Received 24 June 2023, accepted 11 August 2023, date of publication 22 August 2023, date of current version 30 August 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3307357

Enhancing Digital Image Forgery Detection Using

Transfer Learning
ASHGAN H. KHALIL 1 , ATEF Z. GHALWASH1 , HALA ABDEL-GALIL ELSAYED1 ,
GOUDA I. SALAMA2 , AND HAITHAM A. GHALWASH3
1 Computer Science Department, Faculty of Computers and Artificial Intelligence, Helwan University, Cairo 4271184, Egypt
2 Department of Computer Engineering, MTC, Cairo 4393010, Egypt
3 Ethical Hacking and Cyber Security, Coventry University–Egypt Branch, New Cairo, Egypt

Corresponding author: Ashgan H. Khalil ([email protected])

ABSTRACT Nowadays, digital images are a main source of shared information in social media. Meanwhile,
malicious software can forge such images for fake information. So, it’s crucial to identify these forgeries.
This problem was tackled in the literature by various digital image forgery detection techniques. But most
of these techniques are tied to detecting only one type of forgery, such as image splicing or copy-move
that is not applied in real life. This paper proposes an approach, to enhance digital image forgery detection
using deep learning techniques via transfer learning to uncover two types of image forgery at the same time,
The proposed technique relies on discovering the compressed quality of the forged area, which normally
differs from the compressed quality of the rest of the image. A deep learning-based model is proposed to
detect forgery in digital images, by calculating the difference between the original image and its compressed
version, to produce a featured image as an input to the pre-trained model to train the model after removing its
classifier and adding a new fine-tuned classifier. A comparison between eight different pre-trained models
adapted for binary classification is done. The experimental results show that applying the technique using
the adapted eight different pre-trained models outperforms the state-of-the-art methods after comparing it
with the resulting evaluation metrics, charts, and graphs. Moreover, the results show that using the technique
with the pre-trained model MobileNetV2 has the highest detection accuracy rate (around 95%) with fewer
training parameters, leading to faster training time.

INDEX TERMS Deep neural network (DNN), image compression, image forgery detection (IFD), pretrained
model, transfer learning.

I. INTRODUCTION image security especially when the original content is not

The tampering of a digital image is called digital image available [3].
forgery, these forged images cannot be detected by the naked Digital image forgery means adding unusual patterns to
eye. Such images are the primary sources of spreading the original images that create a heterogeneous variation
fake news and misleading information in the context of in image properties and an unusual distribution of image
society with the aid of diverse social media platforms features [3]. Figure 1 shows the classification of digital image
like Facebook, Twitter, etc. [1]. The editing software tools forgery.
that can make these forgeries are available for free with Active approaches require essential information about the
some advanced features that are used for image tampering image for the verification process. The inserted information
such as GNU, GIMP, and Adobe Photoshop [2]. Such within the picture is employed to observe the modification in
forgeries can be detected using digital image forgery that picture. The active approach consists of two types: digital
algorithms and techniques, these algorithms are used in signatures which insert some additional data obtained from
an image by the end of the acquisition process, and digital
The associate editor coordinating the review of this manuscript and watermarking which is inserted into images either during the
approving it for publication was Rajeeb Dey . acquisition phase or during the processing phase.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 91583
 A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

of tampering are Copy Move, Image Splicing, and Image

Retouching [4].
Digital Image Forgery Detection is a binary classification
task, to classify the image as either forged or authentic.
Recently, deep learning has become a promising tool for
enhancing digital image forgery detection. In any Deep
learning model, feature extraction is an important phase
that affects the performance of the algorithm [6], where
the database size is considered a significant factor. Transfer
learning presents a viable alternative solution when dealing
with limited sample size problems that supports taking
FIGURE 1. Digital image forgery classification. the knowledge acquired from a previously trained model
including features, weights, and other relevant information
The passive image forgery detection methods benefit that was trained on a large dataset such as the ImageNet
from the features retained by the image allocation processes database, that contains 1.2 million images grouped into
achieved in different stages of digital image acquisition 1000 classes to solve the problem of small size dataset in
and storage. Passive methodologies do not require past the new target domain [7]. By utilizing a pre-trained model,
information about the image. These approaches exploit that significant amounts of time spent on training can be saved,
the tampering actions modify the contents of information of and the model can be adapted to work with smaller datasets
the image that can facilitate tampering detection [4]. through retraining [8].
Copy move forgery involves duplicating a section or The motivation behind image forgery detection is to check
object within an image and pasting it again in a different the authenticity of digital images, especially when images are
location within the same image to replicate (or move) used as evidence in court and forensics, news, or historical
a specific scene in the image. Copy-move forgery is data, or in the military, and medical diagnosis systems,
the most common technique used to manipulate images, it prevents the distribution of misinformation and fake news,
it is also the most challenging type of forgery to detect particularly in social media and online platforms, these forged
due to the complexity of copying and replicating an images can be used to destroy someone’s reputation or
object or section of the image with identical properties mislead public opinion, or for distorting the truth in news
and feature distributions and pasting it within the same reports, they can also be used to exaggerate the capabilities
image [3]. some post-processing techniques can be added of the countries army.
after CMF processes such as rotation, scaling, JPEG com- Image forgery detection has several challenges due to the
pression, etc. which makes the detection further difficult and nature of image manipulation techniques. These challenges
complex [2]. can be concluded as:
Splicing forgery can be generated by adding or blending • Computational Complexity and the limitation of the
two images or set of images to produce an unprece- CPU and memory is the main challenge, which takes
dented image [3]. The source images used to generate a a large training time and most of the time runs out of
spliced image may include dissimilar color temperatures, memory even with high memory specifications.
illumination conditions, and noise levels based on various • Detecting more than one type of image forgery at the
factors. Average filtering or some other related image same time affects the accuracy rate, so there is a need to
processing operation can be applied as postprocessing like improve its accuracy rate.
resizing, cropping, rotating, and retouching each of the • There is a need to solve the problem of the accuracy-
source images to match the visual attributes, shape, and speed trade-off.
size of the target image so that the forged image can look • Most image forgery detection techniques that have high
realistic [5]. detection accuracy are very complex, there is a need
Retouching forgery involves modifying an image to hide for a simpler technique with high detection accuracy
or highlight particular features such as brightness, color, rate.
contrast, or other visual attributes and altering background • Most image forgery detection techniques suffer from
coloring. It includes the visual quality enhancement of detecting images that lie under post-processing opera-
the image. Resampling Forgery is the act of altering the tions like image rotation, scaling, blurring, brightness
dimensionality of a particular object or section within an adjustment, and adding noise.
image to present a distorted or misleading view. Morphing
forgery involves merging two scenes from different images This paper presents the following contribution:
to create an entirely new scene, this can be done through • Detecting two types of passive image forgeries like
the use of graphic software to create a completely artificial image splicing and copy-move at the same time to be
image with no basis in reality [3]. The three major types suitable for real-life scenarios.

