Unified frequency and graph learning framework for accurate and robust copy‐move image forgery detection

Abstract

Copy‐move forgery (CMF) is one of the most prominent challenges to be addressed by digital image forensics experts today due to the ease with which forged objects can be copied and placed within the same image to conceal or replicate objects. The conventional approaches to detect CMF frequently suffer from a number of challenges, including increased computational complexity, poor localization accuracy, and reduced efficacy to detect minor forgeries. To mitigate the challenges associated with the conventional approach to detect copy‐move forgeries, the current study aims to introduce a novel Dual‐Phase Forgery Identification Network (DPFIN) that is capable of improving the efficacy of the detection process while exhibiting better resilience to complex forgeries. The proposed approach is based on a novel dual‐phase preprocessing mechanism that combines the efficacy of frequency domain analysis with adaptive visual enhancement to detect hidden forgery artifacts. The proposed approach utilizes a YOLOv5‐based backbone to extract multi‐scale features for the input images. Furthermore, a learning mechanism based on the Zebra Optimization Algorithm is also introduced to ensure stable training and better generalization performance for the model. The performance evaluation of the proposed DPFIN is carried out to achieve 99.89% accuracy and 98.89% F1‐score with a false positive rate of 0.21% on the CoMoFoD dataset. Also, 99.99% AUC is achieved with 96.8% F1‐score on the CASIA v2 dataset. This proves that the proposed DPFIN is a reliable solution for image forgery detection.

Highlights.

• Dual‐domain CMFD framework using both frequency and spatial feature analysis.

• Graph convolution improves the localization of duplicated regions.

• Achieves 99.89% accuracy along with a false positive rate of 0.21% on CoMoFoD.

• Robust performance on CASIA v2 with an AUC of 99.99%.

• Forensic image authentication and evidence analysis can reliably be performed using proposed DPFIN.

1. INTRODUCTION

Digital image manipulation is an issue of great concern in modern digital publishing, journalism, and forensic science. With the advent of advanced image editing software, it has become very easy to edit images while retaining their realistic appearance. Image counterfeiting is usually defined as a malicious modification of image content with the help of image editing software like Photoshop and other digital image manipulation tools, wherein the semantic content of an image is modified to provide false information [1, 2]. Image editing is usually carried out with the help of various image editing operations like object insertion/removal, blending, geometric transformations, shifting, and addition of artificial noise to conceal image editing [3]. Although visually identical to real images, statistical and structural traces of manipulations are often left behind by operations in forged images, which can be utilized for analysis. For instance, feature‐based methods like Scale‐Invariant Feature Transform (SIFT) have been extensively employed to identify cloned areas in tampered images by using invariant local descriptors [4]. Moreover, splicing detection methods try to identify inconsistencies between real and tampered areas by analyzing texture, compression, and lighting patterns. Among various kinds of digital image manipulations, CMF has gained prominence in digital image forensic research. In this kind of image manipulation, an image part is copied and moved to another part of the same image to hide or duplicate objects without leaving any signs of external manipulations [5]. Hence, Copy‐Move Forgery Detection (CMFD) is defined as an image forensic technique to automatically detect duplicated parts of an image without any prior knowledge of the kind of forgery. Over the years, various kinds of CMFD methods have been developed to solve this problem [6]. Traditional methods have proven to successfully detect duplicated parts of an image even after going through various image processing operations such as compression, rotation, or noise addition [7]. However, it has been found that these methods have higher complexity and require more time to process images compared to other methods [8].

Digital image manipulation can be achieved through different image processing operations such as image merging, enhancement, transformation, and resampling operations [9]. Resampling images, for example, involves changing the geometric characteristics of objects within images, such as scaling or rotating images, for the purpose of creating convincing image modifications [10]. In the process of creating CMF, the manipulated image may be further altered through different image transformation operations to evade detection. These image transformation operations may include scaling, rotating, JPEG compression, color reduction, Gaussian noise addition, image brightness and contrast adjustment, and blurring images [11, 12, 13]. In this regard, the proposed CMFD algorithms must be able to detect the duplicated images despite these image transformation operations. It is observed that traditional CMFD techniques can be broadly classified into two types: block‐based and keypoint‐based methods. Keypoint‐based methods aim at identifying areas in the image with high entropy values, such as edges and corners, referred to as keypoints in image processing [14]. These keypoint areas in images tend to have certain attributes like invariance to rotation and scale change, allowing for better matching even after undergoing certain transformations [15]. On the other hand, in block‐based methods, images are divided into blocks and descriptors are used to identify areas with similarities in the image through pairwise comparisons [16, 17]. SIFT, SURF, and ASIFT techniques have often been used in conjunction with traditional methods for better accuracy in cloned region detection. Recently, these techniques have been combined with convolutional kernel networks for better efficiency in terms of processing time [18]. Nevertheless, keypoint‐based methods have shown limitations in detecting cloned regions in smooth areas of images where keypoint density is low or even null [19].

To overcome the limitations of traditional feature‐based approaches, recent studies have shown significant interest in exploring the potential of deep learning technology for the development of CMFD systems. CNN and transfer learning models have shown remarkable potential in learning discriminative features automatically from the image data [20]. Most CNN‐based CMFD systems use segmentation and classification approaches for detecting and locating manipulated parts of the image data [21]. By utilizing the feature learning capability of deep learning models, researchers have developed CMFD systems capable of detecting subtle forgery artifacts, which might not be detectable with traditional feature‐based approaches [22]. Moreover, the use of pre‐trained models has also enabled the development of real‐time forgery detection systems, which can process large volumes of image data efficiently [23]. Overall, the use of deep learning technology has improved the performance, robustness, and flexibility of CMFD systems, which automatically learn representative features from the image data and detect even subtle manipulations with high reliability [24].

The rapid advancement of digital image processing tools has increased the threat of image manipulation, especially CMF, in which areas of an image are copied and moved to conceal or duplicate objects in an image. Existing methods of CMFD are often limited in that they are computationally expensive, lack precise localization accuracy, and are not robust enough to detect even minor manipulation of an image. Such limitations point to the need to develop a robust image forgery detection mechanism that is capable of detecting structural and contextual forgery features. In this regard, this research seeks to develop a novel image forgery identification mechanism known as the Dual‐Phase Forgery Identification Network (DPFIN).

The main contributions of this research can be briefly outlined as follows:

• The development of a dual domain forgery detection framework, where the spatial and spectral domains are jointly analyzed to enhance the detection of subtle copy‐move operations and small tampered areas.

• The design of a multi‐scale artifact extraction mechanism, where both local texture and global structural patterns are effectively identified to detect manipulated images.

• The proposal of a relational modeling approach to effectively handle the context relationships between different areas of the image, thereby improving the robustness of forgery detection, especially for structurally similar forged areas.

• The integration of an adaptive optimization and learning process to improve the generality, stability, and reliability of the forgery detection.

