Deep Learning Methods for Copy Move Image Forgery Detection: A Review

In recent years, the importance of detecting image forgery has increased significantly in the real world due to the ease with which a particular image can be altered and shared on social media, leading to the spread of fake news and rumors worldwide [1-3]. Detecting image forgery has become a significant challenge for image forensics due to the availability of sophisticated image manipulation software. Although there have been several traditional approaches for detecting image forgery described in the literature, they tend to focus on extracting simple traits and are designed to address specific types of fraud. However, with the advancement of deep learning techniques, deep learning-based algorithms have gained popularity for detecting fake images. These methods involve building a model that automatically extracts critical features for classification rather than relying on statistical or geometrical computations. Compared to traditional methods, deep learning techniques have demonstrated excellent results in detecting image modifications.

The research on deep learning-based copy-move image forgery detection has important applications in cyber security, computer vision, and digital forensics. For example, CMFD can improve digital evidence processing in forensics, helping law enforcement prosecute cybercrimes and increasing accuracy. Also, it may be used for computer vision problems such as picture identification and object detection, which will result in stronger algorithms. Further, CMFD can help build sophisticated cyber security procedures and solutions that guard sensitive data and guarantee reliable communications by detecting and stopping digital picture manipulation. These cross-disciplinary effects demonstrate the importance and wider application of CMFD.

The purpose of this study is to give researchers an overview of how copy move image forgery research has developed over the last several years and to indicate potential directions for future research. With the use of some insightful comparisons between deep learning-based classifiers and conventional approaches, this study will present performance evaluation, with a focus on the most popular passive detecting copy-move forgeries with deep learning techniques.

The majority of the existing surveys on picture forensics [3-13] centered on traditional feature-based techniques. Recent studies [4, 10] have a specific perspective or scope than ours. Yang et al. [10] categorize the proposed techniques in the literature to two categories: unsupervised and supervised techniques with a special focus on “Deep fake detection techniques”. In fact, this paper employs a wide range of the most recent deep learning techniques to detect forged copies in images. We try also to have a larger coverage than other previous surveys, which have devoted their evaluations to describe and analyze the solutions for one or more specific issues, such as image source identification [10], camera identification [13], and computer-generated picture detection [14].

The remainder of the paper is structured as follows: Section II provides a classification of CMFD techniques and types of picture forgery detection. Section III follows, which presents a description of the copy move image manipulation. Section IV gives an overview of the principal of deep learning based CMFD. Section V presents the traditional and the deep learning methods. Section VI gives keys concepts used in deep learning for CMFD. Section VII reviews the deep learning methods. In section VIII, gives a conclusion that summarize the paper.

Copy-move image forging is a widely used and uncomplicated method of altering images. This technique involves copying a part of an image and pasting it elsewhere in the same picture. Since the pasted portion comes from the original image, its significant characteristics like color, noise, and texture remain unaltered, making it more difficult to detect this type of image forgery.

The objective of detecting copy-move forgery is to recognize any replicated portions in an image being analyzed, which could suggest an attempt to commit fraud. Copy-move alterations can be classified as "simple", "affine", or "complicated" forgeries, depending on the level of difficulty involved in the copying process.

1. Simple cloning: simple or plain cloning involves copying and pasting a section of an image to a different location without any modifications. This type of cloning can be done using basic picture editing software and is straightforward.

2. Affine cloning: is a type of image tampering that involves altering a portion of an image through scaling and rotation using affine transformations. This results in a new portion of the image that is transformed in a way that is different from the original portion. Affine cloning is similar to plain copy-move tampering and can be easily carried out using image editing software that supports affine transformations.

3. Complex cloning: refers to a more intricate method of altering a replicated section of an image, often involving further processes like blending the edges, estimating diffusion, altering colors, or employing other advanced techniques for image manipulation. Accomplishing complex cloning requires the use of advanced image editing software. Figure 2 illustrates a case of copy-Move forgery, where the left image depicts a manipulated version of the original image on the right, which initially featured three missiles. The forged image on the left has four missiles, which is an example of the type of alteration that can occur in this type of forgery. Figure 3 gives the number of research papers produced in Springer and IEEE between the years 2016 and 2024.

2.png

Figure 2. Copy-move forgery example: (left) original image with 3 missiles (right) forged image with 4 missiles

3.png

Figure 3. The number of research papers in the field of image forgery detection between 2016 and 2024 in Springer and IEEE

In the literature the CMFD methods are classified in two categories:

5.1 Traditional methods

In Traditional or image processing based method, there are three main types of traditional methods used for detecting copy-move forgery, namely key points-based (pixel based), overlapping block-based and object based. The overlapping block-based method involves breaking up suspect images into overlapping blocks in order to analyze the similarity of their features and detect any tampered regions. This can be done using techniques such as PCA [17], DCT [18], Zernike [19], and dense-field [20]. While effective, these algorithms are computationally expensive and not suitable for detecting geometric modifications.