91584 VOLUME 11, 2023

A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

• Achieving a high accuracy rate compared to the state- the original image [9]. Deep learning technique has proven
of-the-art results found in the literature. Moreover, to be effective in resolving many activities or issues that
using a pre-trained model and taking the power of machine learning algorithms were previously unable to
transfer learning, with a small number of parame- address [8].
ters, the developed lightweight model is well-suited When considering splicing detection, a scheme was
for environments with memory and CPU limitations. proposed in [10] based on the local feature descriptor
This is an added value in favor of the proposed which is learned by a DNN. An improved initialization
architecture. based on the (SRM) was proposed and developed a splicing
• Evaluating the performance of eight different pre-trained localization scheme based on the proposed CNN model
models such as VGG16, VGG19, ResNet50, Reset101, and fully connected conditional random field (CRF) with
ResNet152, MobileNetV2, Xception and DenseNet are SVM which is robust against JPEG compression. In [11], a
considered. (CNN) model was developed using a relatively small number
• A comparative analysis of the eight forementioned pre- of parameters that can be used as an on-time detection
trained models and state of art is presented. model.
• Using the CASIAV2 dataset which is one of the best For splicing and copy-move separately, an end-to-end
benchmark datasets that is considered as the main fully CNN that combines multi-resolution hybrid features,
challenge itself, it contains two main types of image from RGB and noise streams was introduced in [12], where
forgery (splicing and copy-move) with different sizes a tamper-guided dual self-attention (TDSA) module was
and contains many types of image formats (TIFF, designed to capture the difference between tampered and
JPEG, and BMP ) and also the cropped parts in the non-tampered areas and segments them from the image.
forged images underwent some processing including A proposed hybrid features and semantic reinforcement
distortion, rotation, and scaling, to create an image network (HFSRNet) for IFD at the pixel level was proposed
that seems to be real, involving blurring the spliced in [13], where the network employs an encoding and
region’s edge, which makes the detection process decoding approach and utilizes Long-Short Term Memory
challenging. (LSTM) technology.
The paper is organized as follows: A literature review For copy move, [14] introduced a copy-move forgery
is covered in section II. Section III discusses the proposed detection and localization model based on super boundary-
approach and presents the proposed architecture in detail. to-pixel direction (super-BPD) segmentation and deep CNN
Section IV outlines the experimental results and discussion, (DCNN). Starting with employing the segmentation tech-
along with the experimental setup and dataset structure. nique that is used to enhance the connection among identical
Section V has the conclusion and future work. image blocks, thereby improving the accuracy of forgery
detection, the DCNN is used to extract image features, ending
II. LITERATURE REVIEW by using image BPD information to optimize the edges
In image forgery detection field, various approaches were of the rough detected image and obtain the final detected
proposed. Traditional techniques mostly extract a set of hand- image. [15] developed a deep learning CNN model which
crafted based features, followed by a classifying technique used multi-scale input and multiple stages of convolutional
like feature matching to differentiate between the authentic layers, with two different parts, encoder, and decoder. In [16],
and forged images. In the machine learning approach, a set of a simple and lightweight convolutional neural network
classifiers can be used in the classifying process like Support (CNN) has been proposed for the automatic detection of
Vector Machine and Naïve Bayes classifier. While more copy-move forgery detection, which has a high detection
recent techniques employ convolutional neural networks accuracy rate.
(CNNs) and deep neural networks (DNN) methods, others For copy-move and splicing together, [9] used a new image
employ the network with the help of pre-trained models and segmentation model U-Net by adding L2 regularization.
the power of transfer learning. CNN and deep learning-based Reference [17] introduced a system for IFD using double
techniques will be discussed moving over the use of different image compression, in which the difference between an
pre-trained models. original image and recompressed one was used in training the
model, the method is capable of detecting both image splicing
A. DEEP NEURAL NETWORK-BASED IMAGE FORGERY and copy-move together.
DETECTION TECHNIQUES
DNNs can autonomously learn an extensive number of
features. Over the past few years, a variety of image B. PRETRAINED NETWORK-BASED IMAGE FORGERY
forgery detection methods have been proposed, for detecting DETECTION TECHNIQUES
image forgery, where many of which relied on deep Different IFD techniques based on transfer learning will be
learning [5]. By constructing an appropriate neural network, discussed in this section. For splicing, [18] presented multiple
deep learning networks can identify complex hidden patterns image-splicing forgeries using Mask R-CNN and MobileNet-
in data and effectively distinguish the forged parts from V1 backbone. A novel approach utilizing ResNet50v2 was