The remainder of this paper is structured as follows: In Section 2, the DPFIN framework and datasets used in our experiments are presented. In Section 3, quantitative and qualitative results of our experiments are discussed. In Section 4, interpretation of our results and their potential use in digital forensics is discussed. Finally, in Section 5, the contributions are concluded and future research directions are outlined.

2. MATERIALS AND METHODS

2.1. Workflow overview

The existing CMFD techniques mostly depend on spatial domain feature learning and frequency domain handcrafted descriptors. However, the single domain‐based detection techniques may not be effective in the presence of secondary processing operations on the image. Spatial domain‐based detection may miss the forgery evidence, and frequency domain‐based detection may not be effective in learning the contextual relationships between the image regions. Most of the CNN‐based detection techniques consider the regions of the image independently and may not be effective in analyzing the inconsistencies between the duplicated and original regions.

To address these drawbacks, a novel solution is proposed in the form of a comprehensive and robust framework known as the DPFIN. Initially, the input RGB image $I$, is subjected to a preprocessing step to normalize its visual characteristics and enhance potential forgery clues. Subsequently, the enhanced image is converted from the spatial domain to the frequency domain by applying the Fast Fourier Transform (FFT) to detect irregularities in the spectral domain, indicative of a CMF. In the meantime, deep spatial features are learned from the input image by a YOLOv5 model with a CSPDarkNet backbone.

The spatial representations are integrated with frequency domain features using the dual domain fusion mechanism within the Deep Graph YOLO CMFD Network (DGY‐CMFD) module. The integrated representations are divided into nodes based on regions and arranged in the form of a graph. This graph structure allows the detection of relationships between duplicated regions. A Graph Convolutional Network (GCN) is employed to propagate contextual information between connected nodes. This helps in the detection of inconsistencies between the cloned and authentic regions. Finally, the parameters of the model are optimized using an optimization‐driven method of learning in order to ensure better convergence and reliability of the forgery detection process. It is through the hierarchical integration of these steps of preprocessing, spatial learning using YOLOv5‐CSPDarkNet, fusion of the two domains of the model, and graph‐based reasoning of the relations in the image that the DPFIN framework is able to ensure effective CMF detection (Figure 1).

FIGURE 1.

Block diagram of the proposed DPFIN framework illustrating the overall workflow, including preprocessing, feature extraction, graph‐based feature modeling, and final CMF detection and localization.

2.2. Preprocessing

The preprocessing phase seeks to enhance the subtle indications of tampering before the process of deep feature learning. Conventional spatial‐domain preprocessing techniques directly process the pixel values, which makes them vulnerable to changes in illumination, noise, and compression. This makes these approaches incapable of detecting subtle indications of CMF, which cause minor structural inconsistencies. To overcome the limitations of conventional preprocessing approaches, the suggested framework utilizes a two‐domain preprocessing technique, which combines the use of the frequency domain and color space.

Let the input RGB image be represented as

$$I\left(x,y\right)∊R^{\mathit{\text{MXNX}}3}$$ (1)

where $M$ and $N$ denote spatial dimensions and $x,y$ denote pixel coordinates. The spatial image is transformed into the frequency domain using the FFT

$$F\left(u,v\right)=\sum _{x=0}^{M-1}\sum _{y=0}^{N-1}I\left(x,y\right)e^{-j2\Pi \left(\frac{\mathit{ux}}{M}+\frac{\mathit{vy}}{N}\right)}$$ (2)

where $F\left(u,v\right)$ represents the complex frequency coefficients and $j=\sqrt{-1}$. The magnitude spectrum highlighting spectral energy distribution is obtained as

$$\left|F\left(u,v\right)\right|=\sqrt{\mathit{Re}{\left(F\left(u,v\right)\right)}^2+\mathit{Im}{\left(F\left(u,v\right)\right)}^2}$$ (3)

High‐frequency components correspond to edges, discontinuities, and abrupt intensity transitions, which often arise from cloning operations. To emphasize such artifacts, a weighted spectral representation is computed as follows:

$$\widehat{F}\left(u,v\right)=W\left(u,v\right).\left|F\left(u,v\right)\right|$$ (4)

where $W\left(u,v\right)$ denotes a frequency‐band weighting function that prioritizes tampering‐related spectral regions. The spatial saliency representation highlighting anomalous regions is recovered through the inverse transform

$$S\left(x,y\right)={\mathcal{F}}^{-1}\left(\widehat{F}\left(u,v\right)\right)$$ (5)

where $S\left(x,y\right)$ represents the spectral saliency map, which highlights possible manipulation areas. Although structural inconsistencies are identified through frequency analysis, there is also a possibility of color and illumination inconsistencies caused during manipulation. Hence, the RGB image is converted into HSV.

$$\left(H,S,V\right)=f_{\mathit{HSV}}\left(R,G,B\right)$$ (6)

where $H,S,$ and $V$ represent hue, saturation, and brightness, respectively. To enhance local contrast without amplifying noise, Adaptive Contrast Limited Histogram Equalization (ACLAHE) is applied to the $S$ and $V$ channels. The local histogram of tile $t$ is computed as

$$H_t\left(k\right)=\sum _{\left(x,y\right)∊t}\delta \left(I\left(x,y\right)-k\right)$$ (7)

where k denotes intensity levels and δ (·) is the Dirac delta function. Histogram clipping prevents noise amplification.

$$H_t^{\prime }\left(k\right)=\min \left(H_t\left(k\right),T\right)$$ (8)

where $T$ is the clip limit. The CDF used for intensity remapping is

$$C_t\left(k\right)=\sum _{i=0}^k{H}_t^{\prime }\left(i\right)$$ (9)

The enhanced saturation and brightness components are obtained as follows:

$$S^{\prime }\left(x,y\right)=\frac{C_S\left(S\left(x,y\right)\right)}{L},V^{\prime }\left(x,y\right)=\frac{C_V\left(V\left(x,y\right)\right)}{L}$$ (10)

where $L$ denotes the number of intensity levels. Finally, the enhanced HSV image is reconstructed as,

$$I^{\prime }\left(\mathit{HSV}\right)=\left(H^{\prime },S^{\prime },V^{\prime }\right)$$ (11)

This preprocessing strategy simultaneously highlights structural and chromatic inconsistencies, providing a robust input representation for the subsequent deep learning‐based forgery detection framework.

2.3. YOLOv5 with CSP‐DarkNet backbone

The traditional methods of copy–move forgery detection using sliding window or fixed‐size block matching have high computational costs and may have low accuracy in terms of localization. These methods may not accurately pinpoint the forgery in images with complex textures or small forgery areas, leading to false negatives in forgery detection. Besides, these methods are found to be sensitive to changes in illumination and geometric transformations.

To address these drawbacks, the proposed architecture for the DGY‐CMFD utilizes deep spatial feature extraction and relational reasoning. The architecture is based on YOLOv5 with a CSPDarkNet backbone to obtain multi‐scale spatial and semantic features from input images. Figure 2 shows the architecture of the DGY‐CMFD framework.