On the other hand, key point-based techniques such as SIFT [21], SURF [22], triangle [23] and ORB [24], are more resistant to geometric deformation, but may not work well on areas of the image that are smooth. These methods extract robust key point features and use similarity matching to identify tampered regions.

In the object-based copy-move image forgery detection locates and examines duplicate objects in a picture. Using object detection algorithms, it begins with object detection and then uses techniques to extract characteristics from an object, such as texture, color, and form. After that, related areas in the image are identified by comparing the retrieved characteristics across it. This object-focused method works well for identifying forgeries where entire objects are duplicated and repositioned inside the picture.

The different CMFD categories are illustrated in Figure 5. Also a comparative table between Traditional and DL approached for CFMD is given in Table 1.

5.png

Figure 5. CMFD categories

Table 1. Traditional methods vs deep learning methods

5.2 Deep learning methods

In the CMFD literature, three main approaches are used in deep learning based CMFD: image-based, sequence-based, and region-based. In the image-based approach CNN is used to analyse the full image to extract and match features, suitable for detecting large duplicated regions. Whereas in the sequence-based approach RNNs or Long Short Term Memory networks to capture temporal relationships and identify forgeries with sequential patterns. Finally in the region-based method, which works well for intricate forgeries involving several very small copied segments, divides the image into smaller patches, takes features from each segment, and matches comparable segments.

From another point of view, the proposed approaches base on deep learning for copy-move detection are classified into two subcategories:

5.2.1 Approaches using custom models

These approaches are based on the creation and training of custom deep learning models specifically for copy-move detection tasks. Techniques such as CNN, DNN, and RNN are the most used for this purpose. For example, CNNs are good at spotting replicated regions in faked photos because they are good at collecting spatial hierarchies of characteristics. CNNs and DNNs have powerful capabilities in feature extraction, but they may need a significant amount of processing power. RNNs are mostly made for sequential data, however because they concentrate on temporal links, they might not be the best option for copy-move detection.

5.2.2 Approaches using a model with transfer learning

On the other hand, these methods use transfer learning and pre-trained deep learning models to tackle copy-move detection tasks. Transfer learning is the process of optimizing previously learned models usually CNNs built on substantial datasets like ImageNet to cope with particular objectives like copy-move detection. This method gives excellent performance in CMFD. Transfer learning makes successful feature extraction and classification possible by utilizing the information embodied in pre-trained models. This improves the identification of duplicated portions in forged photos.

The success of computer vision techniques based on deep learning [10, 25], along with advancements in GPU technology, has led researchers to explore the use of Deep Learning models for detecting image tampering. Deep Learning involves two main stages - feature extraction and classification - and is capable of automatically learning abstract and complex patterns required to identify manipulated areas in images. This approach offers several advantages, such as reducing the time and effort needed to detect hand-crafted features in altered photos. However, training Deep Learning models can be challenging due to the high computational requirements and the need for large amounts of data. Convolutional Neural Networks (CNNs), Deep Neural Networks (DNNs) [26] and Recurrent Neural Networks (RNNs) are some examples of the various Deep Learning models available for this task.

Barad and Goswami [27] carried out a study utilizing deep learning techniques to identify manipulated images. They also evaluated the methods used to verify the authenticity of photographs using publicly available databases.

Wu et al. [28] introduced a deep learning-based architecture called BusterNet for detecting copy move picture fraud. This method is end-to-end trainable and consists of two branches. The first branch identifies manipulation areas using visual artifacts, while the second branch finds copy/move locations based on visual similarities. The authors also provided simple techniques to use out-of-domain datasets and an effective training process for BusterNet. Their extensive research revealed that BusterNet outperformed traditional copy move algorithms by a significant margin.

The importance of utilizing deep learning-based algorithms on publically accessible datasets to identify manipulation in photos was covered by Manjunatha and Patil [29]. They discussed the techniques for passive picture forensic analysis and emphasized upcoming difficulties in creating a system for the identification of altered photos.

Rao and Ni [30] suggested a CNN-based architecture in a different research for the identification of fake digital images. They suggested that the pre-processing phase is directly affected by the first layer of the CNN model. It looks for problems that result from manipulation. SVM was used for the detection phase.

Zhan et al. [31] proposed a transfer learning method that utilizes pre-trained weights from the AlexNet model and an SVM classifier to reduce the time needed for training. The performance of the model was good. Another approach based on transfer learning was developed by Doegar et al. [32], which benefited from previous knowledge obtained through the steganalysis model. With this method, they achieved an average accuracy of 97.36%.