VOLUME 11, 2023 91585

 A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

introduced in [19], that considered image batches as an detection with a high detection accuracy rate. Image splicing
input and used YOLO CNN weights with ResNet50v2 seems to be the easiest type to detect, Meanwhile, a lot
architecture. of efforts were made to detect copy-move which seems
For splicing and copy-move separately, [20] proposed a to be difficult to detect. It is also worth to be noticed
multi-task learning network called FBI-Net based on (DCT). that there were a few research studies done for detecting
The network employs a fully convolutional encoder-decoder both splicing and copy-move at the same time where less
architecture, and the Dilated Frequency Self-Attention Mod- detection accuracy rate was recorded compared to the other
ule (DFSAM) in the bridge layer adjusts fused features. techniques.
Reference [21] introduced a lightweight model using mask
R-CNN with MobileNet to detect copy-move and image- TABLE 1. Summary of deep learning-based image forgery detection
splicing forgeries. techniques.
For copy move, [22] used SmallerVGGNet and MobileNet-
V2, time- and memory-saving deep learning models. In [23]
an Optimal Deep Transfer Learning based Copy Move
Forgery Detection (ODTLCMFD) technique was presented
that derived a DL model for the classification of target
images and then localized the copy moved regions. They
used the MobileNet model with a political optimizer (PO)
for feature extraction and the least square support vector
machine (LS-SVM) model with an enhanced bird swarm
algorithm (EBSA) for classification. They utilized the
EBSA algorithm to modify the parameters in the Multiclass
Support Vector Machine (MSVM) technique to enhance
the classification performance. Reference [24] provided an
automated deep learning-based fusion model for detecting
and localizing copy-move forgeries (DLFM-CMDFC), that
combined models of generative adversarial networks (GANs)
and densely connected networks (DenseNet). The two
outputs were merged in the DLFM-CMDFC technique to
create a layer for encoding the input vectors with the first
layer of an extreme learning machine (ELM) classifier. Recently, image forgery detection techniques relied on
The ELM model’s weight and bias values were modified deep learning only that needs the availability of large
using the artificial fish swarm algorithm (AFSA). The data sets for training which is not available in the real.
networks’ outputs were supplied into the merger unit This problem can be solved with the help of pre-trained
as input. models and the power of transfer learning. In addition,
For splicing and copy-move together, a multimodal system previous researches seems to be very complex to reach a
was proposed in [25], which covers classification and high detection accuracy rate. Moreover, different evaluation
localization, forgery detection through a deep neural network matrices were considered in the studies which add difficulties
followed by part-based image retrieval classification. The when comparing such techniques. Also, not all studies
localization of manipulated regions was accomplished using are concerned with image pre-processing such as rotation,
a deep neural network. InceptionV3 was employed for feature scaling, and blurring, which adds difficulties to the detection
extraction. The Nearest Neighbor Algorithm was used to process. The above-mentioned considerations trigger the
retrieve Potential donors and nearly duplicates. In [26] a motivation to consider the transfer learning technique to build
novel approach to detect copy move and splicing image the proposed model.
forgery using a Convolutional Neural Network (CNN), with In [17], the authors focused on the fact that CNNs can
three different models was presented, namely, ELA (Error be utilized to detect image forgery, which is difficult for the
Level Analysis), VGG16, and VGG19. The proposed method human eye to detect, due to artifacts left by the forgeries.
applied the pre-processing technique to obtain the images Also, the source of the forged region and the background
at a particular compression rate. These images were then images are different. This makes it easy to allocate the
utilized to train the model, where the images were classified forged region by compression differences between both. This
as authentic or forged. difference was utilized to train the CNN-based model to
TABLE 1 summarizes the image forgery detection tech- identify image forgery. The experimental results of [17]
niques based on deep learning, and TABLE 2 summarizes showed that the CNN model achieved a 92.3% detection
the image forgery detection techniques based on transfer accuracy rate which still needs improvement. In addition,
learning. Both tables show that previous research reveals the model has a large number of parameters that need to be
that some efforts were made for image splicing forgery reduced to save CPU and memory consumption. Also, the

91586 VOLUME 11, 2023

A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

TABLE 2. Summary of transfer learning-based image forgery detection the accuracy rate that will get better than that which was
techniques.
recorded when using CNN in [17]. This will be elaborated
and discussed in the following section.
In a forged image, if the image is compressed, the forged
section of the image will be compressed differently than the
rest of the image. This is because the source of the original
image differs from the source of the forged section. When
analyzing the difference between the original image and its
compressed version, the forgery component becomes more
distinguished. Therefore, this aspect can be utilized to train a
DNN-based model for detecting image forgery.

