A multi-scale cross-domain interaction network for image tampering localization
By constructing the multi-scale cross-domain bidirectional interactive network MSCDI-Net, the problem of balancing semantic generalization and localization accuracy in existing methods is solved. Robust generalization detection of unknown objects and tampering types is achieved, improving the accuracy and robustness of image tampering localization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGSHA UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-01-22
- Publication Date
- 2026-06-09
AI Technical Summary
Existing image tampering localization methods lose key context by suppressing semantic information, resulting in limited generalization ability and difficulty in achieving accurate tampering detection in complex scenes.
We construct a multi-scale, cross-domain, bidirectional interactive network MSCDI-Net. By establishing a bidirectional cross-attention and adaptive gating fusion mechanism between different feature levels and spatial and noise domains, high-level semantic information guides the extraction of low-level evidence-gathering features, and low-level tampering artifacts correct high-level semantic understanding, forming a closed-loop learning paradigm in which semantics and evidence-gathering clues reinforce each other.
It significantly reduces the false detection rate and false negative rate in tamper detection in complex scenes and on unseen objects, and improves the accuracy, robustness and generalization ability of image tamper localization.
Smart Images
Figure CN122176048A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to computer vision technology and image segmentation, and more specifically, to a multi-scale cross-domain interactive network for image tampering localization. Background Technology
[0002] Image tampering localization (IML) faces a fundamental paradox in digital forensics: tampering operations are inherently semantic objects (such as adding, removing, or modifying specific entities), while effective detection models must exhibit strong generalization capabilities to unknown semantic categories outside the training distribution. This dichotomy requires a delicate balance—extracting tampering patterns intrinsically related to the semantic context while maintaining the semantically independent representations needed for cross-category generalization.
[0003] Traditional IML methods primarily rely on manually designed low-level clues or physical consistency analysis to model general forensic fingerprints that are independent of image content. While theoretically possessing strong generalization capabilities, these methods overlook the fact that tampering traces often depend on the semantic structure of the tampered region, thus easily leading to a high false detection rate in complex scenarios.
[0004] Some methods utilize pre-trained backbone networks to extract features containing high-level semantic information. While these methods perform well on in-distribution data, their generalization ability is highly dependent on the semantic distribution of the training data. This dependence makes the model prone to "semantic shortcuts," i.e., memorizing spurious relevances such as "which objects are usually tampered with," rather than learning universally applicable tampering patterns. Ultimately, this semantic overfitting manifests as false positives for real targets and false negatives for unseen targets, such as... Figure 1 As shown.
[0005] To address the generalization dilemma, recent studies such as MVSS, Sparse-ViT, and Mesorch have attempted to extract semantically irrelevant features for IML. MVSS-Net captures content-irrelevant inconsistencies through multi-view learning from both the noise and edge domains; SparseViT structurally breaks semantic connections through sparse attention; and Mesorch separates meso-scale tampering traces from macro-level semantics. While these methods have achieved some improvements, their strategies of suppressing, bypassing, or filtering semantics inadvertently discard crucial contextual cues for accurate localization, such as… Figure 2 As shown. This limitation prompts us to rethink this problem. Compared to filtering semantics, a more promising direction lies in promoting full interaction between semantic information and low-level forensic clues, thereby dynamically learning standardized tampering patterns that are not constrained by specific semantics. Based on this insight, this invention proposes a cross-scale, cross-domain collaborative framework to construct a closed-loop paradigm in which semantic guidance and forensic forgeries mutually reinforce each other.
[0006] Based on the above analysis, this invention proposes a Multi-Scale Cross-Domain Bidirectional Interaction Network (MSCDI-Net) for image tampering localization. This network establishes a hierarchical interaction mechanism between different feature levels and domains through bidirectional cross-attention and adaptive gating fusion. High-level semantics provides attention priors for low-level feature extraction, guiding the model to focus on semantically relevant regions; low-level features act as "forensic experts," feeding back pixel-level inconsistencies to refine and correct the high-level semantic understanding. Through iterative interactions across levels and domains, MSCDI-Net can learn semantically guided tampering patterns with good generalization ability, maintaining robustness to unseen objects and tampering types. Summary of the Invention
[0007] Objective: To address the problem that existing image tampering localization methods lose key contextual information by suppressing semantics, thus limiting their generalization ability, this invention proposes a multi-scale, cross-domain, bidirectional interactive network, MSCDI-Net. This network constructs a bidirectional cross-attention and adaptive gating fusion mechanism across different feature levels and spatial and noise domains. This allows high-level semantic information to guide low-level forensic feature extraction, while low-level tampering artifacts correct high-level semantic understanding, forming a closed-loop learning paradigm where semantics and forensic clues mutually reinforce each other. This enables robust generalization detection of unknown objects and tampering types.
