modeling in online
news by using two approaches named as, Biterm Topic Model (BTM) and Word
Network Topic Model (WNTM), which give efficient results for long as well as
short news articles, by tackling the sparsity and imbalance problems simultane-
ously. Also compare the results of both approaches in terms of time, accuracy, and
memory. Further, apply Document Clustering algorithm to cluster the news related
to a particular topic from different sources (news websites) together.
3.2 Proposed System Overview
In this paper, a lucid but generic explanation for topic modeling in online news has
been presented. System presents a word co-occurrence network-based model named
WNTM, which works for both long as well as short news articles by managing the
sparsity and imbalance issues simultaneously. Dissimilar from existing methods,
WNTM is modeled by assigning and reassigning (according to probability calcu-
lation) a topic to every word in the document rather than modeling topics for every
document. It has better time and space efficiency, compared to previous approaches.
Also, the context information saved in the word–word space likewise ensures to
detect new and uncommon topics with convincing quality (Fig. 1).
In the proposed system, we retrieve real-time data from various news websites as
an input to the system. The extracted data is forwarded to preprocessing where
stemming and stop word removal takes place. As an output of preprocessing, we get
documents represented as a bag of words which will be provided to the topic
modeling algorithms (BTM and WNTM), in which topic of the news document is
identified. In the case of BTM (Biterm Topic Model), [1] biterms are formed from
the bag of words. These biterms go as input to the Gibbs Sampling algorithm which
infers the topics of the documents.
In the case of WNTM (Word Network Topic Model), it takes lesser time and space
as compared to BTM. This is because the input of BTM is in the form of biterms,
while, input of WNTM is in the form of pseudo documents. In BTM approach, each
time a word is scanned, it forms a new biterm with the word which occurs with it,
irrespective of its previous occurences and then probability is re-evaluated or
updated. However, in case of WNTM, a network or graph is formed containing
nodes, representing words and edges with weights, representing relationships. In this
case, each time a word is scanned, all the words which occurred with it in the entire
document, along with their relationships are mapped or saved in the form of a graph
or network or pseudo document. In this way, less number of tuples or rows are
required, and, WNTM saves the time and space required for execution.
The output of the topic modeling step is given as input to the document clus-
tering step, which groups or clusters the documents coming under specific topic
together. Likewise, all the documents are clustered under corresponding topics.
Further, a summary is generated for the news documents, and, we intend to present
the summarized news along with the respective topics to the user as an output.
Fig. 1 System architecture with preprocessing, Topic modeling, document clustering, and
summarization
Recommendation system: We also intend to create a news recommendation

system, in which according to user preference we would try to prioritize the news
for the user. User preference can be captured by the news which user browses more.
We intend to create a newsbox (kind of a mailbox), in which the news related to
topics which interest the user will be stored, the newsbox will be updated from time
to time, so that whenever user wants to read the news, he/she can directly check the
newsbox, and does not have to browse from all the available news. For example,
suppose a particular user is interested in news related to sports and entertainment,
then all the news associated with or in the field of sports and entertainment
throughout the day would be stored in his/her newsbox, and he/she can directly
check the newsbox, instead of browsing through the entire news.
3.3 Algorithm
Algorithm 1: WNTM Re-weighting Algorithm Required: the original word

network G = (N, E, V), where E is the set of edges and V is the set of weights for
edges.
Ensure: the re-weighted word network G’ = (N, E, V’)
1. compute degree D(n) and activity A(n) of each node n e N
1. for all e = (n1, n2) e E do
2. set Ve = [ve/A(ni)], argmin{D(ni), i = 1, 2}
3. end for
This re-weighting is used to increase the time or space efficiency and improve the
performance of WNTM effectively. After topic modeling step, we will get the set of
topics, T, as an output. These topics will go as input along with the news data to the
Document Clustering step.
Number of topics = Number of clusters
Algorithm 2: Document Clustering Algorithm Input: N: the number of clusters
(number of topics)
Di: Size of documents (i.e. total no. of documents) in a database
Output: A set of n clusters which contain documents.
According to the output of Topic Modeling step, we group or cluster or cate-
gorize the documents under generalized topics such as Politics, Sports, Education,
Entertainment, etc. In our system, instead of centroids, we consider the topics (from
previous step), and calculate similarity from topics to documents. After this step, we
will get clusters of documents, where each cluster consists of documents belonging
to or related to the same topic. Cluster may contain documents from different
newspapers, summarized together.
Let S be a system, Such that:

S = {Input Process, Output}
Input: Online Real-Time Data from News Websites
Output: All News Summarization Related to Specific Topic Together.
Process:
1. Preprocessing
In preprocessing, system performs stop words removing and stemming.
Step 1: Stop words removal
Step 2: Stemming
2. Collection of Words into bags

After preprocessing, preprocessed words are collected into bag of words.
Bag ¼ fw1; w2; . . .; wng
where, Bag is the collection on n number of words (w)

3. Topic Modeling
For topic modeling, BTM or WNTM is used. Common topics are identified in
this step.
T ¼ ft1; t2; . . .:; tng
where T is the set of topics extracted form n number of documents.

4. Document Clustering
In this step, documents are clustered on the basis of detected topics. K-means
clustering algorithm is used for clustering of documents.
DC ¼ fc1; c2; . . .; cng
DC is the set of n number of documents clusters.

Each cluster cn contains similar kind of documents based on relative topic.
Cosine similarity measure is used to measure the similarity between topics and
documents. It is given as:

Sim Ci ; Cj ¼ Ci Cj =jCi jCj ;
where Ci and Cj are the clusters.
4 System Analysis
The system built using Java framework on Windows platform. The Net beans IDE
(version 8.1) are used as a development tool. The system does not require any
specific hardware to run; this application can run on any standard machine.
4.2 Dataset
System extracts real-time news data from news websites (e.g., Indian Express,
Hindustan times). We extract the data from the website links; links containing news
data selected by passing appropriate attributes. On an average, at a time, a news

website contains about 130 news documents, storylines, or articles, which are
extracted. Every news extraction from website is stored in separate documents or
files.
4.3 Performance Measures
The following are the performance measures of topic modeling algorithms (BTM or
WNTM):
Time Efficiency: Amount of time taken for the execution of the algorithm.
Memory: Amount of space of memory used or required for the execution of the
algorithm.
4.4 Result
Table 1 represents the comparison between existing system and proposed system
on the basis of time efficiency and memory efficiency of topic modeling. Proposed
system includes WNTM and Document clustering, it is more efficient than existing
system which includes BTM and Document clustering, it finds out the topic
modeling and document clustering.
Figure 2 shows the time efficiency comparison graph of the proposed system
versus the base system. Time required to identify topics in existing system with
Table 1 Time and accuracy comparison of segmentation time

System Time in ms Memory in bytes
Existing system using BTM and document clustering 1294.7 898517444
Proposed system using WNTM and document clustering 338.2 484382868
Fig. 2 Time efficiency

comparison, X-axis = Topic
modeling with BTM and
WNTM, Y-axis = Time
required in milliseconds
BTM is more than the time required in proposed system with WNTM. Here, we
have shown the average result of about 20 sample cases (20 executions).
Figure 3 represents the memory comparison of BTM and WNTM for topic
modeling in percentage. WNTM outperforms BTM. Here, we have shown the
average result of about 20 sample cases (20 executions).
Figure 4 represents the variation in the results of time comparisons between
WNTM and BTM of the 20 executions or 20 sample cases.
Figure 5 represents the variation in the results of memory comparisons between
WNTM and BTM of the 20 executions or 20 sample cases.
Fig. 3 Memory Comparison,

X-axis = System, Y-
axis = Memory of topic
modeling with BTM and
WNTM
Fig. 4 Time efficiency comparison, X-axis = Topic modeling with BTM and WNTM, Y-
axis = Time required in milliseconds
Fig. 5 Time efficiency comparison, X-axis = Topic modeling with BTM and WNTM,
Y-axis = Time required in milliseconds
Topic modeling in online news is an increasingly essential job because of the

existence of abundant unrecognized news data on the internet. When compared to
normal texts, online short news data texts bring sparsity issue for existing topic
models. In order to overcome these problems, in this paper, a topic model is
proposed for general short texts, namely Word Network Topic Model, and apply it
on online news data. Unlike the previous models, this model efficiently mines or
detects topics for both long as well as short news data texts. WNTM can capture the
topics in news texts by explicitly modeling word co-occurrence patterns in the
entire corpus. Then, document clustering algorithm is used for clustering the news
data related to specific topic together. Next, summarization of the news coming
from various sources takes place, and, we intend to present the summarized news
along with the topic to the user as the output. Further, we intend to create a news
recommendation system according to user preference. In future, we can try to work
with different languages as the news data may contain information in other lan-
guages as well.
References
1. Xueqi Cheng; Xiaohui Yan; Yanyan Lan; Jiafeng Guo. BTM: Topic Modeling over Short
Texts. IEEE Transactions on Knowledge and Data Engineering, Year: 2014, Volume: 26,
Issue: 12 Pages: 2928–2941.
2. M. Sahami and T. Heilman. A web-based kernel function for measuring the similarity of short
text snippets. In Proceedings of the 15th international conference on World Wide Web, pages
377–386. ACM, 2006.
3. X. Phan, L. Nguyen, and S. Horiguchi. Learning to classify short and sparse text & web with
hidden topics from large-scale data collections. In Proceedings of the 17th international
conference on World WideWeb, pages 91–100. ACM, 2008.
4. O. Jin, N. Liu, K. Zhao, Y. Yu, and Q. Yang. Transferring topical knowledge from auxiliary
long texts for short text clustering. In Proceedings of the 20th ACM international conference
on Information and knowledge management, pages 775–784. ACM, 2011.
5. D. Ramage, S. Dumais, and D. Liebling. Characterizing microblogs with topic models. In
International AAAI Conference on Weblogs and Social Media, volume 5, pages 130–137,
2010.
6. W. Zhao, J. Jiang, J. Weng, J. He, E. Lim, H. Yan, and X. Li. Comparing twitter and traditional
media using topic models. Advances in Information Retrieval, pages 338–349, 2011.
7. Y. Wang, E. Agichtein, and M. Benzi. Tm-lda: efficient online modeling of latent topic
transitions in social media. In Proceedings of the 18th ACMSIGKDD, pages 123–131, New
York, NY, USA, 2012.ACM.
8. Pramod Bide, Rajashree Shedge, “Improved Document Clustering using K-means Algorithm”
IEEE, 2015.
9. Mushfeq-Us-Saleheen Shameem, Raihana Ferdous, “An efficient Kmeans Algorithm integrated
with Jaccard Distance Measure for Document Clustering”, pp 1–6, IEEE, 2009.
Spam Mail Detection Using Classification
Techniques and Global Training Set
Vishal Kumar Singh and Shweta Bhardwaj
Abstract Emails are Internet-based services for various purposes like sharing of
data, sending notices, memos, and sharing data. Spam mail are emails that are sent
in bulk to a large number of people simultaneously, while this can be useful for
sending same data to a large number of people for useful purposes, but it is mostly
used for advertising or scam. These spam mails are expensive for the companies
and use a huge amount of resources. They are also inconvenient to the user as spam
uses a lot of inbox space and makes it difficult to find useful and important emails
when needed. To counter this problem, many solutions have come into effect, but
the spammers are way ahead to find these solutions. This paper aims at discussing
these solutions and identifies the strengths and shortcomings. It also covers a
solution to these spam emails by combining classification techniques with knowl-
edge engineering to get better spam filtering. It discusses classification techniques
like Naïve Bayes, SVM, k-NN, and Artificial Neural Network and their respective
dependencies on the training set. In the end of this paper, the global training set is
mentioned which is a way to optimize these training sets and an algorithm has been
proposed for the same.

Keywords Naïve Bayes k-nearest neighbor Support vector machine classifier

method Artificial Neural Network Knowledge engineering Global training set
V. K. Singh (&) S. Bhardwaj

Computer Science and Engineering Department, Amity University,
Noida, Uttar Pradesh, India
e-mail: vs48023@gmail.com
S. Bhardwaj
e-mail: shwetabhardwaj84@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_61
624 V. K. Singh and S. Bhardwaj
1 Introduction
Classification is based on the concept of data mining. Data mining is gathering

meaningful data from a pool of information which is built using logical training sets
and observation in related fields [1]. This helps in building a better system as we
already have all the information of previous failures, success, and current problems
[2]. We take the data from this pool, sort it according to our needs and then select
the data which will help us in our current research. Data mining comprises of
various methods of research and learning like artificial intelligence or machine
learning. It is used in almost every fields like teaching, research, medical, defense,
etc., to optimize the results.
If a spam mail is controlled, it can save a lot of resources and expenditure of the
companies. Previously, a method called “Knowledge Engineering” was used to
separate spam mail from important emails, but its success rate was not as much as
that of spammers as they were able to find a way around it by changing a letter in
the keywords which is used to differentiate spam mail from legitimate mail [3]. To
overcome this, various methods for segregating spam mail using classification was
introduced. These methods used machine learning, artificial intelligence, and var-
ious databases to develop a system to counter spam mail. These methods were more
successful than knowledge bank as the data was updated more frequently, and were
capable of learning on their own and had better algorithms than knowledge banking
[4]. This paper will elaborate these techniques and compare them with previously
used techniques, thus suggesting changes that can be made to spam mail filtering.
1.1 Why We Need Spam Filtering and Advantages

of Classification Techniques for Spam Filtering
Why we need spam filtering Advantages of classification techniques for spam

filtering
• Spam mail use lots of resources • They are more efficient than knowledge engineering
• They contribute a lot to network • Reduces consumer effort in spam filtering
traffic • Adaptive to change on their own, save time and
• They take inbox space which is a other resources required for the spam filter
nuisance
• It becomes hard to find useful mails
among these spam emails
• Service provider suffers loss
because of these spam emails
Spam Mail Detection Using Classification Techniques and Global … 625
1.2 Type of Spam Mail and Their Composition
Type of spam mail Percentage (%)

Leisure/Travel 2
Education 1
Business/Investment 18
Adult 18
Finance 17
Product/Services 19
Health 9
Computer services 7
Others 9
See Fig. 1.
2 Related Works
Mails were introduced to make communication easier and faster. With time its use
for wrong purposes also increased and one of the major problems is sending bulk
emails for advertising or other such purposes [5]. These bulk emails commonly
known as “spam mail” were becoming a huge problem and had to be dealt with.
First, it was dealt with the users manually by classifying or marking mail as spam
mail and later when such similar mail came they were moved to spam folder. The
biggest reason for the rise of spam mail is that it is faster and easier to use mail for
advertising worldwide rather than using posters or TV advertisement for same [6].
Composition of spam mail
Leisure/ Travel EducaƟon Business/ Investment

Adult Finance Product/Services
Health Computer services Others
Fig. 1 Composition of type of spam mail and their composition

They provide a cheaper solution. According to a survey, several billion emails are
exchange in a year and out of which a good amount approximately 50% turns out to
be spam. But soon counters to this were found by spammers and new techniques
had to be developed. In 1900s, classification techniques were introduced for spam
filtering. These techniques were much better than the previous techniques as they
were faster and more accurate in spam mail filtering. They could handle a large
amount of mails, but the solution to these was also found [7]. The United State of
America (USA) was the first country to take legal action against the spammers and
passed a law for the same. Since then, several arrests have been made and many
spammers were caught and punished for their crime. Some were sentenced to jail
while some spammers were asked to compensate the ISP for the losses. These
actions were a little successful in bringing down the number of spam mail but the
problem still persists.
This constant battle between spammers and email service providers is still going
on. Each comes up with a solution better than other [6]. But these classification
techniques are very efficient and keep spammers at bay. New algorithm which uses
classification was introduced to counter this problem. Each technique has its own
advantages and disadvantages. Some of the techniques which were introduced are
Naïve Bayes, SVM, Decision Tree, Neural Network, k-NN, etc. There are still
being developed to be more effective and accurate.
3 Problem Statement
Whenever the email service provider comes up with a solution for spam filtering,
the spammers come up with a way to get around that filter and it is a constant
competition between these two. This is very costly for the companies, a huge
amount of resources being used, and create a nuisance for the customer by filling up
the inbox, wasting their time as it becomes hard to find useful mail among those
spam emails. We need to find a way which is very effective against these spammers
and can evolve with the attacks, which the spammer adapts. The training set which
is used to train these techniques needs to be focused on more success.
4 Implementation
This paper suggests that we use both machine learning and data mining techniques
along with previously used knowledge engineering approach to create a filter which
is more robust and efficient than machine learning and knowledge bank techniques
alone and make this learning technique global. Some providers provide this service
while others either provide only knowledge-based spam filter or machine learning
and are not very efficient in combining these two for a better filter.
5 Knowledge Engineering
In Knowledge Engineering, user input creates a set of rules for the incoming emails
to be classified as a spam mail or uses binary value where 1 stands for spam and 0
stands for not spam. In this, the users define a specific email id and any incoming
emails from this email id were automatically moved to the spam folder. They can
also define text words, which can be used as a classification of spam mail. If the
incoming mails contain these keywords, they were moved to spam filter [3].
Although this technique was successful for some time, the spammer found a way
around it by changing the text. This was done by adding or omitting one or two
alphabets in the email id or the main message and thus passing the filter settings set
by the user.
6 Classification Techniques
Here, we discuss some of the classification techniques with their accuracy and
precision obtained by passing a set of test spam data through the classifier. It can be
depicted like as follows:
Training spam dataset Classifier Performance Analysis
6.1 Naïve Bayesian Classifier
It was one of the first classification techniques used for spam mail filtering. It works
on Bayes Theorem of probability to check if the incoming mail is a spam or not.
The filter in this classifier first has to be trained to check for spam mail. Training a
data set means that the filter is given a set of words the user provides by manually
identifying the mail as spam or not. Through the input from a user, the classifier is
now trained and can check incoming emails. This classifier checks the probability
of the words in the training set in incoming messages and with the result obtained it
can filter spam mail. It creates a different folder for the spam and moves those
emails directly to the folder. Although it is pretty old, it is still preferred over more
sophisticated classifiers [8, 9].
6.2 k-Nearest Neighbor
In this technique of classification, a training set contains a sample of messages

which can help in identifying whether the incoming mail is spam or not. The
incoming mail is compared with the training set to find its k-nearest member found
by comparing it with the training set and its k most similar document is found and
then identified as spam, based on which group its k most similar document was
found [10].
6.3 Support Vector Machine Classifier Method
In this method, a decision plane is formed to separate spam and legit emails. They
are separated by a decision boundary which has certain conditions to separate these
emails. A training set is formed for the sorting and incoming mail is compared with
the training set. Like k-nearest, neighbor the incoming mail is compared with the
training set to find similarity between the incoming mail and training set. A kernel
function, K is used to determine the similarity and based on this the emails are
sorted in decision plane [11].
6.4 Artificial Neural Network
Artificial Neural Network or commonly known as neural network acts like an

artificial human brain. Artificial neurons are interconnected to form this network
and data is passed through it for learning. Like the human brain, it learns by
example and during training, the data is passed through the network so that it can
learn and adapt according to the examples [12]. They change their structure based
on information from the examples so that a better system of classification can be
formed [13] (Fig. 2 and Table 1).
These comparisons may vary if the number of selected feature is changed, as in
classification techniques each technique is dependent on how it is trained [14].
Among the discussed technique Naïve Bayes is the simplest technique which can be
implemented and is comparatively faster too.
These were some of the classifier-based techniques for spam mail classification
which used a training set for detection of spam mail. All techniques have a success
rate of around 90%, neural network gives the best result. These techniques are
Spam Precision (%)
100
95
90
85
80
Naïve Bayes k-NN SVM Neural Network
ClassificaƟon Techniques
Fig. 2 Graphical representation of classification techniques and their precision

Table 1 Tabular comparison Classification technique Spam precision (%)

of classification techniques
Naïve Bayes 96
k-nearest neighbor 88
Support vector machine 91
Neural network 94
dependent on the training set and examples and how they can be optimized so that
the techniques can learn better and can give better results [15]. If we can combine
knowledge engineering technique with these classification techniques, we can
optimize it further.
7 Proposed Work
Data is collected from each user training set and then compared to find which mail
is characterized as spam mail and then a global training set based on these inputs
can be created to optimize the classification technique. Each technique is dependent
on the training set for setting the criteria for spam mail identification. The user can
add words to the training set of their personal devices or email accounts to extend
the global training set and get better results. The service provider provides the user
an option to choose the global training set or not. This global set can arrange the
mails in the training according to their frequency. The spam mail which is sent to a
large number of customers will be included in this training set while the spam mails
which have a low-frequency rate, meaning the mails which are sent only to a few
number of customers can be optional for the training set. The service provider can
add a sensitivity controller to this training set with high sensitivity having all the
spam mails for training the global training set irrespective of its frequency. Low
sensitivity will only have spam mail in the training set with high frequency, thus
making it easier for users to classify mail for the training set.
Proposed Algorithm
Keywords
Hf—high frequency
Mf—medium frequency
Lf—low frequency
Ef—email frequency
Gtb—global training set
For service provider
Create global training set ! Gtb

{Set threshold for high, medium, and low frequency;
Get frequency >> hf, mf, lf

If (ef >= hf)
{Save email feature in all frequency training set}
else if (ef < hf && ef >= mf)
{Save email feature in medium- and low-frequency training set}
else
{Save email feature in low-frequency training set}
}
For the user
Provide option for global training set to user

{
If (Gtb == 1)
{Provide option between high-, medium-, and low-frequency training data set
if (user chooses high frequency training data set)
{(Combine local training set with high-frequency global training set)
else if (user chooses medium training data set)
(Combine local training set with medium-frequency training data set)
else
(Combine local training set with the low-frequency data set)}
8 Result
This algorithm is supposed to enhance and make the training set better for the
classification techniques. With the increase in a number of data for training set
which was submitted by users, it includes most of the keywords and phrases of
recent spam mail as well. This helps in classification of spam mail better and
increase the efficiency of the classification techniques. They can increase the pre-
cision of each technique by at least 2% and this can vary from technique to tech-
nique. A comparison of spam mail technique before using global training set and
after using global training set is shown below which depicts the expected efficiency
of proposed algorithm (Table 2).
Table 2 Comparison of precision of classification technique before and after using proposed
algorithm for global training set
Classification technique Spam mail precision before Expected spam mail precision
using algorithm for global after using algorithm for global
training set (%) training set (%)
Naïve Bayes 96 97–98
k-nearest neighbor 88 91–94
Support vector machine 91 93–95
Neural network 94 95–97
9 Conclusion
All the classification techniques have to be trained first in separating spam emails
from other emails before they are actually used. A data set called training set is used
to train these techniques. Thousands of samples are used in these training set to
make the classifier able to separate the spam mail. But even after this much work
spam mail still persists. They persist because every day a new kind of spam mail is
introduced. Thus, even if we get old spam mail sorted and marked, new one keep
coming in. One of the solutions is to make the training set up-to-date by gathering
information about the new kind of spam mail. The fastest way to do that is to make
the user report the spam mail as soon as they encounter it and contribute to the
global training set because it will take time if the service provided has to monitor
each and every mailbox on their own to search for any new spam mail. This
algorithm is expected to raise the efficiency of other techniques by some margin
depending on the technique. If it is successful in doing so we will have the spam
mail dealt with before it reaches our mailbox. This will also save our time and inbox
will be less crowded thus making it easier to find useful emails.
References
1. Parhat Parveen, Prof. Gambhir Halse, “Spam mail detection using classification”,
International Journal of Advanced Research in Computer and Communication Engineering
Vol. 5, Issue 6, June 2016.
2. Megha Rathi, Vikas Pareek, “Spam mail detection through data mining-A comparative
performance analysis”, I.J. Modern Education and Computer Science, May 2013.
3. W. A. Awad, S. M. ELseuofi, “Machine learning method for spam email classification”,
International Journal of Computer Science & Information Technology, Vol 3, No 1, Feb
2011.
4. R. Malarvizhi, K. Saraswathi, Research scholar, PG & Research, Department of Computer
Science, Government Arts College, “Content-Based Spam Filtering and Detection
Algorithms- An Efficient Analysis & Comparison”, International Journal of Engineering
Trends and Technology, Volume 4, Issue 9, Sep 2013.
5. Vinod Patidar, Divakar Singh, “A Survey on Machine Learning Methods in Spam Filtering”,
International Journal of Advanced Research in Computer Science and Software Engineering,
Volume 3, Issue 10, October 2013.
6. Rekha, Sandeep Negi, “A review of spam mail detection approaches”, International Journal of
Engineering Trends and Technology (IJETT) – Volume 11 Number 6 - May 2014.
7. S. Roy, A. Patra, S. Sau, K. Mandal, S. Kunar, “An Efficient Spam Filtering Techniques for
Email Account”, 1,2,3Computer Science & Technology & 4Mechanical Engineering, NITTTR,
Kolkata, American Journal of Engineering Research, Volume 02, Issue 10, pp-63–73.
8. Tianhao Sun, “Spam Filter Based on Naïve Bayes Classification”, May 2009.
9. Rasim M. Alguliev, Ramiz M. Aliguliyev, and Saadat A. Nazirova, “Classification of Textual
E-Mail Spam Using Data Mining Techniques”, Institute of Information Technology of
Azerbain National Academy of Sciences, September 2011.
10. Konstantin Tretyakov, “Machine Learning Techniques in Spam Filtering”, Institute of
Computer Science, University Of Tartu, May 2004.
11. Akshay Iyer, Akanksha Pandey, Dipti Pamnani, Karmanya Pathak and Prof. Mrs. Jayshree
Hajgude, “Email Filtering and Analysis Using Classification Algorithms”, IT Dept, VESIT,
International Journal of Computer Science Issues, Vol. 11, Issue 4, No 1, July 2014.
12. Vinod Patidar, Divakar Singh, Anju Singh, “A Novel Technique of Email Classification for
Spam Detection”, International Journal of Applied Information Systems, Volume 5 – No. 10,
August 2013.
13. Abhale Babasaheb Annasaheb, Vijay Kumar Verma, “Data Mining Classification Techniques:
A Recent Survey”, International Journal of Emerging Technologies in Engineering Research
(IJETER) Volume 4, Issue 8, August (2016).
14. C. Neelavathi1, Dr. S. M. Jagatheesan, Research Scholar, Department of Computer Science,
Gobi Arts & Science College, T.N., India, “Improving Spam Mail Filtering Using
Classification Algorithms With Partition Membership Filter”, Issue: 01, Jan 2016.
15. Ahmed Obied, “Bayesian Spam Filtering”, Department of Computer Science, University of
Calgary.
Smart Gesture Control for Home
Automation Using Internet of Things
Sunil Kumar Khatri, Govind Sharma, Prashant Johri

and Sachit Mohan
Abstract Internet of Things (IoT) is a system where the machines involved in a

system or an infrastructure can be monitored and certain activities can be controlled
with the help of sensors and actuators. This research presents the model for con-
trolling remote devices with the help of hand gestures. The gestures are recognized
using template matching algorithms to identify the hand gestures and control
remote things accordingly using the Internet of Things. The microcontrollers and
the remote units which are involved in the architecture are connected to the Internet
either via LAN or Wi-Fi module. The proposed system will help those who gen-
erally forget to switch off the power when not in use, as an example. It is a
contribution to the body of knowledge in the field of home automation using a
microcontroller.

