See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/371957250

IMAGE FORGERY DETECTION

Research Proposal · April 2023

DOI: 10.13140/RG.2.2.32571.59680

CITATIONS READS
0 690

5 authors, including:

Gayatri Hegde
Bharati Vidyapeeth Deemed University Department of Engineering and Technology, Kharghar, Navi Mumbai Gayatri does research in Computer Engineering
25 PUBLICATIONS 34 CITATIONS

SEE PROFILE

All content following this page was uploaded by Gayatri Hegde on 30 June 2023.

The user has requested enhancement of the downloaded file.

Journal For Basic Sciences ISSN NO : 1006-8341

IMAGE FORGERY DETECTION

Sankalp Patekar Sumaiya Khan Diksha Bhusare

Information Technology Information Technology Information Technology
PCE New Panvel, Navi Mumbai PCE New Panvel, Navi Mumbai PCE New Panvel, Navi Mumbai

Email:[email protected]. Email:[email protected] Email:[email protected].

ac.in .ac.in ac.in

Manish Bhujbal Prof.Gayatri Hegde

Information Technology Information Technology
PCE New Panvel, Navi Mumbai PCE New Panvel, Navi Mumbai

Email:[email protected]. Email:[email protected]
ac.in

Abstract: Image forgery is a serious problem that I. INTRODUCTION

can have severe consequences in various domains.
The use of deep learning algorithms, such as the It is true that image forgery has become a prevalent
convolutional neural network (CNN), has shown issue in today's society, and the increasing
promising results in detecting such forgeries. availability of image processing tools has made it
CNNs are particularly suited for tasks which are easier for anyone to manipulate and share images
image related, given their ability to extract relevant online. This has led to a need for more sophisticated
features from image data. The proposed system techniques for detecting manipulated images.
involves using a CNN to extract residual Machine learning, and specifically convolutional
noise-based features from images to detect neural networks, have shown promise in this area.
forgeries. This technique involves identifying the
noise pattern left behind by the forgery process, The Error Level Analysis (ELA) method and Peak
which can then be used to distinguish between Signal to Noise Ratio (PSNR) are commonly used
authentic and tampered images. One of the key techniques for detecting manipulation and
advantages of using a CNN for image forgery measuring the quality of compressed or
detection is its ability to handle unseen forgeries. reconstructed images. These methods can be applied
As image forgery techniques become more to images to identify any discrepancies or
sophisticated, it can be challenging to detect alterations, which can then be flagged for further
forgeries using traditional methods. However, analysis.
CNNs can learn to recognize patterns that are not
explicitly defined, allowing them to detect new and In the recognition process, each character in the
previously unseen types of forgeries. Overall, the image can be segmented and identified using
use of CNNs for image forgery detection shows machine learning algorithms. This can be
great potential in combating the problem of image particularly useful in forensic investigations or
tampering. With further research and biomedical research, where accurate identification
development, this technology could be used to of images and their components is crucial.
enhance the reliability and trustworthiness of
digital images in various applications, from Overall, the use of machine learning and image
medical reports to crime scene investigations. processing techniques can help to improve the
efficiency and accuracy of image forgery detection,
Keywords- Image Forgery, Error Level Analysis, and enable better law enforcement and forensic
Convolutional Neural Networks, Deep Learning, investigations.
Forgery Detection of image, Peak Signal to Noise
Ratio.