Bi et al. [33] proposed a technique called the Ringed Residual U-Net (RRU-Net) for detecting image forgeries. Their approach involves using an end-to-end image segmentation network to identify manipulations without the need for pre- or post-processing. The RRU-Net architecture is designed to mimic the processes used by the human brain for consolidation and recollection, which can improve the ability of a CNN to learn. By using residual propagation to remember input feature information in a CNN, the authors were able to overcome the problem of gradient degradation. The final output of the RRU-Net combines the remaining answer with the response characteristic, enabling it to distinguish between real and fake areas in an image. Experimental results showed that the proposed technique outperformed existing conventional methods.

In the research [31], another study was introduced that offered a comprehensive solution based on deep neural networks for identifying copy-move fraud. To accomplish this, a convolutional feature extractor was utilized to analyse the input image and extract block-like characteristics, with the VGG16 feature extractor being used. In order to enhance the image resolution, bilinear up-sampling was used. The main limitation of the proposed model was that it did not perform well on images with only textures.

Another recent DL technique for identifying and locating forged areas in compressed pictures is presented in study [34]. The proposed approaches use the EfficientNet model. An effective feature extraction and classification are made using optimization algorithms such as SPOA, CDT, and TAS together with quality criteria. The effectiveness of the proposed approaches is demonstrated by evaluation on CASIA datasets.

The work [35] presents CAMU-Net, an inventive CMFD technique that successfully detects copy-move fraud inside images. Using the VGG16 architecture, the approach use four-step, the first stage, Hierarchical Feature Extraction (HFE_Stage), extracts multi-scale features that capture both global semantics and local details. To improve the accuracy of forgery detection, the ensuing Hierarchical Feature Matching (HFM_Stage) hierarchically matches features. By concentrating on key regions at various sizes through the use of a coordinate attention mechanism, the Coordinate Attention-based Resource Allocation (CARA_Stage) improves matching maps. Lastly, to improve the detection of copy-move forgeries, the Multi-scale Feature Fusion-based Up-sampling (MFFU_Stage) combines data from several scales. During these phases, CAMU-Net outperforms both conventional and deep learning-based techniques in identifying locations in photos where copy-move forgeries are present.

An end-to-end trainable copy-move fusion strategy for CMFD is presented in the work [36], which combines CNN architecture with the benefits of CenSurE keypoint detection. By employing a data-driven approach, the technique is able to detect and localize copy-move forgeries in a variety of picture contexts by continuously updating its learning through training data. The method enables forgery detection across many copy-move attack types, such as basic, post-processed, and geometrically modified forgeries, alos it cane cope with other image processing modifications, by combining CNN features with CenSurE keypoints. The method has strong performance in a range of textures for pictures and gives accurate results under various attack condition.

In the paper [37], skip connections, cycle learning rates (CLR), and a deep CNN-based model named ResNet-101 are coupled to address the issue of exploding and vanishing gradients in CMFD. To demonstrate the efficacy of the model, the work is trained and assessed using datasets such MICC-F600, MICC-F2000, MICC-F220, and CoMoFoD v2. With accuracy rates of up to 97.75% for CoMoFoD v2 after just 5 epochs and 96.09%, 97.63%, and 96.87% for MICC-F220, MICC-F600, and MICC-F2000, respectively, after 10 epochs, the suggested method outperforms the state-of-the-art models currently in operation.

A cmapraison between various deep learning-based methods for detecting copy-move forgery is presented in Table 2.

Table 2. Comparative analysis of deep learning approaches for copy-move detection

Deep learning showed a powerful and effective tool for CMFD. In the flowing points some achieved results and future challenge:

9.1 Results

Best accuracy: Deep learning based solutions can give best accuracy in detecting copy-moved manipulations compared to traditional methods. All that is due to their capacity to recognize complex patterns and relations in pictures that are difficult to extract using manually created features.

Robustness: Deep learning models can be more robust to various manipulations like rotation, compression, and noise addition, which forgers often employ to disguise their edits. This is due to the model's ability to learn these variations during training.

Automation: Deep learning approaches can automate the copy-move detection process, reducing the need for manual analysis by forensic experts. This saves time and resources.

9.2 Challenges

Deep learning-based copy-move detection still has several challenges and issues despite the encouraging outcomes:

Big datasets: deep learning models training need a significant amount of data comprising both authentic and fake pictures. Collecting such big datasets can be very expensive.

Computational cost: running deep learning model need powerful and sophisticated hardware which can be costly. The accessibility to such hardware may be not easy.

Manipulation techniques: As image manipulation tools advance, CMFD becomes increasingly challenging and deep learning model-based solutions must be updated and improved on a regular basis to keep up with forgers' inventive ways of creating ever-more-complex forgeries.

Interpretability: It might be challenging to comprehend how deep learning models arrive at judgments because to their complexity. This might provide a challenge in forensic environments when it's crucial to offer context for a finding.