ORCA – Online Research @ Cardiff

Thi s is a n O p e n Acc e s s d o c u m e n t d o w nlo a d e d fro m ORCA, C a r diff U niv e r si ty' s

in s ti t u tio n al r e p o si t o ry: h t t p s://o r c a. c a r diff.a c. u k/i d/ e p ri n t/ 1 4 6 8 5 1/

This is t h e a u t h o r’s v e r sio n of a w o r k t h a t w a s s u b mi t t e d t o / a c c e p t e d fo r

p u blic a tio n.

Cit a tio n fo r fin al p u blis h e d ve r sio n:

Ali, S y e d S a d af, G a n a p a t hi, Iyy a k u t ti Iy a p p a n , Vu, N g o c-S o n, Ali, Sy e d D a ni s h,

S ax e n a , N e e t e s h a n d We r g hi 2 0 2 2 . I m a g e fo r g e ry d e t e c tio n u si n g d e e pl e a r ni n g by
r e c o m p r e s si n g t h e i m a g e s . El e c t r o nic s 1 1 (3) , 4 0 3 . 1 0 . 3 3 9 0/ el e c t r o nic s 1 1 0 3 0 4 0 3
file

P u blis h e r s p a g e : h t t p s:// doi.o r g/ 1 0. 3 3 9 0/ el e c t r o nic s 1 1 0 3 0 4 0 3

Pl e a s e n o t e:
C h a n g e s m a d e a s a r e s ul t of p u blis hi n g p r o c e s s e s s u c h a s c o py-e di ti n g, fo r m a t ti n g
a n d p a g e n u m b e r s m a y n o t b e r efl e c t e d in t hi s v e r sio n. Fo r t h e d efi nitiv e v e r sio n of
t hi s p u blic a tio n, pl e a s e r ef e r t o t h e p u blis h e d s o u r c e . You a r e a d vis e d t o c o n s ul t t h e
p u blis h e r’s v e r sio n if yo u wis h t o ci t e t hi s p a p er.

This ve r sio n is b ei n g m a d e a v ail a bl e in a c c o r d a n c e wi t h p u blis h e r p olici e s. S e e

h t t p://o r c a . cf. a c. u k/ p olici e s. h t ml fo r u s a g e p olici e s. Co py ri g h t a n d m o r al ri g h t s fo r
p u blic a tio n s m a d e a v ail a bl e in ORCA a r e r e t ai n e d by t h e c o py ri g h t h ol d e r s .
 electronics
Article
Image Forgery Detection Using Deep Learning by
Recompressing Images
Syed Sadaf Ali 1, *,† , Iyyakutti Iyappan Ganapathi 2,† , Ngoc-Son Vu 1 , Syed Danish Ali 3 , Neetesh Saxena 4
and Naoufel Werghi 2

1 ETIS, CY Cergy Paris Université, ENSEA, CNRS, UMR 8051, 95000 Cergy, France; [email protected]
2 C2PS & KUCARS, Khalifa University, Abu Dhabi 127788, United Arab Emirates;
[email protected] (I.I.G.); [email protected] (N.W.)
3 Machine Intelligence Research (MIR) Labs Gwalior, Gwalior 474001, India; [email protected]
4 School of Computer Science and Informatics, Cardiff University, Cardiff CF10 3AT, UK; [email protected]
* Correspondence: [email protected]
† These authors contributed equally to this work.

Abstract: Capturing images has been increasingly popular in recent years, owing to the widespread
availability of cameras. Images are essential in our daily lives because they contain a wealth of
information, and it is often required to enhance images to obtain additional information. A variety of
tools are available to improve image quality; nevertheless, they are also frequently used to falsify
images, resulting in the spread of misinformation. This increases the severity and frequency of
image forgeries, which is now a major source of concern. Numerous traditional techniques have
been developed over time to detect image forgeries. In recent years, convolutional neural networks
(CNNs) have received much attention, and CNN has also influenced the field of image forgery
detection. However, most image forgery techniques based on CNN that exist in the literature are
limited to detecting a specific type of forgery (either image splicing or copy-move). As a result, a




 technique capable of efficiently and accurately detecting the presence of unseen forgeries in an image
Citation: Ali, S.S.; Ganapathi, I.I.; Vu, is required. In this paper, we introduce a robust deep learning based system for identifying image
N.-S.; Ali, S.D.; Saxena, N.; Werghi, N. forgeries in the context of double image compression. The difference between an image’s original
Image Forgery Detection Using Deep and recompressed versions is used to train our model. The proposed model is lightweight, and its
Learning by Recompressing Images. performance demonstrates that it is faster than state-of-the-art approaches. The experiment results
Electronics 2022, 11, 403. https:// are encouraging, with an overall validation accuracy of 92.23%.
doi.org/10.3390/electronics11030403

Academic Editors: Huy Kang Kim Keywords: convolutional neural networks; neural networks; forgery detection; image compression;
and Gemma Piella image processing

Received: 15 November 2021

Accepted: 24 January 2022
Published: 28 January 2022
1. Introduction
Publisher’s Note: MDPI stays neutral
Due to technological advancements and globalization, electronic equipment is now
with regard to jurisdictional claims in
widely and inexpensively available. As a result, digital cameras have grown in popularity.
published maps and institutional affil-
There are many camera sensors all around us, and we use them to collect a lot of images.
iations.
Images are required in the form of a soft copy for various documents that must be filed
online, and a large number of images are shared on social media every day. The amazing
thing about images is that even illiterate people can look at them and extract information
Copyright: © 2022 by the authors.
from them. As a result, images are an integral component of the digital world, and they play
Licensee MDPI, Basel, Switzerland. an essential role in storing and distributing data. There are numerous tools accessible for
This article is an open access article quickly editing the images [1,2]. These tools were created with the intention of enhancing
distributed under the terms and and improving the images. However, rather than enhancing the image, some people exploit
conditions of the Creative Commons their capabilities to falsify images and propagate falsehoods [3,4]. This is a significant threat,
Attribution (CC BY) license (https:// as the damage caused by faked images is not only severe, but also frequently irreversible.
creativecommons.org/licenses/by/ There are two basic types of image forgery: image splicing and copy-move, which are
4.0/). discussed below:

Electronics 2022, 11, 403. https://doi.org/10.3390/electronics11030403 https://www.mdpi.com/journal/electronics

Electronics 2022, 11, 403 2 of 17

• Image Splicing: A portion of a donor image is copied into a source image. A sequence
of donor images can likewise be used to build the final forged image.
• Copy-Move: This scenario contains a single image. Within the image, a portion of the
image is copied and pasted. This is frequently used to conceal other objects. The final
forged image contains no components from other images.
The primary purpose in both cases of image forgery is to spread misinformation by
changing the original content in an image with something else [5,6]. Earlier images were
an extremely credible source for the information exchange, however, due to image forgery,
they are used to spread misinformation. This is affecting the trust of the public in images,
as the forging of images may or may not be visible or recognizable to the naked eye. As a
result, it is essential to detect image forgeries to prevent the spread of misinformation as
well as to restore public trust in images. This can be done by exploring the various artifacts
left behind when an image forgery is performed, and they can be identified using various
image processing techniques.
Researchers have proposed a variety of methods for detecting the presence of image
forgeries [7–9]. Conventional image forgery detection techniques detect forgeries by concen-
trating on the multiple artifacts present in a forged image, such as changes in illumination,
contrast, compression, sensor noise, and shadow. CNN’s have gained popularity in recent
years for various computer vision tasks, including image object recognition, semantic
segmentation, and image classification. Two major features contribute to CNN’s success
in computer vision. Firstly, CNN takes advantage of the significant correlation between
adjacent pixels. As a result, CNN prefers locally grouped connections over one-to-one
connections between all pixel. Second, each output feature map is produced through a
convolution operation by sharing weights. Moreover, compared to the traditional method
that depends on engineered features to detect specific forgery, CNN uses learned features
from training images, and it can generalize itself to detect unseen forgery. These advan-
tages of CNN make it a promising tool for detecting the presence of forgery in an image.
It is possible to train a CNN-based model to learn the many artifacts found in a forged
image [10–13]. Thus, we propose a very light CNN-based network, with the primary goal
of learning the artifacts that occur in a tampered image as a result of differences in the
features of the original image and the tampered region.
The major contribution of the proposed technique are as follows:
• A lightweight CNN-based architecture is designed to detect image forgery efficiently.
The proposed technique explores numerous artifacts left behind in the image tamper-
ing process, and it takes advantage of differences in image sources through image
recompression.
• While most existing algorithms are designed to detect only one type of forgery, our
technique can detect both image splicing and copy-move forgeries and has achieved
high accuracy in image forgery detection.
• Compared to existing techniques, the proposed technique is fast and can detect the
presence of image forgery in significantly less time. Its accuracy and speed make it
suitable for real-world application, as it can function well even on slower devices.
The rest of the paper is organized as follows. Section 2 provides a literature review of image
forgery detection methodologies. Section 3 introduces the proposed framework for detecting
the presence of forgeries in an image. Section 4 contains a discussion of the experimentation
and the results achieved. Finally, in Section 5, we summarize the conclusions.

2. Literature Review
Various approaches have been proposed in the literature to deal with image forgery.
The majority of traditional techniques are based on particular artifacts left by image forgery,
whereas recently techniques based on CNNs and deep learning were introduced, which
are mentioned below. First, we will mention the various traditional techniques and then
move on to deep learning based techniques.
Electronics 2022, 11, 403 3 of 17

In [14], the authors’ proposed error level analysis (ELA) for the detection of forgery
in an image. In [15], based on the lighting conditions of objects, forgery in an image is
detected. It tries to find the forgery based on the difference in the lighting direction of the
forged part and the genuine part of an image. In [16], various traditional image forgery
detection techniques have been evaluated. In [17], Habibi et al., use the contourlet trans-
form to retrieve the edge pixels for forgery detection. In [18], Dua et al., presented a JPEG
compression-based method. The discrete DCT coefficients are assessed independently for
each block of an image partitioned into non-overlapping blocks of size 8 × 8 pixels. The sta-
tistical features of AC components of block DCT coefficients alter when a JPEG compressed
image tampers. The SVM is used to classify authentic and forged images using the retrieved
feature vector. Ehret et al. in [19] introduced a technique that relies on SIFT, which provides
sparse keypoints with scale, rotation, and illumination invariant descriptors for forgery
detection. A method for fingerprint faking detection utilizing deep Boltzmann machines
(DBM) for image analysis of high-level characteristics is proposed in [20]. Balsa et al. in [21]
compared the DCT, Walsh–Hadamard transform (WHT), Haar wavelet transform (DWT),
and discrete Fourier transform (DFT) for analog image transmission, changing compression
and comparing quality. These can be used for image forgery detection by exploring the
image from different domains. Thanh et al. proposed a hybrid approach for image splicing
in [22], in which they try to retrieve the original images that were utilized to construct the
spliced image if a given image is proven to be the spliced image. They present a hybrid
image retrieval approach that uses Zernike moment and SIFT features.
Bunk et al. established a method for detecting image forgeries based on resampling
features and deep learning in [23]. Bondi et al. in [24] suggested a method for detecting
image tampering by the clustering of camera-based CNN features. Myung-Joon in [2]
introduced CAT-Net, to acquire forensic aspects of compression artifact on DCT and RGB
domains simultaneously. Their primary network is HR-Net (high resolution). They used
the technique proposed in [25], which tells us that how we can use the DCT coefficient
to train a CNN, as directly giving DCT coefficients to CNN will not train it efficiently.
Ashraful et al. in [26] proposed DOA-GAN, to detect and localize copy-move forgeries in
an image, authors used a GAN with dual attention. The first-order attention in the generator
is designed to collect copy-move location information, while the second-order attention
for patch co-occurrence exploits more discriminative properties. The affinity matrix is
utilized to extract both attention maps, which are then used to combine location-aware and
co-occurrence features for the network’s ultimate detection and localization branches.
Yue et al. in [27] proposed BusterNet for copy-move image forgery detection. It has
a two-branch architecture with a fusion module in the middle. Both branches use visual
artifacts to locate potential manipulation locations and visual similarities to locate copy-
move regions. Yue et al. in [28] employed a CNN to extract block-like characteristics from
an image, compute self-correlations between various blocks, locate matching points using
a point-wise feature extractor, and reconstruct a forgery mask using a deconvolutional
network. Yue et al. in [3] designed ManTra-Net that is s a fully convolutional network
that can handle any size image and a variety of forgery types, including copy-move,
enhancement, splicing, removal, and even unknown forgery forms. Liu et al. in [29]
proposed PSCC-Net, which analyses the image in a two-path methodology: a top-down
route that retrieves global and local features and a bottom-up route that senses if the image
is tampered and predicts its masks at four levels, each mask being constrained on the
preceding one.
In [30] Yang et al., proposed a technique based on two concatenated CNNs: the coarse
CNN and the refined CNN, which extracts the differences between the image itself and
splicing regions from patch descriptors of different scales. They enhanced their work in [1]
and proposed a patch-based coarse-to-refined network (C2RNet). The coarse network is
based on VVG16, and the refined network is based on VVG19. In [31] Xiuli et al., proposed
a ringed residual U-Net to detect the splicing type image forgery in the images. Younis et al.
in [32] utilized the reliability fusion map for the detection of the forgery. By utilizing the
Electronics 2022, 11, 403 4 of 17