Volume 23, Issue 4, 2023 PAGE NO: 114

Journal For Basic Sciences ISSN NO : 1006-8341

II. LITERATURE SURVEY suggests a novel method for identifying copy-move

forgeries in digital photos.The approach uses
Image Forgery Detection Using Recompressing statistical features to represent the image and
Images, carried out by Syed Sadaf Ali [1] The employs an optimized technique based on iterative
techniques used are adapted to the individual needs, voting to detect the forgeries.The suggested
interests, and preferences of the user or society. approach is evaluated using different benchmark
Image compression involves reducing the pixels, datasets, and the findings reveal that it detects
size, or colour components of images in order to copy-move forgeries with high accuracy.
reduce the file size for forgery detection. Advanced Digital Image Forgery Detection Based on the
image optimization techniques can detect the more Expectation Maximization Algorithm, Executed by
important image components and discard the less M. H. Alkawaz [9] It proposes a new approach for
vital ones. detecting digital image forgeries using an
expectation-maximization (EM) algorithm. The
Image Forgery Detection by using Support Vector approach models and the probability distribution of
Machine developed by J.Malathi [2] Forgery the forgery and the original image were used to
detection technique that uses illuminant color estimate the parameters of the distribution and
inconsistency and machine learning classifiers such detect the forgery.
as Support Vector Machine (SVM). SVM is a
supervised classification algorithm that is used to Image Forgery Detection Using Image Similarity,
differentiate between two separate categories by carried out by S. al-Zahir and R. Hammad [10] The
drawing a line between them. In this technique, the approach compares the similarities between
illuminant color of input images is estimated, and different regions of an image and uses a clustering
illuminant maps are created for each image. algorithm to identify forged regions. The suggested
Furthermore, all faces present in one image and approach was evaluated using multiple benchmark
corresponding faces of other individual images are datasets, and the findings reveal that it detects image
extracted for investigation. However, it seems that forgeries with high accuracy.
this technique has some drawbacks, such as Copy-move forgery detection based on keypoint
requiring clear textural and inclination highlighting clustering and similar neighborhood search
and affecting the acknowledged substance of the algorithm, executed by H. Chen, X. Yang and Y.
image entirely.It is worth noting that there are Lyu [12] The algorithm uses a clustering technique
several other forgery detection techniques available to group similar keypoints based on scale and color,
that use different approaches, such as image and then matches them to identify tampered regions.
forensics, watermarking, and deep learning-based To locate the tampered regions accurately, a novel
methods. Each technique has its advantages and localization algorithm is employed, which compares
limitations, and the selection of an appropriate the close neighborhoods of matching pairs using
technique depends on various factors such as the two similarity measures and marks the tampered
type of forgery, the available data, and the required regions in the pixels of the images iteratively.
level of accuracy. Overall, this algorithm seems to be designed to
identify tampered regions in images with high
A Full-Image Full-Resolution End-to-End-Trainable accuracy and efficiency.
CNN Framework for Image Forgery Detection, [6]
carried out by F. Marra It proposes a framework for Table 1 Summary of related work on Image Forgery
detecting image forgery using a convolutional Detection
neural network (CNN).The framework includes a
feature extraction module and a classification IMAGE FORGERY DETECTION
module, both using CNNs, and it operates on Literature
full-resolution images.The dataset used is authentic Copy
DNN SVM GBI
Move
and forged images, including various types of
forgeries, to train and test the framework. It also J.Malathi, et al.
Yes Yes
proposes a data augmentation method to improve 2019 [2]
the framework's robustness. F. Matern, et al.
Yes
2020 [3]
Statistical Features based Optimized Technique for
Anushka Singh and
Copy Move Forgery Detection, carried off by S. B.
Jyotsna Singh, Yes
G. T. Babu and C. S. Rao [8] The technique 2022 [5]

Volume 23, Issue 4, 2023 PAGE NO: 115

Journal For Basic Sciences ISSN NO : 1006-8341

S. B. G. T. Babu
and C. S. Rao, Yes To implement an image forgery detection system
2020 [8] using ELA and CNN, the following steps can be
H. Chen, et al.
taken:
2020 [12]
Yes Yes Convert the input image into an ELA image.
Preprocess the ELA image and prepare it for input
into the CNN. Use the CNN to classify the ELA
Collectively, the summary of various techniques image as either real or manipulated.
used as per the recent literature for image forgery If the image is classified as manipulated, further
detection as shown in Table 2. analysis can be performed to determine the type of
forgery that was used.
Table 2 Literature survey summary with techniques It is important to note that while ELA can be a
useful technique for detecting image forgery, it is
Paper Technique not foolproof and can produce false positives or
false negatives. Therefore, it is important to
Syed Sadaf Ali, et al. 2022 [1] Recompression of Images combine ELA with other techniques and methods
for a more accurate and robust detection system.
J.Malathi, et al. 2019 [2] SVM