FIGURE 2.

Architecture of the Deep Graph‐YOLO copy–move detection network used in DPFIN, showing the integration of convolutional feature extraction, graph convolution layers, and detection modules for identifying forged regions.

Let the input image be defined as follows:

$$I∊R^{\mathit{\text{HXWX}}3}$$ (12)

An initial convolution operation extracts the first‐level feature representation, which can be given as follows:

$$F_0=F_{\mathit{\text{conv}}}\left(I\right)$$ (13)

where $F_{\mathit{\text{conv}}}\left(·\right)$ represents a convolution operation parameterized by a kernel K, stride s, and padding p. This transforms the raw image into the first‐level feature map $F_0\in {ℝ}^{H^{\prime }\times W^{\prime }\times C_0}$. The backbone applies a sequence of convolutional operations ${\mathcal{F}}_{\mathit{\text{conv}}}$ and residual connections $\mathcal{R}$ to model hierarchical feature representations,

$$F_l={\mathcal{R}}_l\left({\mathcal{F}}_{\mathit{\text{conv}}}\left(F_l-1\right)\right),l=1,2,\dots ,L$$ (14)

where $F_l$ is the feature map at layer $l$, The residual mapping is defined as

$${\mathcal{R}}_l\left(X\right)=X+{\mathcal{F}}_{\mathit{\text{conv}}}\left(X\right)$$ (15)

Which allows deeper layers to learn identity mappings, preventing gradient vanishing, and improving the ability of the network to extract complex features. Further, the CSP (Cross Stage Partial) mechanism reduces computational redundancy by splitting the feature map into two paths, where one path goes through dense convolution, and other bypasses the transformation. To enhance sensitivity to tampered regions, a Dynamic Kernel Selection Channel Attention mechanism is incorporated. For a feature map $F\in {ℝ}^{C\times H\times W}$, the attention weight on each channel $c$ is computed as Equation (16)

$${\alpha }_c=\frac{\exp \left(W_c^T·g\left(F_c\right)\right)}{{\sum }_{i=1}^C\exp \left(W_i^T·g\left(F_i\right)\right)},F_c^{\prime }={\alpha }_c·F_c$$ (16)

Here, $g\left(.\right)$ represents a global pooling operation, and $W_c$ is a learnable weight vector. This mechanism dynamically emphasizes channels containing crucial features, which improves the sensitivity to small‐scale or low‐contrast forgery artifacts. To extract multi‐scale spatial context, Spatial Pyramid Dilated (SPD) convolution is employed.

$$F_{\mathit{SPD}}=\mathit{\text{Concat}}\left({\mathcal{F}}_{\mathit{\text{conv}}}^{d_1}\left(F\right),{\mathcal{F}}_{\mathit{\text{conv}}}^{d_2}\left(F\right),\dots ,{\mathcal{F}}_{\mathit{\text{conv}}}^{d_n}\left(F\right)\right)$$ (17)

where $d_i$ represents the varying dilation rates. This allows the network to focus on both minute details and more contextual patterns simultaneously, which is required to find subtle tampering in small regions.

Finally, region of interest (ROI) pooling is employed, which converts a variable‐sized feature map $R\in {ℝ}^{H_r\times W_r\times C}$ into fixed‐size vectors $V_r\in {ℝ}^{P\times P\times C}$ for uniform downstream as in Equation (18):

$$V_r=\mathit{RO}{I}_{\mathit{\text{Pool}}\left(R\right)}=\underset{i,j\in {\mathit{bin}}_{\mathit{pq}}}{\max }{R}_{i,j}$$ (18)

Such an operation maintains spatially discriminative information and standardizes the dimensions of the features for subsequent graph analysis. The CSP‐DarkNet backbone network, attention mechanism, SPD convolution, and ROI pooling operation work together in the process of extracting multi‐scale features and accurately locating subtle copy‐move operations.

2.4. Dual‐domain feature extraction

After pre‐processing and spatial feature extraction, further frequency domain analysis is performed on the improved HSV image, and manipulation‐induced spectral irregularities are further emphasized. The frequency domain representation of the improved image is computed using Equation (19).

$$F_{\mathit{HSV}}\left(u,v\right)=\mathcal{F}\left(I_{\mathit{HSV}}^{\prime }\right)$$ (19)

This dual domain representation combines information from these two domains. Although the features learned from the spatial domain through the YOLOv5 backbone represent structural and semantic information about image regions, features from the frequency domain represent periodic artifacts and inconsistencies introduced during cloning operations.

With the integration of spectral enhancement and contrast amplification from color space, the proposed framework provides a strong feature representation that reveals structural and chromatic anomalies. These enriched feature representations are then used to feed the graph construction and relational reasoning modules to ensure reliable detection and localization of forgery regions from CMF.

2.5. Graph construction for regional relationship modeling

After the preprocessing and the multiple‐scale feature extraction steps, the proposed framework uses a graph structure to capture the structural dependencies between the regions in the image. Let the image $I$ be divided into $N$ visually coherent regions through the superpixel segmentation technique. Then the graph $G_i=\left(V_i,E_i\right)$ is defined, where $V_i$ represents the set of nodes {$v_1,v_2$, …, $v_N$}, and $E_i$ represents the edges describing spatial or feature similarity relationships between regions. Each node $v_k$ is associated with a feature vector $x_k$ composed of multiple descriptors, including pixel intensity $I_k$, texture features $T_k$, and color histogram $H_k$. The overall node feature matrix is represented as,

$$X={\left[x_1,x_2,\dots ,x_N\right]}^T$$ (20)

where $X\in R^{N\times d}$ represents the entire feature representation of the regions. The edges are constructed based on spatial adjacency and feature similarity. The adjacency matrix $A\in R^{N\times N}$ is defined by the Equation (21).

$$A_{\mathit{ij}}=\left\{\begin{array}{cc}\exp \left(-{‖x_i-x_j‖}^2/{\sigma }^2\right),& \text{if}{v}_i~v_j\\ 0& \text{otherwise}\end{array}\right.$$ (21)

where $v_i~v_j$ means neighboring or visually similar regions. The propagation of information among nodes in the graph is conducted through a GCN, in which the node embeddings are updated to enable contextual aggregation among regions for accurate CMF localization.