Keywords IoT Embedded microcontroller Gesture control Z-wave protocol
Template matching
S. K. Khatri (&) G. Sharma

Amity Institute of Information Technology, Amity University Uttar Pradesh,
Noida, UP, India
e-mail: skkhatri@amity.edu
G. Sharma
e-mail: govindsharma00709@gmail.com
P. Johri
Galgotias University, Greater Noida, India
e-mail: johri.prashant@gmail.com
S. Mohan
InventingThoughts Technologies, Ghaziabad, India
e-mail: er.sachit@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_62
634 S. K. Khatri et al.
1 Introduction
Internet of Things, more commonly referred to as IOT, is a system where we have a

number of devices connected to the Internet that is made capable of interacting with
each other through the use of sensors and actuators. The complete functioning of
the devices can be controlled and monitored. The growth of the IoT has an influ-
ence on everyone and everything and will bring a great difference to lifestyle and
environment in the next decades [1].
There have been a number of hand gesture-based systems using IoT that have
been proposed by researchers globally. In our model, we have proposed the use of a
camera to recognize the gestures. The previous methods have seen the use of
accelerometers, which were fixed on the hand. This model is highly accurate but
then cannot be used for controlling home appliances [2].
The use of gestures has emerged to be one of the most emerging ways to control
devices. Wearable devices are the considered the best tool to carry out this
practice [3].
Hand gestures tagged along with visual display is a popular choice in predicting
sign language [4]. When it comes to the same concept for appliances, it is different
when compared to the one for sign language. Controlling appliances has become a
very complex and challenging task to execute.
Recently, hand gesture systems are becoming widely used in many products.
Z-wave is used as a protocol in hand gesture system. It is a more effective, reliable
mode of communication for the flow of small data packets. In this research paper,
the main aim is to develop a hand gesture system with Z-wave protocol in order to
make it cost effective and efficient in its working [5].
2 Literature Review
A real-time system developed by Madhuri and Kumar [6] was able to recognize
different gestures made by an individual through a web camera. They used an
algorithm called Haar Classifier. To get the most accurate data, a large amount of
data and also images were required. The whole process took about a week to
complete. This algorithm was deemed to be more flexible and applicable. The
system had the ability to capture and track the gestures with high accuracy. The
drawback to this system was that the distance between the camera and the user
could not be too much.
Choondal and Sharavanabhavan [7] came up with a system which was for the
natural user interface. This system depended upon how you place your fingers, and
it kept track of the same by locating it frequently. Color markers were placed on
different fingers. The whole system was dependant on the color recognition. The
restriction to the system was when the same color appeared in the background too
which would then end up confusing the system.
Smart Gesture Control for Home Automation Using Internet of … 635
3 Analysis
During the initial stages of home automation, the whole system required the
installation and updates to be done by a technician. The manual work that had to be
put in during the initial stages was very high. As time passed by, the manual work
got replaced by installation through Wi-Fi. A couple of examples would be Z-wave
and ZigBee [4] These are wireless technologies that are used for monitoring and
control. Even though their key features are similar, they differ in their specification
and application. The use of such technologies is becoming more widespread as its
ideal for home automating systems.
There are few aspects to differentiate between ZigBee and z-wave for home
automation system using gesture control (Table 1).
4 Gesture Recognition
Gesture recognition refers to the interpretation of hand signals using a particular

algorithm. It is something that is looked into as having the potential for the future
market. The gestures that are interpreted are mainly from movements caused by the
body, more commonly face or the hand. The person making gestures would be able
Table 1 Features of Z-wave and ZigBee [9]

Z-wave ZigBee
1. Compatibility 375 organizations and 1500 items A wide range of organizations
comprise the Z-wave Alliance make gadgets that speak with
ZigBee’s innovation
2. Security Each Z-wave arrange, and the ZigBee uses AES-128 encryption
gadgets inside each system, are and make use of keys to secure it
appointed extraordinary IDs that
speak to your center point
3. Drawback Z-wave items commonly go from ZigBee is mainly renowned for its
$40 to $100 every, which can use in the retail and utility business.
include rapidly on the off chance ZigBee was created for use in the
that you need to associate utility and retail businesses
numerous gadgets
4. Advantage Z-wave is easy to understand and The main plus point to it is that it
gives a straightforward framework requires very less battery and so the
that clients can set up themselves gadgets can last up to 7 years on a
single battery
5. Uses Z-wave is for an essential ZigBee is ideal for the innovation
comprehension of innovation that master who needs a framework
keeps home mechanization secure, they can tweak with their
proficient, easy to utilize, and inclinations and introduce
simple to keep up themselves
to control the whole system to which he is interacting with just by body move-
ments. It allows an individual to interact with a system without the involvement of a
mechanical instrument to connect the two. Human input-based interfaces can be
developed that has great potential. This system, would propose a model for gesture
recognition to control the various home appliances put in place. With the use of
IOT, we will be able to monitor the performance of the device and also its
efficiency.
1. Gestures for the templates
Few gestures which will be stored in the database for operating android devices
using template matching technique are shown in the Fig. 1.
2. Template Matching
In the process of template matching, a template of the image is compared with the
actual image to find the match. After receiving the input image, this method
compares the received image with the template stored in the database. Template
matching is generally done based on pixel or feature. Templates are frequently used
for recognition of characters, numbers, objects, etc. The searching for the template
depends upon the size of the image. More the pixels, more time it will take for
appropriate template match.
Fig. 1 Gestures depicting functions [7]

5 Proposed Work
5.1 Methodology Adopted
In order to apply the smart gesture control for home automation with IOT devices, a
model is constructed as shown in the Fig. 2. The method that has been used is a
pattern matching process, where in templates are set to understand the different
gestures portrayed by an individual. The gestures are captured at 30 Frames per
Second (fps) and are compared with the templates to derive a meaningful output.
The end user can set up his\her own personalized gestures as a template to perform
a particular action, but this can only be done during the initial stages when the
system is being set up. Microcontrollers are responsible for handling the functional
elements, and are developed using various programming languages. The model uses
following devices to function:
Camera—Camera is situated on every android appliance so that it will easily
capture our hand gesture and work as an input device to our gadget.
Microcontroller—The input received will be processed on the microcontroller,
where each gesture is programmed with embedded c language using a particular
function.
Templates—Templates are basically predefined gestures that will be used by the
template matching algorithm.
Z-wave Transmitter—This is the IoT device that uses AES-128 encryption to
send inputs over Wi-Fi. It has 900 MHz frequency to transmit data.
Z-wave Receiver—It receives the input from the z-wave transmitter process (and
uses) the received input to perform a particular action for each device.
Fig. 2 Proposed operation

flow
Camera Microcontroller
Zwave
Templates
Transmitter
Zwave Receiver Devices

5.2 Proposed Model
For the application of gesture recognition in smart homes using IOT protocols, a
prototype needs to be built. Users make various gestures to perform different
actions [8]. To detect these gestures camera(s) are installed, it is used on the top of
each Android device so that the input is taken from the user. The microcontroller is
used to process the input that has been captured by the camera. The microcon-
trollers can be programmed using Embedded C. After the microcontroller has
processed the input, the templates are used to map it to the in-built gestures, i.e.,
gesture’s already present in the database. In Fig. 3, templates are basically the
pre-set understanding of each action that the user performs. Each time a person
performs an action (which is a template), it makes the system generate the desired
output. If the gesture provided by the user is not present in the system or is
unrecognizable, a beep sound can be heard, helping the user understand that he\she
has provided a wrong gesture.
There are a few aspects that the user has to take care of while performing the
gestures. The color pattern should be easily distinguishable from the system, which
means that background color should not be the same as the skin color of the user.
The gestures can only be detected in the presence of a certain amount of light, or
else the system would not be able to capture the image. For example, gestures
carried out in a dark room cannot be recorded by the system. Another key aspect
that the user has to take care of is the speed at which the gesture is performed. If it is
performed at a high speed, the camera would not be able to understand what action
Mapping Compare
Packages Reading Perform
of received
Start are template Action Stop
received image to
imported images on device
image template
Fig. 3 Flow chart of proposed model

the user is trying to perform. Taking these points into consideration would help the
user perform the various actions successfully.
The technique of finding a particular figure on a still image is called active shape
models. Each template is stored with a number of frames in the database, on each
frame. Active shape model is used to identify the figure as an initial approximation
for the next frame. The number of templates to be used depends upon the number of
gestures that is decided to perform on each device. As nine number of gestures are
defined to use in the system and on the basis of the proposed model a certain time is
shown for each gesture in Table 2.
The gestures are performed by a certain frame per second to analyze the per-
formance time. The average time for performing these gestures is 1.34 s as shown
in the graph Fig. 4, which is preferable time for performing such gesture in the
system and also it is better than any other gesture recognition algorithms. To
calculate the timing of each gesture, we conducted 15 trails to perform same gesture
and took an average time of it.
Each gesture has its own priority for the availability and this priority is calcu-
lated according to the number of users for each device. As per every device has its
On and Off state, thus its priority is high as it should be available for the users, that
means high priority gesture should get recognized at any case for each device.
The quality of image should also be appropriate and should not below 480
pixels. For storing templates database is created which include 189 images, i.e., 21
images for each gesture (9). Each part of the template showing particular gesture is
called as a model thus there are total 9 models have been proposed to store in the
database. Each model is processed and compared with the gathered data and
matching template is selected to perform the action on the device. Preprocessing of
the data is needed before performing any gesture to the device. Preprocessing of
data will remove noise, unwanted errors, make data effective and reliable.
Table 2 Number of templates performing gesture [10]

S. Name of the Time of performing the gesture Availability (High/Med/
No. gesture (Sec) Low)
1 Pause/Stop 1.49 High
2 Help 1.42 Medium
3 On 1.25 High
4 Off 1.55 High
5 Sleep 1.22 Low
6 Go/Play 1.40 High
7 Down 1.30 Medium
8 Up 1.19 Medium
9 Next 1.28 Medium
Fig. 4 Analysis of performance
6 Conclusion
This paper presents an effective and efficient home automation system, where
devices can be controlled by hand gesture. It holds all the promises and has shown
great signs to play a key role in home automation. The system is based on RF
communication technology that shows low latency in forwarding data packets
through devices. It enables the system to control up to 232 devices locally or
remotely. Z-wave protocol of IoT is widely used for home automation. A couple of
key advantages of this protocol are that it has low power consumption and supports
mesh network. The work can be extended to control more machines by including
particular Internet-empowered appliances station. This work proposed an approach
to recognize nine kinds of gesture by counting the number of fingers. Each gesture
is related to a particular task that will be performed on the device.
Acknowledgements The authors express their deep sense of gratitude to The Founder President
of Amity Group, Dr. Ashok K. Chauhan, for his keen interest in promoting research in Amity
University and have always been an inspiration for achieving great heights. I would also like to
thank Mr. Sachin Mohan, who is IoT Architect at Inventingthoughts Technologies for providing
vital information regarding various Android appliances. Mr. Sachit Mohan has provided his
consent for the paper for which I am highly obliged. Inventingthoughts Technologies has been of
great help in my research paper providing the base for my paper contents. Mr. Mohan’s knowledge
in the field of Android devices and their working has been of great help.
References
1. M A Rashid, X. Han, “Gesture control of ZigBee connected smart home Internet of Things”,
2016 5th International Conference on Informatics, Electronics and Vision (ICIEV), pp. 667–
670.
2. J. Anderson and L. Rainie, “The Internet of Things Will Thrive by 2025”, 2014 Pew Research
Center. Available at: http://www.pewinternet.org/2014/05/14/internet-of-things/.
3. G. R. S. Murthy, and R. S. Jadon, “A Review of Vision Based Hand Gestures Recognition”,
International Journal of Information Technology and Knowledge Management,
July-December 2009, pp. 405–410.
4. Shah J., Darzi A., “Simulation and Skills Assessment”, IEEE Proc. Int. Workshop on Medical
Imaging and Augmented Reality 2001, pp. 5–9.
5. Jun-Ki Min; Bongwhan Choe; Sung-Bae Cho, “A selective template matching algorithm for
short and intuitive gesture UI of accelerometer-builtin mobile phones” 2010 Second World
Congress on Nature and Biologically Inspired Computing, pp. 660–665.
6. Assad Mohammed, Amar Seeam, Xavier Bellekens, “Gesture based IoT Light Control for
Smart Clothing”, 2016 Emerging Technologies and Innovative Business Practices for the
Transformation of Societies (EmergiTech), IEEE International Conference, pp. 139–142.
7. J. J. Choondal and C. Sharavanabhavan, “Design and Implementation of a Natural User
Interface Using Hand Gesture Recognition Method”, 2013 International Journal of Innovative
Technology and Exploring Engineering, vol. 2, pp. 21–22.
8. K. Madhuri and L. P. Kumar, “Cursor Movements Controlled By Real Time Hand Gestures,”
2013 International Journal of Science and Research, vol.2, issue 2, pp. 4–5.
9. ZigBee Vs Zwave, Available: http://www.safewise.com/blog/zigbee-vs-zwave-review/.
10. Application of Template Matching Algorithm for Dynamic Gesture Recognition, Available:
https://www.youtube.com/watch?v=fbeLUfTED00.
A Proposed Maturity Model for Himachal
Pradesh Government e-Services
Alpana Kakkar, Seema Rawat, Piyush Gupta

and Sunil Kumar Khatri
Abstract With new leaps in the advancement of technology on a daily basis, it

becomes quite essential that basic and necessary governmental services are pro-
vided through the internet as e-services. These are small steps to fulfil the dream of
a “digitalized India”. This paper is an aim to provide a maturity model for the
existing e-services of Himachal Pradesh (HP) government. Its central focus is on
how to make the services easily accessible for the masses and improve their
functionality by giving a detailed and sophisticated maturity model. The paper also
lists the major concerns of information security in the e-service portal of Himachal
Pradesh (HP) government and an attempt has been made to provide a solution for
the same. The proposed solution herewith will help the government to improve the
level of its e-services and the security of the portal.
Keywords Information security Maturity model e-services HP government
1 Introduction
Technology, today, has become the part and parcel of each individual’s life. Most
of the governments have gone online or are in the process of going online, thus
providing the public services to its citizens through the Internet [1]. Both the state
and central governments, in India, have taken the e-governance seriously and have
been continuously endeavouring to provide citizen services in a better manner [2].
A. Kakkar (&) S. Rawat P. Gupta S. K. Khatri

Amity Institute of Information Technology, Amity University, Sec-125, Noida, UP, India
e-mail: akakkar@amity.edu
S. Rawat
e-mail: srawat@amity.edu
P. Gupta
e-mail: piyushhgg23@gmail.com
S. K. Khatri
e-mail: skkhatri@amity.edu

https://doi.org/10.1007/978-981-10-7245-1_63
644 A. Kakkar et al.
Fig. 1 Protection layers of e-services
Himachal Pradesh (HP) government have currently 16 existing e-services on their

portal and this paper focuses on five major services of the government and suggests
some improvements and changes that can be incorporated using a self-developed
maturity model. These changes will make the experience of the user a simple and a
straightforward process. The e-services currently offered by the Himachal Pradesh
(HP) government are effort centric rather than result centric [3].
The information security maturity model offers a new approach to safeguard the
individual’s information over the Internet by improving the existing e-portal for the
services of Himachal Pradesh (HP) government. The major lapse behind the
security issue being that the Himachal Pradesh (HP) government site is on hypertext
transfer protocol (http) rather being on hypertext transfer protocol secure (https),
which allows any hacker an easy access to the sensitive information of the state
government and its citizens. This issue needs to be considered by the state gov-
ernment IT department as a priority. To help resolve this issue, a multi-level
authentication portal can be created for e-services, which has been described under
information security model section.
Figure 1 displays the protection layers of e-services. The maturity level of
e-services for an e-governance portal and proposed information security model is
created and described in Sects. 4 and 5 respectively.
The e-services of Himachal Pradesh (HP) government included in the present
paper are the ones that are providing essential services over the Internet to its
residents. The e-services are (Fig. 2) as follows.
All these e-services need major improvements which can be achieved by the
proposed maturity level in the Sect. 4.
2 Literature Survey
Governments have been slow in realizing the potential benefits of information

technology and rather snail paced in providing e-services [4]. e-services are
cost-effective services, which can drive the growth of economy and government
productivity to new heights in brief period of time. Out of the existing 16 e-services
of Himachal Pradesh (HP) government, the 5 e-services mentioned in the paper are
the ones which provide essentials services to Himachal Pradesh (HP) citizens. It has
A Proposed Maturity Model for Himachal Pradesh Government … 645
Fig. 2 e-services of HP government
been seen that since the emergence of e-services, the cost of data storage and
hardware costs have also taken an upward ride. Fortunately, our technologists have
discovered a solution to this as cloud computing. Cloud computing provides a new
service consumption and delivery model inspired by consumer internet services.
Cloud computing cuts down costs and increases the cost benefits [5]. Himachal
Pradesh (HP) government is using a government initiative MeghRaj cloud for their
e-services.
The existing e-services are currently not known to the citizens of the rural areas
of the Himachal Pradesh (HP) state in comparison to the urban areas of the state. To
do this an awareness campaign is needed to initiate by the government to educate
the population of the remote areas of the state. There is also a major concern over
trust, protection and safety of information which demands a high level of security
within e-government organizations [6]. For this the state government must update
its most critical security issue of not being a hypertext transfer protocol server
(https) site, this is highly essential to popularize the use of e-services as it upgrades
the security level of the portal to new level. The paper has the following sections
which describe the work done and solutions proposed:
• Methodology Adopted
• Maturity model
• Information security model
The e-services portal is a great initiative by the Government of India and Himachal
Pradesh government is concerned with making the initiative a success [7].
However, in Himachal Pradesh, there are challenges faced by government in
effectively delivering the e-services to its citizens. These are majorly reaching out to
its citizens and making them aware of the existence of these services. Challenges
like this and others have been explained in different levels of the proposed model.
3 Methodology Adopted
• A set of questionnaires were created for NIC engineers working at the head-
quarters of IT department of Shimla, Himachal Pradesh (HP) to gather the
information about how the e-services work and how well are they deployed.
• The questionnaires were instrumental in defining the working of IT department
and their role in the maintenance of e-services of Himachal Pradesh
(HP) government.
• NIC engineers, of IT department Himachal Pradesh (HP) government answered
the questions effectively and imparted the information helping in the formation
of the maturity model of the e-services.
These answers help to understand a lot about e-services and their working and
maintenance over the internet. The government is using data centres to host all
government applications and is also using a cloud service provider known as
MeghRaj cloud to provide its services. The sites are using VPN to update the
content and secure the information. The operating system used by Himachal
Pradesh (HP) government for hosting their sites on the cloud is Red Hat Linux
system. The technology currently in use by the government is MS platform and is
working on other platforms such as Android, iOS, and windows applications. The
sites’ data is maintained and updated by NIC developer for any site and application.
NIC has designated nodal officers for the purpose to keep the content up to date
using CMS (Content Management System) available over VPN (Virtual Private
Network).
The questionnaires answer helped me gain an understanding as to where the
e-portal of Himachal Pradesh (HP) government can be enhanced. This made me
develop a maturity model for the e-services portal listing the enhancement that can
be made to the major e-services of the government. The model proposed is generic,
i.e. it can be used to enhance any government e-services’ portal which may need
improvements.
4 Proposed Maturity Model
The maturity of e-services depends upon what kind of maturity level is used to
describe their processes [8]. A maturity model is a method for judging the maturity
of the processes of an organization and to identify the key practices that are required
to increase the maturity of these processes.
A maturity model guides a government in selecting process improvement
strategies by determining current process capability and its maturity and identifying
the few issues that are most critical to e-governance quality and process
improvement [9].
The paper proposes a new maturity model, first one for e-services of Himachal
Pradesh (HP) government. It can provide the basis for developing a better e-services
portal for the citizens to have access. Along with the proposed maturity model, the
current level of e-services has also been presented.
Figure 3 describes the proposed levels for maturity model of e-services from
which any state government can benefit to provide better and effective services to
its citizens. These levels are explained as follows:
Level 1
• Web Presence: The websites for the e-services must be up and present on the
Internet, it can be static or dynamic website.
• Accessibility: The e-service site and the services provided by site’s data should
be accessible 2 7.
Level 2
• Usability: The sites of e-services must be easy to use and user friendly.
• Dynamic: Site of services and their database should be updated regularly by an
administrator.
• Self-Service and Financial Transaction: The e-services sites must have
self-services with the possibility of electronic payments.
• Transparency: The information processing and data transparency should be
there on the sites. Every recent update related to the service should be uploaded
to the site.
Level 3
• Customer Support: The e-services sites need to provide a good customer
support to attend queries of its citizens.
Fig. 3 Proposed levels of maturity

• Knowledge Management: A tab providing all the information related to the site
and FAQs where all the questions should be answered related to all the sites.
• Integrity: Involves maintaining consistent accuracy and trustworthiness that
data will not be misused over an entire life cycle. Data must not change in
transit. Steps must be ensured that unauthorized people do not get access to the
data.
• Awareness: The citizens must be made aware of the e-services and its content
they have access to
• Language: The e-services are for all the sections of citizens, therefore it is
mandatory that every e-service must be in English and Hindi.
Level 4
• Visualisations: They can be added to e-services to make the interaction with site
simpler and impressive
• Security: The e-services portal needs to provide great security features to keep
the content safe and secure.
• Digital Democracy: This maturity level specifies the option of online voting,
public forums and opinion surveys to make the e-service delivery a more better
prospect.
• Setting up Feedback: A tab where the user can give the feedback of the
interaction and impart their valuable suggestions to make site and services more
friendly for the user
• Continuous Improvements: The e-services must be looked upon as to what
improvements can be made so to make the interaction with user more simpler
and easy.
The maturity levels of current e-services of Himachal Pradesh (HP) government
are tested under the proposed maturity model to define the present level of maturity
these services have (Table 1).
Test Results
Employment Portal: The service only has web-presence and accessibility of data
at all times is not available, so the e-services are not up to level 1.
HRTC Ticket Booking: The e-service fulfils all the maturity levels of level 1
and level 2 of the model, however, it can be matured to level 4 with most ease in
comparison of other services.
Table 1 Current level of e-services Current level of maturity

e-services
Employment portal –
HRTC ticket booking Level 2
Lok Praman –
Hpsebl Level 2
MC, Shimla Level 1
Lok Praman: The e-service only has web presence, the data content of the site
is not accessible to its citizens. Maharashtra government public e-services portal can
serve as an example to improve the maturity level of this e-service.
HPSEBL: This e-service qualifies to be of level-2 service as its portal meets all
the required levels.
MC, Shimla: The e-service fulfils all the maturity levels of level 1 of the model.
It is a level 1 e-service, which can be easily moved up to level 2 by following
Uttarakhand online water portal.
Following this model, the maturity of the services can be improved for any state
government e-services.
The levels described in proposed maturity model are to classify the e-services of
the portal as level 1, 2, 3 or 4 services. This model would help understand in
enhancing the services and making them more developed. The test results describe
the current level of maturity in the e-services.
5 Information Security Model
To ensure security, it is important to build-in security in both the planning and the
design phases and adapts a security architecture which makes sure that regular and
security-related tasks are deployed correctly [10]. Currently, the e-services of
Himachal Pradesh (HP) government have a very basic level of security or if very
honest no level of security. The major lapse behind the security issue is that their
site is on hypertext transfer protocol (http) rather being on hypertext transfer pro-
tocol server (https), which allows any hacker and easy access to the sensitive
information of the state government and its citizens. This issue needs to be con-
sidered by state government IT department as a priority.
Proposed e-Portal for the services
The paper proposes a multi-level authentication portal for all the existing e-services
of Himachal Pradesh (HP) government. This portal will help the citizens to register
once for all the e-services rather than registering individually for all e-services. This
portal will partially help solve the issue of data security until the government
decides to move from hypertext transfer protocol (http) to hypertext transfer pro-
tocol server (https). The following steps will result in creating the desired portal.
Registering on the portal