Gradient-Based B. System Architecture

F. Matern, et al. 2020 [3] The proposed system architecture for image fraud
Illumination
detection consists of several steps, starting with
Anushka Singh and Jyotsna ResNet dataset preparation. The open image dataset's
Singh, 2022 [5] annotations are converted into a format accessible
by the model during the training process. The
F. Marra, et al. 2020 [6] ResNet
testing process involves converting the image into
S. B. G. T. Babu and C. S. Copy-move an ELA image format, calculating the noise and
Rao, 2020 [8] signal ratio, denoising the image, and converting it
to a black-and-white format.
M. H. Alkawaz, et al. 2020 Expectation Maximization
[9]
The model is split into two datasets using the
S. alZahir and R. Hammad, et Clustering Algorithm
train/test method, with 80% used for training and
al. 2020 [10] 20% used for testing the model. The CNN model is
applied to high-scoring regions within the image
H. Chen, et al. 2020 [12] Copy-move that is considered forgeries. A confusion matrix
technique is used to summarize the performance of
the classification algorithm. A table is plotted of all
III. SYSTEM METHODOLOGY the predicted and actual values of the classifier, and
a confidence score is calculated as an evaluation
A. Methodology standard.
ELA (Error Level Analysis) is a technique used for
detecting image forgery. It involves compressing an The confidence score represents the probability of
image to a low quality, then re-saving it at a higher the image being detected correctly by the algorithm
quality, and then calculating the difference between and is given as a percentage. If the confidence score
the two versions of the image. The resulting image is not above a sufficient threshold (i.e., 0.9), it may
is known as an ELA image, and it highlights the be prudent to hold back from making decisions. By
parts of the image that have been manipulated or making fewer predictions, the model's accuracy can
edited. be significantly improved. Each label is assigned a
numerical value called Confidence, while Predict is
Convolutional neural networks (CNNs) are a evaluating an Issue.
popular choice for image forgery detection because
they can learn to recognize patterns and features in Overall, the proposed system architecture appears to
images. The CNN is trained on a dataset of real and be a comprehensive approach to detecting image
manipulated images to learn the characteristics of fraud, with multiple steps to prepare and test the
forged images. Once the CNN is trained, it can be model's accuracy. The use of a confusion matrix
used to classify new images as either real or technique and confidence scores adds an extra layer
manipulated. of evaluation, ensuring that the algorithm's

Volume 23, Issue 4, 2023 PAGE NO: 116

Journal For Basic Sciences ISSN NO : 1006-8341

predictions are reliable before making any In the feature-extraction part of the CNN
decisions. architecture, the input image is passed through a
series of convolutional layers, which apply a set of
learnable filters to extract features from the image.
Each filter produces a feature map that highlights a
particular pattern or feature in the input image.
These feature maps are then passed through
activation functions like ReLU to introduce
non-linearity and avoid the vanishing gradient
problem.

After the feature extraction process, the output of

the last convolutional layer is flattened into a
1-dimensional vector and fed into a series of fully
Fig.1 System architecture for Image Forgery Detection connected layers for classification. The fully
connected layers use the extracted features to make
C. Convolutional Neural Network predictions about the class of the input image. The
Convolutional Neural Networks (CNNs) have final output layer usually employs the softmax
indeed become a popular tool for detecting forgery function to generate a probability distribution over
images. CNNs are a type of deep learning algorithm the classes, indicating the most likely class for the
that can be trained to extract features from images input image.
and classify them into different categories. They are
inspired by the human visual system and consist of
multiple layers of interconnected neurons that
perform convolution operations on the input image
to extract features.

One of the advantages of using CNNs for image

forensics is their ability to detect subtle artifacts that
may not be visible to the naked eye. For example,
when an image is manipulated, such as by
copy-pasting a fragment from one image to another,
there may be slight variations in the pixel values or
texture that are indicative of the manipulation.
CNNs can learn to detect these differences and Fig. 2 CNN Architecture
classify the image as either genuine or fake.
That's a good summary of the main functions of
each layer in a convolutional neural network. Here's
Overall, CNNs have shown great promise in a
a bit more detail on each layer:
variety of computer vision and image processing
Convolutional layer: This layer applies a set of
applications, including image forensics. With the
filters to the input image or feature map, generating
increasing prevalence of digital manipulation in
a set of output feature maps. Each filter looks for a
today's world, the ability to detect forged images has
specific pattern or feature in the input, and the
become more important than ever, and CNNs
output feature maps highlight where those features
provide a powerful tool for this purpose.
are present in the input. By stacking multiple
convolutional layers, the network can learn
increasingly complex and abstract features.
The input layer in a convolutional neural network
Pooling layer: This layer downsamples the output
(CNN) is where the images from the dataset are fed.
feature maps from the convolutional layer, reducing
The images are usually in the form of 3-dimensional
their spatial size and number of parameters. The
arrays, with the first two dimensions representing
most common type of pooling is max pooling,
the height and width of the image (the number of
which takes the maximum value within a small
pixels), and the third dimension representing the
region of the feature map. This helps to capture the
red, green, and blue (RGB) colors present in each
most salient features while discarding redundant
pixel.
information, and also makes the network more
robust to variations in input position and scale.

Volume 23, Issue 4, 2023 PAGE NO: 117

Journal For Basic Sciences ISSN NO : 1006-8341

Fully-connected layer: This layer takes the

flattened output of the previous layer and applies a
set of weights to produce a vector of class
probabilities. The weights are learned during
training using backpropagation and gradient
descent. The softmax activation function ensures
that the output probabilities sum to 1, allowing the
network to make a single prediction for the input
image. The number of neurons in the
fully-connected layer corresponds to the number of
output classes.