FIGURE 2. Set of images created in the proposed work.

The set of images created in the proposed work can be

shown in Figure 2. The first image, (a) represents the original
image without forgery, (b) represents the forged image that
is denoted as F, (c) represents the compressed version of (b)
that is denoted as Fcomp, (d) represents the mathematical
difference between F and Fcomp denoted as Fdiff.

performance parameters e.g., F1 score, recall, precision, TPR,

and TNR need to be improved by increasing their values.
In addition, the model has large FPR and FNR which need to
be decreased. All these evaluation matrices will be discussed
in the following section.

III. PROPOSED APPROACH

The proposed approach considers the fact shown in [17], that
copying a part of an image from one to another may impose
some changes in the image properties due to the different
sources of the images. Although these changes may not be
detectable to the human eye, they can be detected by CNNs
in manipulated images.
The proposed model aims to avoid all of the forementioned
drawbacks, by adapting the idea of calculating the difference
in compression qualities to produce the featured image as an
input to a deep neural network with the assistance of a pre-
trained model to benefit from the power of transfer learning.
As a result, the evaluation matrix will be improved including FIGURE 3. Flowchart of the proposed system (System Architecture).

VOLUME 11, 2023 91587

 A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

In the proposed model, the preprocessing phase starts with

the forged image (input image), denoted as F, as shown
in Figure 3, which is compressed to get a compressed
version of the input image, denoted as Fcomp. The difference
between the forged image and its compressed version is then
calculated by mathematical subtraction, denoted as (Fdiff),
as shown in Equation (1).

Fdiff = F − −Fcomp (1)

As a result, the forged part of the image appears in (Fdiff)

due to the difference between the source of the forged and
original parts. Fdiff is then reshaped to 160 × 160 pixels
to fit as an input feature image for training a pre-trained
model (M), which is then used to classify images as forged
or authentic. Figure 3 shows the overall architecture of the
proposed system.
In Figure 3, the pre-trained model, shown as block
(M), is used to extract features from input images (Fdiff) FIGURE 4. Detailed view of the proposed model classifier.
and classify them as authentic images or forged images.
In this block ( pre-trained model), eight different pre-trained
models are considered (one at a time) namely, VGG16 [27], IV. EXPERIMENTAL RESULTS AND DISCUSSION
VGG19 [28], ResNet [29], DenseNet [30], Xception [31], and This section will discuss the training and testing environ-
MobileNet [32] for fine training with input images (Fdiff), mental setup used in the proposed approach and compare its
to nominate the model with the best performance among performance with state-of-the-art techniques.
them.
Each model of the forementioned eight pre-trained models A. EXPERIMENTAL SETUP
has its own architecture which consists of a set of con- 1) ENVIRONMENT
volutional layers with activation functions and ends with For the experiment’s environment, Google Colab Pro was
a set of fully connected layers that can classify up to used with premium GPUs and 100 compute units, with 25 GB
1000 classes of images. So, each model architecture has to system RAM and 16 GB GPU RAM.
be modified to fit the binary classification problem with
only two classes (authentic or forged images) as in the 2) DATASET
case of image forgery detection problems. Therefore, the The CASIA 2.0 dataset is used in the experiments, the
native fully connected layers in each model are replaced same dataset used in [17]. As described in [33], CASIA
with a new set of fully connected classification layers able 2.0 has a total of 12614 images, with 7491 original images
to handle the binary classification problem at hand. The and 5123 forged images, including 3274 copy-move images
convolutional layers in every model should remain untouched and 1849 spliced images. The images are in JPEG and
since they contain all the trainable parameters used in transfer TIFF formats. The pixel dimensions of the images range
learning. from 320 × 240 to 900 × 600 [18]. TABLE 3 provides all
Figure 4 shows the detailed architecture of the proposed information about the CASIA 2.0 database.
model classifier with the newly added layers. After removing
the fully connected layers of the pre-trained model, a flatten TABLE 3. CASIA.2.0 image forgery database specification.
layer is added to convert the input data, which is typically a
multi-dimensional array, into a one-dimensional vector that
can be fed to the next layers. The next two (new) layers
are fully connected layers added with the ReLU activation
function. The two layers have 1024 and 256 neurons,
respectively. After each layer, a dropout 0.5 was added to
prevent overfitting by randomly dropping out (setting to zero)
about 50% of the output values of the previous layer will Tampered images in CASIA v2 were created by combining
be randomly set to zero during the training phase. The last two different authentic images or using the same authentic
fully connected layer with a sigmoid activation function is image. Cropped parts underwent some processing including
added, which is the common activation function used in distortion, rotation, and scaling, to create an image that seems
binary classification problems. to be real, involving blurring the spliced region’s edge [19].