For simple access to e-services, a service portal can be created with a default page
setup to register and login. Users can register with Aadhaar card no. or a working
phone number. In case, the user does not possess an Aadhaar card, they can register
in the portal by creating a username, password and working phone no. This will
enable the generation of a unique ID of eight digits on the subsequent page. Once
this unique ID is created, login tab shall appear.
Logging into services:

Users will be required to have two credentials to log in for to have access to all the
services:
1. USER_ID
2. OTP (which will be sent to the registered phone.no)
Fig. 4 Flowchart of the proposed e-portal

Once this is created, the user can have access to all the services provided by the
portal and are saved from registering himself/herself at each service separately. This
user id can also serve as the user’s personal digital space where they can upload the
needed documents and access all governmental documents, related to each service.
This feature can be incorporated on the same page that indicates a list of all the
services of the portal, but in a separate section and by the name of documents. This
portal can be like a one-stop shop for the users as they will be able to access all the
e-services at one go and will be saved from the pain of registering separately for
each e-service (Fig. 4).
6 Conclusion
Through all the research done for the paper, it can be safely concluded that the
proposed model can be of immense help to the Himachal Pradesh (HP) government
in improving their e-services. Since this paper proposes the first ever maturity
model for both information security and services of the e-services of Himachal
Pradesh (HP) government, it can be appropriated as a standard base model for the
improvement of the portal.
Acknowledgements I would like to express my gratitude to the Founder-President of Amity

Group, Dr. Ashok K. Chauhan, for his keen interest in promoting research in Amity University.
I have had the opportunity of interacting with several people whose inputs have been extremely
vital. I would also like to thank the NIC engineers Mr. Sandeep Soodand for answering my
questionnaires and providing help as to better understand the existing e-portal of the Himachal
Pradesh (HP) government. Mr. Sandeep Sood has provided his consent for the paper for which I
am highly obliged. NIC Engineer has been of great help in my research paper providing the base
for my paper contents. Mr. Sood’s consent in using the details provided for the paper have been of
great help.
References
1. M. D. Waziri, Z. O. Yonah, “A Secure Maturity Model for Protecting e-Government Services:

A Case of Tanzania”, ACSIJ Advances in Computer Science: An International Journal, Vol.
3, Issue 5, No.11, pp. 98–106, 2014
2. M. R. Perumal, “E-governance in India: A Strategic Framework”, International journal for
Infonomics, Vol-1, pp. 1–12, 2014
3. S. Gupta, S. Kaushal, “E-Governance Infrastructure-Status and Challenges Case Study on
Himachal Pradesh”, National Informatics Centre Department of Electronics and Information
Technology, Vol-1, pp. 1–6, 2013
4. V. Varma, “Cloud Computing for E-Governance”, International Institute of Information
Technology (White paper) Vol-1, pp. 10–23, 2010
5. R. Larbi, “A Proposed Cloud Security Framework for Service Providers in Ghana”,
International Journal of Computer Applications (0975 – 8887), Vol 158, pp. 17–22, 2017
6. V. A. Canal, “Information Security Management Maturity Model”, www.isecom.org, Vol-2,

pp. 11–24,2011 Lok Praman Patra, http://admis.hp.nic.in/epraman/
7. M.P. Gupta, “E-government evaluation: A framework and case study”, Indian Institute of
Technology, Department of Management Studies, Vol-1, pp. 365–387, 2010
8. T. N. Raja, “A Maturity Model Framework for e-Gov Applications – EGSARMM”,
International Journal of Computer Applications® (IJCA), Vol-1, pp. 20–25, 2012
9. A. F. Allah, L. Cheikhi, “E-Government Maturity models: A Comparative study”,
International Journal of Software Engineering & Applications (IJSEA), Vol.5, No.3,
pp. 71–85, 2014
10. B. Stevanović, “Maturity Models in Information Security”, International Journal of
Information and Communication Technology Research, Vol 4, No. 2, pp. 15–19, 2011
Malaria Detection Using Improved Fuzzy
Algorithm
Mukul Sharma, Rajat Mittal, Tanupriya Choudhury,

Suresh Chand Satapathy and Praveen Kumar
Abstract Malaria is one of the most life-threatening diseases, which need a serious
attention in today’s scenario. This disease is estimated to be having 3–6 billion
infected cases worldwide annually having the mortality rate of 1–3 million people.
Malaria should be diagnosed on time and treated precisely as it can lead to death of
a person. The main objective of this paper is to design and describe an algorithm
that can diagnose malaria in the early stage only so that a person cannot go up to the
serious and hazardous stages or complications and also the mortality rate of malaria
is reduced to an extent. Fuzzy logic is an approach to implement expert systems,
which are portable and can diagnose malaria accurately as compared to the other
systems.
Keywords Fuzzy sets Fuzzy algorithm Malaria Fuzzy rules

Fuzzy values
M. Sharma (&) R. Mittal T. Choudhury P. Kumar

Amity School of Engineering & Technology, Amity University, Noida, Uttar Pradesh, India
e-mail: mukulrock38@gmail.com
R. Mittal
e-mail: rajatmittal100@gmail.com
T. Choudhury
e-mail: tchoudhury@amity.edu
P. Kumar
e-mail: pkumar3@amity.edu
S. C. Satapathy
P.V.P. Siddhartha Institute of Technology, Kanuru, Vijayawada, India
e-mail: sureshsatapathy@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_64
654 M. Sharma et al.
1 Introduction
Malaria is disease caused by the group of single-celled microorganism called

parasitic protozoan. It is an infectious disease caused by liver parasites or spleen of
genus plasmodium. Once the parasite is transmitted, they start getting multiplied in
the liver before destroying red blood cell. Symptoms of malaria that are considered
in this paper are fever, vomiting, headache, weakness, chills, nausea, sweating,
diarrhoea and enlarged liver. Soft computing technique involves various methods
like approximate reasoning and functional approximation. In approximate reason-
ing, collection of imprecise premises is used to deduce imprecise conclusion.
Probabilistic reasoning and fuzzy logic are the methods involved in approximate
reasoning. In functional approximation, a function among well-defined class is
selected that closely matches a target function in a different way. Evolutionary
computation, artificial neural networks and machine learning are the methods
involved in functional approximation. There are several methods that exist for the
malaria diagnosis. The methods can be classified on the basis of performance and
cost. Polymerase chain reaction (PCR) and third harmonic generation (THG) are
high cost methods. PCR method is used to detect nucleic acid sequences. According
the studies these methods are best and are highly sensitive to malaria diagnosis. But
due to their high cost, handling difficulties and infrastructures, they are rarely used
in developing countries. Another method called RDT is relatively faster methods
than other in diagnosis and can be performed by unskilled people but results are not
reliable. Since RDT kits can be used for the single species, so four kits are used
which make it expensive. Another method called conventional microscopy is used
for the diagnosis but this time consuming and results are difficult to reproduce.
From the above discussion, we found that more the sophisticated technique is,
results are more reliable but very expensive and unaffordable in many countries. On
the other hand, less sophisticated techniques are cheap and affordable but the results
are not reliable.
2 Literature Survey
Chandra et al. [1] presents detection of malaria disease through soft computing. For
this, image processing technique is used which is done by following some stages
like image filtering, image pre-processing, image segmentation, feature extraction
and artificial neural network (ANN). Mohamed et al. [2] presents a method to
diagnose malaria parasite using the fuzzy logic systems. A proper methodology is
mentioned for this approach of diagnosis of malaria. A certain number of data is
taken from a database and then this data is received by the system through input
technology. Fuzzification and defuzzification of the data are done and a desired
output is obtained by the system. Djam et al. [3] describe (FESMM), i.e. fuzzy
expert system for the management of malaria and giving a decisive as well as
Malaria Detection Using Improved Fuzzy Algorithm 655
supportive platform to the scientists who are engaged in the research of the malaria
disease, the specialist doctors and the healthcare experts. This FESMM comprises
of 4 blocks which are (1) Data Collection (2) Fuzzification (3) Inference Engine and
(4) Defuzzification. The basic idea behind the designing of the fuzzy expert system
is based on the clinical perceptions, medical analysis and healthcare practitioners.
Onuwa et al. [4], presents an expert system called fuzzy expert system for the
detection of malaria disease. The paperwork also focuses on that the algorithm
which is built to diagnose malaria rapidly and accurately is very simple to use,
cheap and portable. These qualities of algorithm make it perfect to use in the
diagnosis of malaria. This algorithm not only supports medical practitioners, but
also helps malaria researchers to manage the imprecision and provides accuracy in
the final output on the basis of given input. Sharma et al. [5], in this paper the
authors have proposed a decision support system for malaria and dengue disease
diagnosis for the remote (rural) areas where medical support is not available to the
people easily. The decision support system for malaria diagnosis and treatment was
designed, developed and simulated by using MATLAB, GUI function with the
implementation of fuzzy logic. Duodu et al. [6] presents a brief description about
malaria and its consequences, this also tells us about how fuzzy logic can be
implemented in the various steps like collecting the data of the patients suffering
from malaria disease, collected data is used as an input, fuzzification is done on the
input, fuzzified data is matched with the pre-defined set of fuzzy logic in the
inference engine and finally defuzzification is done which gives a representative
output from the inference engine.
3 Fuzzy Logic System
In 1965, Prof. LotfiZadeh introduced a fuzzy logic system with the help of a fuzzy
set of theory. In this theory, a group of objects having unclear boundaries can relate
in which membership is just a matter of degree. By making the use of fuzzy
approach, transition in terms can be binary, gradual or null; selection must be at the
extreme end of the continual. Fuzzy logic system identifies and uses a real-world
phenomenon. This method makes use of the concept of fuzziness in computa-
tionally better manner. It is basically associated with the reasoning that mainly uses
approximate value rather being a fixed value.
• Fuzzy sets can be seen as an expansion and speculation of the basic concepts of
the crisp set, which tells us that either an element is a member of set or not but
on the other hand fuzzy sets gives partial memberships to the elements in a set
with the degree of membership ranging from 0 to 1. 0 (not an element of a
particular set) and 1 (element of a particular set). Different types functions such
as triangular, trapezoidal, Gaussian and sigmoid membership functions are
there, but in this paper, the membership function which is used is triangular
membership function whose calculation can be done with the help of the
below-mentioned functions (Fig. 1).
This Triangular Function is defined by a lower limit ‘a’, an upper

limit ‘c’ and a value ‘b’ lying between ‘a’ and ‘c’
Fig. 1 Triangular membership function
A fuzzy set B in X can be expressed as a set of ordered pairs:
B ¼ fðx; lBðÞÞjx 2 Xg ð1Þ
where µB = membership function,
lB : X ! P ð2Þ
where P is the membership space in which each element of X is mapped.

If P = {0, 1}, B is called crisp set.
However, if {0 P 1}, B is called fuzzy set.
Fuzzy logic has been applied in many fields such as artificial intelligence and
control theory. It has various applications from customer end such as cameras,
washing machines, ovens to various industrial processes.
4 Methodology
Structure of the Fuzzy System

Fuzzy system has an AMDFLT (algorithm for malaria diagnosis using fuzzy logic
for treatment) structure shown in the figure, which consists of four major parts, i.e.
fuzzification, knowledge base, fuzzy inference engine and defuzzification. Other
components include input and output data.
4.1 Data Collection
In the literature survey, various observation and various other patient studies are
used to collect the information to make a relation between input and output vari-
ables, create various functions and design fuzzy logic decision matrix.
Knowledge base and database are maintained in the data store. Malaria reference
review and medical experts are major sources of data store. Knowledge base
consists of decision matrices, which has a set of condition called IF and THEN
which is formed using logic gate operator and forward chaining and storing
information for references with the help of fuzzy inference engine.
4.2 Input Data
Mainly keyboard is used as input resource to enter various data into the system.
Exact numbers are taken for patient’s pulse rate, body temperature, weight and
height from the different instruments in BPM, Celsius, kilogram and metres.
Intensity values that are converted are determined in 100% and entered.
4.3 Fuzzification
Member functions (fuzzifier) are used to convert crisp input values into grades of
membership degrees. In fuzzification, we can obtain a fuzzy value from real scalar
value. This all can be obtained from different types of fuzzifiers. Three types are as
follows: Are Singleton fuzzifier, Gaussian fuzzifier, Triangular fuzzifier.
Mathematically,
x ¼ fuzzifierðx0 Þ ð3Þ
Where x is fuzzy set, x0 is called crisp input value and fuzzifier represents fuzzi-
fication function. Mainly for changing the scalar values to fuzzy sets (range 0–1),
triangular fuzzifier is used. Later crisp value is converted to a linguistic variable. To
obtain degree of symptom, a formula is used, i.e.
Xi =Xn ð4Þ
Xi is no. of linguistic variable and Xn is total no. of linguistic variable. Through

this formula, we can make the table for triangular fuzzifier. For a case, if a patient is
complaining of very high fever then the degree of symptom will be 4/5 = 0.80
(Table 1).
4.4 Data Store
One of the main key factors responsible for the intelligent system is the knowledge.
Knowledge consists of mainly database. Knowledge base acts as one of the major
Table 1 Ranges of fuzzy Linguistic variables Fuzzy values

value
Very low 0.00 x < 0.20
Low 0.20 x < 0.40
Medium 0.40 x < 0.60
High 0.60 x < 0.80
Very high 0.80 x 1.00
source for the information to be searched collected, organised, shared and utilised.
The knowledge base for the malaria diagnosis is composed of mainly structured and
organised information. Knowledge base was taken from various books on malaria
and medical experts. In knowledge base, forward chaining and logical AND
operator is used to construct the if-then decision matrices and fuzzy inference
engine is used to store data for references. This data is stored for the future
references.
Representation of the above-mentioned rules from A to J is as follows:
A IF Fever = Very High, Vomiting = High, Headache = High, Weakness = High,
Chills = Medium, Nausea = Medium, Sweating = Low, Diarrhoea = High and
Enlarged Liver = Medium THEN Malaria = High
B IF Fever = Medium, Vomiting = Low, Headache = Very Low, Weakness =
Low, Chills = Very Low, Nausea = Very Low, Sweating = Very Low,
Diarrhoea = Very Low and Enlarged Liver = Very Low THEN Malaria = Very
Low
C IF Fever = Low, Vomiting = Very Low, Headache = Very Low, Weakness =
Low, Chills = Very Low, Nausea = Very Low, Sweating = Low, Diarrhoea =
Very Low and Enlarged Liver = Very Low THEN Malaria = Very Low
D IF Fever = High, Vomiting = High, Headache = Low, Weakness = Medium,
Chills = Low, Nausea = Very Low, Sweating = Medium, Diarrhoea = Low
and Enlarged Liver = Low THEN Malaria = Medium
E IF Fever = Medium, Vomiting = Low, Headache = Very Low, Weakness =
Medium, Chills = Very Low, Nausea = Very Low, Sweating = Low,
Diarrhoea = Very Low and Enlarged Liver = Very Low THEN Malaria = Low
F IF Fever = Medium, Vomiting = High, Headache = High, Weakness = Medium,
Chills = Low, Nausea = Very Low, Sweating = Medium, Diarrhoea = Medium
and Enlarged Liver = Very Low THEN Malaria = Medium
G IF Fever = Very High, Vomiting = Very High, Headache = Very High,
Weakness = Very High, Chills = Very High, Nausea = Very High, Sweating =
High, Diarrhoea = High and Enlarged Liver = High THEN Malaria = Very High
H IF Fever = Very High, Vomiting = High, Headache = High, Weakness = Very
High, Chills = Very High, Nausea = Very High, Sweating = Very High,
Diarrhoea = Very High and Enlarged Liver = Very High THEN
Malaria = Very High
I IF Fever = Very High, Vomiting = Very High, Headache = High,

Weakness = High, Chills = Medium, Nausea = Medium, Sweating = High,
Diarrhoea = High and Enlarged Liver = High THEN Malaria = High
J IF Fever = Low, Vomiting = Low, Headache = Very Low, Weakness = Low,
Chills = Very Low, Nausea = Very Low, Sweating = Very Low,
Diarrhoea = Low and Enlarged Liver = Low THEN Malaria = Low
4.5 Fuzzy Inference Engine
Fuzzy inference system uses knowledge base to derive the answer. It is also
sometimes referred as the brain of expert system that is used to reason the infor-
mation in knowledge base and make new conclusions. In fuzzy inference engine,
the degree of membership is determined by mapping the fuzzy inputs into their
linguistic variables and weighting factors. The degree of the firing strength is
determined by aggregation operator. And for this fuzzy logic AND is used. In fuzzy
rule sets, combination many fuzzy logical operators such as AND, OR and NOT are
used to have several predecessor. AND operator—used for minimum weight for all
predecessor. OR operator—uses maximum value for all predecessors. NOT oper-
ator—used to find complementary function by subtracting membership function
from 1. RSS also called as Root Sum Square is used in this research work whose
formula is given by
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn
RSS ¼ i¼1 i
r2 ð5Þ
RSS = value of firing rule and r2i are the values of different rules
Fuzzy inference engine has many uses.
Input data file and FIS structure file can be used to perform fuzzy inference.
Fuzzy inference engine can be customised accordingly to include our own mem-
bership function. Executable codes can be embedded into other application. Fuzzy
inference mainly consists of two c code source files, i.e. fis.c and fismain.c in fuzzy
folder. In fisman.c file, it has only one main function which can be easily modified
for our own application as it is ANSI C compatible.
4.6 Defuzzification
The process of converting a fuzzy output to a crisp value from inference engine is
called defuzzification. Defuzzification is a method used to obtain a value that will
represent the output obtained from the inference engine. That is, the level of the
illness can be determined from the result which we get from the inference system
using defuzzified root sum square. The technique developed by Takagi and Sugeno
in 1985 called centroid of the area is also called centre of area or centre of gravity,
which is used for defuzzification. Since it is very simple and accurate, it becomes
one of the most commonly used techniques. The centroid defuzzification technique
is
R xmax
f ðxÞdx
X ¼ Rxmin
xmax ð6Þ
xmin f ðxÞdx
In the above equation, x* refers to defuzzified output where f(x) dx refers to

aggregated membership function and the variable x indicates the output variable
(Fig. 2).
5 Algorithm
The algorithm for the fuzzy logic system is as follows:

Step 1 Enter the details such as symptoms and indications of the patient into the
system, where the variable m represents a number of symptoms and
indications of the patient.
Step 2 Now, we have to rectify the signs and symptoms of the patient for which
we have to explore the knowledge base for the disease which consists
those symptoms.
Fig. 2 Architectural structure of fuzzy system

Step 3 Now get the weighing factors (the associated degree of intensity) wf = 1,
2, 3, 4, 5 where 1 refers to very low, 2 refers to low, 3 refers to medium,
4 refers to high and 5 refers to very high.
Step 4 Now, fuzzy rules are applied.
Step 5 In the next step, the degree of membership is determined for which we
have to map the fuzzy inputs into their particular weighing factors.
Step 6 Now the rule base is to be determined for which the non-minimum
values can be evaluated.
Step 7 Evaluate the strength of firing (conclusion) for the rules R.
Step 8 The non-zero minimum values are evaluated to calculate the degree of
truth R.
Step 9 Now the intensity of the disease is calculated.
Step 10 Now we get the output as fuzzy diagnosis.
6 Research Experimentation
In this research experiment, a sample of 5 patients who have already gone through
medical doctor diagnosis and laboratory diagnosis has been taken consideration.
Fuzzy system is used to diagnose the patient, based on the complaint made by him
in order to get the knowledge of how the system works when implemented. Grades
such as very low, low, medium, high and very high are used. A patient called
Manmohan with patient ID AAS002 complains the doctor having symptoms of
malaria as shown in Table 3. Fuzzy rules are shown in Table 4. About 10 rule bases
are used for the patient in Table 2. Root sum square is used to compute the output
membership function in the range of 0–1 for every linguistic variable. Now, the
centroid of the area is used to defuzzify the outputs from the inference engine to
obtain a crisp value.
7 Comparing Results of Malaria Diagnosis
The basic aim is to have a comparison between the algorithm’s diagnosis with the
doctor’s diagnosis using laboratory diagnosis as accurate and standard diagnosis
(Table 5).
662
Table 2 Fuzzy rule composition for the detection of malaria

Rule No Fever Vomiting Headache Weakness Chills Nausea Sweating Diarrhoea Enlarged liver Linguistic variables
A 0.9 0.7 0.7 0.7 0.5 0.5 0.3 0.7 0.5 High
B 0.5 0.3 0.1 0.3 0.1 0.1 0.1 0.1 0.1 Very low
C 0.3 0.1 0.1 0.3 0.1 0.1 0.3 0.1 0.1 Very low
D 0.7 0.7 0.3 0.5 0.3 0.1 0.5 0.3 0.3 Medium
E 0.5 0.3 0.1 0.5 0.1 0.1 0.3 0.1 0.1 Low
F 0.5 0.7 0.7 0.5 0.3 0.1 0.5 0.5 0.1 Medium
G 0.9 0.9 0.9 0.9 0.9 0.9 0.7 0.7 0.7 Very high
H 0.9 0.7 0.7 0.9 0.9 0.9 0.9 0.9 0.9 Very high
I 0.9 0.9 0.7 0.7 0.5 0.5 0.7 0.7 0.7 High
J 0.3 0.3 0.1 0.3 0.1 0.1 0.1 0.3 0.3 Low
M. Sharma et al.
Table 3 Symptoms of Symptoms Linguistic variables Given values

Malaria in MANMOHAN
Fever Very high 0.9
Vomiting High 0.7
Headache Very high 0.9
Weakness Very high 0.9
Chills High 0.7
Nausea Medium 0.5
Sweating High 0.7
Diarrhoea Medium 0.5
Enlarged liver High 0.7
8 Results
By applying Eq 5, i.e. the formula of root sum square, we will get the actual fuzzy
values corresponding to the respective linguistic values.
• Very Low—0.1414, Low—0.3162, Medium—0.4242, High—0.7071, Very
high—0.9899
OUTPUT X
ð0:1414 0:1Þ þ ð0:3162 0:3Þ þ ð0:4242 0:5Þ þ ð0:7071 0:7Þ þ ð0:9899 0:9Þ
¼
0:1411 þ 0:3162 þ 0:4242 þ 0:7070 þ 0:9899
¼ 0:6619
Hence, the patient is having 66% malaria which is ‘High’ according to our
linguistic variable which is at a complicated stage.
9 Conclusion
This research paper provides the algorithm of fuzzy logic for the diagnosis of the
disease ‘Malaria’. It also tells us that fuzzy logic is a very helpful technique which
can deal precisely with the issues that occur in malaria. By following the proper
methodology of fuzzy logic that is mentioned above, one can accurately detect
malaria without going through the second opinion. Mamdani Fuzzy Inference
System is a better system for the diagnosis of human disease. A sample of 5 patients
was selected and the result of diagnosis by the doctor, the laboratory and the
algorithm was compared with each other which proves that the fuzzy logic algo-
rithm shows the best results amongst the doctor diagnosis and the laboratory
664
Table 4 Fuzzy rules for patient MANMOHAN

Rule Fever Vomiting Headache Weakness Chills Nausea Sweating Diarrhoea Enlarged Linguistic Non-Zero
No liver variables minimum
value
A 0.9 0.7 – 0.7 0.5 – – – 0.5 High 0.5
B – – 0.1 0.3 – 0.1 0.1 – 0.1 Very low 0.1
C 0.3 – – – – – – 0.1 0.1 Very low 0.1
D – 0.7 – – 0.3 – – – – Medium 0.3
E – – 0.1 0.5 – – 0.3 0.1 – Low 0.1
F 0.5 – 0.7 – 0.3 – 0.5 – – Medium 0.3
G – 0.9 – 0.9 – 0.9 – 0.7 0.7 Very high 0.7
H 0.9 – 0.7 – 0.9 – – 0.9 – Very high 0.7
I – – 0.7 0.7 – 0.5 0.7 – 0.7 High 0.5
J – 0.3 – – – – – – 0.3 Low 0.3
M. Sharma et al.
Table 5 Compared results of malaria diagnosis

Patient id Doctor diagnosis Laboratory diagnosis Algorithm diagnosis
AAS001 Uncomplicated malaria Uncomplicated malaria Uncomplicated malaria
AAS002 Free from malaria Uncomplicated malaria Complicated malaria
AAS003 Uncomplicated malaria Free from malaria Free from malaria
AAS004 Uncomplicated malaria Uncomplicated malaria Uncomplicated malaria
AAS005 Free from malaria Complicated malaria Uncomplicated malaria
diagnosis, some of the results were same in the laboratory and the algorithm
diagnosis but the accurate results were shown by the algorithm diagnosis.
References
1. Upendra Kumar Chandra, Yogesh Bahendwar, Volume 5, Issue 10, October-2015,“Detection

of Malaria Disease through Soft Computing” International Journal of Advanced Research in
Computer Science and Software Engineering [ISSN: 2277 128X].
2. Khalid Alsir Mohamed, Eltahir Mohamed Hussein, Index Copernicus Value (2013): 6.14 |
Impact Factor (2015): 6.391 Volume 5 Issue 6, June 2016, “Malaria Parasite Diagnosis using
Fuzzy Logic” International Journal of Science and Research (IJSR) [ISSN (Online): 2319–
7064].
3. Djam, X. Y, Wajiga, G. M, Kimbi Y. H, and Blamah, N. V (2011).A Fuzzy Expert System for
the Management of Malaria. International Journal of Pure and Applied Sciences and
Technology (5(2) (2011), pp. 84–108.
4. “Fuzzy Expert System for Malaria Diagnosis” by Ojeme Blessing Onuwa et. all, An
International Open Free Access, Peer Reviewed Research Journal, Published By: Oriental
Scientific Publishing Co., India. June2014, Vol.7, No. (2):Pgs. 273–284 [ISSN: 0974-6471].
5. “Decision Support System for Malaria and Dengue Disease Diagnosis (DSSMD)” by Priynka
Sharma, DBV Singh, Manoj Kumar Bandil and Nidhi Mishra, International Journal of
Information and Computation Technology.Volume 3, Number 7 (2013), pp. 633–640, [ISSN:
0974-2239].
6. Quashie Duodu, Joseph Kobina Panford, James Ben Hafron-Acquah, Volume 91 – No.17,
April 2014, “Designing Algorithm for Malaria Diagnosis using Fuzzy Logic for Treatment
(AMDFLT) in Ghana” International Journal of Computer Applications [ISSN: 0975 – 8887].
Feature Extraction Techniques Based
on Human Auditory System
Sanya Jain and Divya Gupta
Abstract Feature extraction is a relevant method in the performance of the ASR

system. A good technique not only removes irrelevant characteristics, but also
represents important attributes of a speech signal. This paper intends to concentrate
on the comparison between feature extractions techniques of speech signals based
on human auditory system for better understanding and to enhance its further
applications. In this review, we explain three different techniques for feature
extraction. The main emphasis is to show how they are useful in processing signals
and extracting features from unprocessed signals. The human auditory system is
explained which combines this study altogether. The aim is to describe techniques
like Zero Crossing with Peak Amplitude, Perceptual Linear Prediction and Mel
Frequency Cepstral Coefficient. As each method has its own merits and demerits,
we have discussed some of the most important features of these techniques.