D. Image Forgery Detection

The detection of fake images is done by
convolutional neural networks. CNN helps in
recognising tempered images, and it is mainly used
to find tampered images accuracy.
Image Processing: The image obtained undergoes
Fig.3 A CNN which acts a backbone for the model
the following steps: ELA conversion, grayscale
conversion, thresholding, and calculation of
IV. SYSTEM IMPLEMENTATION
confidence.
A. Software and Hardware
● The Error Level Analysis (ELA) method is
one way to identify the areas of the image that The system requirements for running an image
have been modified. It works by creating a forgery detection system using a CNN model
difference map of the image by compressing implemented in Python 3.7.X (IDLE) and the
and decompressing it with a low-quality JPEG CASIA dataset.It is recommended that the computer
algorithm. The parts of the image that have system has at least:
been modified will have a different ● RAM: 8GB or more
compression rate and will appear as bright ● Hard Disk Drive (HDD): 80GB or more
spots in the difference map. ● Processor: i5 or higher
These system requirements are necessary to handle
● Grayscale conversion is often used to simplify large amounts of data and the constant nature of the
the image and reduce its complexity. It environment. It is important to note that the specific
involves converting the image into a hardware and software requirements may vary
black-and-white or gray-scale image, where depending on the size of the dataset and the
each pixel's value represents its intensity. complexity of the CNN model being used.
Therefore, it is always a good idea to check the
● Thresholding is a technique used to convert specific requirements of the software and datasets
the grayscale image into a binary image, being used before implementing an image forgery
where each pixel is either black or white. This detection system.
process helps eliminate noise in the image and Additionally, it is recommended to have sufficient
can improve the accuracy of the detection cooling and power supply to ensure that the system
algorithm. can run smoothly and avoid any unexpected
shutdowns or errors.
● Confidence: A confidence score is a number
between 0 and 1.The confidence value can be B. Dataset
calculated for only one input, which gives the The CASIA v2.0 database contains a total of 10,000
meaning of the algorithm confidence for this images, divided into two subsets: a training set of
class. 5,000 images and a testing set of 5,000 images.
Each subset includes eight categories of images:
animal, architecture, article, character, nature, plant,
scene, and texture. The images are in JPEG format
and have a size of either 256 x 384 or 384 x 256
pixels.CASIA V2.0 dataset is used for image
forgery detection.

Volume 23, Issue 4, 2023 PAGE NO: 118

Journal For Basic Sciences ISSN NO : 1006-8341

Two classes make up this dataset: actual photos and positive outcome, and true negatives occur when the
tampering detection. There are 7354 images, which model correctly predicts a negative outcome. False
are classified into real images and altered images in positives occur when the model predicts a positive
JPG format. outcome, but the actual outcome is negative, and
false negatives occur when the model predicts a
negative outcome, but the actual outcome is
Dataset Size Categories Format positive.
The above table has the following cases:
CASIA 5 GB 8 categories JPEG
V2.0 of images

Table 3 Details of CASIA Dataset

Fig.6 Confusion matrix of image forgery detection

A confusion matrix is a table used to evaluate the

performance of a classification model on a set of
test data whose true class labels are known.

V. RESULTS AND ANALYSIS

The original image and its ELA-converted

counterpart are shown in Fig. 7 and Fig. 8 of the
Fig.4 CASIA Dataset dataset, respectively. And the fake image and its
corresponding ELA-converted image are shown in
C. Confusion Metrics Figs. 9 and 10, respectively.
A confusion matrix is a table that is commonly
used to evaluate the performance of a classification In Fig. 11, the red line represents the model's
algorithm by comparing the predicted labels to the training loss and training accuracy, while the blue
true labels of a set of test data. The matrix displays line represents the model's validation loss and
the number of true positive, false positive, true validation accuracy. The model is iteratively trained
negative, and false negative predictions made by the and has an accuracy of 78.08%.
algorithm

Fig.5 Confusion Matrix Fig.7 Original image from dataset

The above table is a confusion matrix that
summarizes the performance of a binary
classification model, and it includes four possible
outcomes: true positives (TP), true negatives (TN),
false positives (FP), and false negatives (FN). True
positives occur when the model correctly predicts a

Volume 23, Issue 4, 2023 PAGE NO: 119

Journal For Basic Sciences ISSN NO : 1006-8341

Fig.12 Accuracy of the model

VI. CONCLUSION

Fig.8 ELA conversion of original image Image Forgery Detection System is developed and
implemented using Convolutional Neural Networks
for detecting manipulation in the images using the
model over the CASIA V2.0 dataset. These images
are converted into black-and-white format using the
ELA method, then PSNR is applied to calculate
noise and denoise the images, which are then passed
to the detection system where recognition of
manipulated images takes place. Once the forged
images are recognized, they are displayed as output.
A confusion matrix is used to evaluate performance,
and the findings are displayed in a table that takes
into account all of a classifier's anticipated and
actual values. The confidence score is then
Fig.9 Fake image from the dataset computed as an evaluation standard. The model's
accuracy after iterative training is 78.08%.