2.6. Graph convolutional network

After obtaining the regional relationship graph, a GCN is used to extract the contextual dependencies between image regions and detect the inconsistencies caused by copy‐move operations. The node feature matrix is given by $X∊R^{\mathit{NXd}}$, where $N$ is the number of regions and $d$ is the feature dimension of each region, which is extracted from the above dual‐domain representation. The adjacency matrix $A∊R^{\mathit{NXN}}$ is used to represent the similarity relationship between each node. To stabilize training and incorporate self‐connections, a normalized adjacency matrix is computed as follows:

$$\tilde{A}=A+I$$ (22)

where $I$ is the identity matrix. The corresponding degree matrix is given by Equation (23) as,

$${\tilde{D}}_{\mathit{ii}}=\sum _j{\tilde{A}}_{\mathit{ij}}$$ (23)

The graph convolution is performed at each layer $l$ which is formulated as,

$$H^{\left(l+1\right)}=\sigma \left({\tilde{D}}^{-1/2}{\tilde{A}\tilde{D}}^{-1/2}{H}^{\left(l\right)}{W}^{\left(l\right)}\right)$$ (24)

where $H^{\left(l\right)}$ is the node embedding for the $l$‐th layer, and $W^{\left(l\right)}$ is the corresponding weight matrix, while the activation function is given by $\sigma \left(.\right)$. Through the iterative aggregation of neighborhood information, the GCN model is able to transfer the information of related regions, and thus detect the relational inconsistencies between duplicated and original regions, enhancing the robustness of the CMF detection.

2.7. Zebra Optimization Algorithm

Hyperparameter tuning is one of the most important concerns for the optimum performance of the proposed framework. Poor choices of hyperparameters, including excessively high learning rates causing divergence or very low rates that keep the convergence slow, may degrade accuracy. The Zebra Optimization Algorithm (ZOA), for balancing exploration and exploitation, tunes important hyperparameters such as learning rate ($\eta )$, dropout rate ($d)$, $L_2$ regularization ($\lambda )$, and epoch count (e). Minimizing validation loss and enhancing generalization lead the ZOA to accurately identify splicing and copy‐move frauds. ZOA is a bioinspired metaheuristic algorithm imitating the foraging and defense behaviors of zebra herds, showing cooperative motion patterns and predator evasion through collective defense strategies. Each zebra represents one candidate solution, where a set of hyperparameters is given in Equation (25):

$$Z_i=\left[\eta ,d,\lambda ,e\right]$$ (25)

The objective function evaluates each configuration based on validation loss is given as Equation (26):

$$F\left(Z_i\right)=L_{\mathit{val}}\left(Z_i\right)$$ (26)

where $L_{\mathit{val}}$ represents the validation loss function, commonly computed as cross‐entropy loss in the context of classification tasks. Lower values of $F\left(Z_i\right)$ signify better‐performing hyperparameter sets. The optimization procedure in ZOA consists of two primary phases: foraging and defense, which correspond to exploration and exploitation, respectively. In the foraging phase, zebras move toward promising regions in the search space, guided by the pioneer zebra, which denotes the current best solution. The update rule is expressed in Equation (27) and (28).

$$x_{i,j}^{\mathit{new},P1}=x_{i,j}+r·\left(P{Z}_j-I·x_{i,j}\right)$$ (27)

$$X_i=\left\{\begin{array}{c}{X}_i^{\mathit{new},P1},F_i^{\mathit{new},P1}<F_i\\ X_i,\text{else}\end{array}\right.$$ (28)

where $x_{i,j}$ is the location that $i^{\mathit{th}}$ zebra in the $j^{\mathit{th}}$ dimension (corresponding to a hyperparameter), $P{Z}_j$ represents the best hyperparameter value found so far in dimension $j$, $r\in \left[0,1\right]$ is a uniformly scattered number generator controlling stochasticity, and $\mathrm{I}=\text{round}\left(1+\mathrm{rand}\right)$, where $\mathrm{rand}$ is an unknown number within the range [0; 1] $I\in \left\{1,2\right\}$ adjusts the step scaling factor. where $X_i^{\mathit{new},P1}$ is the new status of the $i^{\mathit{th}}$ zebra based on the first phase, $x_{i,j}^{\mathit{new},P1}$ is its $j^{\mathit{th}}$ dimension value, and $F_i^{\mathit{new},P1}$ is its objective function value. This mechanism ensures that the population of zebras converges gradually toward high‐quality solutions while preserving diversity in candidate solutions.

The defense phase simulates predator–prey interactions, where the zebras either scatter to evade the predator (diversification of search) or cluster in defense, intensifying near‐promising solutions. The mathematical formulation of this phase is given below as Equation (29) and (30):

$$x_{i,j}^{\mathit{new},P2}=\left\{\begin{array}{c}{x}_{i,j}+R·\left(2r-1\right)\left(1-\frac{t}{T}\right)·x_{i,j},P_s\le 0.5\\ x_{i,j}+r·\left(A{Z}_j-I·x_{i,j}\right),\text{otherwise}\end{array}\right.$$ (29)

$$X_i=\left\{\begin{array}{c}{X}_i^{\mathit{new},P2},F_i^{\mathit{new},P2}<F_i\\ X_i,\text{else}\end{array}\right.$$ (30)

where $X_i^{\mathit{new},P2}$ corresponds to ith zebra's new status depending on the following phase. $x_{i,j}^{\mathit{new},P2}$ is its $j^{\mathit{th}}$ dimension value, $F_i^{\mathit{new},P2}$ is its objective function value, $R$ is a scaling coefficient, $t$ is the current iteration, $T$ is the highest amount of repetitions, $A{Z}_j$ denotes the hyperparameter position of a neighboring zebra, and $P_s$ is the probability of choosing between exploration (escape) and exploitation (defense). The term $\left(1-\frac{t}{T}\right)$ balances the exploration at the initial phases and exploitation in the later stages of optimization, refining the search toward the optimal hyperparameter regions.

Iteratively, over successive iterations, ZOA continuously assesses each candidate configuration, updates the positions of zebras using the above‐mentioned rules, and traces the best‐ever configuration across the population. Mathematically, the global best solution can be expressed as follows:

$$Z^{*}=\arg \underset{Z_i\in \mathcal{P}}{\min }F\left(Z_i\right)$$ (31)

where $\mathcal{P}$ indicates the entire population of zebras. This iterative process converges closer to a near‐optimal hyperparameter set (${\eta }^{*},d^{*},{\lambda }^{*},e^{*}$), which guarantees stable training with reduced overfitting and fast convergence and also yields improved classification accuracy in detecting both copy‐move and splicing forgeries. The Zebra Optimization Algorithm is explained in a systematic way using a straightforward method: the parameters inherently change from exploitation to exploration based on heuristics to increase performance, and description of the Deep Graph YOLO Copy‐Move Detection Network (Algorithm 1).

ALGORITHM 1. Zebra Optimization Algorithm (ZOA) for hyperparameter tuning.