CNNs, Younis et al. in [33] classify an image as the original one, or it contains copy-move
image forgery. In [34] Vladimir et al., train four models at the same time: a generative
annotation model GA, a generative retouching model GR, and two discriminators DA and
DR that checks the output of GA and GR. Mayer et al. in [35] introduced a system that
maps sets of image regions to a value that indicates if they include the same or different
forensic traces.
In [36] Minyoung et al., designed an algorithm that leverages the automatically
recorded image EXIF metadata for training a model to identify whether an image has
self-consistency or if its content may have been generated from a single image. In [37]
Rongyu et al., proposed a UNet that consists of a dense convolutional and deconvolutional
networks. The first is a down-sampling method for retrieving features, while the second is
an up-sampling approach for recovering feature map size. In [38] Lui et al., introduced the
CNN segmentation-based approach to find manipulated regions in digital photos. First, a
uniform CNN architecture is built to deal with various scales’ color input sliding windows.
Then, using sampling training regions, they meticulously build CNN training processes.
In [39], an unfixed encoder and a fixed encoder are used to build a Dual-encoder U-
Net (D-Unet). The unfixed encoder learns the image fingerprints that distinguish between
genuine and tampered regions on its own. In contrast, the fixed encoder offers direction
data to facilitate the network’s learning and detection. In [40] Francesco et al., tested the
efficiency of several image forgery detectors over image-to-image translation, including
both ideal settings and even in the existence of compression, which is commonly performed
when uploading to social media sites. Kadam et al. in [41] Proposed a method based on
multiple image splicing using MobileNet V1.Jaiswal et al. in [42] proposed a framework
in which images are fed into a CNN and then processed through several layers to extract
features, which are then utilized as a training vector for the detection model. For feature
extraction, they employed a pre-trained deep learning resnet-50.
Hao et al. in [43] proposed using an attention method to analyze and refine feature
maps for the detection task. The learned attention maps emphasize informative areas
to enhance binary classification (false face vs. genuine face) and illustrate the altered
regions. In [44], Nguyen et al., developed a CNN that employs a multi-task learning
strategy to detect altered images and videos while also locating the forged areas. The
information received from one work is shared with the second task, improving both
activities’ performance. To boost the network’s generability, a semi-supervised learning
strategy is adopted. An encoder and a Y-shaped decoder are included in the network. Li
et al. introduced a deepfake detection method in [45]. The DeepFake techniques can only
create fixed-size images of the face, which must be affinely warped to match the source’s
face arrangement. Due to the resolution disparity between the warped face area and the
surrounding context, this warping produces different artifacts. As a result, DeepFake
Videos can be identified using these artifacts. Komodakis et al. in [46] suggested a method
for learning image features by training CNNs to recognize the two-dimensional rotation
that is applied to the picture that it receives as input. The method proposed in [47] is
composed of three parts: single image super-resolution, semantic segmentation super-
resolution, and feature affinity module for semantic segmentation. In [48] Yu et al., used
dual attention upon pyramid visual feature maps to fully examine the visual-semantic
relationships and enhance the level of produced sentences. For more details about image
forgery and media, forensics readers may refer to [5–13].
The state-of-the-art techniques available for detecting the presence of tampering in
the images generally take a very long time to process the images. Most of them can detect
either image splicing forgery or copy-move type of forgery, not both. Another major issue
with them is that they detect the forgery with low accuracy. Hence, there is a need for a
better framework that is fast and more accurate. To address this, we presented a novel
image recompression-based system. Apart from achieving better image forgery detection
accuracy, our proposed framework has also achieved faster response time. This makes it
Electronics 2022, 11, 403 5 of 17

suitable for real-life applications, as it is more accurate and can be utilized even by slower
machines. The proposed framework is detailed in the next section.