Keywords Feature extraction Perceptual linear prediction Mel frequency

cepstral coefficient Zero crossing with peak amplitude
1 Introduction
Speech is the primary source of communication with people around us. If there is no
speech, contacting people surely is difficult. With the emergence of digital technol-
ogy, highly resourceful digital processors were invented with relatively high power,
low cost and high speed. With their development, the researchers were able to
transform analog speech signal to digital speech signal, which was undergone a study
S. Jain (&) D. Gupta

Department of Computer Science and Engineering, Amity School of Engineering
and Technology, Amity University, Noida Uttar Pradesh, India
e-mail: sannyajain@gmail.com
D. Gupta
e-mail: dgupta1@amity.edu

https://doi.org/10.1007/978-981-10-7245-1_65
668 S. Jain and D. Gupta
for future applications. In the field of digital signal processing, speech recognition or
processing has versatile relevance so it is an intensive field of research.
Automatic Speech Recognition (ASR) has emerged in digital speech signal
processing applications [1]. In this process, the spoken words are automatically
transformed into the written language by computer systems and smartphone also
nowadays. The applications of speech processing are increasing day by day, for
example in information searching on websites, data entry in fields and even in
biometric recognition technologies [2].
Different data and algorithms are combined together such as statistical pattern
recognition, linguistics, signal processing, etc., for speech recognition [3]. The front
end of signal processing is feature extraction, which also has been proved to be of
greater importance for parametric representation of signals.
A slight amount of data is extracted from the voice signal which is used to build
a separate model for each speech utterance; this is known as the parametric rep-
resentation [4]. It is then used for analysis for representation of specific speech
utterance. Figure 1 depicts the speech recognition process. The successful tech-
niques use the psychological process of human hearing in the analysis which is
discussed in further sections in this paper.
There are many feature extraction techniques, which are as follows:
• Linear predictive analysis (LPC);
• Zero Crossing Peak Amplitude (ZCPA);
• Linear predictive cepstral coefficients (LPCC);
• Mel frequency cepstral coefficients (MFCC);
• Mel scale cepstral analysis (MEL);
• Perceptual linear predictive coefficients (PLP);
• Relative spectra filtering of log domain coefficients (RASTA);
• Power spectral analysis (FFT);
• First-order derivative (DELTA).
Linguistic Acoustic
features features
Feature
Voice Extraction Decoder Words
(Front end)
Language
Model
Fig. 1 Speech recognition

Feature Extraction Techniques Based on Human Auditory System 669
2 Human Auditory System
Technological development in speech recognition system presently focuses on

providing trusted and robust interfaces for functional deployment. One of the
greater challenges is achieving robust behavior of these techniques in noisy areas in
the applications which use human–computer dialog system, dictation and
voice-controlled devices [5]. Voice acquisition, its processing and its recognition
are a complicated task, because of the presence of unwanted noises and variables.
These additive noises from the environment in which speech is processed degrades
the quality of recorded voice and makes it unfit for processing.
At the very first level of communication, the idea or thought of the message is
developed in the mind of the speaker which is also known as the linguistic com-
munication level. That thought is then converted to some significantly meaningful
letters, words, phrases and finally sentences according to the grammatical rules of
that particular language in which the speaker speaks in. The electric signal which
activates muscles in vocal tract and vocal cord is first created by the brain that
moves along the motor nerve [6]. This vocal tract movement creates pressure
changes within vocal tract and lips, initiates sound waves that is to propagate in
space. At last at the linguistic level brain performs speech recognition and under-
standing of words and sentences.
3 Issues in ASR System
There are many aspects of variability of the voice signal and these can be segre-
gated with the help of a comparatively better ASR system. The system’s robustness
allows dealing with these variable characteristics.
For achieving a better overall result of voice recognition, it is relevantly
important to train and test the system according to the environment it is going to be
deployed in [7]. The differences between training and testing condition in a noisy
environment often lead to the degradation in the ASR’s system’s recognition
accuracy.
The issues in ASR system are shown in Table 1. When the data which is
acquired in noisy condition is tested against the data recorded in clear environment,
the quality and accuracy of the data get degraded over time, which is not possible in
actual human hearing. To reduce the mismatch between the above-mentioned
conditions and degradation in accuracy, various voice signal improvement,
extraction and feature normalization techniques are applied one by one.
Table 1 Issues of ASR design

Conditions Types of noises; Signal-to-noise ratio; training and testing environment
conditions
Transducer Microphone; mobile
Speech Voice pitch; production of sound, e.g. continuous speech or isolated words;
styles pace of speech
Speakers Speaker dependence/independence; physical state
Channel Amplitude; reverberation
Vocabulary Aspects of available training and testing data; vocabulary
4 Overview of Feature Extraction in Speech Recognition
The front end of signal processing is feature extraction, which also has been proved
to be of greater importance for parametric representation of signals. A relatively
small detail of information is extracted from the signal that is used to build a
separate model for each speech expression; this is known as the parametric rep-
resentation. It is then used for analysis for the representation of specific speech
utterance. It separates speech patterns from one another [8]. Figure 2 shows the
feature extraction.
Speech signals are quasi-stationary as they are slow varying time signals. This is
the reason it is better to perform feature extraction in short-term interval which
would reduce these variabilities. Therefore, the examined period of time for the
speech signals is 10–30 ms, where the characteristic of speech signal becomes
stationary [9]. The features extracted from the speech signals should portray certain
qualities such as it should not be subject to mimicry, it should be stable over a time
period, does not deviate in different environments, etc. [3]. Parameters which can be
determined through processing of speech signals are known as features.
Feature extraction aims to produce a perceptually significant representation of
speech signal. Extraction of the feature vector is done through extraction methods
Speech Parametric
Spectral
data transformations
Analysis
Extracted Statistical
features Modeling
Fig. 2 Feature extraction process

like Mel Frequency Cepstral Coefficient (MFCC), Linear Predictive Coding (LPC),
Perceptual Linear Prediction (PLP), Zero Crossing Peak Amplitude (ZCPA),
Wavelet Transform, etc.
Feature extraction process can be divided into 3 operations:
• Spectral Analysis:
In time-varying signals, the characteristic properties of speech is represented by
parameterization of spectral activity when speech is produced. Six classes in
spectral analysis algorithms include Fourier transform, linear prediction, digital
filter bank in speech processing system. From the mentioned classes, the linear
prediction tends to give the best results.
• Parametric Transforms:
The two basic operations from which the parameters of the signal are generated are
differentiation and concatenation. The resultant of both of these stages include the
parameter vector of the raw signal.
Differentiation characterizes the temporal variations of speech signals and the
signal model’s higher order derivative. Concatenation is creating an individual
parameter vector per frame that consists of all the signals which are desired.
• Statistical Modeling:
This is the third and last step of feature extraction. It is presumed here that some
multivariate process generated the signal parameters. To understand this process, a
model is imposed on the data, then the data is optimized for training and finally, the
approximation quality of data is measured. The observed output (signal parameter)
which we get from this process is the only information. Signal observation is
another name given to the resultant parameter vector.
The main intention to perform statistical analysis is to find out if the spoken
words are actually meaningful or just noise. Being the fundamental process of the
speech recognition, it is very important to use a well-versed statistical model.
Vector Quantization (VQ) is certainly the successful and sophisticated algorithm of
statistical analysis model.
5 Feature Extraction Techniques
The various methods which are used for feature extraction based on the human
auditory system are available, and some of them are discussed below:
5.1 Perceptual Linear Prediction (PLP)
The PLP representation was developed by Hermansky in 1990. As the name sug-
gests, the main aim of the PLP approach is describing the psychophysical human
hearing with accuracy throughout the whole feature extraction procedure [10].
Figure 3 depicts the steps to compute PLP coefficient. Using transformations relied
on psychophysics; the spectrums are modified in PLP with contradiction to pure
LPC analysis [11]. The PLP speech features are as follows: The PLP parameters of
18 filters cover the frequency of range 0–5000 Hz and its dependence are based on
Bark spaced Filter bank [8]. The method which computes the PLP coefficients is:
• The x(n) signal which is distinct in the time domain is subjected to the N—point
DFT.
• The downsampled is applied with the pre-emphasis of equal loudness.
• Intensity-loudness compression is executed.
• An inverse DFT is performed on the result obtained so far to result in the
equivalent functions which are correlated.
• At last, the coefficients of this PLP technique are calculated after autoregressive
modeling and conversion of these autoregressive coefficients to cepstral
coefficients.
The most often applied acoustic features in speech recognition are mel frequency
cepstral coefficients (MFCC) and PLP features. As and when there is found a
mismatch in training and test data, PLP features are found to be much more robust
[12].
However, when the acoustic conditions are variable over whole data set in many
applications, for example: in broadcast news transcription, segments with back-
ground music or noise are often intermixed with segments with clean speech [12].
Sample quantized Hamming Bark filter bank

signal window
Linear Intensity- Equal loudness

prediction >loudness pre-emphasis
Recursive PLP cepstral

cepstrum coefficient
Fig. 3 Computational steps of PLP

The following changes in PLP are considered: (i) Mel filter Bank replaces the Bark
filter- bank; (ii) The speech signal which is applied by the pre-emphasis substitutes
the equal loudness weighting of spectrum; Finally, (iii), a very large number of
filters are contained in new filter bank and there is no significant loss as they have
the same bandwidth as the old ones.
5.2 Zero Crossing with Peak Amplitude (ZCPA)
The method ZCPA of extraction of features is based on an auditory system of

humans. Zero crossing interval presents information of frequency signals and
intensity information represented by amplitude value. Finally, the complete feature
output is resulted by combining amplitude information and frequency information.
This model is used in cochlear sound processing. Significant immunity is provided
to white noise as compared to PLP and MFCC. Figure 4 portrays ZCPA block
diagram for extraction of features. The diagram contains zero crossing detection
block, Band Pass Filters, peak detection block, the frequency receiver and com-
pression of nonlinear amplitude. The filters cover the range of 200–4000 Hz fre-
quency and are based on the auditory system, which contains 16 BPF. The
unprocessed speech information passes through the filter bank. These are then
transformed to 16 different information processing paths.
Every signal is distributed in frames and each and every frame goes to detection
of zero crossing interval and also peaks the detections in every gap. The equation
which is a monotony function is used by nonlinear compression of peak values and
is represented by x in peak upward going position in zero crossing intervals and
when it is compressed logarithmically it results in g(x).
gðxÞ ¼ logð1:0 þ 20xÞ: ð1Þ
Zero crossing Interval

FIR Filter
detector histogram
Nonlinear
Peak detector compression ZCPA
Fig. 4 Principal diagram of ZCPA

Each band called frequency bin forms a result when the frequency receiving block
is divided into some sub-bands by frequency band. The ZCPA feature is the 16 path
output together formed. The pitch and intensity features of speech are represented
by the output feature vector as these are the basic features of speech.
5.3 Mel Frequency Cepstral Coefficient
The mel frequency cepstrum coefficient (MFCC) technique usually creates the basic
fingerprint of the sound files. Human ear’s critical bandwidth frequency is the basis
of MFCC with filters which are usually spaced in linearity with typically less
frequencies and logarithmically with more frequencies which captures the salient
aspects of speech. To compute MFCC coefficients, the signal is divided into
overlapping frames [12]. Every overlapped frame contains N samples and
M samples separates adjacent frames. Hamming window is multiplied with every
frame and is given by [13]:
WðnÞ ¼ 0:54 0:46 cos½ð2pnÞ=ðN1Þ: ð2Þ
Then the signal is gone through Fourier transform and converted to frequency
domain from time domain by [13]:
X
Xk ¼ xi eðj2pki=N1Þ : ð3Þ
Now for a frequency, Mel is calculated by [13]:
M ¼ 2595 log10 ð1 þ ðf =700ÞÞ: ð4Þ
DCT converts log Mel scale cepstrum to the time domain by [13]:
X
Xk ¼ a xi cosfð2i þ 1Þpk=2Ng: ð5Þ
The result so calculated is mel frequency cepstral coefficient.

In MFCC, all input statements are converted in a series of acoustic vectors.
A diagram of the MFCC procedure is depicted in Fig. 5.
The last step includes conversion of Mel spectrum scale to the standard scale of
frequency. This gives a good presentation of some spectral properties which are the
key to recognizing relevant characteristics of the speaker.
The MFCC is the most evident cepstral analysis based on feature extraction
method for speech and speaker identification tasks [14, 15]. MFCC is popularly
used as it is the most evident cepstral analysis based feature extraction technique
and it portrays the human response better than any other technique. Table 2 depicts
the summarized features of all these techniques.
Speech Signal Pre-Emphasis Framing
MFCC Hamming
Features Windowing
Discrete Cosine Mel Filter Fast Fourier

Transform Bank Transform
Fig. 5 Block diagram of MFCC
Table 2 Comparative analysis of feature extraction techniques

Techniques Features
Perceptual linear By choosing a particular model, suppression of
prediction speaker-dependent information is allowed
Robust to noise
Uses engineering approximations for psychophysical hearing
Zero crossing peak Used in cochlear implant sound processing
amplitude Provides immunity to white noise
Zero crossing of speech stimuli is the main basis of speech feature
extraction
Mel frequency with peak Good distinction
amplitude Required phonetics data can be accumulated
Less association among coefficient
6 Conclusion
The principal aim of this research is, to review basic feature extraction techniques
which use the human auditory system for functioning. The three techniques dis-
cussed are Zero Crossing with Peak Amplitude, PLP and Mel Frequency Cepstral
Coefficient. We compare the features of these three techniques and show that all the
techniques have distinguished features. We come to a conclusion that PLP is
entirely based on psychophysics of human hearing and is robust to noise as com-
pared to the other two techniques. ZCPA provides immunity to white noise and its
coefficient is calculated by zero crossing of speech signal whereas MFCC provides
good distinction in the speech variables.
References
1. J. W. Picone, “Signal modelling technique in speech recognition,” Proc. Of the IEEE, vol. 81,
no.9, pp. 1215–1247, Sep. 1993.
2. Rabiner L.R. and Shafer R.W. Digital processing of speech signals, Prentice Hall, 19, 1978.
3. Vimala. C and Radha. V, “Suitable feature extraction and speech recognition techniques for
isolated tamil spoken words” (IJCSIT) International Journal of computer Science and
Information Technologies, Vol. 5(1), 2014378-383.
4. Hagen A., Connors D.A. & Pellm B.L.: The Analysis and Design of Architecture Systems for
Speech Recognition on Modern Handheld-Computing Devices. Proceedings of the 1st IEEE/
ACM/IFIP international conference on hardware/software design and system synthesis,
pp. 65–70, 2003.
5. Thomas F. Quatieri, Discrete-Time Speech Signal Processing: Principles and Practice,
Pearson Education, Inc. 2002.
6. Hermansky, H.: Should recognizers have ears? Speech Communication, Vol.1998, No. 25,
pp. 2–27, 1998.
7. Lee Y. & Hwang K.-W.: Selecting Good speech Features for Recognition. ETRI Journal, Vol.
18, No. 1, 1996.
8. Mark D. Skowronski and John G. Harris, “Improving The Filter Bank Of A Classic Speech
Feature Extraction Algorithm”, IEEE Intl Symposium on Circuits and Systems, Bangkok,
Thailand, vol IV, pp 281–284, May 25 - 28, 2003, ISBN: 0-7803-7761-31.
9. Urmila Shawankar and Dr. Vilas Thakare, Techniques for feature extraction in speech
recognition system: a comparative study.
10. J. Mason and Y. Gu, “Perceptually-based features in ASR,” in IEEE Colloquium on Speech
Processing, 1988, pp. 7/1–7/4.
11. H. Hermansky, B. A. Hanson, and H. Wakita, “Perceptually based processing in automatic
speech recognition,” Proc. IEEE Int. Conf. on Acoustic, speech, and Signal Processing,”
pp. 1971–1974, Apr.1986.
12. H. Hermansky, “Perceptual Linear Predictive Analysis of Speech,” The Journal of the
Acoustical Society of America, vol. 87, no. 4, pp. 1738–1752, 1990.
13. Pratik K. Kurzekar, Ratnadeep R. Deshmukh, Vishal B. Waghmare, Pukhraj P. Shrishrimal,
“A Comparative Study of Feature Extraction Techniques for Speech Recognition System”,
IJIRSET Department of Computer Science and Information Technology, Dr. Babasaheb
Ambedkar Marathwada University, Vol. 3, Issue 12, December 2014.
14. Fang Zheng, Guoliang Zhang, and Zhanjiang Song “Comparison of Different
Implementations of MFCC”, The journal of Computer Science & Technology, pp. 582–
589, Sept. 2001.
15. Hossan, M.A. “A Novel Approach for MFCC Feature Extraction”, 4th International
Conference on Signal Processing and Communication Systems (ICSPCS), pp. 1–5, Dec 2010.
Movie Recommendation System: Hybrid
Information Filtering System
Kartik Narendra Jain, Vikrant Kumar, Praveen Kumar

and Tanupriya Choudhury
Abstract The movie recommendation system is a hybrid filtering system that

performs both collaborative and content-based filtering of data to provide recom-
mendations to users regarding movies. The system conforms to a different approach
where it seeks the similarity of users among others clustered around the various
genres and utilize his preference of movies based on their content in terms of genres
as the deciding factor of the recommendation of the movies to them. The system is
based on the belief that a user rates movies in a similar fashion to other users that
harbor the same state as the current user and is also affected by the other activities
(in terms of rating) he performs with other movies. It follows the hypothesis that a
user can be accurately recommended media on the basis others interests (collabo-
rative filtering) and the movies themselves (content-based filtering).
Keywords Hybrid filtering system Recommender K-means

Pearson correlation coefficient
1 Introduction
Movie recommendation systems are quite prevalent in today’s market as people

tend to spend a lot of money when they go to the movies or rent a movie, so they
need to make an informed decision about it.
K. N. Jain (&) V. Kumar P. Kumar T. Choudhury

Amity School of Engineering and Technology, Amity University
Uttar Pradesh, Noida, India
e-mail: knjain1995@gmail.com
V. Kumar
e-mail: vikrantpositive@gmail.com
P. Kumar
T. Choudhury

https://doi.org/10.1007/978-981-10-7245-1_66
678 K. N. Jain et al.
Over the past decade, a large number of recommendation systems for a variety of
domains have been developed and are in use. These recommendation systems use a
variety of methods such as content based approach, collaborative approach,
knowledge based approach, utility based approach, hybrid approach, etc.
Lawrence et al. (2001) described a purchase behavior personalized product
recommendation system that suggests new products to shoppers that shop in the
supermarket. This system was adopted by IBM research as a PDA based remote
shopping system called SmartPad. The recommendation was processed by content
based and collaborative filtering for refinement.
Jung et al. (2004) improved people’s search results by using previous ratings
from the people to provide current users who are having similar interests, the proper
recommendations. Their system works for a variety of items and collected their
rating regarding the validity of their search results. This encouraged people to
involve themselves in longer searches and use of more in-depth queries for their
requirements and the system used these new rating to refine further ratings.
MovieLens is an online movie recommendation system created by the
GroupLens research group that allows users to rate movies and utilizes rating from
other sources along the user rating to collaboratively filter the input and generate
ratings for recommendation to other people who have not seen said movies. They
allow for account management services which allows people to rate movies and
create profiles to receive recommendations based on their established tastes. This
system utilizes multiple streams of data to provide comprehensive and personalized
recommendations. This system can also be used over a mobile device online.
Websites and web services such as Netflix, IMdB, MovieLens etc. are some of
the popular recommendation systems that harness the power of data mining and
data analytics to analyze various factors associated with their users and the movies
and make recommendations based on them.
2 Literature Review
Michael Fleischman et al. equated the problem of limited information about user
preferences to one of content similarity measurement with the help of Natural
Language Processing. They described a naïve word-space approach and a more
sophisticated approach as their two algorithms and evaluated their performance
compared to commercial, gold and baseline approaches [1]. A. Saranya et al.
implemented a system based on probabilistic matrix factorization and provide users
to solve cold start problems. Cosine similarity is used to judge similarity among
users. They also implemented NB tree which is used to generate the user link
formation. Their system considers various ranges of movie ratings which are given
by review experts [2]. Hans Byström demonstrated a system in their paper by
implementing k-means clustering and SoftMax regression classification. The
baseline predictor showed an RMSE value as 0.90 and the system best achieved
result was with RMSE value as 0.884 which is 1.81% improvement to baseline
Movie Recommendation System: Hybrid Information Filtering System 679
value [3]. Eyrun A. Eyjolfsdottir et al. introduced a system implemented using

clustering and machine learning over a hybrid recommendation approach. Their
system inputs user’s personal information and with the help of the well-trained
support vector machine (SVM) models predicts user’s movies preferences.
Using SVM prediction their system selects and clusters the movies from the dataset
and generates questions. Then user answers those questions and their system refines
its movie dataset and outputs movies recommendation for the users [4]. Arnab
Kundu et al. developed a system by using collaborative filtering approach. In their
system clustering is achieved among users and their preferences with the help of
Expectation Maximization Algorithm [5]. Manoj Kumar et al. used collaborative
filtering approach in their paper MOVREC. This approach makes use of the
information provided by users. Their system then analyses the information and
movies with highest rating using K-means algorithm is recommended to the user. In
their system, user is provisioned to select attributes for the recommended movies
[6]. Kaivan Wadia et al. implemented recommendation system based on
self-organizing map (SOM). A SOM is defined as a neural network technique
which comes in the domain of supervised learning. They used content based
approach as well as SOM in their system for the categorization and retrieval of
information [7]. Karzan Wakil et al. implemented their system which is based on
human emotions. They used a hybrid approach which comprises of content based
filtering, collaborative filtering and emotion detection algorithm. The output of their
system provides better recommendation as it enables users to understand connec-
tion between their emotional states and the recommended movies [8]. Roberto
Mirizzi et al. implemented a Facebook application that semantically recommends
movies with the knowledge of user’s profile. To detect similarity among movies
Semantic version of Vector Space Model (SVM) is implemented by them [9]. Zan
Wang et al. described improvements of the typical collaborative filtering methods in
their paper. Their system uses collaborative filtering for extracting like-minded
user’s preferences. Then they used hybrid model of clustering which is a combi-
nation of Genetic Algorithms k-means (GA-KM) and Principal Component
Analysis (PCA). They implemented PCA to reduce the sparseness in the feature
space and GA-KM to reach to the global optimal solution [10]. Rupali Hande et al.
implemented using hybrid filtering approach in their paper MOVIEMENDER.
Various techniques like clustering, similarity and classification are implemented in
their system to result in better recommendation. Their system resulted with high
precision and accuracy as well as reduction in MAE score [11].
3 Proposed System
3.1 Architecture
The project is a web application that facilitates client server architecture, where the
user of the application as a client through a web-browser on their system and the
system resides on the server that responds to user’s request with appropriate
outputs.
The front end of the project is HTML 5 with CSS 3 with functioning and
interactive bits driven by JavaScript and JQuery. Bootstrap served as the provider of
the CSS and templates for the UI design. The Django web application framework
was used to design the application along with Jinja performing the functions of
dynamic HTML components.
The server is in charge of handling the client requests and authenticate clients by
utilizing the CSRF token provided by the Django authentication services. Our
application follows the Model-View-Template architecture wherein the model
represents database schemas, View contains the user interface and the Templates
are a combination of HTML and Jinja. The URL Dispatcher in Django handles the
mapping of templates to appropriate URLs.
The recommendation engine is designed using Python language as python
provides a plethora of libraries and data structures to make functioning of the
system viable and efficient (Fig. 1).