91588 VOLUME 11, 2023

A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

TABLE 4. Details division of CASIA-V2 dataset in the experiments. pre-trained model experiment is drawn, and three samples
from them are displayed in Figures 5,6 and 7. In each
figure, (a) displays the relationship between the training and
validation accuracy, and (b) displays the relationship between
the training and validation loss for each model.
These graphs are useful in many directions, the training
accuracy curve shows how well the model is learning from
the training data over time. As shown the curve generally
In the experiments, the dataset is divided into two sets, increases as the model gets better at fitting the training
training, and testing sets with ratios of 80% and 20%, data.
respectively. As TABLE 4 shows, the testing set is used
as a testing and validation set as done in the paper [17].
The training set contains 10091 images, which are divided
into 5993 authentic images and 4098 forged images, and the
testing set contains 2523 images divided into 1498 authentic
images and 1025 forged images.

B. EVALUATION METRICS
The performance of the proposed model is evaluated using
the metrics specified in [16].
•Accuracy: The accuracy is determined by dividing the
total number of correctly classified instances from both
classes by the total number of instances in the dataset.

Accuracy = [(TP + TN)/(TP + FN + FP + TN)] × 100

(2)
•Recall: is the percentage of tampered images that were
correctly classified out of the total number of images that
were actually tampered.
Recall = TP/(TP + FN) (3)
•Precision: is the proportion of images identified as forged FIGURE 5. VGG19 training and validation curves.
and that are truly forged.
Precision = TP/(TP + FP) (4)
•F1 score: is a measure of the accuracy of a test, which is
defined as the harmonic mean of the precision and recall.

F1 score = [(2 × Recall × Precision)/(Recall + Precision)]

× 100 (5)

C. MODEL TRAINING AND TESTING EVALUATION

In order to fairly evaluate the training and testing phase for
the eight different pre-trained models, a set of initial value
parameters should be fixed all over the eight experiments.
These parameters are as follows: The size of the input images
is 160 × 160, with initial weight ‘ImageNet’, the number of
epochs =100 with early stopping condition monitoring the
minimum validation loss with patience = 10. The optimizer
used is the ‘‘Adam’’ optimizer with learning rate = 1e-5 and
loss ‘binary cross entropy’.
In the experiments, the relation between the training and
validation curve for the accuracy and the loss for each FIGURE 6. ResNet50 training and validation curves.

VOLUME 11, 2023 91589

 A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

decisions during the training process. By monitoring the

validation performance, the training process can stop at the
point where the validation error is minimized, preventing
overfitting and improving generalization. During the exper-
iments’ training processes, an early stopping condition was
put in to monitor the training and validation loss, when the
validation loss increases or diverges from the training loss,
it indicates that the model is overfitting the training data and
is not generalizing well to new data, so it will stop the training
process immediately. This appeared in Figures 5(b), 6(b),
and 7(b), which display the training and validation loss curves
in the selected models.
The evaluation metrics used to compare the performance
of the binary classification models for the proposed eight
techniques with another state-of-the-art technique are shown
in TABLE 5.

TABLE 5. Evaluation metrics of the proposed eight techniques with

state-of-the-art technique.

FIGURE 7. MobileNetV2 training and validation curves.

On the other side, the validation accuracy curve shows how

well the model is performing on a separate set of testing data
that has not been seen during the training. The curve generally
follows the training accuracy curve, but it may not increase as
quickly or may plateau earlier. When the validation accuracy
curve starts to decrease or diverge from the training accuracy For all evaluation matrix, all deep pre-trained models
curve, it indicates that the model is overfitting the training achieve high values compared to the [17] which indicate
data, and is not generalizing well to new data. that using deep pre-trained models improve the overall
The training loss curve shows how well the model is performance of the model.
minimizing the training loss function over time. The curve Figure 8 shows the accuracy rate [17] of all eight
generally decreases as the model gets better at fitting the techniques that relied on pre-trained models compared with
training data. On the other hand, the validation loss curve the state-of-the-art accuracy in that employed CNN only, [25]
shows how well the model is minimizing the loss function that employed the pre-trained model InceptionV3, and [26]
on a separate set of testing data that it has not seen that employed pre-trained models VGG16 and VGG19.
during training. The curve generally decreases but does not Generally, as shown in Figure 8, all the techniques that used
decrease as quickly as the training loss curve or may plateau pre-trained models recorded better detection accuracy rates
earlier. When the validation loss curve starts to increase or than the accuracy given in [17] which used the CNN network
diverge from the training loss curve, it indicates that the only, regardless of [26] which has a lower accuracy rate.
model is overfitting the training data and is not generalizing The proposed model with VGG19 recorded the best detection
well to new data, so the training process should stop accuracy rate since it is classified as one of the best pre-
immediately. trained models in image classification problems. It is worth
The accuracy and loss curves provide insights into how to be noted that MobileNet and ResNet50 come next in order
well the model is learning from the data, and whether it is respectively.
generalizing well to new data. By monitoring these curves It was shown in [28] and [29], that VGG19 and ResNet
during the training process, the performance of the model can are known for their ability to learn rich hierarchical features
be improved that are robust to image transformations, such as scale and
In the experiments, the dataset is divided into a training rotation, which can be useful in detecting different types of
set and a validation set by a ratio of 80:20 respectively, the image manipulations. Moreover, MobileNet [32] is designed
validation set serves as an independent dataset that allows to be lightweight and efficient, making it suitable for mobile
for evaluating the network’s performance, optimizing hyper- devices and embedded systems, while still achieving good
parameters, preventing overfitting, and making informed accuracy on various computer vision tasks. In addition,

91590 VOLUME 11, 2023

A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

these models often include advanced techniques such as

batch normalization, residual connections, and depth-wise
convolutions, which can help to improve their performance
on image classification tasks.

