2023 3rd International Conference on Intelligent Technologies (CONIT)

Karnataka, India. June 23-25, 2023

Image Forgery Detection using CNN

Meet Patel Kartikay Rane Niyati Jain
Information Technology Information Technology Information Technology
St. Francis Institute Of Technology St. Francis Institute Of Technology St. Francis Institute Of Technology
Mumbai, India Mumbai, India Mumbai, India
meetpatel14101@gmail.com kartikayrane1852001@gmail.com nitijain167@gmail.com
2023 3rd International Conference on Intelligent Technologies (CONIT) | 979-8-3503-3860-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/CONIT59222.2023.10205377

Praneel Mhatre Shree Jaswal

Information Technology Information Technology
St. Francis Institute Of Technology St. Francis Institute Of Technology
Mumbai, India Mumbai, India
praneelmhatre14@gmail.com shreejaswal@sfit.ac.in

Abstract—With the increasing use of digital images in advantages of CNNs is their ability to learn spatial
various applications, the problem of image forgery has become hierarchies of features, enabling them to detect patterns at
more prevalent than ever. In this paper, we propose a novel different scales and orientations within an image. This makes
image forgery detection system based on Convolutional Neural them well-suited for tasks such as object recognition, face
Networks (CNNs) that can detect various types of image detection, and even medical image analysis. Additionally,
manipulations, including copy-move, splicing, and retouching. CNNs are highly adaptable and can be trained on large
Our proposed system integrates Error Level Analysis (ELA) datasets to improve their accuracy and generalization ability.
with deep learning techniques to provide a more accurate and
reliable solution to the problem of image forgery detection. We II. LITERATURE REVIEW
evaluated the proposed system on a dataset of real-world
images and achieved a high detection accuracy of 93%. Our A. Literature review related to exiting paper
system outperformed existing methods for image forgery Ranjan et al [1] created a Graphical User Interface (GUI)
detection and demonstrated its potential for various for detection of digitally tampered images. This method has
applications, including forensics, security, and digital image accuracy of 96.4% and has proven to be efficient and handy.
analysis. Overall, the proposed CNN-based image forgery
Foundation on investigation of digitally manipulated
detection system offers a robust and effective solution to the
documents and provides a solution to distinguish original
growing problem of image manipulation and forgery in today's
visual media landscape. documents from a digitally morphed document.
Ali et al [2] mentioned in their study the comparison of
Keywords—Convolutional Neural Network(CNN), Error differ- ent techniques. Experimental output with accuracy
Level Analysis(ELA), Copy-Move, Slicing was shown. The output was shown visually for better
I. INTRODUCTION understanding.
Image forgery, the practice of manipulating digital Amit Doegara et al[3] used CNN algorithm for detection
images to deceive or mislead viewers, has become a growing of forgery in the images as the self learning capabilities of
concern in today's world of visual media. With the increasing CNN is very useful. Two modes were used that is active and
use of digital images in fields such as journalism, social passive. In the active mode approach images can be pre-
media, and e-commerce, the need for effective image forgery embed with some digital signature or watermarks. In the
detection systems has become more pressing than ever. In passive mode approach the images need not to be pre-embed
forensics, these systems can help detect tampering in crime with any digital signature or watermark.
scene images, while in security, they can assist in identifying Boubacar Diallo et al[4] presented an approach in their
fraudulent documents or images. Moreover, image forgery study that strengthens the robustness of image forgery
detection systems can also be used in e-commerce and detection. They worked on a convolutional neural network-
advertising to ensure that the product images are authentic based model for camera recognition. They tested the
and not manipulated.For making this type of model different approach on this modification since lossy compression, such
deep learning algorithms can be used. With deep learning, as JPEG, is thought to be the most common method of
machines can identify patterns and relationships in data that purposeful or unintentional concealment of picture forgeries.
would be impossible for humans to detect, leading to A mixture of different qualities of compressed and
breakthroughs in image and speech recognition, natural uncompressed pictures are fed to trained CNN.
language processing, and more. In recent years,
Convolutional Neural Networks (CNNs) have emerged as a B. Literature review related to existing websites
powerful tool for image forgery detection. These deep Forensically [10] is very simple and user-friendly which
learning models are capable of analyzing complex visual allows clone detection and error level analysis with other
features in images and detecting even subtle signs of features. Its functions include ELA, providing MetaData,
manipulation. CNN-based forgery detection systems can detecting clones, etc. Forensically may be compared to a
identify various types of image manipulations, including magnifying glass. It demonstrates users the true details that
copy-move, splicing, and retouching, and provide reliable are concealed from everyone else. Sub features for clone
results even in the presence of noise or compression artifacts. features are a little hard to understand. No percentage of
CNN provides multiple layers of neurons for processing the forgery is shown on the website.
complex features of the image.[8] One of the main