3.2 Working of the Hybrid Filtering System
The hybrid filtering system is the information filtering system that takes inputs of
ratings on movies from users and recommends movies by application of a content
based filter over a collaborative filter. The collaborative filter utilises the Pearson
Coefficient Correlation and the content based filter is our proposed algorithm that
utilises the genres associated with the active users’ profile and those associated with
the movies.
• The system first loads the movie-lens dataset into arrays for access by the
various functions.
• Upon loading the data set, the utility matrix is created which contains all the
ratings of each user to each movie. A similar test matrix is created for the test
data.
• Post the creation of matrices, the movie genres are used to create clusters for
clustering on items of the utility matrix the results of which are stored in matrix
‘clustered utility matrix’. This matrix is dumped to a pickle file for access of the
application.
• Following the clustering, user’s average ratings are found and stored in user’s
object for accessing later.
• Collaborative filtering is performed in the form of Pearson Correlation coeffi-
cient which is calculated between two users. This is done for the full range of
users and stored in the ‘PCS matrix’.
• After the creation of the similarity matrix, the system proceeds to the testing part
and employs guessing of rating of test movies given by users on a particular
movie by incorporating a set of similar users:
– First, the system access the similarity matrix to find the top certain number of
users to the active user and normalise their scores by their average scores
considering they have rated the movie in question.
– The system computes the score for the movie by adding the average of the
scores of the top users to the average of the current users if there were any
similar users.
– Following this computation, the proposed algorithm is applied and compu-
tations are made according the active user and movie to be rated to compute
a new score.
– The new score is compared with the actual score and after repeating process
for the complete test set, the root mean square error is computed (Fig. 2).
3.3 Proposed Algorithm
The proposed algorithm is the calculation of score of movie for a particular user
which utilizes the preference shown by users to movies of a range of genres and the
Fig. 2 Working of proposed system
association of those genres to the genres of the movie to be recommended to the

user. The algorithm considers weights for different genres that all movies conform
to for a user. Based on these observations a new score is computed for each movie.
n Xn
x X z z0 n
þ þ ¼ x0 : ð1Þ
y 0
2y 2
0
2y 02 m
where,
x′ the new calculated score of movies.
x the score derived by computing similar users and altering the movie average.
z weight of genre for the user that matches with his genre preference and movie.
z′ weight of genre for the user that does not match with his genre preference.
y number of genres matched between user’s profile and movie’s profile.
y′ number of genres not matched between user’s profile and movie’s profile.
n number of genres matched with preferred movies.
m number of genres that a user prefers based on his past rating.
The proposed algorithm employs a concept similar to genre correlations but does
so between the user and the movie items instead of performing item-item collab-
oration approach.
4 Data
The system works on the movie-lens 100k dataset which consists of 9125 movies,
100,005 ratings provided by 671 users. The movies are classified with 20 genres.
The ratings file consists of the user-id, the movie-id’s they are rating along with
the rated score and timestamp of scoring.
The test and train data consists of 10 and 90% distributed data of the ratings file
respectively.
The items file consists of the movies with their movie-ids, names along with the
year of release, and the genres they are attached with which are represented as 1s for
true and 0s for false. They correspond to a sequence of that is consistent throughout
the data.
5 Results
The application results in a set of genres that are recommended to the user based on
his ratings on the random movies. This factors in the genres attached to the random
movies and weighted according to the rating the user gave it along with the pref-
erence of genres by users that have similar taste to the active user. The number
along the genre so specifies the degree that the genre is recommended to the user.
The more times a genre appear, the more it has been recommended to the active
user.
The movies are the top movies that have been generated upon the computation
of similarity to the other users and the proposed algorithm on the original ratings to
generate new ratings that are specific for the user. The recommended movies need
not be the highest rated as evidenced by the presence of ‘Oliver Twist’ which is a
4-starred movie but since the recommended genres according to the user is heavily
inclined to ‘adventure’ and a bit of “children’s” theme and ‘Oliver Twist’ happens
to be a ‘great adventure movie’, hence it has been recommended to the user instead
of a much higher rated movie (Fig. 3).
The recommender engine shows Mean Squared Errors in the range of 0.96–1
based on different sets of test data within the movie set. The resultant graph is based
on an iteration with a mean squared error of 0.980763 and it shows the ratings
originally given by user to movies and the ratings predicted by the system for the
same (Fig. 4).
Fig. 3 Recommendation output page, genres and movies recommended

Fig. 4 Graph showing actual ratings (red) and predicted ratings (blue)
The graph below shows the rating provided by the users and the original ratings
for the first 100 user-movie items of the test set (Fig. 5). Here we can see that
considering rating 3 to be a central point (i.e. normalizing about 3), the ratings
provided by the system are in the correct direction and differs by the magnitude
only. This implicates the fact that since the system is predicting a new rating for
every movie, although it does not match with the original rating, the variation
between the movie set is consistent and hence it approximately makes the correct
predictions. Another thing to note is that the initial movie-user sets, upon reflection
on the data are heavily populated by ratings and well defined features (varied
number of genres). In comparison to this the graph displaying the same data for the
last 100 set (Fig. 6) shows predictions in the correct direction but with a more
varied magnitude. This can be attributed to the fact that the last set of movies were
Fig. 5 Graph showing actual ratings (red) and predicted ratings (blue) for first 100 ratings
Fig. 6 Graph showing actual ratings (red) and predicted ratings (blue) for last 100 ratings
Table 1 Movielens 100 k Method RMSE MAE

database worked on using
different techniques UserItemBaseline 0.96754 0.74462
UserKNNCosine 0.937 0.737
ItemKNNCosine 0.924 0.727
less populated with ratings and also had a lack of refinement in terms of descriptors
(lesser or no genres defined).
This observation leads us to believe that the system should work much better
when a better quality of data in terms of sparsity would result in a better RSME
value and better predictions (Table 1).
The table above specifies the mean squared error and root mean squared error on
applying user-item filtering, cosine similarity between users and KNN and Item
similarity using Cosine and KNN. The results show that our method having an
average mean squared error of 0.980763 is comparable to these techniques con-
sidering they might have worked on a similar iteration of the dataset as similar to
ours and that our system should work better with data having more user ratings.
6 Conclusion
The result shows that the hybrid system works within a respectable margin of error
for recommending movies to users. The user can rate some movies to get recom-
mendations of movies that suits the taste of the user instead of getting the most
trending movies. The proposed algorithm accomplishes in this regard and couples
over the similarity scores to add genre weighting in the mix. The fact that most of
the predictions are at least in the correct direction are indicative towards the
accuracy of the system. The results show that employing more dense data is bound
to increase the accuracy of the recommender and increasing the number of genres
will help in making more specialized recommendations. The system also establishes
the viability of hybrid recommenders as good and effective recommender.
Movie recommendation has always been to a certain extent, a subjective opin-
ion. Users tend to exhibit responses to recommendations that is at great odds from
their past activities. This makes it considerably hard to evaluate recommenders
mathematically and always involve a large margin of error. The system needs to be
deployed in real-life scenario to receive manual responses of accuracy to judge the
full extent of the system.
References
1. Michael Fleischman, Eduard Hovy. “Recommendations without User Preferences: A Natural

Language Processing Approach”. In proceedings of the 8th international conference on
Intelligent user interfaces, pages 242–244. Miami, Florida, USA, 2003.
2. A. Saranya, A. Hussain. “User Genre Movie Recommendation System Using NB Tree”. In
proceedings of the International Journal of Innovative Research in Science, Engineering and
Technology. Vol. 4, Issue 7, July 2015.
3. Hans Byström. “Movie Recommendations from User Ratings”. Stanford University, 2013.
4. Eyrun A. Eyjolfsdottir, Gaurangi Tilak and Nan Li. MovieGEN: “A Movie Recommendation
System”. Technical report, Computer Science Department, University of California Santa
Barbara, 2010.
5. Debadrita Roy, Arnab Kundu. “Design of movie recommendation system by means of
collaborative filtering”. Int. J. Emerg. Technol. and Adv. Eng, 2013.
6. Kumar, M., Yadav, D.K., Singh A., Gupta and V.K. “A movie recommender system:
Movrec”. In International Journal of Computer Applications, ACL 2002 Conference on
Empirical Methods in Natural Language Processing, vol. 124, pp. 0975–8887, 2015.
7. Kaivan Wadia, Pulkit Gupta. “Movie Recommendation System based on Self-Organizing
Maps”. The University of Texas at Austin, Austin, Texas, 2011.
8. Karzan Wakil, Rebwar Bakhtyar, Karwan Ali and Kozhin Alaadin. “Improving Web Movie
Recommender System Based on Emotions”, International Journal of Advanced Computer
Science and Applications, vol. 6, no. 2, 2015.
9. Roberto Mirizzi, Tommaso Di Noia, Azzurra Ragone. “Movie recommendation with
dbpedia”. In CEUR Workshop Proceedings, vol. 835, 2012.
10. Wang Z, Yu X, Feng N and Wang Z. “An improved collaborative movie recommendation
system using computational intelligence”. J Vis Lang Comput 25:667–675, 2014.
11. Rupali Hande, Ajinkya Gutti, Kevin Shah, Jeet Gandhi and Vrushal Kamtikar.
“MOVIEMENDER - A MOVIE RECOMMENDER SYSTEM”, International Journal of
Engineering Sciences & Research Technology (IJESRT) (Vol. 5, No. 11), 2016.
Eco-Friendly Green Computing
Approaches for Next-Generation Power
Consumption
Seema Rawat, Richa Mishra and Praveen Kumar
Abstract The term Green Computing is combination of two words which basically
means usage of electronic devices in an efficient way. Green computing is a trend/
study which came lately and is about the building, designing and operating elec-
tronic devices to maximize the efficiency in terms of energy. Green Computing
includes the manufacturing and disposing of electronic devices including the
computer resources such as the monitor, storage devices, CPU and associated
subsystems etc. without causing any ill effect to the environment. Also, one of the
important aspect of green computing is recycling of the devices which are used
worldwide in IT industry. This research paper highlights the advantageous use of
green computing and its associated approaches through a survey conducted over
Power Consumption in Amity University.

Keywords Green computing Eco-friendly Resource utilization

Power consumption Recycling Benefits
1 Introduction
The idea of Green Computing was given by Environment Protection Agency in

year 1992, when they launched a program called “Energy Star”, which promoted
hardware efficiency. All the operations which need sustainability of environment
comes under the scope of green computing [1, 2]. Till date, the IT industry has
delivered a lot of innovative products but on the other hand, it has also been a topic
of major concern if we consider global issues such as global warming, climatic
S. Rawat (&) R. Mishra P. Kumar

Amity University Uttar Pradesh, Noda, India
e-mail: srawat1@amity.edu
R. Mishra
e-mail: richa.m18@gmail.com
P. Kumar

https://doi.org/10.1007/978-981-10-7245-1_67
688 S. Rawat et al.
changes etc. A lot of factor are there to prove that Green Computing is an important
solution to be followed as it minimizes the cost related to power management.
Minimizing the electricity consumption which is risky for environment is the main
aim of green IT. This model aims towards the reduced usage of electronic material
that have negative impact on the environment. Increased knowledge and awareness
about the environment has caused the industries all around world to recheck their
environmental credentials.
2 Investigating the Power Consumption
This article focuses on calculating the total amount of power consumed by the
desktops on an early basis [3]. Different costs including the cost of computing, cost
of cooling process etc. has been provided in this research paper which also keep
track on the amount of heat generated through the desktops. Further, this study
throws light on the impact which the increased computing power has on the
environment, it emits gases like Sulphur, CO2 etc. The survey was conducted over
the power consumption in Amity University, Noida which approximately has
10,000 desktops. With the help of meters, we were able to gauge the power con-
sumption which was around $150,000, excluding the cooling process of these
desktops. On the other hand, including the price of cooling process it would go
around $220,000. To operate the desktop, we require power plant to generate power
which in turn produces harmful gases like CO2, mercury etc. This huge quantity of
electricity initiates the downfall of environment. Here we are giving solutions to
reduce the power consumption.
2.1 Advantage of Using Thin Clients
Currently organization have multiple servers and majority of them not even being
used completely. Typically, server uses around 5–20% of their full capacity, even
then they consume full energy. An efficient solution to the problem is using Thin
Clients which consume only 8–20 W of energy in comparison to a typical PC
which uses about 150 W. Thin clients maintain the environmental decorum by
minimizing the consumption of energy. Disposing a PC is way too high as it
contains heavy materials used in the hardware of system. A thin client uses a
common or shred terminal which reduces the power consumption and provides high
efficiency. Some of the pros which makes thin clients reliable [4].
• Low energy consumption.
• In comparison to PCs, they require lesser cost in transportation as they are made
up of lightweight material.
• More availability.
Eco-Friendly Green Computing … 689
• Little disposals cost.

• Cost of terminals are relatively low.
• Operational cost is low.
One of the examples of thin client is Sunray. It is a thin client which requires 4–8
watts, whereas intake for a video card is 12–15.
2.2 Consideration for General Changes
Green computing is not something which can be easily achieved. A collaborated

attitude is needed to attain the same. The toxic substances must be urgently replaced
to prevent environment from degradation [5, 6].
• Avoiding of plastic cases made up of brominated flame retardant (BFR).
• Using low capacity flash based drives instead of solid state drives.
• Controlling of landfills through periodic upgrading and repairing of resources.
• Organic light diode and Green LED should replace the CRTs displays.
• Substituting harmful chemicals like copper, lead etc.
• Avoiding the nonproductive mechanism to save energy resources.
• Instead of CPUs, GPUs can be considered as better option as it reduces power
consumption while increasing the efficiency [7]. CPUs and GPUs data in idle
state and 100% load state (Table 1).
2.3 Converting a Company into Green Company
Converting a company into sustainable business or green company can be achieved

if a successful plan is followed. A company can be turned into a green company by
Table 1 Power consumed by CPUs and GPUs [7]

S. No. CPU/GPU used Idle state (W) 100% load state
1. GPU 26 45
ATI RADEON X1900XT
2. CPU 206 318
Pentium4XE@ 2.30 GHz
3. GPU 43 0
ATI RADEONHD 6990 CF(4)
4. CPU 32 65
Intelcorei3@ 2.40 GHz
5. GPU 23 64
ATI RADEON HD 6950 OC(900/1440)(32)
6. CPU 38 95
Intelcorei5@ 3.10 GHz
690 S. Rawat et al.
Fig. 1 Formation of green company [9]
some simple and easy guidelines [8]. A sustainable development procedure should
of following processes—reusing, environmental purchase decisions, energy con-
servation and reviewing the paper consumption (Fig. 1). These processes cope up
with each and in result form a green company which has silent and green infras-
tructure [9]. There are two major concerns attached to green computing—pollution
control and reduction in energy consumption. While the first can be achieved
through efficiently using electronic goods, later can be achieved by their reduced
usage and proper recycling.
3 Feasible Green Computing Trends
Several researchers are trying to tend towards Green Computing. Research on

Green computing leads to concepts that allow the usage of green concept in
computing domain. Some practices adopted from green computing such as use of
small screen size devices, using computer systems having ENERGY STAR label on
them etc. [10]. Following topics are further discussed in the paper.
3.1 Virtualization
One of the main concept which fall under green computing is Virtualization.
Virtualization is basically about the extraction of computer resources like using two
or more logical systems on a single set of physical hardware. Virtualization is a
green computing trend which offers both the management and the virtualized
software. Virtualization is performed at different levels like storage, desktop and
server level [10, 11]. In the virtualized environment, a single server can be used
up-to its full utility. Through the virtualized datacenters, hardware costs can be
reduced up to the half of original. But if we talk about virtualized servers, a server
room is there which requires minimum maintenance services and minimizes real
estate savings. In datacenters, the virtualized servers have been replaced by the
blade servers which result in more than 70% reduction in consumption of energy
[12]. It allows complete utilization of the resources and provides various benefits
like
1. Reduced amount of hardware.
2. Reduced amount of space, rent and ventilation required etc.
3.2 PC Energy Usage
Before handing over the systems to the end users, companies should be wholly
responsible to enforce recommended power settings. Allowing the monitor to sleep
mode after it has been idle for long is a good practice in energy saving. Also, a lot
of experiments have been performed as to check the performance of PC while in
sleep mode and in idle mode (Graph 1).
Graph 1 Power distribution

of different modes [14]
692 S. Rawat et al.
3.3 Recycling Devices
Green computing should consider the product life cycle, from producing it to its
operation and lastly recycling it. For sustaining environment recycling of cadmium,
lead etc. is vital. The computer systems which have completed their lives can be
re-purposed, reused or can be given to the needy one’s (Fig. 2). Parts of systems
which are no more in use or those which are outdated can be recycled. The com-
puting supplies like batteries, paper, cartridges etc. can be recycled and could be
converted from a total waste into useful objects. There is a constant requirement for
new technologies and high tech products which results in huge amount of e-waste
on a regular basis [1].
Such practices would help us in fewer costs, space saving low carbon emission,
reduced power supplies etc. (Fig. 3). Some practices which promote green com-
puting has been introduced in our education system as well and are discussed below
[13]:
• Online learning methodology.
• Software based on intranet portals for attendance system, fee payment etc.
• E-admissions process.
• College e-brochures.
Fig. 2 Reusing of electrical components [15]

Fig. 3 Advantages of green

computing [16]
4 Conclusion
This paper highlights a lot of loopholes of the electronic industries which are highly
hazardous to the environment as well as to human being. Technology isn’t a passive
observer, rather it is achieving green computing goals by being an active observer.
Thin client cables, has reduced energy consumption up to half of the original
consumption. Virtualization is another concept covered in this paper, it provides
maximum utilization of the resources which results as a major advantage. Another
aspect which has been covered here in this paper is about recycling devices.
Recycling can improve the environment as well as the ecosystem on a huge basis.
How reusing a system helps in saving a lot of energy and maintaining the ecology
has been covered in this paper. There are enormous number of challenges to
achieve the same. Also, the government is putting appreciable amount of effort. But
it’s a matter of huge concern as it cannot be achieved in a day or two. The overall
approach of promoting Green Computing is to control environmental deterioration,
enhanced life expectancy of devices etc. Implementing Green Computing trends is
not the responsibility of a single person, instead it is something which can be
achieved only when all of us work together in a positive way. The needs of present
and future generation can only be fulfilled if we provide Eco-friendly environment.
Therefore, I hope this paper works in support of development and enhancement of
green computing based organizational infrastructure.
References
1. Gingichashvili, S., “Green Computing” 2007.

2. Tebbutt D., Atherton M. and Lock T., “Green IT for Dummies”, John Wiley and Sons Ltd.
2009.
3. Mujtaba Talebi, “Computer Power Consumption Benchmarking for Green Computing”,
Villanova University, April 2008.
694 S. Rawat et al.
4. Aruna Prem Biazino, Claude Chaudet, Dario Rossi and Jean-Louis Rougier, “A Survey of
Green Networking Research”, IEEE Communications surveys and tutorials, vol. 14, no., First
quarter 2012.
5. Praveen Kumar, Dr. Vijay S. Rathore “Improvising and Optimizing resource utilization in Big
Data Processing “in the proceeding of 5th International Conference on Soft Computing for
Problem Solving (SocProS 2015) organised by IIT Roorkee, INDIA (Published in Springer),
Dec 18–20, 2015, pp. 586–589.
6. Seema Rawat, Praveen Kumar, Geetika, “Implementation of the principle of jamming for
Hulk Gripper remotely controlled by Raspberry Pi “in the proceeding of 5th International
Conference on Soft Computing for Problem Solving (SocProS 2015) organised by IIT
Roorkee, INDIA, Dec 18–20, 2015, pp 199–208.
7. Y. Navneeth Krishnan, Vipin Dwivedi, Chandan N. Bhagwat, “Green Computing using
GPU”, IJCA, vol. 44, no. 19, April 2012.
8. http://www.berr.gov.uk/files/file35992.pdf.
9. Yan Chen, Shunqing Zhang and Shugong Xu, “Fundamentals Trade Offs on Green Wireless
Networks”, IEEE Communications Magazine June 2011.
10. Payal Mittal, Seema Rawat, Praveen Kumar “Proof Based Analytical Comparison of Energy
Parameters between Android and iOS Platform for Heterogeneous Cloud Based Application
in International conference on Computational Intelligence & Communication Technology
(CICT), ABES Ghaziabad UP on 12–13 Feb. 2016.
11. S.K. Fayaz Ahamad, P.V. Ravikant, “Green Computing Future of liveliness”, IJCER ISSN:
2250–3005.
12. Annapurna Patil, Abhinandan Goti, “Green Computing in Communications Networks”, IJCA
vol. 68, no. 23, April 2013.
13. Shalabh Agarwal, KaustuviBasu, Asoke Nath, “Green Computing and Technology Based
Teaching and Learning Administration in Higher Education Institutions”, IJACR vol. 3, no. 3,
issue. 11, September 2013.
14. J.V. Mashalkar, S.K. Kasbr, M.B. Bhatade, “Green Computing: A Case Study of Rajarshi
Shaher Mahavidyalaya, Latur, India”, IJCA vol. 62, no. 2, January 2013.
15. Riyaz A. Sheikh, U.A. Lanjewar, “Green Computing- Embrace a Secure Future”, IJCA vol.
10 no. 4, November 2010.
16. S. Murugesan, “Going Green With IT: Your Responsibility towards Environmental
Sustainability” Cutter Business- IT Strategies Executive Report, vol. 10, no. 8, 2007.
Implementing Test Automation
Framework Using Model-Based Testing
Approach
Japneet Singh, Sanjib Kumar Sahu and Amit Prakash Singh
Abstract There are various stages present in the process of software testing life-
cycle which starts mainly with the requirement review process, proceeds with test
planning, test designing, test case development, and execution of test cases and
ends with test reporting phase. This paper will emphasize mainly on the application
of model based testing with test automation frameworks for automating the test
designing and test development phases of software testing lifecycle. In the test
development phase it mainly provides information about using model based testing
in the activity related to the development of test automation scripts for execution of
tests. This mainly tells about the steps involved in the implementation of test
automation frame work using the approach of model based testing. Test develop-
ment and test designing both can be automated to decrease the amount of human
effort and cost for the same in the project. The development of test scripts which
includes both automated and manual tests can be automated using the concept of
model based testing. In this approach firstly the model is created for capturing the
behaviour of the system which is under test. After this the model based testing tool
parses the same to create the manual testing scripts. The model based testing tool
then in addition generates the automated test scripts for automated test execution
and can also be integrated with popular tools and test automation frameworks. So,
using a model for automating the creation of both automated and manual test scripts
not only saves cost and thereby effort, but also increases the amount of coverage
and reduces significant amount of time for product to go to market.

Keywords Test automation framework Model based testing Software

requirement specification Quality assurance Graphical user interface
J. Singh (&) S. K. Sahu A. P. Singh

USICT, Guru Gobind Singh Indraprastha University, Dwarka, Delhi, India
e-mail: japneetheyer@yahoo.in
S. K. Sahu
e-mail: sahu_sanjib@rediffmail.com
A. P. Singh
e-mail: amit@ipu.ac.in

https://doi.org/10.1007/978-981-10-7245-1_68
696 J. Singh et al.
1 Introduction
The process of model based testing can be defined as an application of model based
design used for execution of automation and manual based test cases to perform
software testing and also used for test designing purposes. It can be stated that the
models can be used for the representation of desired behaviour of the system which
is under testing and also to represent the test environment and test strategies.
The general approach of model based testing has been discussed in Fig. 1. In this
approach the model which is created is a partial description of the system which is
under testing. The test cases which are derived from such a model will always be
functional test cases. These test cases are also known as abstract test cases. These
abstract test cases suite cannot be directly executed on the system which is under
testing. So an executable test suite has to be derived from the corresponding abstract
test suite cases. These tests then can directly be executed on the system under
testing. This can be achieved by associating the abstract test cases to the specific test
cases which will be suitable for execution.
This paper mainly focuses on the application of model based testing approach
for automating the test development and test designing phases of software testing
lifecycle. In the test development phase the model based testing is mainly imple-
mented for the activity of developing test automation scripts. This approach mainly
helps us in identifying steps to build the test automation framework using the model
based testing.
Also, the test designing process can be called as the foundation phase for the
software testing as this phase involves the analysis of the requirements or specifi-
cation for the system and then generating the test cases using the same and vali-
dating them against the system that documented requirements or the specification
have been met. This requires human effort so the model based testing approach can
be used for automating the process of test designing as well to reduce the amount of
effort required to write these test cases manually.
Fig. 1 Components of model

based testing
Implementing Test Automation Framework Using Model … 697
Figure 1 depicts the general approach of model based testing.