FIGURE 10. Precision chart for the eight experiments.

FIGURE 8. Accuracy chart for the eight experiments.

Overall, the combination of factors, such as deep architec-

ture, large dataset training, robustness to image transforma-
tions, and advanced techniques, make MobileNet, VGG19,
and ResNet well-suited for detecting image forgeries and
achieving high detection accuracy rates. FIGURE 11. Recall chart for the eight experiments.

The F1 score has the same meaning of precision and

recall, commonly used to evaluate the performance of a
classification process and how well the model can clearly
classify the two classes. F1 score, precision, and recall
should have a higher score. As shown in Figures 9,10
and 11, it is worth to be noticed that the models VGG19,
reference [9], ResNet50, and MobileNetV2 achieved the best
results in F1 score, and the models VGG19, ResNet50, and
MobileNetV2 achieved the best results in precision and recall,
respectively.

FIGURE 12. AUC chart for the eight experiments.

respectively, which means that these three models possess a

better ability to correctly classify the images.
As known, a model with a smaller number of parameters
can lead to faster training, lower computational costs, and
reduced risk of overfitting the model to the training data.
Complex models with many parameters relative to the
FIGURE 9. F1 Score Chart for the eight experiments. available data can lead to overfitting, this is because of
memorizing the training data instead of generalizing to new
Area Under Curve, (AUC) is another performance measure data.
for a binary classification model. A higher AUC value From the experimental results and Figure 13, MobileNet,
means that the model can differentiate between positive DenseNet121, and VGG16 have the lowest number of param-
and negative classes accurately with less error. In other eters that lead to faster training time, and lower computational
words, the model can correctly identify true positives and costs. However, as mentioned earlier, a model with too
true negatives, while minimizing false positives and false many parameters can lead to overfitting. Therefore, it is
negatives. From the experimental results, shown in Figure 12, important to achieve a balance between model complexity
VGG19, MobileNet, and ResNet have the highest AUC value, and generalization.

VOLUME 11, 2023 91591

 A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

recorded in F1 score, recall, precision AUC, and TPR,

which makes them the best choice in any image forgery
detection system. In addition, weighing the number of
network parameters as a measure of the system computational
cost, MobileNet, DenesNet, and VGG16 have the minimum
values. When compromising between the computational
costs of the network and its accuracy, we found that the
MobileNet pre-trained model is the best choice for an image
forgery detection system that satisfies maximum detection
accuracy rate and minimum computational costs and training
time.
FIGURE 13. Total number of model’s parameter Chart for the eight
experiments.
V. CONCLUSION AND FUTURE WORK
Image forgery detection techniques have become essential
with the increased availability of image editing tools that
can create forged. The paper presented an image forgery
detection technique based on deep learning via a pre-
trained model and transfer learning. The proposed tech-
nique considered the difference between an image and its
compressed version to produce a featured image as an
input to a pre-trained model that improved the detection
accuracy rate. The technique with a given data set was
applied to eight different pre-trained models adapted for
binary classification. The recorded experimental results
were compared with the state-of-the-art method. The results
showed that using pre-trained models help achieve a higher
detection accuracy rate than the state of the art which used
FIGURE 14. TPR and TNR Chart for the eight experiments.
CNN model.
Moreover, comparing the resulting evaluation metrics,
charts, and graphs, for the eight pre-trained models, it was
found that MobileNetV2 had the highest detection accuracy
rate (around 95%) with a smaller number of training
parameters which led to faster training, and lower computa-
tional costs, and lower system complexity and low memory
consumption. So, it is highly recommended as a backbone
with the image compression technique that effectively detects
image splicing and copy-move at the same time with highly
encouraging results.
Although the proposed model recorded high performance,
compared to the performance of other studies, still there exist
a set of limitations such as data generalization since the model
performed well with training data and failed to generalize
FIGURE 15. FPR and FNR Chart for the eight experiments. with unseen data. Forgery types generalization that the model
was unable to detect all image forgery techniques including
the novel forgery techniques. Localization is another issue
As shown in Figures 14 & 15, MobileNet, ResNet50, since the model did not consider localizing the forged parts.
and VGG19 have the highest TPR and TNR, indicating that Computational complexity and resource requirements for
these models can correctly detect the actual positive and deep neural networks can be computationally intensive that
negative cases that are identified as positive and negative, require significant computational resources and time for
respectively. training and inference.
As mentioned before, when considering the resulting In the future, an enhancement to the proposed technique
evaluation metrics, charts, and graphs, the highest detection can be added to increase the detection accuracy rate,
accuracy rates for the three models, VGG19, MobileNet, keeping in mind the minimization of the training time
and ResNet50 are 94.77%, 94.69%, and 94.6%, respec- and computational cost. Additionally, image forgery type
tively. Comparable results for the three models are also detection, splicing, or copy move, can be extended with