3. Proposed Technique
CNNs, which are inspired by the human visual system, are designed to be non-linear
interconnected neurons. They have already demonstrated extraordinary potential in a
variety of computer vision applications, including image segmentation and object detection.
They may be beneficial for a variety of additional purposes, including image forensics.
With the various tools available today, image forgery is fairly simple to do, and because it
is extremely dangerous, detecting it is crucial. When a fragment of an image is moved from
one to another, a variety of artifacts occur due to the images’ disparate origins. While these
artifacts may be undetectable to the naked eye, CNNs may detect their presence in faked
images. Due to the fact that the source of the forged region and the background images
are distinct, when we recompress such images, the forged is enhanced differently due to
the compression difference. We use this concept in the proposed approach by training a
CNN-based model to determine if an image is genuine or a fake.
A region spliced onto another image will most likely have a statistically different
distribution of DCT coefficients than the original region. The authentic region is compressed
twice: first in the camera, and then again in the fake, resulting in periodic patterns in the
histogram [2]. The spliced section behaves similarly to a singly compressed region when
the secondary quantization table is used.
As previously stated, when an image is recompressed, if it contains a forgery, the
forged portion of the image compresses differently from the remainder of the image due to
the difference between the source of the original image and the source of the forged portion.
When the difference between the original image and its recompressed version is analyzed,
this considerably emphasizes the forgery component. As a result, we use it to train our
CNN-based model for detecting image forgery.
Algorithm 1 shows the working of the proposed technique, which has been explained
here. We take the forged image A (images shown in Figure 1b tamper images), and then
recompress it; let us call the recompressed image as Arecompressed (images shown in Figure 1c
are recompressed forged images). Now we take the difference of the original image and
the recompressed image, let us call it Adi f f (images shown in Figure 1e are the difference of
Figure 1b,c, respectively). Now due to the difference in the source of the forged part and
the original part of the image, the forged part gets highlighted in Adi f f (as we can observe
in Figure 1d,e, respectively). We train a CNN-based network to categorize an image as a
forged image or a genuine one using Adi f f as our input features (we label it as a featured
image). Figure 2 gives the pictorial view of the overall working of the proposed method.
To generate Arecompressed from A, we use JPEG compression. Image A undergoes JPEG
compression and produces Arecompressed as described in Figure 3. When there is a single
compression, then the histogram of the dequantized coefficients exhibits the pattern as
shown in Figure 4, this type of pattern is shown by the forged part of the image. Moreover,
when there is a sort of double compression then, as described in Figure 5, there is a gaping
between the dequantized coefficients as shown in Figure 6, this type of pattern is shown by
the genuine part of the image.
We constructed a very light CNN model with minimal parameters in our proposed
model (line number 5 to 13 of Algorithm 1). We constructed a model consisting of 3 convo-
lutional layers after which there is a dense fully connected layer, as described below:
• The first convolutional layer consists of 32 filters of size 3-by-3, stride size one, and
“relu” activation function.
• The second convolutional layer consists of 32 filters of size 3-by-3, stride size one, and
“relu” activation function.
• The third convolutional layer consists of 32 filters of size 7-by-7, stride size one, and
“relu” activation function, followed by max-pooling of size 2-by-2.
Electronics 2022, 11, 403 6 of 17

• Then we have the dense layer that has 256 neurons with “relu” activation function,
finally which is connected to two neurons (output neurons) with “sigmoid” activation.

Figure 1. Various sample images and their processed forms: (a) the original images (RGB colour
format); (b) the images with forgery (RGB colour format); (c) the recompressed forged images (RGB
colour format); (d) ground truth of forgery (Binary format); (e) difference of the tampered image with
its recompressed image (RGB colour format).

Table 1. Details of the CASIA.2.0 image forgery database.

Genuine Images Tampered Images Total Images

CASIA.2.0 7491 5123 12,614
Training (80%) 5993 4098 10,091
Testing (20%) 1498 1025 2523

The feature image (Adi f f ) is resized to 128 × 128 (Areshaped_di f f ) and then fed to the
network. The network learns the presence of any tampering present through the feature
images (images shown in Figure 1e). During training, the proposed model learns the
existence of the forgery in an image through the numerous artifacts left behind during
image forgery. The trained model can identify tampering with high accuracy, discussed in
the next section.
Electronics 2022, 11, 403 7 of 17

Figure 2. Flowchart of the proposed work.

Figure 3. JPEG compression on image pixels, first DCT is applied on the image pixel blocks fol-
lowed by the quantization. Then decompression of the compressed image is done with through
de-quantization followed by IDCT, to obtain the image in pixel format.
Electronics 2022, 11, 403 8 of 17

Figure 4. The histogram of DCT coefficients of the forged region in the image, which behaves as
singly compressed.

Figure 5. The recompression of the genuine region of the image.

Figure 6. The histogram of DCT coefficients of the genuine region of an image which gets
double compressed.
Electronics 2022, 11, 403 9 of 17

Algorithm 1: Working of the proposed technique for the image forgery detection.
1: /* Model Training (line 2 to 23) */
2: Input: Image ‘Ai ’ (i = 1 to n), with labels ‘Li ’ (Li = 1 if Ai is tampered image, else Li = 0).
3: Output: Trained Model: Image_Forgery_Predictor_Model()

4: /* Prediction Model Description */

5: Image_Forgery_Predictor_Model(image with size 128 × 128 × 3)
6: {
7: First convo. layer: 32 filters (size 3 × 3, strid size one, activation: “relu”)
8: Second convo. layer: 32 filters (size 3 × 3, strid size one, activation: “relu”)
9: Third convo. layer: 32 filters (size 3 × 3, strid size one, activation: “relu”)
10: Max-pooling of size 2 × 2
11: Dense layer of 256 neurons with “relu” activation function
12: Two neurons (output neurons) with “sigmoid” activation
13: }

14: for epochs = 1 to total_epochs do

15: training_error = 0
16: for i = 1 to n do
17: Arecompressed_i = JPEGCompression ( Ai , Q)
18: Adi f f _i = Ai − Arecompressed_i
19: Areshaped_di f f _i = reshape( Adi f f _i , (128, 128, 3))
20: training_error = (Li − Image_Forgery_Predictor_Model(Areshaped_di f f _i )) + training_error
21: end for
22: modify_model(training_error, Image_Forgery_Predictor_Model(), Adam_optimizer)
23: end for

24: /* Image forgery prediction (line 25 to 32) */

25: Input: Image ‘Input_Image’
26: Output: ‘Input_Image’ labelled as tampered or untampered
27: Input_Imagerecompressed = JPEGCompression ( Input_Image, Q)
28: Input_Imagedi f f = Input_Image − Input_Imagerecompressed
29: Input_Imagereshaped_di f f = reshape( Input_Imagedi f f , (128, 128, 3))
30: Predicted_label = Image_Forgery_Predictor_Model(Input_Imagereshaped_di f f )
31: /* If Predicted_label [0][0]>Predicted_label [0][1], then Input_Image is tampered
32: /* If Predicted_label [0][1]>Predicted_label [0][0], then Input_Image is untampered

4. Experimental Results and Discussion

This section describes the training and testing environment for the proposed approach.
Aside from that, we’ll examine and contrast its performance with that of other techniques.