The roadmap for the remainder of the paper will be as follows. In Sect. 2 we
have discussed about the related work being done in the area of implementing
model based testing in software testing lifecycle process. Section 3 talks about the
proposed solution and features of implementing test designing using model based
testing. Section 4 talks about the proposed approach of test automation using model
based testing and integration of test automation frameworks [1]. Finally the paper
has been concluded and the future scope has been defined in Sect. 5 of the paper.
2 Related Work
There have been large number of advancements happening in the field of software
testing. The process of capturing the requirements by creating models was intro-
duced around twenty years back which is around two decades ago. But, the talk
around using models for the process of software testing have been there around the
last decade. There are good amount of tools like Test optimal, ConformIQ, Math
Works and many more available today in the market which helps users in the
creation of models and also further help in generating the test cases and test scripts
by processing these models.
The test automation practises have also developed over the last ten years. As the
linear scripts were developed earlier but now these have been replaced by the
structured test scripts which use the test libraries, modifying and automating these
test cases also require programming knowledge. There are new frameworks
available today like keyword driven and data driven frameworks which have helped
the testers to reduce the complexities in the test scripting [2]. All these advance-
ments has helped in reducing the time, effort and cost required in the process of
software testing. These advancements have also allowed the various software
vendors to become aggressive and pretty much competitive in the market.
Model based testing usage is being done in various software processes, devel-
opment organizations, application stacks and application domains. These includes
commercial providers, transaction processing and open source providers. These also
include implementation of programming languages like Python, C, C++, and Java
and also include software development lifecycle models like V-model, Agile and
Incremental model.
It is also true that testing itself is a noteworthy process of software development
lifecycle in terms of both time and cost. All the features of a system need to be
tested not only when the feature is introduced for the first time in the release but
also whenever the next release comes up with the same set of features with some
enhancements in it. So, the regression test cycles are very long and they do get
repeated monthly or quarterly which increases the cost and effort required to do this
testing manually.
So, the companies have been looking for an alternative solution in terms of
performing automation testing to save the time and cost of the project. The model
698 J. Singh et al.
based testing process helps in the same way by making two specific advancements
of automating the test design and test automation processes helping software
vendors to automate the process of testing the software products [3].
3 Solution for Test Designing Using Model Based Testing
Test designing along with the process of test automation are very much specialized
skills. These are always in less amount of supply. The test designing process
involves the creation of test cases and also requires good amount of domain
knowledge along with the general information related to software development.
The model based testing has now been accepted as a methodology for the
purpose of test designing and test development as illustrated in Fig. 2. In the test
development process the model based testing can be used for developing the
manual test scripts as well as the automation based test scripts. From this statement
it can be infer that model based testing can be used to build test automation
frameworks.
The steps involved in the software test lifecycle can be represented in the form of
a Pictorial representation implementing model based testing as illustrated in Fig. 2.
In a traditional test designing approach test cases are generated or written by
going through the software requirement specification document which helps us in
determining what all needs to be tested. Then tester design the test cases to test each
of the listed requirements.
3.1 Features of the Solution
The user can start the process of implementing model based testing using the
software requirement specification document by modelling the specified charac-
teristics of the software under test. The model can be created in the form of a
Fig. 2 Test automation across software test lifecycle using model based testing
markov chain or a state chart or in the form of finite state machine and are also
represented using the UML diagrams.
The model can be used for representing the various requirements to a particular
extent depending upon the focus of testing. It may also be very closer to the
pictorial representation of requirements or a form of replica at a very small scale.
Due to higher focus on the requirements it looks like a description of the
requirements. The model helps in representing the key requirements. Then Model
based testing tool is used for carrying out parsing on the model and helps in
generating the test scripts and test cases for purpose of manual execution [4].
The model based testing tools additionally supports various coding languages
like Python, Java and many more object oriented languages and model based
languages such as Unified Modelling language which is used as a basis for the
model based testing [5].
There are variety of model based testing tools available in the market today
which can be used for implementing this type of testing which are listed as
Conform IQ creator, JSXM, Model J unit and many more. These tools also support
the code based model, helps in parsing of the code, creation of test scripts and test
cases for the purpose of manual execution [6].
Figure 3 represents a model displaying the functionality of login page for web
based application.
4 Proposed Approach for Test Automation Using Model

Based Testing
The model based testing technique can also be used for developing the automation
test scripts. The manual test scripts which have been generated using model based
testing as a part of test designing process can easily be converted to automated test
scripts for automatic execution. It is a type of fourth generation based test
automation. It also supports requirement and defect identification, field defect
prevention and automated generation of tests from the models which in turn helps in
reducing the cost, effort and eliminates manual test designing [7].
Model based testing consist of majorly three steps in it which will be explained
by using the model taken for the login functionality of web page in Fig. 3.
1. Creating a model suitable for execution: This step is required for the diagram
based models generated in the form of finite state machine. Many model based
testing tool can be used for the same purpose [8]. If the data is correctly
associated and all the dependencies and constraints have been defined then it is
an executable form of model. The coding is the best way to implement the
association and it can easily be simulated as shown in Fig. 4.
Test data associated with the model:
700 J. Singh et al.
Fig. 3 Model representing

the functionality of the login
page of web based application
2. Furnishing the test cases from the model to create the test scripts for the purpose
of test execution: Generally most of the model based testing tools support the
process of furnishing or rendering using popular execution tools like Unified
Functional Testing or Selenium and also supports unit testing based frameworks
like JUnit.
The Model based testing tool also helps in creating the executable script with
good amount of code for few of the steps and placeholder like providing additional
graphical user interface details for the same. There is always a scope of further
Fig. 4 Representation of the association of test data with the model
Fig. 5 Representation for the stub written for test automation
modification in the stubs created before they can be executed as a complete auto-
mated test. This is further illustrated in Fig. 5.
Stub written for test automation:
3. Creating the scripts which are automatically executable and are complete tests:
The small automation stubs which have been generated using the model based
testing tool can be further developed for automated execution using automation
702 J. Singh et al.
tools like unified functional testing, Selenium and many more available in the
market today. This thing is achieved by using the record and play back feature
of the right model based testing tool which is based up on the domain of the
system or software under testing, also depends up on the capability of the tool to
integrate with popular test automation frameworks and helps in generating the
automated test scripts to build an automated test framework.
4.1 Integration with Test Automation Frameworks
1. Keyword driven framework: It is a type of software testing framework which

separates the work of test automation involving test programming from the
given test designs of a product. This technique is mainly based up on the usage
of keywords for the basic functions or operations performed on the system
which can include closing, opening, and button clicking, enter and many more
[9]. Then the listed keywords are converted into automated test scripts for an
automation tool to execute them. The process can be better understood by taking
the example of web based login page functionality and determining the key-
words for the same (Table 1).
2. Data driven framework: It is a type of framework which mainly revolves around
data which is stored in a database or in the form of spreadsheets. It is a very
useful approach as a single test can be tested multiple times with different test
data which avoids huge number of test cases to be written for testing the same
type of scenario which only varies with data [10]. Model based testing tool
integrates with this framework by doing association of test data with different set
of states and actions. It is really useful when the system under test needs to be
tested with good number of input values (Table 2).
The data set represented above is then read in the script to test the similar
scenarios by providing different set of values and leads to good coverage.
Table 1 Keywords or test directives generated for testing login page script
Graphical user interface Locator Action based Parameters
window keyword
Login web page Browser Open Loginwebpage.com
Login web page Browser Enter Admin, P@ssw0rd
Login web page Login Click Login Button
button
Login web page Browser Verify Home Page
Table 2 Test data generated LoginUsername LoginUserPassword Result

for login page script
administrator admin False
administrator P@ssw0rd True
admin admin False
admin123 password False
Model based testing can help in the establishment of test automation framework
around it which begins with the process of test design automation and is followed
by the manual test scripts development. This type of testing also allows us to
integrate with different automation frameworks to achieve test script designing and
also integrates with variety of tools for supporting automated execution of tests.
Such a framework helps the organization to build there infrastructure for automated
test designing and automated test execution [11]. There are many benefits of model
based testing which mainly involves the exhaustive coverage of test scenarios in the
form of a model, ease of maintenance of test cases and test scripts, expected shorter
cycle time which helps in the faster execution of test cases and also helps in the
elimination of need to modify the code which prevents the occurrence of human
error in larger scripts.
It also helps the organizations in delivering the test automation at faster rate
which can be utilized during the regression cycles and helps in saving time, effort
and cost of the project. There are huge amount of benefits of model based testing
which are derived by the organizations at smaller and larger level in terms of easy
maintenance of automation and exhaustive coverage of test cases. The future work
mainly lies around the area of analyzing the relative importance of model based
testing and how it can be implemented more efficiently in the industry. Also, the
future work can also involves the work of analysing the model based testing and its
integration with more number of automation frameworks to efficiently test a soft-
ware product.
References
1. Fei Wang and Wencai Du: A Test Automation Framework based on Web. IEEE/ACIS 11th
International Conference on Computer and Information Science (2012)
2. Pekka Laukkanen: Data Driven and Keyword Driven Test Automation Frameworks (2006)
3. Rashmi and Neha Bajpai: A Keyword Driven Framework for Testing Web Applications.
(IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 3,
No. 3 (2012)
4. Model-Based Testing, https://en.wikipedia.org/wiki/Model-based_testing
5. Yan Li and Li Jiang: The research on test case generation technology of UML sequence
diagram. IEEE 9th International Conference on Computer Science & Education (2014)
704 J. Singh et al.
6. Sebastian Wieczorek, Vitaly Kozyura and Matthias Schur: Practical model-based testing of
user scenarios. IEEE International Conference on Industrial Technology (2012)
7. Test Automation, http://en.wikipedia.org/wiki/Test_automation
8. Naveen Jain: Model-Based Testing, HCL Whitepaper. http://www.hcltech.com/white-papers/
engineering-services/model-based-testing (2013)
9. Anuja Jain, S. Prabu, Swarnalatha Muhammad Rukunuddin Ghalib: Web-Based Automation
Testing Framework. International Journal of Computer Applications (0975 – 8887) Volume
45– No. 16 (2012)
10. Susanne Rӧsch, Sebastian Ulewicz, Julien Provost, Birgit Vogel-Heuser: Review of
Model-Based Testing Approaches in Production Automation and Adjacent Domains
Current Challenges and Research Gaps. Journal of Software Engineering and Applications
(2015)
11. Ambica Jain: MBT, A superior alternative to traditional software testing (2012)
Improved Exemplar-Based Image
Inpainting Approach
Hitesh Kumar, Shilpi Sharma and Tanupriya Choudhury
Abstract Image inpainting is a very common technique which is widely used to

bring back or recover an image when it gets destroyed, or when there is an intention
to perform a morphological operation on an image. It is a simple process to fill up
the pixels of a particular region. The task is very similar to that of a skilled painter
who has to draw or remove an object from its painting on its canvas. Inpainting’s
application may vary from its usage which includes object removal, object
replacement etc. The aim of this process is to change the image properties by
removing or adding objects. It has been also used to recover lost pixels or block of
pixels of an image during a transmission of image through a noisy channel. It has
been a useful method for red-eye removal and default stamped date from pho-
tographs clicked through primitive cameras. In this paper, we have built an
exemplar-based image inpainting tool using basic functions of MATLAB. We have
discussed the results by comparing the output generated by the image inpainting
tool with Adobe Photoshop results to compare the results and check how much
efficient the tool is.

Keywords Inpainting Image restoration Canny edge detection

Object replacement Object removal Exemplar-based inpainting
H. Kumar (&) S. Sharma T. Choudhury

Computer Science and Engineering Department, ASET, Amity University,
Noida, Uttar Pradesh, India
e-mail: hitesh2194@gmail.com
S. Sharma
e-mail: ssharma22@amity.edu
T. Choudhury

https://doi.org/10.1007/978-981-10-7245-1_69
706 H. Kumar et al.
1 Introduction
In computer science, this similar task is done by filling the damaged or corrupted
pixels by the process of inpainting. There are multiple techniques and processes to
do this task [1]. The techniques come from the area field of image inpainting and
artificial intelligence. During taking a photograph, some problems are encountered:
i. Sometimes while sending or receiving an image, the pixels get corrupted or
damaged
ii. While taking a photograph, there are sometimes unwanted objects that come in
the frame which was not meant to be there.
To cope up with this problem, inpainting process comes into play and the desired
image is obtained by removing the object and filling the cavity with desired pixel
values. After the process, an image obtained must be plausible and the viewer must
not able to detect any changes made to the output image [2].
In Fig. 1, the left image is corrupted and the right image is being inpainted to its
original form. The red squares have been removed by the inpainting algorithm and
the original image has been restored. Various Inpainting techniques include the
following:
1.1 Texture Synthesis
Texture synthesis or TS inpainting takes into account the structural definition of the
image. It tries to construct a large image from a small digitally defined image. It
helps in expanding the image small image, filling cavities/holes for inpainting
purpose and to fill the background of an image. Here, the texture can be either
‘regular’ or ‘stochastic’ [3]. Regular texture comprises similar patterns over an
image. The example is as follows (Fig. 2).
Fig. 1 Inpainting example

Improved Exemplar-Based Image Inpainting Approach 707
Fig. 2 Regular and

near-regular texture [4]
Following points must be taken care while using this technique:

(a) the output size of the image must be mentioned by the user
(b) the input image size should be similar to the output image size
(c) the image should be free from complex structures and artifacts.
1.2 Partial Differential Equation Based
PDE or Partial Differential Equation is an iterative algorithm that takes the infor-
mation of the region by moving forward in those regions where the change is less.
This algorithm is very effective where the corrupted regions are smaller in size [5].
When dealt with larger regions, it will take a longer time to process the image. Also,
the results calculated will not be satisfactory. The total variation using the concept
of differentiation is calculated for change in information in the image. This algo-
rithm fails to restore those image which is partially degraded [6].
1.3 Exemplar and Search Based
Exemplar and search-based inpainting were initially introduced by Harrison [7]. The
empty regions of the image were filled with the texture around the patch being created
after the object removal. In other words, in exemplar and search based, the locality of
the empty patch being created is taken into consideration and for the pursuit of
information, the locality is being explored for the desired information. This is suitable
for images having a linear structure. This technique works in the following manner.
708 H. Kumar et al.
2 Literature Review
Inpainting is a technique to regain or recover lost information in an image. This is done

by gaining information related to the background of the image, surrounding of the
image or manually as per requirements. The word ‘inpainting’ comes from ancient
times where the skilled and specialist painters were used to re-paint a lost or degraded
painting using some pre-requisites about the painting. Major researches have been
done in this field. Many approaches have been implemented to do inpainting. Deep
neural networks have been also implemented where the machine is trained to perform
the desired results [8]. Some other fascinating works have been in done in [9], where
the bifurcated image which is corrupted and separates out those corrupted into smaller
sub-images. Then they feed those piece of images into a neural network using
Euclidean loss function also calculated the MSE. The plus point for using this
approach is that is efficient in eradicating small distortions but each corruption
required a new and a separate neural layer. Criminisi et al. [10] have used the
exemplar-based technique using texture synthesis. Their approach’s result is quite
impressive when it comes to region filling and background inpainting. The authors of
this research paper [11] has proposed a new method to determine an enhanced ‘priority
factor’ to fill the patches. The patch filling here is done from propagation from outer
boundaries of the patch image to inwards of the image. The results then are compared
with existing inpainting techniques by using PSNR as a metric to do the same. In this
paper [12], the author has used a novel method by using exemplar-based inpainting as
the base method. Here, initially the author has removed big objects from the image and
has used both the combined benefits of inpainting and texture synthesis method. The
main motto of the author here is to reduce the processing time and achieve better
results. Their results showed that their computational time got reduced to 0.25 times to
the existing inpainting techniques. In this paper [13], the target is a video on which the
inpainting has been done. Author’s aim here is to regenerate the part of the video that
has been either deleted or destroyed. The workflow follows that the cross-correlation
method to get the video frame by frame and the objects being analyzed in each frame.
The source frame is obtained and inpainting technique is applied to these frames only.
The order of filling the patch is done by determining the priority factor.
3 Methodology
The tool has been developed using MATLAB R2013a. The basic functions of
MATLAB and those related to digital image processing have been used to get the
desired result. The tool’s working has been tested using various images. To check the
quality of the image, Mean Square Error (MSE) and PSNR ratio have been used to
compare the result generated by the tool and the Adobe Photoshop. The complete
result has been shown in this paper by using three images and calculating their MSE
and PSNR in tabular form and showing the result figures in graphical form (Fig. 3).
Fig. 3 Workflow of exemplar-based technique

710 H. Kumar et al.
4 About the Tool
The inpainting tool is developed using MATLAB 2013a. Basic functions of image
processing have been used to achieve the goal. This tool is a simple GUI-based tool
that fetches any image format in the working area of the tool.
Intially, preprocessing of the image has been done. The input RGB image is
converted to grayscale using function ‘rgb2gray()’. To differentiate the objects from
the background, ‘Canny edge detection’ algorithm has been used. When there a
change in the intensity of pixels and to deal with object boundaries, Canny edge
detection is the best algorithm that comes into play. Canny labels the pixels
depending on the gradient value larger or lesser than it [14]. Canny is one of the
best methods to find the edges without producing any effect to the features of the
image. Initially, to remove noise, smoothening is done and then the gradient is
found which calculates the derivates of the pixel intensity values. Then the region
highlighting is done with the help of gradient value. A threshold value is set, above
which the pixel is labelled as edge and below it, the pixel value is set to zero [15]
(Fig. 4).
Fig. 4 a Synthetic image, b obtaining edge, c minor and major edges, d final edges [16]
To find the gradient:

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
G¼ G2x þ G2y : ð1Þ

Gy
h ¼ arctan : ð2Þ
Gx
4.1 Structuring Element
A Structuring Element (SE) is used to define neighbourhood structures or to

identify some kind of pattern. It is a binary image or also called as mask or kernel
(Fig. 5).
In our coding, we have created a square-shaped structuring element. The code
snap is as follows (Fig. 6).
4.2 Dilation
In morphological operation on the image, dilation is also known as growing of the

image. It is usually applied to binary as well as grayscale image. It is called
‘Growing’ because it makes an object’s pixel grow in size. The pixels present at the
boundary tend to increase their size at the boundary. This method decreases the
empty cavity/holes in the image. This is also used for region filling. Let M be a set
of pixels for the input image. Let N be the structuring element matrix. The formula
is as follows (Fig. 7):
M N ¼ fxjðNÞx \ M 6¼ /g
Fig. 5 Four-neighbourhood
structuring element
Fig. 6 Use of strel() in

MATLAB
712 H. Kumar et al.
Fig. 7 Dilation results
5 Results
The inpainting tool has been developed on the MATLAB 2013a software. The
working, GUI and the results have been described in this paper. The results have
been compared with the same operation performed on ADOBE PHOTOSHOP
software. The results of the tool have been compared by calculating PSNR and
MSE values of each output. The tool’s GUI is a follows (Fig. 8).
The above image shows the functioning of the tool. The image is fetched and the
objects are detected using canny edge detection. The objects count is also men-
tioned by the tool.
The working of the tool has been represented in the following flowchart (Fig. 9).
In the below image, a middle face has been removed and has been replaced by
the background. This inpainting technique is used to remove objects and replace
with the background texture (Fig. 10).
Fig. 8 Tool’s GUI with image fetched and objects detected

Fig. 9 Working of the tool
In the below image, the object has been replaced with the other object in the
picture (Fig. 11).
In the below image, one face has been replaced with the other taken from the
other image only manually by the user (Fig. 12).
The metric used to compare the output generated by the tool and the
ADOBE PHOTOSHOP are MSE and PSNR. These two metrics are used to check
the quality of the image generated by the tool with the same image being generated
by ADOBE PHOTOSHOP software. The results are as follows (Fig. 13; Table 1).
714 H. Kumar et al.
Fig. 10 Image 1: a original image, b object removed by our tool, c object removed by Adobe
Photoshop software
Fig. 11 Image 2: a original image, b object replacement by our tool, c object replacement by
ADOBE PHOTOSHOP software
Fig. 12 Image 3: a original image, b face replacement by our tool, c face replacement by
ADOBE PHOTOSHOP software
Fig. 13 Graph showing the comparison of MSE and PSNR of different operations on different
images
Table 1 Table showing PSNR MSE (dB)

experiment results of three
images Image 1 37.9711 6.8229
Image 2 28.9338 83.1187
Image 3 23.4150 296.1951
716 H. Kumar et al.
6 Conclusion
The tool works almost perfectly with simple images and has a pretty low MSE.
PSRN ratio calculated for the images are low except for the images which have
object overlapped and complex in structure. The three results mentioned have good
PSNR ratio. The developed tool comes out to be good for object replacement. It is
also a helpful tool for object removal as well. As per results, the graph of image 1
and image 2 suggests that the MSE is less but when it comes to image 3, the MSE
is high. Hence, this tool works perfectly for simple texture images with low MSE
and a good PSNR value. Well, its MSE value increase when complex images are
fed to the tool. As per Image 1 experiment results; the MSE is very less. This
suggests that for images of type Image 1, the tool works comparatively well.
References
1. J. Electron., Image completion based on statistical texture analysis, Imaging. 24(1), 013032
(Feb 23, 2015)
2. Tao Zhou, Brian Johnson, Patch-based Texture Synthesis for Image Inpainting, 1605.0157
(May 5, 2016)
3. Survey on Image Inpainting Techniques: Texture Synthesis, Convolution and Exemplar
Based Algorithms, IJIRST ‘“International Journal for Innovative Research in Science and
Technology| Volume 1 | Issue 7 | December 2014
4. https://upload.wikimedia.org/wikipedia/commons/0/02/Texture_spectrum
5. Rajul Suthar et al Int. Journal of Engineering Research and Applications, ISSN: 2248-9622,
Vol. 4, Issue 2(Version 1), February 2014, pp. 85–88]
6. T. Chan and J. Shen, ‘Local in painting models and TV in painting, SIAM Journal on Applied
Mathematics, Vol. 62, 2001, pp. 1019–1043]
7. P. Harrison. A non-hierarchical procedure for re-synthesis of complex texture, In Proc. Int.
Conf. Central Europe Comp. Graphics, Visua. and Comp. Vision, Plzen, Czech Republic, Feb
2001
8. Gupta, Kushagr, Suleman Kazi, and Terry Kong. “DeepPaint: A Tool for Image Inpainting”
9. Blind inpainting using the fully convolutional neural network, Nian Cai, et al., Springer 2015
10. Criminisi, Antonio, Patrick Prez, and Kentaro Toyama. ‘Region filling and object removal by
exemplar-based image inpainting.’ Image Processing, IEEE Transactions on 13.9 (2004):
1200–1212
11. K. Sangeetha, P. Sengottuvelan and E. Balamurugan, “A Novel Exemplar based Image
Inpainting Algorithm for Natural Scene Image Completion with Improved Patch Prioritizing”,
International Journal of Computer Applications (0975 – 8887), Volume 36– No. 4, December
2011
12. Waykule J.M., “Modified Image Exemplar-Based Inpainting”, International Journal of
Advanced Research in Computer and Communication Engineering, Vol. 2, Issue 9,
September 2013
13. Sreelekshmi Das and Reeba R., “Robust Exemplar based Object Removal in Video”,
International Journal of Scientific Engineering and Research (IJSER), ISSN (Online): 2347–
3878, Volume 1 Issue 2, October 2013
14. Lijun Ding and Ardeshir Goshtasby,On the Canny edge detector,Computer Science and
Engineering Department
15. Nisha, Rajesh Mehra, Lalita Sharma, Comparative Analysis of Canny and Prewitt Edge
Detection Techniques used in Image Processing, International Journal of Engineering Trends
and Technology (IJETT), Volume 28 Number 1 - October 2015
16. On the Canny edge detector, Lijun Ding and Ardeshir Goshtasby Computer Science and
Engineering Department
The Hidden Truth Anonymity
in Cyberspace: Deep Web
Saksham Gulati, Shilpi Sharma and Garima Agarwal
Abstract The main objective of this paper is to study and illustrate the workings of
the invisible Internet and practically determining the Onion Routing relays, bridges
and exit nodes. A detailed research has been done on the working of exit nodes and
privacy over the deep web through vulnerable nodes. This paper also illustrates a
practical study of the depth of data available on the deep web. Emphasis is laid
down towards the safe access to deep web without compromising one’s privacy.
A survey was conducted in the paper to show the awareness among the technically
sound public and results were shown in pie charts.
Keywords Darknet Onion router Tor Exit nodes Privacy
1 Introduction
The deep web is that part of Internet which is not indexed by the usual search
engines [1].
Although deep web has very common uses, it is most notoriously coined for
illegal activities such as buying and selling of drugs, illegal armoury, anonymous
chat platforms, child pornography and black hat hacking services. The websites
which are generally accessed through deep web are not registered with regular
search engines. The deep web is generally accessed through an anonymous network
known as Tor, which completely shields an individual’s identity. Although Google
is known as the largest holder of indexed information, it is believed that information
in deep web may be 2–3 times larger than that in the surface web [2]. In this paper,
S. Gulati (&) S. Sharma G. Agarwal