1. Input: Population Size N (number of zebras), Maximum Iterations T, Hyper‐parameter dimensions D (i.e., learning rate, dropout, $L_2$ regularization), objective function $F\left(Z_i\right)$  Output: Optimal Hyperparameter set $Z^{*}$

2. Initialization:

Randomly generate initial zebra population $X=\left\{Z_1,Z_2,\dots \dots ,Z_N\right\}$, where each zebra $Z_i=\left[x_{i,1},x_{i,2},\dots \dots ,x_{i,D}\right]$, evaluate objective function using Equation (24) for each zebra and identify pioneer zebra with best fitness

3. Repeat for $t=1\mathit{\text{to}}T$ (iterations)

Foraging Phase (Exploration & Leader Guidance):

For each zebra $Z_i$ in the population

Update position for each dimension j using Equation (28)

Evaluate new fitness using Equation (29)

Defense Phase (Exploitation and Predator‐Avoidance):

For each zebra $Z_i$ in the population

Generate probability

Update position based on predator type using Equation (30)

Evaluate new fitness using Equation (31)

4. Update Pioneer Zebra:

Identify the zebra with the best fitness in the population using Equation (32)

5. Termination:

If t = T or convergence criteria met, stop

Return the best hyperparameter set $Z^{*}$

In summary, hyperparameter tuning using ZOA balances exploration in the search space with focused exploitation of the most promising regions to avoid convergence to a local minimum. ZOA then adaptively forages and defends an optimal combination of learning rate, dropout, regularization, and epoch count. The optimized configuration further strengthens the robustness of the proposed Deep Graph YOLO framework to accurately detect tiny tampering artifacts in a variety of difficult forgery scenarios.

2.8. Output classification

The final classification step aims at determining whether an image region is genuine or forged based on the learned graph embeddings. Let $H^{\left(L\right)}\in R^{Nxd}$ represent the node representations of the last GCN layer, where $N$ represents the total number of regions and $d$ represents the dimension of the features. These graph embeddings are used in a fully connected network to obtain the prediction scores as follows:

$$Z=H^{\left(L\right)}{W}_c+b_c$$ (32)

where $W_c$ and $b_c$ represent the classification weights and bias. The probability of forgery is computed using the softmax function.

$$P\left(y_i|z_i\right)=\frac{e^{z_i}}{{\sum }_{k=1}^K{e}^{z_k}}$$ (33)

Here, $K$ represents the total number of classes. Regions with high probability are marked as forged regions in the image.

The proposed DPFIN framework ultimately results in a binary classification output that indicates whether a given input image is a victim of CMF while simultaneously localizing the forgery. The proposed approach is able to effectively capture inconsistencies as well as contextual relationships between different image regions by incorporating dual‐domain features, YOLOv5‐based spatial features, a CSPDarkNet backbone, and relational reasoning using a graph convolutional layer. The experimental setup and results are provided in the next section to show the efficiency of the proposed approach by comparing its performance.

3. RESULTS

In this section, the experimental results for the proposed DPFIN are presented based on the evaluation of the CoMoFoD and CASIA v2 datasets. The analysis discusses the dataset description, simulation environment, and experiment setup for the evaluation of the proposed CMFD framework. The performance of the proposed framework is measured using well‐accepted evaluation metrics for the evaluation of a forensic system. The evaluation metrics for the proposed system are Accuracy, Precision, Recall, F1‐Score, Dice Score, Intersection over Union (IoU), True Positive Rate (TPR), False Positive Rate (FPR), and Area Under the Receiver Operating Characteristic Curve (AUC).

All experiments are carried out using a common processing pipeline that involves image preprocessing, feature extraction, and region‐level matching. The proposed DPFIN framework is compared against various conventional methods and deep learning‐based baseline methods for performance evaluation of the proposed framework. The proposed framework integrates frequency domain enhancement, multi‐scale spatial feature extraction, and relational modeling using a graph for effective detection of subtle copy‐move manipulations, which are difficult for conventional methods to detect. The experimental outcomes demonstrate that the proposed framework is effective in terms of detection accuracy, localization of manipulated regions, and generalization ability for various forgery types. The experimental outcomes demonstrate the practical viability of the proposed DPFIN framework for digital image forensic analysis.

3.1. Experimental setup and environment

All the experiments were performed on a workstation with an NVIDIA GeForce RTX 3090 GPU (24 GB VRAM), an Intel Core i9 CPU, and 64 GB RAM to perform both training and inference tasks for the models. The implementation of the proposed DPFIN framework was done using the Python 3.10 programming language with the TensorFlow/Keras deep learning framework. The GPU was also enabled for the implementation using CUDA 11.8 and cuDNN 8.6 libraries. Some other libraries used for the implementation of the experimental process are NumPy, OpenCV‐Python, and Scikit‐learn for data preprocessing, image processing, etc. The training and inference experiments were performed on the same environment to measure the computational performance metrics of the models.

The datasets were split using the 80:10:10 ratio for training, validation, and testing, which was done individually for the datasets with stratified sampling to avoid any data leakage between the sets. The hyperparameters such as learning rate, batch size, and dropout rate were tuned using the ZOA, and its training details are shown in Table 1. The network was trained for 25 epochs, which took a total of 3.5 h on the RTX 3090 GPU. The details of the implementation, along with the source code, will be made available for public access upon the publication of the manuscript.

TABLE 1.

Optimized simulation parameters after ZOA hyperparameter tuning.

Parameter	Value
Dataset split	80:10:10
Optimizer	Adam
Batch size	16
Learning rate	0.0001
Loss function	BCE with logits loss
Number of epochs	25
Activation function	ReLU
Dropout rate	0.3
Metric for evaluation	Accuracy, precision, recall, F1‐score, FPR, TPR
Model architecture	DFIN

3.2. Dataset description

The CoMoFoD and CASIA v2 are the selected datasets, which include a wide range of post‐processing operations like JPEG compression, noise addition, rotation, etc. These operations help in testing the robustness of DPFIN effectively.

The CoMoFoD dataset is widely utilized for benchmarking detection of copy‐move forgery methodologies and consists of 260 image sets, divided into small (512 × 512) and large (3000 × 2000) categories. Each set consists of an original image, a forged image, a colored mask showing manipulated regions, and a binary mask for evaluation. Five different transformations are employed to create the forged images: scaling, rotation, movement, distortion, and combination. Furthermore, post‐processing techniques like JPEG compression, addition of noise, blurring, brightness adjustment, color reduction, and contrast adjustment were done on all the images. Images in this dataset have a specific naming convention to identify the type and transformations applied. This gives an extensive and challenging benchmark that discusses techniques to identify copy‐move forgeries. The CoMoFoD dataset is accessible to the public and may be accessed from: (https://www.kaggle.com/datasets/tusharchauhan1898/comofod).

CASIA v2.0 (CASIA Image Tampering Detection Evaluation Database), which has been widely used to evaluate CMFD methods. A total of 12,614 color images (7491 authentic and 5123 tampered samples in JPEG format) are included, with a typical resolution of 320 × 240 to 800 × 600 pixels. The tampered images are then created using manipulation techniques such as copy‐move, splicing, and object removal, usually refined by post‐processing operations like blurring, smoothing, and compression in order to create the editing realistic. To identify copied and moved forgeries, CASIA v2.0 provides diverse and challenging examples where image regions have been duplicated inside a similar picture to hide or duplicate the information. Variability in manipulation complexity and image quality makes this dataset a strong standard for testing the accuracy, robustness, and generalization capability of proposed CMFD algorithms. This dataset can be accessed from (https://www.kaggle.com/datasets/divg07/casia‐20‐image‐tampering‐detection‐dataset/data).