91592 VOLUME 11, 2023

A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

localization. The combination of the proposed approach with [21] K. D. Kadam, S. Ahirrao, and K. Kotecha, ‘‘Efficient approach towards
other known image localization techniques will improve the detection and identification of copy move and image splicing forgeries
using mask R-CNN with MobileNet v1,’’ Comput. Intell. Neurosci.,
accuracy, but it may increase the time complexity so it will vol. 2022, pp. 1–21, Jan. 2022.
need more improvement. The detection of forged videos that [22] M. N. Abbas, M. S. Ansari, M. N. Asghar, N. Kanwal, T. O’Neill,
may be created by merging several videos is an incredibly and B. Lee, ‘‘Lightweight deep learning model for detection of copy-
move image forgery with post-processed attacks,’’ in Proc. IEEE
challenging task. 19th World Symp. Appl. Mach. Intell. Informat. (SAMI), Jan. 2021,
pp. 125–130.
[23] C. D. P. Kumar and S. S. Sundaram, ‘‘Metaheuristics with optimal deep
REFERENCES transfer learning based copy-move forgery detection technique,’’ Intell.
[1] K. D. Kadam, S. Ahirrao, and K. Kotecha, ‘‘Multiple image splicing dataset Autom. Soft Comput., vol. 35, no. 1, pp. 881–899, 2023.
(MISD): A dataset for multiple splicing,’’ Data, vol. 6, no. 10, p. 102, [24] N. Krishnaraj, B. Sivakumar, R. Kuppusamy, Y. Teekaraman, and
Sep. 2021. A. R. Thelkar, ‘‘Design of automated deep learning-based fusion model for
[2] R. Agarwal, O. P. Verma, A. Saini, A. Shaw, and A. R. Patel, ‘‘The advent copy-move image forgery detection,’’ Comput. Intell. Neurosci., vol. 2022,
of deep learning-based,’’ in Innovative Data Communication Technologies pp. 1–13, Jan. 2022.
and Application. Singapore: Springer, 2021. [25] S. Jabeen, U. G. Khan, R. Iqbal, M. Mukherjee, and J. Lloret, ‘‘A deep
[3] M. A. Elaskily, M. H. Alkinani, A. Sedik, and M. M. Dessouky, multimodal system for provenance filtering with universal forgery
‘‘Deep learning based algorithm (ConvLSTM) for copy move forgery detection and localization,’’ Multimedia Tools Appl., vol. 80, no. 11,
detection,’’ J. Intell. Fuzzy Syst., vol. 40, no. 3, pp. 4385–4405, pp. 17025–17044, May 2021.
Mar. 2021. [26] D. Mallick, M. Shaikh, A. Gulhane, and T. Maktum, ‘‘Copy move and
splicing image forgery detection using CNN,’’ in Proc. ITM Web Conf.,
[4] A. Mohassin and K. Farida, ‘‘Digital image forgery detection approaches:
vol. 44, 2022, Art. no. 03052.
A review,’’ in Applications of Artificial Intelligence in Engineering.
[27] K. Simonyan and A. Zisserman, ‘‘Very deep convolutional networks for
Singapore: Springer, 2021.
large-scale image recognition,’’ 2014, arXiv:1409.1556.
[5] K. B. Meena and V. Tyagi, Image Splicing Forgery Detection Techniques: [28] T.-H. Nguyen, T.-N. Nguyen, and B.-V. Ngo, ‘‘A VGG-19 model
A Review. Cham, Switzerland: Springer, 2021. with transfer learning and image segmentation for classification of
[6] S. Gupta, N. Mohan, and P. Kaushal, ‘‘Passive image forensics using tomato leaf disease,’’ AgriEngineering, vol. 4, no. 4, pp. 871–887,
universal techniques: A review,’’ Artif. Intell. Rev., vol. 55, no. 3, Oct. 2022.
pp. 1629–1679, Jul. 2021. [29] K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for image
[7] W. H. Khoh, Y. H. Pang, A. B. J. Teoh, and S. Y. Ooi, ‘‘In-air hand recognition,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR),
gesture signature using transfer learning and its forgery attack,’’ Appl. Soft Jun. 2016, pp. 770–778.
Comput., vol. 113, Dec. 2021, Art. no. 108033. [30] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, ‘‘Densely
[8] Abhishek and N. Jindal, ‘‘Copy move and splicing forgery detection using connected convolutional networks,’’ in Proc. IEEE Conf. Comput. Vis.
deep convolution neural network, and semantic segmentation,’’ Multimedia Pattern Recognit. (CVPR), Jul. 2017, pp. 2261–2269.
Tools Appl., vol. 80, no. 3, pp. 3571–3599, Jan. 2021. [31] F. Chollet, ‘‘Xception: Deep learning with depthwise separable convo-
[9] M. M. Qureshi and M. G. Qureshi, Image Forgery Detection & lutions,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR),
Localization Using Regularized U-Net. Singapore: Springer, 2021. Jul. 2017, pp. 1800–1807.
[10] Y. Rao, J. Ni, and H. Zhao, ‘‘Deep learning local descriptor for image splic- [32] R. Indraswaria, R. Rokhanab, and W. Herulambang, ‘‘Melanoma image
ing detection and localization,’’ IEEE Access, vol. 8, pp. 25611–25625, classification based on MobileNetV2 network,’’ in Proc. 6th Inf. Syst. Int.
2020. Conf. (ISICO), 2022, pp. 198–207.
[33] J. Dong, W. Wang, and T. Tan, ‘‘CASIA image tampering detection
[11] K. M. Hosny, A. M. Mortda, N. A. Lashin, and M. M. Fouda, ‘‘A
evaluation database,’’ in Proc. IEEE China Summit Int. Conf. Signal Inf.
new method to detect splicing image forgery using convolutional neural
Process., Jul. 2013, pp. 422–426.
network,’’ Appl. Sci., vol. 13, no. 3, p. 1272, Jan. 2023.
[12] F. Li, Z. Pei, W. Wei, J. Li, and C. Qin, ‘‘Image forgery detec-
tion using tamper-guided dual self-attention network with multireso-
lution hybrid feature,’’ Secur. Commun. Netw., vol. 2022, pp. 1–13,
Oct. 2022.
[13] C. Haipeng, C. Chang, S. Zenan, and L. Yingda, ‘‘Hybrid features and
semantic reinforcement network for image,’’ Multimedia Syst., vol. 28,
no. 2, pp. 363–374, 2021.
[14] Q. Li, C. Wang, X. Zhou, and Z. Qin, ‘‘Image copy-move forgery detection
and localization based on super-BPD segmentation and DCNN,’’ Sci. Rep.,
vol. 12, no. 1, Sep. 2022, Art. no. 14987.
[15] A. K. Jaiswal and R. Srivastava, ‘‘Detection of copy-move forgery in ASHGAN H. KHALIL received the B.S. degree
digital image using multi-scale, multi-stage deep learning model,’’ Neural in computer science from the Modern Academy,
Process. Lett., vol. 54, no. 1, pp. 75–100, Aug. 2021. Cairo, Egypt, in 2000, the Deplume Program in
[16] S. Koul, M. Kumar, S. S. Khurana, F. Mushtaq, and K. Kumar, management and the M.S. degree in information
‘‘An efficient approach for copy-move image forgery detection using systems from the Sadat Academy, Cairo, in
convolution neural network,’’ Multimedia Tools Appl., vol. 81, no. 8, 2004 and 2011, respectively, and the M.S. degree
pp. 11259–11277, Mar. 2022. in computer science from the Arab Academy
[17] S. S. Ali, I. I. Ganapathi, N.-S. Vu, S. D. Ali, N. Saxena, and N. Werghi, for Science, Technology and Maritime Transport,
‘‘Image forgery detection using deep learning by recompressing images,’’ Cairo, in 2015. She is currently pursuing the
Electronics, vol. 11, no. 3, p. 403, Jan. 2022. Ph.D. degree in computer science with Helwan
[18] K. Kadam, S. Ahirrao, K. Kotecha, and S. Sahu, ‘‘Detection and University, Cairo.
localization of multiple image splicing using MobileNet v1,’’ IEEE Access, From 2000 to 2011, she was a Teaching Assistant with the Computer
vol. 9, pp. 162499–162519, 2021. Science Department, Modern Academy. From 2012 to 2019, she was an
[19] E. U. H. Qazi, T. Zia, and A. Almorjan, ‘‘Deep learning-based digital image Assistant Lecturer with the Computer Science Department, University of
forgery detection system,’’ Appl. Sci., vol. 12, no. 6, p. 2851, Mar. 2022. Wales, Validated Schemes, Faculty of Computer Science and Artificial
[20] A.-R. Gu, J.-H. Nam, and S.-C. Lee, ‘‘FBI-Net: Frequency-based image Intelligence, and the Modern University for Technology and Information
forgery localization via multitask learning with self-attention,’’ IEEE (MTI), Cairo. Her research interests include image processing and security
Access, vol. 10, pp. 62751–62762, 2022. techniques using deep learning.