979-8-3503-3860-7/23/$31.00 ©2023 IEEE 1

Authorized licensed use limited to: The Technology Library. Downloaded on December 17,2024 at 19:09:51 UTC from IEEE Xplore. Restrictions apply.
 FotoForensics [11] has an option for feature for forgery tampered images.The number of authenticated images is
detection. It demonstrates the error level on the picture, thus around 7000 and the number of tampered images is around
identifying whether it was edited using Photoshop or not. In 3000.
case the picture was altered or photoshopped, you will see
color in picture analysis. In case the picture is unedited, you Thus to meet this requirement two datasets – CASIA 2.0
will see a regular white color on it. Not a simple User [13] and MICC F200 [14] is combined and this combination
Interface. Not many features, only forgery feature is there is used to train the model.
which does not show percentage of forgery. C. Working of CNN
Amped Authenticate [12] has many analysis features and In the context of image forgery detection, the CNN
tools.It helps to detect if the images are original or edited. algorithm is used to classify an image as either authentic or
Amped Authenticate combines powerful tests, operations forged based on the features extracted from the preprocessed
and reports in one package to enable users to identify altered images. The algorithm is trained on a large dataset
or entirely original pictures. Vuforia is a popular freemium containing both authentic and forged images, and the
platform where using Vuforia Object Scanner, we can scan features are learned through a process of convolution,
and create object targets pooling, and activation functions.

III. METHODOLOGY The convolution layer applies a set of learnable filters to

the input image to extract relevant features, such as edges
A. System Architecture and shapes. The pooling layer reduces the spatial dimensions
In the below figure 1, the user must upload an image on of the feature maps and helps to extract the most important
the website front-end. The image will be evaluated by the information from the features. The activation function
CNN model. The model will be developed by training and applies a nonlinear transformation to the output of the
testing the dataset. For creating a model, it has to go through convolution and pooling layers, which enhances the ability
the following stages. First is the pre-processing stage where of the network to learn complex features and patterns.
any noisy data present in the dataset is cleaned and null After feature extraction, the output of the network is
values are removed. After the PreProcessing stage, the image passed through a fully connected layer, which maps the
is then passed on to the feature extraction stage. In the learned features to a set of probabilities. The softmax
feature extraction stage, the required features are kept and activation function is used to convert the output into a
the other unimportant features are removed for further stage. probability distribution over the two classes (authentic and
It generates better outputs rather than applying any algorithm forged), and the class with the highest probability is
directly to the raw data. After feature extraction, CNN considered as the final classification result.
algorithm is applied to get the final result. After all the
processes are completed and the final output is generated, it During training, the CNN algorithm learns to optimize
is displayed to the user. the parameters of the network to minimize the loss function
and improve the accuracy of classification. The
backpropagation algorithm is used to propagate the error
from the output layer to the input layer, and the parameters
are updated using an optimization algorithm, such as
stochastic gradient descent (SGD).
The CNN algorithm is effective in detecting various
types of image forgery, including copy-move forgery,
splicing forgery, removal forgery, and retouching forgery.
By analyzing the features extracted from the preprocessed
images, the algorithm can identify the inconsistencies and
artifacts that are associated with image forgeries. The
accuracy of the CNN algorithm in detecting image forgeries
depends on the quality and diversity of the training dataset,
the complexity of the network architecture, and the
optimization algorithm used for training.
IV. RESULTS

Fig. 1. System Architecture

B. Dataset Details
There are several types of image forgery, some of them
are copy-move forgery, slicing forgery, removal forgery,
retouching forgery. To detect these above-mentioned types
of forgeries, the model must be trained with data which
contains these types of tampered images. So a custom dataset
was created with some of these types of tampered images.
Fig. 2. Accuracy of the model
The dataset consists of two folders – authenticated images,

2
Authorized licensed use limited to: The Technology Library. Downloaded on December 17,2024 at 19:09:51 UTC from IEEE Xplore. Restrictions apply.
 As shown in fig 2 we can observe that after training the
model with a small amount of image i.e approximately 2000
images, an accuracy of 99.3217% was achieved. Then after
training and testing the model with 9500 images, an
approximate accuracy of 93% was achieved.