4.1. Experimental Setup

We examined the proposed technique on a popular CASIA 2.0 image forgery database [22,49],
to evaluate how efficient it is. There are a total of 12,614 images (in BMP, JPG, and TIF
format), out of which 7491 are genuine images and 5123 tamper images. CASIA 2.0 includes
images from various categories, including animals, architecture, articles, characters, plants,
nature, scenes, textures, and indoor images. There are different-different sizes of the
images present in the database; the resolution of images varies from 800 × 600 pixels
to 384 × 256 pixels. The details about the CASIA 2.0 database are given in Table 1. A
processor (Intel(R) Core(TM) i5-2400 CPU @ 3.1 GHz) having 16 GB RAM has been used
for the experimentation.
Following terms are initially calculated for the evaluation:
• Total_Images: The total number of images that were tested.
• TP (true positive): Correctly identified tampered images.
• TN (true negative): Correctly identified genuine images.
• FN (false negative): Wrongly identified tampered images, the tampered images which
have been identified as genuine images.
• FP (false positive): Wrongly identified genuine images, the genuine images which
have been identified as tampered images.
Electronics 2022, 11, 403 10 of 17

We calculate the accuracy, precision, recall, and Fmeasure [1] for the evaluation and the
comparison of the proposed method with others. These are calculated as given below:
Now the accuracy is defined as given below:

TP + TN
Accuracy = × 100
TTotal_Images
TP
Recall =
TP + FN
TP
Precision =
TP + FP

2 × Recall × Precision
Fmeasure = × 100
Recall + Precision

4.2. Model Training and Testing

To evaluate the proposed technique, we randomly divided the CASIA 2.0 database
in the ratio of 80% and 20% (Table 1), we used 80% of the images (5993 authentic images,
4099 tampered images, total 10,092 images) for training the model. We used Adam opti-
mizer with an initial learning rate of 1 × 10−5 and a batch size of 64. The remaining 20%
images (1498 genuine images, 1024 tampered images, total 2522 images) are for testing the
proposed model and comparing it with the other existing frameworks. Figure 7 illustrates
the training and testing accuracy of the proposed model when trained on the CASIA 2.0
database with the settings mentioned above.

Figure 7. The training and the testing accuracy of the proposed method for the image forgery detection.

4.3. Comparison with Other Techniques

We compare the proposed technique to the other techniques in terms of accuracy and
time required for image forgery detection.

4.3.1. Accuracy Comparison

Table 2 shows the image forgery detection accuracy by the various techniques. These
techniques and the proposed method have been evaluated on the same set of images of
the CASIA 2.0 database. For the techniques mentioned in [2,3,27], when these techniques
Electronics 2022, 11, 403 11 of 17

process an input image, then, if the mask generated from them reports forgery, then we
categorize the input image as tampered else; it is considered as genuine image. It must be
noted that Buster-Net [27] is basically for copy-move type image forgeries, so we have used
copy-move forged images to report its results. Whereas CAT-Net [2] is basically for splicing
type of image forgeries, so we have used the images with splicing to report its results.
Mantra-Net [3] can handle both image splicing and copy-move type image forgeries. It can
be observed that we chose techniques that can handle either image splicing or copy-move,
but also techniques that can handle both image splicing and copy-move type of image
forgeries. All the techniques have CASIA 2.0 database as a common database for the
evaluation. We have used these techniques’ publicly available trained model for evaluation.
Apart from this, we retrained their models on the same CASIA 2.0 database images, on
which the proposed model has been trained. The results obtained by retraining these
models are also given in Table 2, along with their original models. After retraining, these
techniques’ accuracy has improved; however, the proposed technique still outperforms
them. CAT-Net [2], Buster-Net [27], and Mantra-Net [3] concentrate more on where the
forgery is present (localization, where the output is pixel-level forgery detection) in the
given image rather than focusing on that is the image is tampered or genuine (detection,
where the output is a binary classification). However, the proposed technique focuses on
whether the given image is tampered with or genuine.
The proposed technique achieved better forgery detection accuracy due to the fact
that instead of directly using the original pixel image, it uses the feature image, which is
the difference of the image with its recompressed image. This helps to detect image forgery
better because it can be observed that in the feature image the forged part gets highlighted.
Hence, it has resulted in achieving high accuracy. On the other hand, [2,3,27] show poor
accuracy in image forgery detection as these techniques try to find image forgery at the
pixel level, and due to this there are false positive pixels reported which reduces their
overall forgery detection accuracy at the image level.
As mentioned in the previous section, we used JPEG compression to recompress the
image; now, various quality factors are available while recompressing the image. So we
have evaluated the proposed model for different JPEG quality factors and reported them as
well. It is observed that the accuracy is better if the quality factor is kept at more than 90.
The proposed technique achieved better accuracy as it utilizes better input features rather
than directly using the original image as input features. To verify this we have trained our
model by directly using the original images (instead of the better processed features), and
its results are also reported in Table 2. It can be observed that in such a scenario the accuracy
of the model drops from 92.23% to 72.37%, this shows the effectiveness of the processed
input features (the difference of the original image with its recompressed version). Figure 8
show the comparison of the accuracy and the Fmeasure for the proposed method and the
other techniques.
Figure 9 shows a few visual examples demonstrating the proposed approach’s per-
formance. The codes and the test images used by the proposed technique are available
at codes (https://github.com/sadaf-ali/Image-forgery-detection-using-deeplearning-by-
recompressing-the-images, accessed on 23 January 2022).
We have acquired 92.23% accuracy on the test images when the recompressed images’
quality factor (JPEG compression) is 98. Compared to the other techniques, our technique
achieved much better performance in detecting the presence of the image forgery in the
image. It must also be noted that our method can handle both image splicing and copy-
move types of image forgeries. In contrast, many of the state-of-the-art techniques can
handle only one type of image forgery, and very few can handle both types of image forgery.
Electronics 2022, 11, 403 12 of 17

Table 2. Accuracy comparison of the proposed technique with other techniques on CASIA
2.0 database.