VOLUME 11, 2023 91593

 A. H. Khalil et al.: Enhancing Digital Image Forgery Detection Using Transfer Learning

ATEF Z. GHALWASH received the Ph.D. degree GOUDA I. SALAMA received the B.Eng. and
from the Faculty of Engineering, Computer Engi- M.Eng. degrees from MTC, Cairo, Egypt, in
neering Department, University of Maryland, 1988 and 1994, respectively, and the Ph.D.
USA, in 1988. He is currently a Professor with the degree in electrical and computer engineering
Faculty of Computers and Artificial Intelligence, from Virginia Tech University, USA, in 1999.
Helwan University, Egypt. He is involved in scien- He is currently a Faculty Member with the
tific research in the field of computer science and Department of Computer Engineering, MTC.
his specialty includes data and network security, His research interests include image and video
artificial intelligence algorithms, machine learn- processing, pattern recognition, and information
ing, image processing, and software engineering. security.
He has published more than 100 scientific papers in high quality journals
and international conferences. He was the Head of the Scientific Promotion
Committee, High Ministry of Education, in the computer and information
systems field, from 2016 to 2022.

HAITHAM A. GHALWASH received the Ph.D.

degree in computer science and engineering from
the University of Connecticut, USA, in 2020.
HALA ABDEL-GALIL ELSAYED is currently a He was an Assistant Professor of residence with
Professor of artificial intelligence and the Head the University of Connecticut, in 2020, where
of the Computer Science Department, Faculty he taught computer science courses until 2022.
of Computers and Artificial Intelligence, Helwan He is currently the Acting Course Director of
University. the Ethical Hacking and Cybersecurity Program,
Coventry University–Egypt Branch. His research
interests include software-defined networks,
applied cryptography, PKI systems, A.I, machine learning, and network
security.

91594 VOLUME 11, 2023