VII. REFERENCES

[1] Ali, S.S.; Ganapathi, I.I.; Vu,N.-S.; Ali, S.D.;

Saxena, N.; Werghi, N., “Image Forgery
Detection Using Deep Learning by Recompre
-ssing Images,” Electronics 2022, 11, 403.
[2] J.Malathi, B.Narasimha Swamy, Ramgopal
Musunuri, "Image Forgery Detection by using
Machine Learning, International Journal of
Innovative Technology and Exploring Engine
-ering (IJITEE)ISSN: 2278-3075, Volume-8,
Issue- 6S4, April 2019.
Fig 10 ELA conversion of fake image [3] F. Matern, C. Riess and M. Stamminger,
"Gradient-Based Illumination Description for
Image Forgery Detection," in IEEE Transac
-tions on Information Forensics and Security,
vol. 15, pp. 1303-1317, 2020, doi:10.1109/TIF
S.2019.2935913.
[4] Z. J. Barad and M. M. Goswami, "Image
Forgery Detection using Deep Learning: A
Survey," 2020 6th International Conference on
Advanced Computing and Communication
Systems (ICACCS), 2020, pp. 571-576, doi:
10.1109/ICACCS48705.2020.9074408.
[5] Anushka Singh and Jyotsna Singh,“ Image
Fig. 11: Evaluation between training loss w.r.t. validation forgery detection using Deep Neural
loss and training accuracy w.r.t. validation accuracy Network,” Conference: 2021 8th International

Volume 23, Issue 4, 2023 PAGE NO: 120

Journal For Basic Sciences ISSN NO : 1006-8341

Conference on Signal Processing and

Integrated Networks (SPIN)At: New Delhi,
January 2022 DOI:10.1109 SPIN525336.2021
.9565953.
[6] F. Marra, D. Gragnaniello, L. Verdoliva and G.
Poggi,"A Full-Image Full-Resolution End-to
-End-Trainable CNN Framework for Image
Forgery Detection," in IEEE Access, vol. 8,
pp. 133488-133502, 2020, doi:10.1109/ACCE
SS.2020.3009877.
[7] R. Agarwal, D. Khudaniya, A. Gupta and K.
Grover, "Image Forgery Detection and Deep
Learning Techniques: A Review," 2020 4th
International Conference on Intelligent
Computing and Control Systems (ICICCS),
Madurai, India, 2020, pp. 1096-1100, doi:
10.1109/ICICCS48265.2020.9121083.
[8] S. B. G. T. Babu and C. S. Rao, "Statistical
Features based Optimized Technique for Copy
Move Forgery Detection," 2020 11th Inter-
national Conference on Computing, Communi
-cation and Networking Technology
(ICCCNT), Kharagpur, India, 2020, pp. 1-6,
doi: 10.1109/ICCCNT49239.2020.9225426.
[9] M. H. Alkawaz, M. T. Veeran and R. Bachok,
"Digital Image Forgery Detection based on
Expectation Maximization Algorithm," 2020
16th IEEE International Colloquium on Signal
Processing and Its Applications (CSPA),
Langkawi, Malaysia, 2020, pp. 102-105, doi:
10.1109/CSPA48992.2020.9068731.
[10] alZahir, S., Hammad, R. Image forgery
detection using image similarity. Multimed
Tools Appl 79, 28643–28659 (2020).
[11] K. M. Hosny, A. M. Mortda, M. M. Fouda
and N. A. Lashin, "An Efficient CNN Model
to Detect Copy-Move Image Forgery," in
IEEE Access, vol. 10, pp. 48622-48632, 2022,
doi: 10.1109/ACCESS.2022.3172273.
[12] H. Chen, X. Yang and Y. Lyu, "Copy-Move
Forgery Detection Based on Keypoint
Clustering and Similar Neighborhood Search
Algorithm," in IEEE Access, vol. 8, pp.
36863-36875,2020,doi:10.1109/ACCESS.202
0.2974804.

Volume 23, Issue 4, 2023 PAGE NO: 121

View publication stats