Techniques Accuracy Precision Recall Fmeasure

Mantra-Net [3] 40.62 0.40 1 57.14
Mantra-Net * [3] 56.14 0.48 0.79 59.33
Buster-Net [27] 30.77 0.30 1 46.75
Buster-Net * [27] 49.06 0.35 0.77 47.74
CAT-Net [2] 19.33 0.19 1 31.93
CAT-Net * [2] 87.29 0.62 0.87 72.76
Proposed technique (with normal image) 72.37 0.62 0.79 70.11
Proposed technique (85% quality) 87.71 0.80 0.92 85.99
Proposed technique (90% quality) 91.31 0.83 0.98 90.19
Proposed technique (91% quality) 91.35 0.83 0.97 90.14
Proposed technique (92% quality) 91.39 0.84 0.96 90.15
Proposed technique (93% quality) 90.84 0.83 0.96 89.52
Proposed technique (94% quality) 91.71 0.84 0.97 90.50
Proposed technique (95% quality) 91.83 0.84 0.97 90.68
Proposed technique (96% quality) 92.15 0.85 0.97 90.99
Proposed technique (97% quality) 91.63 0.85 0.96 90.32
Proposed technique (98% quality) 92.23 0.85 0.97 91.08
Proposed technique (99% quality) 91.83 0.86 0.94 90.42
Proposed technique (100% quality) 91.79 0.86 0.94 90.29
Note: “*” means that the model has been retrained on CASIA 2.0 database.

(a)
Figure 8. Cont.
Electronics 2022, 11, 403 13 of 17

(b)
Figure 8. Comparison of the performance of Mantra-Net [3], Buster-Net [27], CAT-Net [2], and the
proposed technique in terms of: (a) accuracy, (b) Fmeasure .

4.3.2. Processing Time Comparison

Table 3 shows the average time taken by the various techniques to process an image
and to predict whether it is a genuine image or a tampered one. In terms of predicting the
presence of forgery in an image, it can be noted that the proposed technique is fast and
more efficient than the other state-of-the-art techniques. Figure 10 pictorially shows the
comparison of the average time taken by the proposed method and the other techniques
for the forgery detection in an image.
This is because we provide an efficient feature image to our model, and the proposed
CNN-based model is relatively light compared to other techniques. As a result, it can
provide predictions in a much shorter amount of time. This makes our model will be
advantageous in real-world scenarios. Table 4 shows the comparison of the proposed
technique with the other techniques.

(a) (b) (c) (d)

Figure 9. A few visual examples demonstrating the proposed approach’s performance: (a) true
negative cases, (b) false negative cases, (c) true positive cases, (d) false positive cases.
Electronics 2022, 11, 403 14 of 17

Figure 10. Average time taken by Mantra-Net [3], Buster-Net [27], CAT-Net [2], and the proposed
technique to process the image for forgery detection.

Table 3. Comparison of the time to process an image by the proposed technique with other techniques
on CASIA 2.0 database (displayed values are in milliseconds).

Techniques Time Taken

Mantra-Net [3] 10,927
Buster-Net [27] 1160
CAT-Net [2] 2506
Proposed technique 34

Table 4. Comparison of the proposed technique with the other techniques [2,3,27].

Sr. No. Proposed Technique Other Techniques

It focuses on whether the given image is They focus more on where the forgery is
1 tampered with or genuine (image level present in the given image (pixel level
forgery detection). forgery detection).
It uses feature image (difference of
original image with its recompressed They directly use the image in the
2 image), in the feature image the forged original pixel format, which makes image
part gets highlighted, which helps to forgery detection difficult.
detect image forgery better.
It is utilizes enhanced features, due to Most of the techniques do not utilize the
which it is able to handle both image enhanced features, due to which they are
3
splicing and copy-move types usually able to handle only one type of
image forgeries. image forgery.
They do pixel level forgery detection.
Due to this, they suffer from false
It attains a high accuracy in classifying
4 positive pixels. Hence, there is
the images as tampered or not.
degradation in classifying the images as
tampered or not.
These techniques are much slower, hence,
It is much faster, hence, it is suitable for
5 they are not suitable for the slower
slow machines as well.
machines.

Hence, from the experimental results, the following observations can be made:
Electronics 2022, 11, 403 15 of 17

• Unlike other techniques, the proposed technique works well for both image splicing
and copy-move types of image forgeries.
• It is highly efficient for image forgery detection and has exhibited significantly better
performance than the other techniques.
• The difference in the compression of the forged part and the genuine part of the image
is a good feature that can be learned by our CNN based model efficiently, which makes
the proposed technique more robust in comparison to the other techniques.
• The proposed model is much faster than the other techniques, making it ideal and
suitable for real-world usage, as it can be implemented even on slower machines.

5. Conclusions and Future Work

The increased availability of cameras has made photography popular in recent years.
Images play a crucial role in our lives and have evolved into an essential means of conveying
information since the general public quickly understands them. There are various tools
accessible to edit images; these tools are primarily intended to enhance images; however,
these technologies are frequently exploited to forge the images to spread misinformation.
As a result, image forgery has become a significant problem and a matter of concern. In
this paper, we provide a unique image forgery detection system based on neural networks
and deep learning, emphasizing the CNN architecture approach. To achieve satisfactory
results, the suggested method uses a CNN architecture that incorporates variations in image
compression. We use the difference between the original and recompressed images to train
the model. The proposed technique can efficiently detect image splicing and copy-move
types of image forgeries. The experiments results are highly encouraging, and they show
that the overall validation accuracy is 92.23%, with a defined iteration limit.
We plan to extend our technique for image forgery localization in the future. We will
also combine the suggested technique with other known image localization techniques to
improve their performance in terms of accuracy and reduce their time complexity. We will
enhance the proposed technique to handle spoofing [50] as well. The present technique
requires image resolution to be a minimum of 128 × 128, so we will enhance the proposed
technique to work well for tiny images. We will also be developing a challenging extensive
image forgery database to train deep learning networks for image forgery detection.