Fig. 3. Loss and Accuracy

Fig 3 shows the training validation loss and training

validation accuracy of our model. As we can observe from
fig 3. there is very minimal loss achieved during training our Fig. 5. ELA
model with a high detection accuracy.
Error Level Analysis (ELA) is a digital image analysis
technique used to identify potential areas of manipulation in
an image. ELA works by analyzing the differences in the
error levels between different regions of an image. Error
levels refer to the amount of compression artifacts present in
an image after it has been compressed and saved in a lossy
format such as JPEG.
A. UI Design

Fig. 6. Uploading Real Image

Fig. 4. Confusion Matrix

Fig 4. shows the confusion matrix achieved while

training our model.
True positive: 393
False Positive: 14
False Negative: 51
True Negative: 375

Fig. 7. Result

3
Authorized licensed use limited to: The Technology Library. Downloaded on December 17,2024 at 19:09:51 UTC from IEEE Xplore. Restrictions apply.
 Fig 7 shows the result of the image we took from the user forgery system is a valuable contribution to fight against
and the result shows the model confidence, that how digital deception.
confident the model is that our image is fake or real. So the
confidence was 99% meaning that the model is 99% sure it is REFERENCES
a real image. [1] Ranjan, Shruti, Prayati Garhwal, Anupama Bhan, Monika Arora, and
Anu Mehra. ”Framework for image forgery detection and
classification using machine learning.” In 2018 2nd International
Conference on Trends in Electronics and Informatics (ICOEI), pp. 1-
9. IEEE, 2018.
[2] Ali, S. S., I. I. Ganapathi, N. S. Vu, S. D. Ali, N. Saxena, and N.
Werghi. ”Image Forgery Detection Using Deep learning by
Recompressing Images. Electronics 2022, 11, 403.” (2022).
[3] Amit Doegara , Maitreyee Duttaa , Gaurav Kumarb. ”CNN based
Image Forgery Detection using pre-trained AlexNet Model” 2018.
[4] Boubacar Diallo, Thierry Urruty, Pascal Bourdon, Christine
Fernandez- Maloigne. ”Robust forgery detection for compressed
images using CNN supervision” IForensic Science International:
Reports,Volume 2,2020,100112.
[5] Sarma, Barnali, and Gypsy Nandi. ”A study on digital image forgery
Fig. 8. Uploading Fake Image detection.” International Journal of Advanced Research in Computer
Science and Software Engineering 4, no. 11 (2014).
[6] Barad, Zankhana J., and Mukesh M. Goswami. ”Image forgery
detection using deep learning: a survey.” In 2020 6th International
Conference on Advanced Computing and Communication Systems
(ICACCS), pp. 571-
[7] 576. IEEE, 2020.
[8] Kuznetsov, A. ”Digital image forgery detection using deep learning
approach.” In Journal of Physics: Conference Series, vol. 1368, no. 3,
[9] p. 032028. IOP Publishing, 2019.
[10] Doegar, Amit, Maitreyee Dutta, and Kumar Gaurav. ”Cnn based
image forgery detection using pre-trained alexnet model.”
International Journal of Computational Intelligence IoT 2, no. 1
(2019)
[11] Ganguly, Shreyan, Aditya Ganguly, Sk Mohiuddin, Samir Malakar,
and Ram Sarkar. ”ViXNet: Vision Transformer with Xception
Fig. 9. Result Network for deepfakes based video and image forgery detection.”
Expert Systems with Applications 210 (2022)
Fig 9 shows the result of the fake image imputed from [12] ”Forensically Beta”, https://29a.ch/photo-forensics/forensic-magnifier
(accessed on : Jan 16, 2023)
the user and shows the confidence of almost 93% meaning
[13] ”FotoForensics”,https://fotoforensics.com/ (accessed on : Jan 17,
the model is 93% sure that the image imputed is fake. 2023)
V. CONCLUSION [14] ”Amped Authenticate”, https://ampedsoftware.com/authenticate (ac-
cessed on : Jan 17, 2023)
In conclusion, the image forgery system using the CNN [15] ”CASIA 2.0”, https://www.kaggle.com/datasets/divg07/casia-20-
algorithm is a promising tool for detecting and localizing image- tampering-detection-dataset/code (accessed on Jan 15, 2023)
image forgeries. Its accuracy, speed, and versatility make it [16] “MICC F200”, http://lci.micc.unifi.it/labd/2015/01/copy-move-
suitable for various applications, and its potential for forgery-detection-and-localization/ (accessed on feb 15, 2023)
enhancing the integrity and credibility of digital images
cannot be overstated. As image manipulation continues to be
a significant concern in today’s digital world, the image

4
Authorized licensed use limited to: The Technology Library. Downloaded on December 17,2024 at 19:09:51 UTC from IEEE Xplore. Restrictions apply.