3.3. Evaluation metrics

Accuracy, precision, recall, F1‐score, AUC, Dice coefficient, IoU, TPR, and FPR are used to support the effectiveness of the suggested CMFD model. The details of these performance measurements and their relevance to assessing the model's effectiveness are presented in the following section.

$$\text{Accuracy}=\frac{\mathit{TN}+\mathit{TP}}{\mathit{TP}+\mathit{TN}+\mathit{FP}+\mathit{FN}}$$ (34)

$$\text{Precision}=\frac{\mathit{TP}}{\mathit{TP}+\mathit{FP}}$$ (35)

$$\text{Recall}=\frac{\mathit{TP}}{\mathit{TP}+\mathit{FN}}$$ (36)

$$F1-\text{Score}=2\times \frac{\text{precision}\times \text{recall}}{\text{precision}+\text{recall}}$$ (37)

$$\mathit{TPR}=\frac{\mathit{TP}}{\mathit{TP}+\mathit{FN}}$$ (38)

$$\mathit{FPR}=\frac{\mathit{FP}}{\mathit{FP}+\mathit{TN}}$$ (39)

$$\mathrm{IoU}=\frac{\mid P\cap G\mid }{\mid P\cup G\mid }$$ (40)

$$\text{Dice coefficient}=\frac{2·\mid P\cap G\mid }{\mid P\mid +\mid G\mid }$$ (41)

where P∩G represents the quantity of accurately forecasted pixels as forged (intersection), P∪G is the total count of forged or actually forged pixels predicted (union), ∣P∣ is the total quantity of pixels anticipated as forged, and ∣G∣ denotes the total quantity of real forged pictures.

3.4. Results obtained from the proposed method

To ensure reproducibility and robustness of the proposed DPFIN, a systematic dataset preprocessing/splitting strategy was employed for both CoMoFoD and CASIA v2 datasets. Firstly, all images are resized to a uniform spatial resolution to ensure consistent input dimensions for the YOLOv5‐based feature extraction module. Before training the model, a dual‐phase preprocessing pipeline is employed to enhance forgery‐related patterns. In the first phase, FFT analysis is employed to detect frequency‐domain inconsistencies that are likely to appear due to copy‐move manipulations. In the second phase, Adaptive Contrast Limited Adaptive Histogram Equalization (ACLAHE) is employed to enhance low‐contrast regions of the image, thereby improving the visibility of duplicated textures. The preprocessed output, along with the original image and its respective ground‐truth mask, constitutes the input sample as depicted in Figure 3 for the CoMoFoD dataset and Figure 4 for the CASIA v2 dataset.

FIGURE 3.

Qualitative detection results on the CoMoFoD dataset showing (A) original image, (B) preprocessed image, (C) ground‐truth mask, and (D) predicted forgery mask generated by the proposed DPFIN framework.

FIGURE 4.

Forgery localization results on the CASIA v2 dataset illustrating (A) original image, (B) ground‐truth forgery annotation, (C) forged image example, and (D) predicted forgery regions detected by the proposed method.

To enhance the model's generalization capacity and avoid overfitting, data augmentation strategies such as horizontal and vertical flipping, as well as random rotations and brightness changes, were carried out. Data partitioning strategies were adopted as in Table 1 to separate the data into the training set, validation set, and testing set. This ensures a balanced distribution of the forged and original data among the data partitions. This data preprocessing pipeline allows the DPFIN model to learn discriminative spatial and frequency features.

ROC curve shown in Figure 5 illustrates the capacity of the suggested DPFIN to differentiate forged regions from genuine ones. The curve rises steeply from (0.0,0.0) to (0.0,1.0) and then goes horizontally to (1.0,1.0), which shows that the TPR is close to 1.0 while maintaining a very low FPR. With an AUC value of 0.9999, the model shows outstanding discriminative capability in accurately detecting subtle forgeries while keeping misclassifications of authentic regions at a minimum.

FIGURE 5.

ROC curve for the CoMoFoD dataset demonstrating the discrimination capability of the proposed DPFIN framework for forged and authentic regions.

Figure 6 depicts the ROC curve that represents the execution of the suggested DPFIN on the CASIA v2 collection. It is evident from the figure that the method produces an extremely high AUC value of 0.9999, which means it can differentiate between real and fake images almost perfectly. The curve also stays in the top‐left corner in most parts, indicating a very high TPR with minimal false alarm cases. This excellent performance verifies that multi‐scale feature extraction based on YOLOv5 and modeling with a GCN enhance feature representation, spatial consistency, and precise forgery localization.

FIGURE 6.

ROC curve for the CASIA v2 dataset illustrating the detection performance of the proposed framework and its ability to distinguish forged regions under diverse image conditions.

Figure 7 presents the progress in performance achieved by the proposed CMFD model, with increased training epochs. Indeed, the plotted metrics of Accuracy, F1‐Score, IoU, Dice coefficient, Precision, and Recall show an upward trend with a rise in the number of time periods, reflecting stable learning and confluence. The Accuracy rapidly reaches close to 100%, indicating the good generalization capability of the proposed model. The F1‐Score and Dice coefficient also retain high balanced values, showing that the proposed model has the ability to detect forged regions without bias to any class. Further confirmation of precise localization of manipulated areas within the images is denoted by increased IoU values.

FIGURE 7.

Training performance trends across epochs for the CoMoFoD dataset, showing the evolution of accuracy, loss, and related evaluation metrics during model convergence.

The quantitative results of the proposed DPFIN framework are presented in Table 2 to show the strong capability of the model for accurate CMFD. The accuracy of the model is found to be 0.9989, which is near perfection. The model is able to correctly classify the regions of the images as genuine or forged. The precision value of 0.9800 and recall value of 0.9979 show the capability of the model to correctly identify the forged regions of the images. The model is also able to achieve a low rate of false detection. The F1‐score value of 0.9889 and Dice value of 0.9889 show the balanced performance of the model in terms of precision and recall. This is a clear indication of the reliability of the model in terms of accurate segmentation of the forged regions. The Intersection over Union value of 0.9780 shows the high level of overlap between the model output and the actual regions. Furthermore, the extremely low loss value of 0.0026 proves that the model converges efficiently during the training phase. Moreover, the evaluation metrics indicate that the model has a high value of TPR at 0.9979 along with a low FPR of 0.0021.

TABLE 2.

Quantitative performance metrics of the proposed CMFD framework.