Amity University, Noida, Uttar Pradesh, India
e-mail: saksham.gulati28@gmail.com
S. Sharma
e-mail: ssharma22@amity.edu
G. Agarwal
e-mail: garima.agarwal14@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_70
720 S. Gulati et al.
the invisible web has been categorized into three major subgroups which are deep
web, dark web and the invisible Internet project.
In the methodology section of this paper, details are provided about the methods
being used to collect the desired data and survey. The main emphasis is towards the
dog command that is used to collect information about DNS [3, 4].
In the result section of this paper, the dataset was collected through systematic
research. Former results contain a theoretical explanation of coexistence of deep
web and surface web which has been explained practically using DIG command.
Later, the section contains a survey which was conducted by over 300 IT experts.
Graphical representation using pie charts were developed according to the aware-
ness among the individuals about the deep web. The major category included
General Public, IT Experts and Information Security Experts to survey user’s
knowledge and awareness of deep web and dark web.
1.1 Dark Web Services
Dark web generally deals with shady sides of Internet including the bitcoin mining
and exchange, terrorism, hoaxes, illegal market [3]. Dark net can be considered as a
subset of deep web. Often pages that are found in dark web are cached within a
domain of the original web but can only be accessed by special browsing strategies
known as onion routing or tor [5–7]. Hence, dark web is sometimes also considered
as the onion world. Dark web is that part of deep web which is kept hidden, private
and protected at source and destination. Examples are as follows:
a. Fake Photo Identity
Fake photo identity like U.S and U.K citizenships cards, driving license, etc., can
be easily bought in through dark web [8].
The seller’s promise for an original passport with information/picture. They
guarantee that the information of the passport holder will be stored in the official
passport database and thus make it possible to travels abroad [9].
b. Renting Services
Hiring a hacker over the web, who is willing to do practically anything for money
[6, 7, 9].
Skilled hackers vary in ages even from 10 years onwards. These hackers are not
identified because they are not able to find a decent job somewhere but because the
pay is good and they get a thrill out of this job [10]. The amount depends on the
type of service.
c. Black Market
No state or country allows buying a professional gun until and unless license and
purpose of defence are clearly given to the required acting committee to do so.
The Hidden Truth Anonymity in Cyberspace: Deep Web 721
Onion dark web is the largest place to buy guns and ammo’s illegally. It happens to
be one of the biggest markets for some of the most deadly and life-threatening
psychedelics and drugs like LSD, COCAINE, METH, ECSTASY, etc. [10, 11].
The Dark net has a huge market for the buying and selling of guns. These sites
do not check for licenses and they have every gun with little descriptions of them
and cost. It is as easy as buying clothes from www.amazon.com [12–14].
d. Counterfeiting
The dark web also provides an easy market to counterfeit small bills in order to
illegally launder money [12].
The money sold is said to be produced from cotton-based paper. It is often UVI
incorporated hence it can pass the UV test. It can also pass the Pen test without any
problems. They also have necessary security features to be spent at most retailers
[12, 15].
1.2 Deep Web Services
Search engines—These search engines, unlike surface search engines—Google and

Bing, are able to crawl much deeper into the peer-to-peer network [3]. Examples
are:
a. CryptoCurrency
It is a virtual form of currency which is harvested using seeding and leaching
methods similar to that of Torrents and holds a very high value in deep web. Almost
anything and everything can be bought on deep web using cryptocurrency [11, 13,
15].
b. Drug Marketplace
One of the most common practices on dark web is the marketing of banned drugs
and crack. Since the dark web is untraceable, drug dealers and their suppliers use
this opportunity to connect to people who desire such substances [9, 14, 16].
c. Illegal Pornography
Dark Web is a major hub for child pornography. Due to its ability to maintain its
user’s anonymity, dark web offers a platform for child molesters, paedophiles and
rapists alike to upload videos and photographs of their victims without the fear of
authority [8]. These categories of porn receive many views from people with the
same mindset and thus in turn encourage their uploaders to make more such videos
[8–10].
2 Related Work
Deep Web and I2P have been around for several years now. Deep Web is that part
of Internet which is generally not accessible by the standard search engines [1].
When talking about Deep Web, it is important to distinguish between traditional
search engines and search engines capable of Deep Web harvesting [1–7]. Unlike
traditional search technologies, like Google, that index links and allow to view the
results, Deep Web engines take it a step further and harvest all of the results. The
traditional web crawlers only focus on the surface web while the deep web keeps
expanding behind the scene. The structure of the deep web pages makes it
impossible for traditional web crawlers to access deep web contents [17]. Curiosity
about Dark Web peaked when the FBI took down the Silk Road Marketplace in
2013 [2, 5, 11]. The media frenzy that took place made many people aware of the
concept [2, 5].
Reference [1] introduces several different networks other than Tor that guarantee
anonymity. It further analyzes malicious activities that take place over the Dark
Web, which has now become like a safe haven for criminals by providing them a
platform to conduct illegal activities without fear of authority. In the paper by
Hardy [18], the process of utilizing a modelling technique, informed by trade
auction theory is used to analyze the reputation of seller’s in the encrypted online
marketplace. This analysis of the seller’s reputation helps in giving insights into the
factors that affect and determine the prices of goods and services. The need for
people to maintain their anonymity from government agencies is helping to expand
the number of users on Darknet [8]. References [2, 19] state “Cybercriminals from
every corner of the world take advantage of the anonymity of the Web, particularly
the Deep Web, to hide from the authorities. Infrastructure and skill differences affect
how far into the Deep Web each underground market has gone”. Although mere
access via TOR is not considered an illegal practice, it can arouse suspicion with the
Law [1–7]. Terrorists use the deep web for their own advantage. Many terrorist
organizations and their supporters use deep web, its anonymity and protection for
recruitment and communication purposes. In their paper [21], H. Chen et al.
developed a web-based knowledge portal, called dark web portal. This portal was
developed to support the discovery and analysis of information discovered on the
dark web. TOR has long been used by journalists, researchers, thrill seekers or in
general curious people in heavily censored countries in order to hide their browsing
histories and exchange information anonymously.
Dark net offers a variety of services [3, 4] including buying and selling of
ammunition, access to child pornography and drugs. In this paper, we tried to
discuss few services along with the awareness of people about the services of deep
web.
3 Methodology
Dig command is a tool used to query DNS name servers for information about host
addresses, mail exchanges, name servers and related information. The tool is widely
used on Linux-based systems and Macintosh (Mac). This dig command is very
useful in troubleshooting and educational purposes. It replaces older techniques like
ns lookup and host programs.
The basic idea of dig is to pretend as a name server and working down from the
root of the server by iterative queries. After it gets access to name servers it then
requests for the specific details matching to the query.
We have implemented the following dig command to gather the information
regarding domains which are part of the deep web but are operated at surface web
along with deep web for maintaining high levels of anonymity.
Dig command connect to the domain while visiting all the intermediate nodes
and tries to collect the information at each node. While looking up at the deep web
domain the following nodes are used to route data.
a. Tor Relay
Tor relays are routers or nodes which are used to transmit data over peer networks.
Since tor is a peer-to-peer service, a working user passes through three nodes before
actually reaching its destination. This is done to mask the identity of the user but
masking leads to slow and unreliable connection. There are three types of tor relays
[22].
b. Middle Relays
These types of relays are used to mask the identity of the user entering the tor
network. They are responsible for routing data from source to the destination while
masking the identity. These are the middlemen in the network and store very little
data. For security purposes, data travels through three nodes before reaching the
actual destination, two of which are middle relays [22].
c. Exit Nodes
These nodes are the last nodes before which data reaches its destination and third
node of the three. Generally, exit nodes are advertised all over the tor web or the
deep web so that an uploader can easily setup a connection. Exit nodes are thor-
oughly protected using encryption so as to avoid eavesdropping and vulnerability
extraction. But recently tor exit nodes have been attacked by various groups and
defence forces.
d. Bridges
Bridges are Tor relays which are not publicly listed as part of the Tor network. Bridges
are essential censorship-circumvention tools in countries that regularly block the IP
addresses of all publicly listed Tor relays. A bridge is generally safe to run in homes, in
conjunction with other services, or on a computer with personal files.
Other than dig, a survey has been conducted to gather the information regarding
the awareness of the mass. The survey was conducted with over 500 individual and
was divided into three major categories namely general, IT and Network Security.
The survey was kept anonymous. It contained no personal question like name,
address or anything involving personal life except for work field which was relevant
to the paper. The work field specifies the relativity of an individual to the hidden
aspects of Internet, for example a network engineer is expected to be more aware of
deep web than an individual in any other field of work.
The following set of questions was asked in survey.
1. Primary work field categorized into General, IT and Network Security
2. Have you heard about deep web?
3. If yes, have you ever used accessed deep web?
4. Do you access deep web on regular basis? If yes, no of days in week.
5. Have you ever come across any illegal activity on the Internet? Was the owner
of this activity traceable?
4 Results
Interaction Between surface and deep web.

According to the current Internet protocols, the deep web can be considered to be
hosted on current visible network with complete protection through overlay web-
sites and hence it is the major reason why it is difficult to completely trace deep
web. Some of the websites even use plain registered DNS, which is directly used to
transfer files through wireless server [14, 22].
Case I—Hostmysite.com, the only single home page claims to upload the file
anonymously but does not specify where it is to be uploaded. This DNS is
accessible through resolved name but it does no information regarding the own-
ership of this page or relocation of the uploaded file can be found. Hence, it seems
to be working on surface web but it is directly related to deep web [22].
Proof 1: It can be accessed without the use of http protocol or the https standard
protocols.
This means that “hostmyshit.com” does not uses http or https before the DNS. It
is unusual in web 3.0 and further to have such kind of anomaly since it was
compulsory to use http or https tag to correctly determine the address of a website
over the internet.
In case of deep web, this is an exception since deep web converts a normal DNS
to the hash equivalent to reach to a desired website.
Proof 2: When this DNS was dug using an open-source software DIG (Figs. 1 and 2).
Fig. 1 DIG command result
Fig. 2 Connection testing using IP address

But the DNS or the website is not traceable from original ip. Even if a host is behind
a firewall, false hop proxy or VPN the Domain is searchable through proxy IP.
In such cases, an event where the host is not traceable from masked IP which
positively replies on ICMP ping (ICMP ping is unreliable hence if it gives and
active connection it means the connection is smooth.) means it has a constantly
changing untraceable IP which is a property of Onion routing [1].
Proof 3: The website uses 4096 UDP port.
A website which is connectable without a packet drop cannot be relies on UDP
port. Peer-to-peer connections generally work on UDP port. Hence this website runs
on a peer-to-peer connection and since this uses FTP protocols (uploading) it has to
spread connection and hence files. This is how the files are shared on peer-to-peer
connection. Tor is entirely based on P2P connection reliability between the nodes.
Case II—In a recent chain of events which took place in the United States a
major torrent site kickass torrent went down and was restricted to access. Torrent is
a major example of coexistence of deep web and surface web. Torrents use
peer-to-peer protocol for data download and upload with a registered domain name.
Deep Web is completely based on peer based connection and Torrent search
engines queries through deep web for accurate result because it is not possible to
trace deep web log files. This makes piracy related crime much easier to execute.
Kickass torrent can be reached easily through surface web but on resolving this
domain name no specific working IP is obtained. The above proof already shows
how a website can be masked in a deepweb and work on surface web [22].
Exit Nodes
The following list gives the description of exit nodes that are used in communi-
cating through the deep web architecture. The exit nodes play the major role since
with only vulnerable exit nodes a deep web website can be deciphered. This can be
very helpful in limiting down the criminal activity over the internet (Fig. 3).
ExitNode 0011BD2485AD45D984EC4159C88FC066E5E3300E
Published 2016-05-22 22:17:50
LastStatus2016-05-22 23:02:51
ExitAddress162.247.72.201 2016-05-22 23:11:01
ExitNode 0091174DE56EFF09A7DAECCAC704F638D6D370F5
Published 2016-05-22 09:38:21
LastStatus2016-05-22 10:03:04
ExitAddress45.35.90.36 2016-05-22 10:07:30
ExitNode0111BA9B604669E636FFD5B503F382A4B7AD6E80
Published 2016-05-22 17:11:36
LastStatus2016-05-22 18:02:57
ExitAddress176.10.104.240 2016-05-22 18:05:00
Fig. 3 List of currently active exit nodes (as on 21 June 2016)

Fig. 4 Deep web awareness survey
These are the list of exit nodes which were accessed and verified as a part of
deepweb data routing mechanism. The PGP signatures are mentioned along with
their exit node addresses
A survey was conducted to check what percent of population amongst 300 data
sets of users is aware of the deep web and black market. The survey was conducted
and results were categorized into the general public, IT related and Network and
Security Associated. The following results were obtained for deep web and dark
web (Figs. 4 and 5).
Fig. 5 Darkweb awareness survey
5 Conclusion
We conclude that there is very less number of people who actually know about the
deep web and thus it raises a major concern. Also as 80% of the web is not visible
and contains the most valuable data is results a great impact towards research area.
The finer concern lies that the people have almost no idea about the dark web and
illegal activities related to the Internet making them a potential victim of many
crimes.
6 Future Work
This paper can be used to spread the awareness about the deep web, dark web and
criminal related activities associated with the general public of internet. It not only
prevents them from getting influenced by such activities but also helps to catch the
criminals.
Internet is easily accessible by anyone at any place in the world and it takes a
little effort to access deep web. It is very easy for anyone to indulge in the criminal
market of deep web, which definitely leads to trouble. Through this chapter, we are
aiming at spreading awareness among the masses to prevent themselves from

getting involved in criminal activities.
This paper also explains the coexisting relationship between deep web and
surface web hence further research studies can be done keeping this in mind that
surface web is just subset of deep web and holds a relation which was not con-
sidered in many of the research work until now.
The exit nodes as listed in the paper can be monitored or attacked and infor-
mation can be collected to perform a strategic way out of criminal activities in the
coming upcoming researches.
References
1. Hamilton, Nigel. “The Mechanics of a Deep Net Metasearch Engine”. CiteSeerX:

10.1.1.90.5847.
2. Owen, Gareth. “Dr Gareth Owen: Tor: Hidden Services and Deanonymisation”. Retrieved 20
June 2015.
3. Devine, Jane; Egger-Sider, Francine (July 2004). “Beyond Google: the invisible web in the
academic library”. The Journal of Academic Librarianship 30 (4): 265–269.
4. http://www.scribd.com/doc/172768269/ Ulbricht-Criminal-Complaintulbr_mirror. Scribd.
“Ulbricht Criminal Complaint.” [Accessed on 10 June 2015].
5. Owen, Gareth. “Dr. Gareth Owen: Tor: Hidden Services and Deanonymisation”. [Accessed
20 June 2015].
6. Vincenzo Ciancaglini. (8 November 2013). TrendLabs Security Intelligence Blog. “The Boys Are
Back in Town: Deep Web Marketplaces Back Online.” Last accessed on 10 June 2015, http://blog.
trendmicro.com/trendlabs-security-intelligence/the-boysare-back-in-town-deep-web-marketplaces-
back-online/.
7. Vincenzo Ciancaglini. (10 March 2015). TrendLabs Security Intelligence Blog. “The Deep
Web: Shutdowns, New Sites, New Tools.” Last accessed on 10 June 2015, http://blog.
trendmicro.com/trendlabs-security-intelligence/the-deep-web-shutdowns-new-sitesnew-tools/.
8. Fake Photo Identity Services, 2ogmrlfzdthnwkez.onion [Accessed on 11 June 2016].
9. Illegal Pronography, Vfqnd6mieccqyiit,onion [Accessed on 11 June 2016].
10. Drug Market Place, S5q54hfww56ov2xc.onion [Accessed on 11 June 2016].
11. Illegal Services, Tuu66yxvrnn3of7l.onion [Accessed on 11 June 2016].
12. Illegal Services, Qkj4drtgvpm7eecl.onion [Accessed on 11 June 2016].
13. Robert McArdle and David Sancho. Trend Micro Security Intelligence. “Bitcoin Domains.”
[Accessed on 10 June 2015].
14. The Invisible Internet Project. [Accessed 11 June 2015] https://geti2p.net/en/.
15. FeikeHacquebord. (5 September 2015). TrendLabs Security Intelligence Blog. “The
Mysterious MEVADE Malware.” [Accessed on 10 June 2015], http://blog.trendmicro.com/
trendlabs-security-intelligence/the-mysterious-mevade-malware/.
16. Trend Micro Incorporated. Trend Micro Security News. “Cybercriminal Underground
Economy Series.” [Accessed 11 June 2015], http://www.trendmicro.com/vinfo/us/security/
special-report/cybercriminal-underground-economy-series/index.html.
17. R.Anita, R. “Deep iCrawl: An Intelligent Vision-Based Deep Web Crawler”. World Academy
of Science, Engineering and Technology, Internationals Science Index 50, International
Journal of Computer, Electrical, Automation, Control and Information Engineering, (2011), 5
(2), 128–133.
18. HARDY, ROBERT AUGUSTUS, and JULIA R. NORGAARD. “Reputation in the Internet
Black Market: an Empirical and Theoretical Analysis of the Deep Web.” Journal of
Institutional Economics, vol. 12, no. 3, 2016, pp. 515–539., doi:https://doi.org/10.1017/
S1744137415000454.
19. http://blog.trendmicro.com/trendlabs-security-intelligence/cybercrime-in-the-deepweb/.
Robert McArdle. (4 October 2013). TrendLabs Security Intelligence Blog. “Cybercrime in the
Deep Web.” [Accessed on 10 June 2015].
20. http://www.trendmicro.com/cloud-content/us/pdfs/security-intelligence/white-papers/wp-
bitcoin-domains.pdf.
21. Qin, Jialun, et al. “The dark web portal project: collecting and analyzing the presence of
terrorist groups on the web.” Proceedings of the 2005 IEEE international conference on
Intelligence and Security Informatics. Springer-Verlag, 2005.
22. https://www.torproject.org The Tor Project, Inc. Tor Project. [Accessed on 11 June 2015].
Test Case Optimization and Prioritization
of Web Service Using Bacteriologic
Algorithm
Gaurav Raj, Dheerendra Singh and Ishita Tyagi
Abstract Regression testing, testing is done on the changes made in existing

software to check whether the existing software is working properly or not after the
changes has been done. Therefore, retesting is performed to detect the new faults
found. This type of testing is performed again and again after the changes have been
made in the pre-existing software. Various methods are used for test case reduction
and optimization for a web service. Regression testing creates a large number of test
suites which consumes a lot of time in testing and many other problems are faced.
Therefore, some technique or method should be used so that number of test cases
are reduced and also test cases can be prioritized keeping in mind the time and
budget constraints. The test case reduction and prioritization need to be achieved
depending on various parameters such as branch coverage and also on basis of fault
coverage etc. Therefore, this paper discusses about the analysis of the code of a web
service and the technique used to analyze a web service based on branch or code
coverage and also the fault detection for test case reduction and prioritization is
bacteriologic algorithm (BA). The test cases generated and also other requirements
are mapped with the branch coverage and fault coverage of the code of the web
service.

Keywords Test case Test suite Bacteriologic algorithm Web service

Regression testing Test case reduction
G. Raj (&)
Punjab Technical University, Kapurthala, Punjab, India
e-mail: graj@amity.edu
D. Singh
Computer Science and Engineering Department, Chandigarh College
of Engineering and Technology, Sector-26, Chandigarh, Punjab, India
I. Tyagi
Computer Science and Engineering Department, Amity University, Uttar Pradesh, India
e-mail: ishita27t@gmail.com

https://doi.org/10.1007/978-981-10-7245-1_71
732 G. Raj et al.
1 Introduction
Main goal of web service is to communicate and exchange information by using a

common platform among different applications built in different programming
languages and operating system. The confidentiality and integrity of a web service
need to be maintained as attacks on web application have increased.
The term software engineering is defined as a branch of engineering which is
related to software development and management with the help of properly defined
methods, principles and steps. The aim to provide the customer a reliable and
efficient product. The software engineering contains various models for the software
development. The models are used according to the requirements and other
parameters required for software development and management. Since, one of the
most important phase of any model is testing phase, so the testing should be done
properly and carefully to get reliable results. It is not practically possible to
implement exhaustive testing. There can be both valid and invalid inputs because it
is very difficult to perform testing on all the test cases available. There are many
reasons like time constraints, budget etc. because of which exhaustive testing is not
possible. It is not easy to completely test because of the design problems. In
regression testing, testing is done on the changes made in existing software to check
whether the existing software is working properly or not after the changes has been
done. Therefore, retesting is performed to detect the new faults found. This type of
testing is performed again and again after the changes has been made in the
pre-existing software. Changes can be of different types such as adding extra fea-
tures, changes in the configuration etc. [1]. But this testing increases the budget of
the software or the product because of the features enhanced and also a lot of
maintenance is required.
Various methods are used for test case reduction and optimization for a web
service. But the technique should be selected in such a way that budget must be
less. Regression testing also creates a large number of test suites which consumes a
lot of time in testing and many other problems are faced [2, 3]. The test case
reduction and prioritization need to be achieved depending on various parameters
such as branch coverage and also on basis of fault coverage etc. so, [4]. The
technique used to analyze the web service based on branch or code coverage and
also the fault detection for reduction of test case and prioritisation is bacteriologic
algorithm (BA). The test cases generated and also other requirements are mapped
with the branch coverage and fault coverage of the code of the web service. This
helps in calculating fitness of the code of the web service based on branch coverage
and fault detection capability.
Test Case Optimization and Prioritization of Web Service … 733
2 Terminologies
There are some terms that are explained under this section which will be used in the
further sections of the paper.
2.1 Test Case
Test case is a collection the preconditions, post-conditions, test data and results
which are documented for a specific test scenario so that compliance can be verified
against the requirement [3]. For execution of the test, test case is the starting point.
Some of the parameters of test case are: Test case ID, Test data, Expected results,
Actual results etc.
2.2 Test Suite
A group of test cases is known as test suite. The test execution status can be
reported by the testers with the help of test suites. Each test case can used in one or
more test suites. Number of test cases are there in a test suite. Before creating test
suites, a test plan is made. According to the scope and the cycle test suites are
generated. There can be number of tests viz-functional or Non-Functional.
2.3 Test Case Minimization
It is defined as process to make lesser number of test cases with the help of making
test suites from test cases. It is used for reducing the cost of resources, time taken in
execution etc. The test suites should be created in such a way that satisfy all the
requirements. The main aim of test suite is remove redundancy that is created
because of the test cases. There are different techniques used for removing
redundancy such as genetic algorithm etc.
2.4 Test Case Prioritization
There are number of test suites available in regression testing [2]. Since, it is not
easy because of the time constraints to test all the test suites over the code, we use
test case prioritization. It can be defined as ordering the test suites that follow one or
more criteria. The aim is to find the faults by selecting least no of test cases. It is
734 G. Raj et al.
used so that the best results are found at the time of the result. There can be many be
reasons for prioritizing the test suites: fault detection rate should be increased. Code
coverage should me more and at a faster rate. Every system has a different archi-
tecture, Therefore, prioritization can also be done on basis depending on the system
architecture. There are some parts of the architecture that have a major effect on the
entire system, they should be tested properly. The components or parts of system
can be tested individually also for test case prioritization.
2.5 Exhaustive Testing
The testing in which the testing is done for the all the test cases that are possible
keeping in mind the quality [5, 6]. Every possible case is made but this type of
testing is impractical. A product is considered to be perfect if the exhaustive testing
of a software is done. It is actually not easy to pass exhaustive testing. There are
strict deadlines because of which this type of testing becomes impossible to use.
Some of the inputs become invalid after particular time period. Therefore, time
constraint is also an issue. It is impossible if real time scenarios need to be tested
like temperature etc. Also, it is impractical to test all the combinations of each user
or many users [7]. All the outputs also cannot be checked.
2.6 Web Service
Goal of web service is to communicate and exchange information by using a