Author Contributions: S.S.A. and I.I.G. designed and performed the experiments, conceptualization
and methodology, and analyzed the data. S.S.A. and I.I.G. wrote the manuscript in consultation
with N.-S.V., N.S., S.D.A. and N.W. All authors have read and agreed to the published version of the
manuscript.
Funding: This research has been funded by CY Cergy Paris Université (CY initiative via contract
number 2019-067-C01-A0) and Khalifa University.
Acknowledgments: We are thankful to CNRS for providing us the GPU Cluster Jean Zay (http:
//www.idris.fr/jean-zay/ accessed on 10 January 2022) to train our network.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Xiao, B.; Wei, Y.; Bi, X.; Li, W.; Ma, J. Image splicing forgery detection combining coarse to refined convolutional neural network
and adaptive clustering. Inf. Sci. 2020, 511, 172–191. [CrossRef]
2. Kwon, M.J.; Yu, I.J.; Nam, S.H.; Lee, H.K. CAT-Net: Compression Artifact Tracing Network for Detection and Localization of
Image Splicing. In Proceedings of the 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), Waikoloa, HI,
USA, 5–9 January 2021; pp. 375–384.
3. Wu, Y.; Abd Almageed, W.; Natarajan, P. ManTra-Net: Manipulation Tracing Network for Detection and Localization of Image
Forgeries With Anomalous Features. In Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern
Recognition (CVPR), Long Beach, CA, USA, 15–20 June 2019; pp. 9535–9544.
4. Ali, S.S.; Baghel, V.S.; Ganapathi, I.I.; Prakash, S. Robust biometric authentication system with a secure user template. Image Vis.
Comput. 2020, 104, 104004. [CrossRef]
5. Castillo Camacho, I.; Wang, K. A Comprehensive Review of Deep-Learning-Based Methods for Image Forensics. J. Imaging 2021,
7, 69. [CrossRef] [PubMed]
Electronics 2022, 11, 403 16 of 17

6. Zheng, L.; Zhang, Y.; Thing, V.L. A survey on image tampering and its detection in real-world photos. J. Vis. Commun. Image
Represent. 2019, 58, 380–399. [CrossRef]
7. Jing, L.; Tian, Y. Self-supervised Visual Feature Learning with Deep Neural Networks: A Survey. IEEE Trans. Pattern Anal. Mach.
Intell. 2020, 43, 1. [CrossRef]
8. Meena, K.B.; Tyagi, V. Image Forgery Detection: Survey and Future Directions. In Data, Engineering and Applications: Volume 2;
Shukla, R.K., Agrawal, J., Sharma, S., Singh Tomer, G., Eds.; Springer: Singapore, 2019; pp. 163–194.
9. Mirsky, Y.; Lee, W. The Creation and Detection of Deepfakes: A Survey. ACM Comput. Surv. 2021, 54, 1–41. [CrossRef]
10. Rony, J.; Belharbi, S.; Dolz, J.; Ayed, I.B.; McCaffrey, L.; Granger, E. Deep weakly-supervised learning methods for classification
and localization in histology images: A survey. arXiv 2019, arXiv:abs/1909.03354.
11. Lu, Z.; Chen, D.; Xue, D. Survey of weakly supervised semantic segmentation methods. In Proceedings of the 2018 Chinese
Control Furthermore, Decision Conference (CCDC), Shenyang, China, 9–11 June 2018; pp. 1176–1180.
12. Zhang, M.; Zhou, Y.; Zhao, J.; Man, Y.; Liu, B.; Yao, R. A survey of semi- and weakly supervised semantic segmentation of images.
Artif. Intell. Rev. 2019, 53, 4259–4288. [CrossRef]
13. Verdoliva, L. Media Forensics and DeepFakes: An Overview. IEEE J. Sel. Top. Signal Process. 2020, 14, 910–932. [CrossRef]
14. Luo, W.; Huang, J.; Qiu, G. JPEG Error Analysis and Its Applications to Digital Image Forensics. IEEE Trans. Inf. Forensics Secur.
2010, 5, 480–491. [CrossRef]
15. Matern, F.; Riess, C.; Stamminger, M. Gradient-Based Illumination Description for Image Forgery Detection. IEEE Trans. Inf.
Forensics Secur. 2020, 15, 1303–1317. [CrossRef]
16. Christlein, V.; Riess, C.; Jordan, J.; Riess, C.; Angelopoulou, E. An Evaluation of Popular Copy-Move Forgery Detection
Approaches. IEEE Trans. Inf. Forensics Secur. 2012, 7, 1841–1854. [CrossRef]
17. Habibi, M.; Hassanpour, H. Splicing Image Forgery Detection and Localization Based on Color Edge Inconsistency using
Statistical Dispersion Measures. Int. J. Eng. 2021, 34, 443–451.
18. Dua, S.; Singh, J.; Parthasarathy, H. Image forgery detection based on statistical features of block DCT coefficients. Procedia
Comput. Sci. 2020, 171, 369–378. [CrossRef]
19. Ehret, T. Robust copy-move forgery detection by false alarms control. arXiv 2019, arXiv:1906.00649.
20. de Souza, G.B.; da Silva Santos, D.F.; Pires, R.G.; Marana, A.N.; Papa, J.P. Deep Features Extraction for Robust Fingerprint
Spoofing Attack Detection. J. Artif. Intell. Soft Comput. Res. 2019, 9, 41–49. [CrossRef]
21. Balsa, J. Comparison of Image Compressions: Analog Transformations. Proceedings 2020, 54, 37. [CrossRef]
22. Pham, N.T.; Lee, J.W.; Kwon, G.R.; Park, C.S. Hybrid Image-Retrieval Method for Image-Splicing Validation. Symmetry 2019,
11, 83. [CrossRef]
23. Bunk, J.; Bappy, J.H.; Mohammed, T.M.; Nataraj, L.; Flenner, A.; Manjunath, B.; Chandrasekaran, S.; Roy-Chowdhury, A.K.;
Peterson, L. Detection and Localization of Image Forgeries Using Resampling Features and Deep Learning. In Proceedings of the
2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Honolulu, HI, USA, 21–26 July 2017;
pp. 1881–1889.
24. Bondi, L.; Lameri, S.; Güera, D.; Bestagini, P.; Delp, E.J.; Tubaro, S. Tampering Detection and Localization Through Clustering
of Camera-Based CNN Features. In Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition
Workshops (CVPRW), Honolulu, HI, USA, 21–26 July 2017; pp. 1855–1864.
25. Yousfi, Y.; Fridrich, J. An Intriguing Struggle of CNNs in JPEG Steganalysis and the OneHot Solution. IEEE Signal Process. Lett.
2020, 27, 830–834. [CrossRef]
26. Islam, A.; Long, C.; Basharat, A.; Hoogs, A. DOA-GAN: Dual-Order Attentive Generative Adversarial Network for Image
Copy-Move Forgery Detection and Localization. In Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), Seattle, WA, USA, 13–19 June 2020; pp. 4675–4684.
27. Wu, Y.; Abd-Almageed, W.; Natarajan, P. BusterNet: Detecting Copy-Move Image Forgery with Source/Target Localization. In
Proceedings of the European Conference on Computer Vision (ECCV), Glasgow, UK, 23–28 August 2020.
28. Wu, Y.; Abd-Almageed, W.; Natarajan, P. Image Copy-Move Forgery Detection via an End-to-End Deep Neural Network. In
Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Munich, Germany, 8–14 September
2018; pp. 1907–1915.
29. Liu, X.; Liu, Y.; Chen, J.; Liu, X. PSCC-Net: Progressive Spatio-Channel Correlation Network for Image Manipulation Detection
and Localization. arXiv 2021, arXiv:2103.10596
30. Wei, Y.; Bi, X.; Xiao, B. C2R Net: The Coarse to Refined Network for Image Forgery Detection. In Proceedings of the 2018 17th
IEEE International Conference On Trust, Security And Privacy in Computing Furthermore, Communication, New York, NY, USA,
1–3 August 2018; pp. 1656–1659.
31. Bi, X.; Wei, Y.; Xiao, B.; Li, W. RRU-Net: The Ringed Residual U-Net for Image Splicing Forgery Detection. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, Long Beach, CA, USA, 15–20 June 2019.
32. Younisand, Y.; Xu, M.; Qiao, T.; Wu, Y.; Zheng, N. Image Forgery Detection and Localization via a Reliability Fusion Map. Sensors
2020, 20, 6668.
33. Abdalla, Y.; Iqbal, M.T.; Shehata, M. Convolutional Neural Network for Copy-Move Forgery Detection. Symmetry 2019, 11, 1280.
[CrossRef]
Electronics 2022, 11, 403 17 of 17