Performance	Values
Accuracy	0.9989
Precision	0.9800
Recall	0.9979
F1‐Score	0.9889
Dice coefficient	0.9889
IoU	0.9780
Loss	0.0026
TPR	0.9979
FPR	0.0021

The performance of the proposed DPFIN framework was further tested on the CASIA v2 dataset to evaluate the effectiveness of the proposed framework for classification and segmentation tasks. As shown in Figure 8, the proposed framework has shown promising results by obtaining a high accuracy rate of 0.997 along with a high F1‐score of 0.986. This clearly indicates that the proposed framework has shown high detection capabilities. Moreover, the high value of precision (0.980) and recall (0.992) clearly indicate that the proposed framework has shown high capabilities for identifying the forged regions without any false alarms. In addition, the high value of the Dice coefficient (0.986) along with a high value of IoU (0.972) clearly indicates that there is a high degree of overlap between the predicted masks and the corresponding ground truth. This clearly indicates that the proposed framework has shown high capabilities for identifying the tampered regions. Therefore, from the obtained results, it is clear that the proposed DPFIN framework has shown high capabilities for handling the diverse manipulation patterns present in the CASIA v2 dataset for digital image forensic analysis.

FIGURE 8.

Quantitative performance evaluation of the proposed DPFIN framework on the CASIA v2 dataset using metrics such as accuracy, precision, recall, F1‐score, Dice coefficient, and IoU.

3.5. Comparative analysis

The following section describes the comparative analysis of various techniques used for CMFD analysis, with F1‐score being a major parameter, which accounts for the overall reliability of the detection techniques. Various techniques are compared with the proposed DPFIN framework, which includes the Robust Copy Move Forgery Detection (RCMFD) technique [25], the Copy–Move Forgery Detection using Two‐Stage Filtering (CMFD‐TSF) technique [25], the Copy–Move Forgery Detection Using Ciratefi (CCMFD) technique [25] Keypoint Copy Move Forgery Detection (KCMFD) technique [26], and the Center Surround Extrema Keypoint detector‐based Convolutional Neural Network (CenSurE KD‐CNN) technique [26]. The comparison of the techniques is carried out to determine the effectiveness of the techniques for detecting the forged regions without incorrectly identifying the genuine regions. The analysis focuses on the TPR as well as the reliability of the techniques for the overall improvement achieved by the proposed DPFIN framework for the detection of copy‐move forgeries.

Figure 9 visualizes the F1‐Score (%) across six detection of fraud methods. The suggested DPFIN attains the greatest F1‐Score achievement at 98.89%, followed by top baselines CenSure KD‐CNN (98.56%) and KCMFD (98.43%). For mid‐performance, RCMFD reached 93.75% and CCMFD attained 90.0%, while CMFD‐TSF had the lowest F1‐Score of 82.0%, proving this technique to be rather inefficient. This proves that DPFIN has the cutting edge performance in detecting forged regions with minimum false positives.

FIGURE 9.

Comparative F1‐score analysis of the proposed DPFIN method and existing CMFD approaches on the CoMoFoD dataset, highlighting improvements in detection accuracy and robustness.

Table 3 summarizes the six methods' performances on Recognition of copying and copying forgeries by comparing the TPR, FPR, and F1‐Score (%). The proposed DPFIN yields the best detection capability, at a TPR of 100%, an FPR of 0.21%, and the highest F1‐Score of 98.89%. Other strong baseline methods like CenSure KD‐CNN have 98.56%, followed by KCMFD with 98.43%, RCMFD with 93.75%, and CCMFD with 90.0%, while CMFD‐TSF has the lowest F1‐Score, at 82.0%, which shows the insufficiency in balancing precision and recall. Overall, DPFIN outperforms existing techniques with a good trade‐off between the detection of forged regions and a low misclassification rate of authentic areas.

TABLE 3.

Comparative performance metrics of CMFD methods.

Methods	TPR (%)	FPR (%)	F1 (%)
KCMFD	–	–	98.43
CenSure KD‐CNN	–	–	98.56
RCMFD	93.75	7.25	93.75
CMFD‐TSF	94	–	82
CCMFD	92.98	1.0	90
Proposed DFIN	100	0.21	98.89

The CASIA v2 dataset was used for a comparative analysis that evaluated several forgery detection models such as Progressive Spatio‐Channel Correlation Network [27, 28], Progressively‐Refined Neural Network [29], Enhanced Multiple Tampering Traces Network [30], Dual Supervision Neural Network [31], RGB‐Network (RGB‐N) [31], Spatial Pyramid Attention Network (SPAN) [31], and Southeast University (SEU)‐U‐Net (SEU‐U‐Net) [31], Attention and multi‐scale learning have been incorporated into these methods to improve the localization of tampering. In contrast, the proposed DPFIN framework, which integrates YOLOv5‐based multi‐scale feature extraction and graph convolutional modeling, was compared with them, exhibiting superior accuracy and robustness.

Table 4 contrasts the suggested DPFIN framework with various innovative forgery detection models on the CASIA v2 dataset in terms of AUC and F1‐score. The conventional methods‐RGB‐N (79.5%, 40.8%), SPAN (83.8%, 38.2%), PSCCNet (87.5%, 55.4%), SEU‐U‐Net (87.3%, 50.9%), PR‐Net (0, 55.1%), EMT‐Net (85.6%, 45.9%), and DS‐Net (86.4%, 56.0%) have performed reasonably, but their capability for complex tampering patterns is limited. On the contrary, the suggested DPFIN paradigm performs noticeably better than the earlier cutting‐edge techniques. With an AUC of 99.99% and a high F1‐score of 96.8%, which indicates stronger discriminative power and more accurate localization of forged regions. Such a large improvement proves the efficiency of the proposed integration of YOLOv5‐based multi‐scale feature extraction and graph convolutional modeling, enabling the enhanced capturing of intrinsic spatial dependencies and boosting the forgery detection precision.

TABLE 4.

Comparative performance metrics of CMFD methods for the Casia V2 dataset.

Methods	AUC (%)	F1 (%)
RGB‐N	79.5	40.8
SPAN	83.8	38.2
PSCCNET	87.5	55.4
SEU‐U‐NET	87.3	50.9
PR‐Net	–	55.1
EMT‐Net	85.6	45.9
DS‐Net	86.4	56.0
Proposed DFIN	99.99	96.8

Table 5 shows a comparative analysis of the computational efficiency and statistical reliability of the proposed DPFIN with various state‐of‐the‐art CMFD techniques. From the table, we can see that our proposed framework has achieved the highest accuracy among all techniques with 99.89% accuracy and a small confidence interval (95%) of ±0.06. On the other hand, classical feature‐based techniques such as SURF‐HOG‐RGMS and Zernike Moments [25] show a high processing time of 32.76 s due to computationally expensive feature matching techniques. Similarly, deep learning‐based techniques such as the CNN‐CenSurE hybrid model [26] and ECMVFD‐FTLTDO [32] achieve high accuracy but are computationally expensive in terms of parameter sizes and inference latency. Although our proposed framework has achieved high accuracy with fast inference time through the D2PRL framework [33], its accuracy is comparatively much lower. The proposed framework has achieved a balanced trade‐off between model complexity (~27.6 M parameters), computational cost (~16.2 GFLOPs), and inference time (~0.035 s), showing high computational efficiency while achieving statistically significant performance over existing techniques.