common platform among different applications built in different programming
languages and operating system. A web service contains UDDI, SOAP and WSDL.
2.7 Genetic Algorithm
The optimization problem that are related to natural selection are solved using the
algorithm i.e. genetic algorithm. It changes each solution of the population again
and again. In each step, the children are generated for the upcoming production
from the parent test cases or test suites. And, as the generation passes, best or
favourable results are seen. This type of algorithm is used in solving the issues that
cannot use optimization algorithms etc. For knowing the nature of chromosome,
fitness is calculated. There are various steps applied in this algorithm:
1. First step is selection in which test suites are selected and then crossover and
also mutation are used over the test suites. The chromosomes or test suites are
chosen generally which are having greater fitness.
2. Second step is crossover in which the successors are combined and then form
new children.
3. The next operator used is mutation in which the current descendants are changed
automatically.
2.8 Bacteriologic Algorithm
Mala et al. [8] this algorithm is the alternative form of the GA algorithm. It has
slightly different steps because it follows the bacteria’s nature. There is no operator
like crossover present in it. It shows the properties of a bacteria. It only uses the
operator like selection and mutation from the above algorithm referred. It is dif-
ferent because it contains a new function i.e. “memorization function” which uses
the bacteria which is the most fit in each population. There is basic diagram
showing the flow of the entire process.
1. First step is selection in which test suites are selected and then crossover and
also mutation are used over the test suites. The chromosomes or test suites are
chosen generally which are having greater fitness.
2. The next operator used is mutation in which the current descendants are changed
automatically.
3. Next step is to find the “fitness” of the code using its formula.
4. Then, best test suite is memorized out of each population for addition to the next
step/generation.
3 Related Works
Krishnamoorthi and Mary [1] in their paper “Regression Test Suite Prioritization
using Genetic Algorithms” said that there are ways to use prioritization method
with the help of genetic algorithm. The parameters checked are the effectiveness
and time consumption. The faults are found which helps in knowing how effective
is the proposed method. Time coverage plays an important role here. It helps in
knowing the problems faced during time-aware prioritization. The paper also says
also tells about the methods which help in reducing time above of prioritization.
Singh et al. [9] in their paper “A Hybrid Approach for Regression Testing in
Interprocedural Program” mentioned that regression testing is performed again
and again after the changes has been made in the pre-existing software. Changes
can be of different types such as adding extra features, changes in the configuration
etc. Paper contains a new algorithm related to variable which is applied on the
variables with the help of hybrid method. The paper gives an algorithm for
regression testing. Various codes have been analyzed to find a conclusion. The
accuracy and performance according to the conclusion is much higher.
736 G. Raj et al.
Jatain and Sharma [2] in their paper “A Systematic Review of Techniques for
Test Case Prioritization” explained that there are many important activities that a
maintenance phase faces like deletion, correction etc. For the retesting of the
software retesting is done as modifications are done in the software time to time.
But it Is expensive to retest again and again. There are many methods present for
regression testing. The tester can choose the method according to their need and can
do testing. Here, various methods used by different researchers for prioritization are
presented for regression testing. Also different algorithms for the prioritization
process is used. It helps in knowing scope of different methods. Also, explains the
difficulties faced in prioritization based on requirements.
Joshi [5] in his paper “Review of Genetic Algorithm: An Optimization
Technique” told about a technique which can be used for the optimization of the
whole. After this there are many methods that come under genetic algorithm are
used for optimization. The paper explains the basic of the algorithm and gives the
proper flow of the “GA algorithm”. Also, the operators of the algorithm are
reviewed properly. It tells about the origin of the algorithm. There are many other
optimization technique other than GA that are explained in the paper. Also it tells
about different techniques of using the operators of the given algorithm.
Garg and Mittal [6] “Optimization by Genetic Algorithm” in their paper said the
optimization problem that are related to natural selection are solved using the algo-
rithm i.e. genetic algorithm. This research paper explains about the algorithm and the
technique of optimization. The optimization problem that are related to natural
selection are solved using the algorithm i.e. genetic algorithm. It changes each
solution of the population again and again. In each step, the children are generated for
the upcoming production from the parent test cases or test suites. The optimization
method is related or present in every field like engineering etc. This paper explains
about the “dejong function” and the experiments have been performed using this
function in genetic algorithm. The paper with the help of this function gives better
results. There is also the use “unimodal and multimodal benchmark functions” for the
optimization. The results are represented using graphs and tables.
Ramesh and Manivannan [7] “Test Suite Generation using Genetic Algorithm
and Evolutionary Techniques with Dynamically Evolving Test Cases” in their paper
explained a new way for making test oracles with the help of evolutionary algo-
rithm. This gave positive results which helped in finding bugs in various classes.
This method is also easy to use. Test oracles take a lot of time and is difficult to
make meaningful test cases if done manually on the data generated. In this paper
coverage parameter is used for optimization of test suites. Also, branch coverage
parameter has been used.
Shahid and Ibrahim [10] “A New Code Based Test Case Prioritization
Technique” in their papers said that test cases and test suites play an important role
in software testing. Validation is done of the software or product that is under
inspection with the help of test cases and suites. It becomes difficult to test all the
cases for all the code. Therefore, prioritization is used which can help in improving
the effectiveness and also help in saving time. Here, a new algorithm or approach
has been used for prioritization method and is done on basis of code covered. The
algorithm is used on a case study for the results which are good and promising. The
test case that cover most part are considered more important.
Mala et al. [8] “A hybrid test optimization framework-coupling genetic algo-
rithm with local search technique” in their paper explained that there are three
different algorithms that are used out of which “HGA algorithm” is used for making
test cases of better quality. For this to happen path coverage and score of mutation
is analyzed of every test case. The best test cases are selected having greater path
coverage and score of mutation. It also takes less time as test cases get reduced. The
test cases analyzed using this method are compared with the other algorithm on
basis of efficiency. The other methods used are genetic and bacteriologic algorithm.
4 Architecture and Algorithm Used
4.1 Flow Diagram
The basic flow of the algorithm that is used to find the results is explained in Fig. 1.
Analyze source code of the Initial test Fault

web service (define branches) detection
START tabulation
Mapping function
Branch coverage
tabulation
Regression testing Test suite's optimal size L
Minimization
algorithm
Test suite's optimal size L Randomly generate test suites of size
L(Initial population for BA/GA)
Calculate BC%,
FD% Selection
TS1 TS2
Mutation Fitness Calculation
Add to next Memorize fittest

generation bacteria
No Yes
SC STOP
Fig. 1 Flow diagram

738 G. Raj et al.
4.2 Algorithm Applied
Source code is the input i.e. P

There is a set/group of initial test cases (given by the user) Q
Reduced and prioritized test suites is the O/P.
Algorithm:
1. Find the branches in the code P which act as requirements from P.
Rb = {B1, B2, B3 … Bn}
2. Initial test cases(Q) are mapped with Rb
BranchCoveragetable = MappingFunction (Q, Rb)
3. In code P, search use of the variables used.
Rc = {z1(line x, line y)… zn(line x, line y)}
4. Q is mapped with Rc, FaultDetectAbilitytable = MappingFunction(Q, Rc)
5. Search for test suite (T) having maximum branch coverage and fault detection
ability.
T = Search(FaultDetectAbilitytable, BranchCoveragetable, Q)
6. Optimal length of test suites is to be searched,
OL = SearchLength (T)
7. Make any random test suites tsi = {ts1, ts2, ts3 … tsn} having length OL from Q
8. Use bacteriologic algorithm over tsi
a. Fitness calculation for every test suite.
fitness = (weight1* branch coverage) + (weight 2* fault detection analysis)
where branch coverage is obtained from Table 1.
fda = fault detection ability (obtained from Table 2)
w1 and w2 are weights ranging between 0 and 1.
b. Arrange test suites present in tsi such that test suites with high fitness comes
first.
c. Apply mutation
d. Fitness to be calculated again
e. Learn the most fit test suite
Table 1 Value to be Test cases Search value

searched
1 2
2 3
3 7
4 10
5 1
6 9
7 4
Table 2 Branch coverage

Test case Search value Branch b1 Branch b2 Branch b3 Branch b4
1 2 1 0 1 1
2 3 1 1 1 1
3 7 1 1 0 1
4 10 1 0 0 1
5 1 1 0 1 1
6 9 1 1 0 1
7 4 1 1 0 1
f. Most fit test suite is must be added to next level/generation

g. Repeat the steps from a and stop when stopping criteria is met. (stopping
criteria here can be number of levels/generations)
h. Rearrange all memorised bacteria in descending order of fitness, to prioritize
them.
4.3 Formula Required
1. “Branch Coverage” is,

(no. of branch covered)/(total no. of branches)*100.
2. “Fault Detection analysis” is, ((Variable(p,q)covered)/(TotalVariable(p/q)cov-
ered))*100
3. “Fitness” is,
(Weight1*BranchCoverage) + (Weight2*Fault Detection analysis) and, the
value of the weight can be in between or 0 or 1.
5 Case Study
To get a more practical idea the code of the web service of binary search is observed
and is passed as the input into the algorithm. The code is given below:
int * BinarySearch (int value)
{
unsigned int A = 0, B = array_length(array), S;
while (A<B)
{
S = (A+B− 1)/2;
if (value = = array [S])
return array+S;
else if (value<array [S])
B = S;
740 G. Raj et al.
else
A = S+1;
}
return null;
}
The variables used in the above code i.e. A, B and S are taken as input for
creating test cases from the user.
This part of paper shows the result. First step is to take input from the user to
generate the test cases. The code is divided into four branches to know the branch
coverage. Each line of the code is assigned a number to find the fault detection
coverage. The first step is to input the array for binary search code and set the
number of test cases to be generated. Then the test cases are formed and values to
be searched are given as input by the user. For example:
Array given: {2,3,4,7,9}
The value to be searched are given in Table 1. For the above array, the branch
coverage is calculated for the values given by the user in Table 1.
The branch coverage analysis is given Table 2 where, 1 represents that branch
has been covered to search the particular number entered by the user and 0 rep-
resents that branch has not been covered to search the particular number entered by
the user.
Table 3 represents the fault coverage where variable(p, q) is represented as, L
(3,6) means L is the variable used in the code and (3,6) is the line in which it
initialized and used respectively. Also, 1 represents that test case is covering line
p and line q for that particular test case.
After the fault coverage analysis the regression testing is performed on the test
cases generated by the user in Table 1. All the possible suits are generated in
Table 4 using regression testing (Fig. 2).
Table 3 Fault coverage table Test L R M M M

case (3,6) (3,6) (6,8) (6,10) (6,12)
1 1 1 1 0 1
2 1 1 1 1 1
3 1 1 1 0 1
4 1 1 0 0 1
5 1 1 0 1 1
6 1 1 1 0 1
7 1 1 1 0 0
Table 4 Table showing the Test suite dD Test case Bc Fda Fit
results of fit calculation for all
possible test suite using 1 1,2 100 100 100
Regression testing 2 1,3 100 100 100
3 1,4 75 80 80
4 1,5 75 80 80
5 1,6 100 100 100
6 1,7 100 100 100
7 2,3 100 100 100
8 2,4 100 100 100
9 2,5 100 100 100
10 2,6 100 100 100
11 2,7 100 100 100
12 3,4 75 80 80
13 3,5 100 100 100
14 3,6 75 80 80
15 3,7 75 80 80
16 4,5 75 80 80
17 4,6 75 80 80
18 4,7 75 80 80
19 5,6 100 100 100
20 5,7 100 100 100
21 6,7 75 80 80
Also, branch coverage, fda(fault detection analysis) and fitness is calculated for
all the possible test suites of pair 2 in Table 4. The formula used for:
1. “branch coverage” is,
(no. of branch covered)/(total no. of branches)*100.
2. “fault detection analysis” is,
((variable(p, q)covered)/(total variable(p/q) covered))*100
3. “fitness” is,
(weight1* branch coverage) + (weight 2* fault detection analysis)
4. and, the value of the weight can be in between or 0 or 1.
5. For example. In Table 4:
• For test suite id 18 i.e. 4,7, 3 out of total 4 branches are being covered.
Therefore, (3/4)*100 = 75%
Similarly, for fault detection analysis, (4/5)*100 = 80%
Therefore, after putting the values in formula of fitness, answer is 80%.
742 G. Raj et al.
Fig. 2 Graph showing bc,

fda and fit calculated for
regression testing
• For test suite id 19 i.e. 5,6, 4 out of total 4 branches are being covered.
Therefore, (4/4)*100 = 100%
Similarly, for fault detection analysis, (5/5)*100 = 100%
Therefore, after putting the values in formula of fitness, answer is 100%.
Next step is to calculate fitness using BA algorithm in which minimization
algorithm is applied and best possible test suites are calculated After the application
of minimization algorithm, the bacteriologic algorithm is used to find the fittest
bacteria and reduces the number of test suites in Table 5 using mutation operator.
The final answers are shown in Table 5; (Fig. 3).
Table 5 Table showing Test suites Test cases Bc Fda Fit

results after applying BA
algorithm 1 5,9 100 100 100
2 5,7 100 100 100
3 2,9 100 100 100
4 2,7 100 100 100
5 6,7 100 100 100
6 6,3 100 100 100
7 2,7 100 100 100
8 2,3 100 100 100
9 8,6 100 100 100
10 8,2 100 100 100
11 8,9 100 100 100
12 2,5 100 100 100
Fig. 3 Graph showing bc,

fda, fit calculated using
bacteriologic algorithm
7 Conclusions and Future Work
The paper is about the web service being developed and its code being tested for
fitness calculation using regression testing. The aim is to reduce and order the test
suites in such a way that all the best results are obtained rather than creating and
testing number of test cases as done in regression testing. The criteria to analyze the
code is branch coverage and fault detection analysis. Therefore, bacteriologic
algorithm is implemented on the code of the web service so that improved results
are obtained rather than using regression testing for fitness calculation of the code.
The usage of this approach is shown with the help of the example and the graphs
showing the results generated using regression testing and then using bacteriologic
algorithm on the same piece of code to get reduced and ordered results. The result
have been shown in the result section of the paper. The results show the reduced set
of test suites having the best fitness. Also, the terms that are used in the paper are
explained in the terminologies section. Further analysis can be done in future by
comparing the results of the various web service’s code using the same algorithm.
Also, the results of one web service can be compared with many other algorithm
derived in future to make improvements and get best result for test case reduction
and prioritization using code coverage and fault detection as parameters.
References
1. Krishnamoorthi, R. and Mary, S.A.S.A, “Regression test suite prioritization using genetic
algorithms”, International Journal of Hybrid Information Technology, Vol. 2, No. 3, pp. 35–
51, 2009.
2. Aman Jatain, Garima Sharma, “A Systematic Review of Techniques for Test Case
Prioritization”, International Journal of Computer Applications (0975–8887), Volume 68,
No. 2, April 2013.
3. Presitha Aarthi. M, Nandini. V, “A Survey on Test Case Selection and Prioritization”,
International Journal of Advanced Research in Computer Science and Software Engineering,
ISSN: 2277 128X, Volume 5, Issue 1, January 2015.
744 G. Raj et al.
4. R. Beena, Dr. S. Sarala, “Code coverage based test case selection and prioritization”,
International Journal of Software Engineering & Applications (IJSEA), Vol. 4, No. 6,
November 2013.
5. Gopesh Joshi, “Review of Genetic Algorithm: An Optimization Technique”, Volume 4, Issue
4, April 2014.
6. Richa Garg, Saurabh mittal, “Optimization by Genetic Algorithm”, International Journal of
Advanced Research in Computer Science and Software Engineering, ISSN: 2277 128X,
Volume 4, Issue 4, April 2014.
7. K. Ramesh and P. Manivannan, “Test Suite Generation using Genetic Algorithm and
Evolutionary Techniques with Dynamically Evolving Test Cases”, International Journal of
Innovation and Scientific Research, ISSN 2351-8014 Vol. 2 No. 2 Jun. 2014.
8. Dharmalingam Jeya Mala, Elizabeth Ruby, Vasudev Mohan(2010), “A Hybrid Test
Optimization Framework-Coupling Genetic algorithm with local search technique”,
Computing and Informatics, Vol. 29, 2010.
9. Yogesh Singh, Arvinder Kaur and Bharti Suri, “A Hybrid Approach for Regression Testing in
Interprocedural Program”, Journal of Information Processing Systems, Vol. 6, No. 1, 2010.
10. Muhammad Shahid and Suhaimi Ibrahim, “A New Code Based Test Case Prioritization
Technique”, International Journal of Software Engineering and Its Applications, Vol. 8,
No. 6, 2014.
Author Index
A Desai, Latika R., 425

Adhil, Mohamood, 217 Deshpande, Yudhishthir, 547
Agarwal, Garima, 719 Dhotre, Prashant S., 291
Agarwal, Mahima, 217 Doja, M. N., 395
Agnihotri, Jidnyasa, 317 Durga Prakash, M., 9
Ahmad, Tanvir, 395
Algonda, Ajinkya R., 405 F
Anand, Deeksha, 515 Fatima, Afroz, 19
Angal, Yogesh, 97
G
B Gade, Anita, 97
Bakal, Jagdish, 447 Gaopande, Sumant, 547
Bania, Ujjen Man, 279 Geeta, R. B., 77
Bansal, Manish, 143 Geetha, S. B., 505
Bedekar, Mangesh, 591 Ghorpade-Aher, Jayshree, 547
Bhardwaj, Shweta, 623 Ghosh, Krittika, 217
Bhargavi, K., 279 Godse, Sachin, 579
Bhateja, Vikrant, 515, 525 Gosavi, Chhaya S., 57
Bhoi, Bandan Kumar, 195 Gulati, Saksham, 719
Bhowmick, Kiran, 37 Guleria, Ajay, 327
Bhutada, Saranga N., 405 Gupta, Ashish, 279
Bindhumadhava, B. S., 491 Gupta, Divya, 667
Biyani, Nitesh, 567 Gupta, Piyush, 643
Bobde, Sharwin P., 557 Gupta, Poonam, 415
Gurav, Uma, 383
C
Chandak, Manoj, 27 H
Chandavale, Anjali A., 273, 467 Hiwale, Anil, 1
Chandra, Namita, 539
Chaudhari, Prajakta, 129 I
Chaudhary, Asha, 105 Ingle, Rajesh B., 265
Chavan, G. T., 247
Chavan, Umesh, 247 J
Chhallani, Rajul, 47 Jagdale, Balaso, 447
Chopade, N. B., 353 Jahagirdar, A. S., 363
Choudhury, Tanupriya, 653, 677, 705 Jain, Kartik Narendra, 677
Jain, Sanya, 667
D Jog, Anita, 467
Dedgaonkar, Suruchi, 567 Johri, Prashant, 633

https://doi.org/10.1007/978-981-10-7245-1
746 Author Index
Joshi, Pranoti, 353 Pamarthy, Gopika Rani Alekhya, 9

Joshi, Yashaswini, 77 Pandey, Manjusha, 457
Parag, Kulkarni, 373
K Parimala, N., 305
Kaalay, Aakankssha, 567 Pathak, Ajeet Ram, 457
Kadam, Krishna, 247 Patil, Abhishek Girish, 547
Kakkar, Alpana, 643 Patil, Aboli, 207
Kapoor, Kajal, 207 Patil, Dipti D., 557
Kaviraj, S., 491 Patil, Venkanagouda C., 505
Khajuria, Samant, 291 Pattewar, Rasika, 47
Khatri, Sunil Kumar, 633, 643 Pawar, Karishma, 457
Khonde, S. R., 435 Phalnikar, Rashmi, 317
Kishore Kumar, K., 341 Phatak, Eeshan, 547
Kohar, Rachna, 305 Pradhan, Manoranjan, 195
Koley, Santanu, 183 Prasad Reddy, P. V. G. D., 77
Kulkarni, D. B., 67 Priyadarshi, Mayank, 533
Kulkarni, Parag, 479
Kulkarni, Prajakta G., 435 R
Kumar, Hitesh, 705 Rajendran, Balaji, 491
Kumar, Pawan, 247 Raj, Gaurav, 731
Kumar, Praveen, 653, 677, 687 Rao, Jyoti, 47
Kumar, Vikrant, 677 Rathod, Neha S., 415
Ratnaparkhi, Vivek, 1
M Rautaray, Siddharth, 457
Mahalle, Parikshit, 579 Rawat, Seema, 643, 687
Makarand, Velankar, 373
Mali, Suresh N., 57, 87, 425 S
Mandal, Kunal Kumar, 183 Sahni, Aashka, 611
Mantri, Shamla T., 557 Sahu, Sanjib Kumar, 695
Manwade, K. B., 67 Saini, Taranpreet Singh, 591
Marale, Prajakta S., 273 Sale, Deepali, 153
Maurya, Shevet Kamal, 533 Salunkhe, Uma R., 87
Misbahuddin, Mohammed, 491 Sane, S. S., 117, 129
Mishra, Richa, 687 Satapathy, Suresh Chand, 653
Misra, Neeraj Kumar, 195 Satapathy, Suresh Chandra, 525
Mittal, Rajat, 653 Saxena, Chandni, 395
Mohan, Sachit, 633 Shah, Rohan, 247
Mohite, Madhuri, 567 Sharma, Govind, 633
Mondal, Saptarshi, 183 Sharma, Mukul, 653
More, Nitin S., 265 Sharma, Priyanka, 257
More, Pournima, 105 Sharma, Rohini, 327
Sharma, Shilpi, 705, 719
N Shinde, Abhijit, 153, 163, 237
Nagmode, M. S., 363 Shinde, Ashish, 153, 163, 237
Naika, Ravika, 603 Shinde, Gitanjali, 227
Nakhate, Sangeeta, 143 Shrestha, Sachit, 279
Narvekar, Meera, 37 Sidnal, Nandini, 383
Singhal, Ashutosh, 525
O Singh, Amit Prakash, 695
Oak, Rajvardhan, 173 Singh, Anil, 525
Olesen, Henning, 227, 291 Singh, Dheerendra, 731
Singh, Japneet, 695
P Singh, Satya P., 533
Palwe, Sushila, 611 Singh, Vishal Kumar, 623
Author Index 747
Singla, R. K., 327 Tyagi, Ishita, 731

Sirsikar, Sumedha, 27
Sonar, Rewati R., 405 U
Sonawane, Bhakti, 257 Urooj, Shabana, 533, 539
Srivastava, Ashita, 515
Sule, Manas, 217 V
Suryawanshi, Renuka, 207 Vidhate, Deepak A., 479
T W
Talukder, Asoke K., 217 Wadhai, Vijay, 557
Tidake, V. S., 117, 129 Waseem, Shaik Mohammed, 19
Tiwari, Deepak Kumar, 515
Totad, Shashikumar G., 77 Y
Trinatha Rao, P., 341 Yadlapati, Avinash, 9
Tuladhar, Jagatjyoti G., 279

Icicc 2017

Uploaded by

Copyright:

Available Formats

Icicc 2017

Uploaded by

Document Information

Original Title

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

Icicc 2017

Uploaded by

Copyright:

Available Formats

Advances in Intelligent Systems and Computing 673

More information about this series at http://www.springer.com/series/11156

Anjali A. Chandavale Anil S. Hiwale

Suresh Chandra Satapathy

Vikrant Bhateja Suresh Chandra Satapathy

ISSN 2194-5357 ISSN 2194-5365 (electronic)

© Springer Nature Singapore Pte Ltd. 2018

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The second International Conference on Intelligent Computing and Communication

Aizuwakamatsu, Japan Subhash Bhalla

35 W GaN Solid-State Driver Power Ampliﬁer for L-Band Radar

A Hybrid Approach for Preprocessing of Imbalanced Data in Credit

Gene Presence and Absence in Genomic Big Data for Precision

Model-Based Design Approach for Software Veriﬁcation Using

Implementation of Credit Card Fraud Detection System with Concept

Biogas Monitoring System Using DS18B20 Temperature Sensor and

Improved Exemplar-Based Image Inpainting Approach . . . . . . . . . . . . 705

Prof. Subhash Bhalla joined as a Faculty School of Computer and Systems

He is Associate Editor in International Journal of Convergence Computing

Vivek Ratnaparkhi and Anil Hiwale

Keywords Class AB GaN HEMT L-band radar PAE

Rapid development in the ﬁeld of gallium nitride (GaN) semiconductor devices

© Springer Nature Singapore Pte Ltd. 2018 1

Table 1 Comparative analysis of different device types

2 GaN Device Technology

Table 2 Material parameters comparison

3 GaN SSPA Design

Fig. 1 Topology of proposed GaN SSPA

4.2 Stability Analysis

Fig. 2 Stability factor and stability measure

4.3 Load Pull Analysis

4.4 Impedance Matching Network Design and Complete

Fig. 3 Results of load pull simulation at maximum power and gain

(b) freq PAE1

Fig. 4 a Output power, b power added efﬁciency

Solid-state driver power ampliﬁer using CGHV40030 GaN HEMT is designed at

Gopika Rani Alekhya Pamarthy, M. Durga Prakash

Abstract Today’s scenario of semiconductor technology is a tremendous inno-

Open core protocol (OCP) is a competent protocol for communication on SoC.

G. R. A. Pamarthy (&) M. Durga Prakash

© Springer Nature Singapore Pte Ltd. 2018 9

OCP is a conﬁgurable protocol that deﬁnes one of the communicating entities as

3.1 OCP Signal Description

Fig. 1 Signal transmission between master and slave

The data transfer between master and slave is communicated non-serially.

Fig. 2 OCP master block

4.2 Slave Block

Fig. 3 OCP slave block

To connect two IP cores together by allowing them to communicate over an OCP

5.1 Basic Mode of OCP

Fig. 4 Timing analysis of basic mode OCP

5.2 Burst Mode of Operation

The implemented design of OCP is veriﬁed by creating the UVM veriﬁcation

Fig. 5 Timing analysis of burst mode OCP

6.1 Simple Mode of Operation (Both Write and Read)

6.2 Burst Mode of Operation (Both Write and Read)

Precise burst mode was implemented and veriﬁed. MBurstLength,

Fig. 6 Simulation results of basic mode OCP