34. Kniaz, V.V.; Knyaz, V.; Remondino, F. The Point Where Reality Meets Fantasy: Mixed Adversarial Generators for Image Splice
Detection. In Advances in Neural Information Processing Systems; Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox,
E., Garnett, R., Eds.; Curran Associates, Inc.: New York, NY, USA, 2019; Volume 32.
35. Mayer, O.; Stamm, M.C. Forensic Similarity for Digital Images. IEEE Trans. Inf. Forensics Secur. 2020, 15, 1331–1346. [CrossRef]
36. Huh, M.; Liu, A.; Owens, A.; Efros, A.A. Fighting Fake News: Image Splice Detection via Learned Self-Consistency. In
Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany, 8–14 September 2018.
37. Zhang, R.; Ni, J. A Dense U-Net with Cross-Layer Intersection for Detection and Localization of Image Forgery. In Proceedings of
the ICASSP 2020—2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, Spain,
4–8 May 2020; pp. 2982–2986.
38. Liu, Y.; Guan, Q.; Zhao, X.; Cao, Y. Image Forgery Localization Based on Multi-Scale Convolutional Neural Networks. In
Proceedings of the 6th ACM Workshop on Information Hiding and Multimedia Security, Innsbruck, Austria, 20–22 June 2018;
Association for Computing Machinery: New York, NY, USA, 2018; pp. 85–90.
39. Bi, X.; Liu, Y.; Xiao, B.; Li, W.; Pun, C.M.; Wang, G.; Gao, X. D-Unet: A Dual-encoder U-Net for Image Splicing Forgery Detection
and Localization. arXiv 2020, arXiv:2012.01821.
40. Marra, F.; Gragnaniello, D.; Cozzolino, D.; Verdoliva, L. Detection of GAN-Generated Fake Images over Social Networks.
In Proceedings of the 2018 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR), Miami, FL, USA,
10–12 April 2018; pp. 384–389.
41. Kadam, K.; Ahirrao, D.S.; Kotecha, D.K.; Sahu, S. Detection and Localization of Multiple Image Splicing Using MobileNet V1,
2021. arXiv 2020, arXiv:2108.09674.
42. Jaiswal, A.A.K.; Srivastava, R. Image Splicing Detection using Deep Residual Network. In Proceedings of the 2nd International
Conference on Advanced Computing and Software Engineering (ICACSE), San Francisco, CA, USA, 13–15 October 2019.
43. Dang, H.; Liu, F.; Stehouwer, J.; Liu, X.; Jain, A.K. On the Detection of Digital Face Manipulation. In Proceedings of the 2020
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 13–19 June 2020; pp. 5780–5789.
44. Nguyen, H.H.; Fang, F.; Yamagishi, J.; Echizen, I. Multi-task Learning for Detecting and Segmenting Manipulated Facial Images
and Videos. In Proceedings of the 2019 IEEE 10th International Conference on Biometrics Theory, Applications and Systems
(BTAS), Tampa, FL, USA, 23–26 September 2019; pp. 1–8.
45. Li, Y.; Lyu, S. Exposing DeepFake Videos By Detecting Face Warping Artifacts. In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR) Workshops, Nashville, TN, USA, 19–25 June 2019.
46. Komodakis, N.; Gidaris, S. Unsupervised representation learning by predicting image rotations. In Proceedings of the
International Conference on Learning Representations (ICLR), Vancouver, BC, Canada, 30 April–3 May 2018.
47. Wang, L.; Li, D.; Zhu, Y.; Tian, L.; Shan, Y. Dual Super-Resolution Learning for Semantic Segmentation. In Proceedings of the 2020
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 13–19 June 2020; pp. 3773–3782.
48. Yu, L.; Zhang, J.; Wu, Q. Dual Attention on Pyramid Feature Maps for Image Captioning. arXiv 2022, arXiv:2011.01385.
49. Dong, J.; Wang, W.; Tan, T. CASIA Image Tampering Detection Evaluation Database. In Proceedings of the 2013 IEEE China
Summit and International Conference on Signal and Information Processing, Beijing, China, 6–10 July 2013; pp. 422–426.
50. Ali, S.S.; Iyappan, G.I.; Prakash, S. Fingerprint Shell construction with impregnable features. J. Intell. Fuzzy Syst. 2019,
36, 4091–4104. [CrossRef]