TABLE 5.

Comparative analysis of computational efficiency and statistical significance on the CASIA v2 dataset.

Proposed method	Accuracy (%)	95% CI (accuracy)	Processing/inference time (s)	Parameters (M)	GFLOPs
SURF‐HOG + RGMS + Zernike Moments	93.75	±0.9	32.76	–	–
CNN‐CenSurE Two‐Stage Hybrid	93.3	±0.8	15.17	~138.0	~15.5
ECMVFD‐FTLTDO	95.26	±0.7	13.5	~45.2	~4.8
Multiscale Feature Fusion + CNN Segmentation	94.6	±0.7	4.5	~28.5	~3.2
D2PRL	83.2	±1.4	~0.048	~31.2	~52.4
Proposed DPFIN	99.89	±0.06	~0.035	~27.6	~16.2

3.6. Ablation study

The ablation analysis aimed at assessing the individual and combined contributions of the major components within the proposed DPFIN framework based on two datasets is shown in Table 6. The baseline configuration is a single configuration that only includes spatial feature extraction by using YOLOv5. The results show that it has moderate detection capability. The accuracy is reasonable but has some limitations in detecting subtle duplication forgery and texture similarity. The configuration that includes frequency preprocessing by using FFT and ACLAHE enhancement shows improved detection capability compared to the baseline configuration. The results show that frequency features are complementary to spatial features for forgery detection, especially for low‐contrast forgery areas created by post‐processing operations. The configuration that includes graph convolutional modeling with spatial features shows improved detection capability compared to other configurations. The results show that relational learning within duplicated areas is improved by using graph convolutional modeling. However, it shows suboptimal detection capability compared to the full configuration. The full configuration of DPFIN shows the best accuracy and F1‐score compared to other configurations for both datasets. The results show that the proposed DPFIN framework is effective for reliable detection of copy‐move forgery.

TABLE 6.

Cross‐dataset ablation analysis of DPFIN architectural variants.

Model variant	Architecture components	CoMoFoD dataset			CASIA v2 dataset
		Acc (%)	F1 (%)	AUC (%)	Acc (%)	F1 (%)	AUC (%)
Baseline (spatial only)	YOLOv5 spatial features	97.8	95.4	96.2	95.8	84.3	88.3
Spatial + frequency (no fusion/GCN)	YOLOv5 + FFT + ACLAHE	98.7	97.1	97.6	97.8	91.8	94.6
Spatial + GCN (no frequency)	YOLOv5 + graph convolution	98.2	96.5	97.1	96.9	88.5	91.4
Full DPFIN (spatial + frequency + GCN + fusion + ZOA)	YOLOv5 + FFT + ACLAHE + GCN + fusion + ZOA	99.89	98.89	99.4	99.99	96.8	99.1

Abbreviations: Acc, accuracy; F1, F1‐score.

4. DISCUSSION

The results obtained from the experimental analysis clearly indicate that the proposed DPFIN outperforms all existing CMFD methods on both the CoMoFoD dataset and CASIA v2 dataset. This improvement in performance is due to the synergy that occurs from the integration of different feature representations in the proposed architecture. For instance, the dual‐domain preprocessing stage that utilizes FFT and ACLAHE improves the detectability of subtle copy‐move forgeries that often get masked by image compression or noise. In most existing CNN‐based methods, the focus is on intensity patterns in the image, which may not be sufficient for detecting copy‐move forgeries since these methods often fail to capture high‐frequency inconsistencies. The addition of frequency‐domain information helps in preserving the feature information corresponding to boundary transitions and texture repetitions. The other important element that plays an important role in the enhanced performance is the GCN‐based relational modeling. Compared to other keypoint‐based approaches like SURF‐HOG, which are based on independent feature matching, the GCN module is capable of modeling the spatial dependencies between candidate regions. This helps to ensure structural consistency for duplicated regions while suppressing false positives for homogeneous textures or backgrounds. Additionally, the ZOA used for hyperparameter tuning helps to stabilize the learning process while optimizing the search space for the optimal parameters, hence improving the F1‐score compared to other recent CNN‐based baselines.

Despite the good performance achieved, some limitations are still associated with the model. The detection capability may not perform well on images that have undergone severe JPEG compression or have undergone severe post‐processing operations. Such operations may destroy the high frequency components necessary to perform the FFT preprocessing. Additionally, the model may fail to detect very small regions of forgery due to the resolution issues associated with the feature aggregation layers. From an operational point of view, DPFIN imposes a moderate computational burden due to the combination of the FFT‐based preprocessing and GCN modeling. While this makes the inference process more complex compared to lightweight CNN‐based solutions, the payoff is in terms of the robustness and accuracy gained in detection and localization. The low false positives obtained in the experiments are also important in digital forensics since minimizing false accusations is critical. The localization capability can help in the verification process by reducing the search space. The possible issues related to unknown image acquisition pipelines, social media compression, and cameras are also recognized, but future work is planned to address these issues for increased robustness under heterogeneous real‐world scenarios. Moreover, other possible directions for future work include the development of lightweight versions of DPFIN for environments with scarce resources, the extension of the framework to detect video forgeries, and the addition of mechanisms for explainable artificial intelligence to provide explainable evidence.

5. CONCLUSION

This study proposed a DPFIN, which is an effective CMFD model that combines dual‐domain preprocessing techniques and multi‐scale spatial feature extraction and graph relational modeling. The model combines the strengths of both the Fourier Transform and ACLAHE techniques in the dual‐domain preprocessing method and the YOLOv5 model in the multi‐scale spatial feature extraction method. The model also employs a graph convolutional neural network for the relational modeling method. The study evaluated the performance of the model using the CoMoFoD and CASIA v2 datasets and obtained a high detection rate of 99.89%, an F1‐score of 98.89%, and an AUC of 0.9999. The model also obtained a very low FPR of 0.21%. The model also has a low inference time of 0.035 s per image and a high number of parameters (~27.6 M) and a high number of FLOPs (~16.2GFLOPs), indicating that the model is efficient in terms of its computational complexity. The study concluded that the DPFIN model is a reliable model that can be effectively used by digital forensic experts to identify forged regions in images and verify evidence. Thus, the proposed DPFIN framework is seen to be a reliable tool for forensic analysts to detect tampered regions. Future work will be to improve the robustness to severe compression artifacts, extend the proposed approach to video forgery detection, and apply explainable AI to improve the interpretability of the proposed approach for forensic purposes.

CONFLICT OF INTEREST STATEMENT

The authors have no conflicts of interest to declare.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.