[0008] The objective of this invention can be achieved through the following technical solutions:
[0009] A multi-scale, cross-domain interactive network for image tampering localization. It includes the following steps:
[0010] Step 1: Construct a dual-branch network structure consisting of an image domain branch and a noise domain branch. Use a cross-domain fusion module to fuse the features from the image and noise domain branches. Input the fused features into the decoder to obtain the final prediction mask. The model is defined as follows: (1);
[0011] in, Indicates the input image. and These represent the image encoder and the noise encoder, respectively. This indicates a cross-domain fusion module. Indicates decoder, This represents the predicted tampering mask.
[0012] Step 2: Based on Step 1, the image encoder extracts features at four different scales from the tampered image and introduces multiple Global-Local Context Aggregation (GL-CA) modules into the high-level features to enhance contextual reasoning ability, defined as Equation (2). (2);
[0013] To efficiently capture both global context information and local tamper artifacts simultaneously, the GL-CA module selectively applies only to high-level features. and This design is based on the fact that these stages have a larger receptive field and richer semantic information, which allows local forgery traces to be more effectively integrated into a unified global understanding.
[0014] Step 3: Based on Step 2, each GL-CA module consists of parallel global attention branches and local attention branches. The global branch retains the original self-attention mechanism and is used to model long-range dependencies, thereby obtaining... At the same time, the local branches divide the feature map into several non-overlapping regions. Window, and perform self-attention operations within each window to obtain This is to capture fine-grained local tampering artifacts. Subsequently, the outputs of the two branches are adaptively fused using a gating mechanism. (3); (4); in, This represents a multilayer perceptron. This represents the Sigmoid activation function. The learned gating weights are used to balance the contributions of global context and local details.
[0015] Step 4: Based on Step 1 and Step 3, in order to incorporate high-level semantic features... With low-level detail features The bidirectional fusion between the two is achieved through a Cross-Attention Bidirectional Interaction (CA-BI) mechanism. This bidirectional interaction constructs two complementary information transmission paths through a gated cross-attention mechanism, forming a mutually reinforcing mechanism: semantic context is used to guide the precise localization of fine-grained tampering traces. Formally, the top-down path uses semantic guidance to enhance local details through cross-attention. (5); The bottom-up approach, through cross-attention to low-level evidence artifacts, in turn refines and corrects high-level semantic understanding. (6); in, express Query projection, and They represent Key and value projection, Representing feature dimension, These are learnable residual coefficients. , . Then a gating mechanism is used to... and Two-way refinement is performed to obtain fusion features .
[0016] Step 5: Based on Step 1, the noise encoder captures statistical inconsistencies caused by manipulation, and then obtains noise features through feature downsampling. Cross-domain integration module Interactive modeling using the cross-domain CA-BI module yields fused features. The calculation method for CA-BI is the same as in step four.
[0017] Step Six: Merge Features The input is fed into a Progressive Prediction Head (PPH), which refines the model of the fused features through multi-level progressive decoding, ultimately generating a high-resolution tamper region localization mask. .
[0018] Step 7: Considering that high-level semantics focuses attention on semantically reasonable regions and that pattern representations are invariant to specific object categories, we add a cross-level interaction loss term to the loss function to guide the model in learning semantically independent patterns: (7); in Indicated in scale The tampering pattern distribution learned in school It was derived from the actual altered image. Finally, the total loss function is: (8); in For binary cross-entropy loss, This is the boundary loss. , These are learnable hyperparameters used to balance the contributions of the three loss functions.
[0019] Based on the above steps, the specific steps are as follows:
[0020] In step two, the tampered image to be detected is input into the image encoder to extract tampering features at four different scales. .in Enhanced by the GL-CA module, it enables the discovery of fine-grained local tampering artifacts in high-level semantic features and the modeling of global long-range dependencies.
[0021] In step three, GL-CA uses two parallel attention methods, which are calculated as follows: (9); (10); in, , and These represent the query, key, and value projection of the input feature map within each local window, respectively. , and These represent global feature queries, key-value projections, and value projections, respectively. The dimension of the key is represented. Subsequently, the outputs of the two branches are adaptively fused using gating mechanism formulas (3) and (4) to obtain the features. .
[0022] In step four, a cross-attention bidirectional interaction (CA-BI) module is introduced to achieve high-level semantic features. With low-level detail features Two-way integration between them. Through and Alternating between formulas (6) and (7) as queries for cross-attention, the key and value projection calculation yields enhanced features. , To better integrate the enhanced features, we employ multi-scale adaptive fusion to integrate bidirectional refined features: (11); (12); (13); in, Indicates passage The channel-wise gated weights are obtained through convolution and learning via the Sigmoid activation function. ⊙ represents element-wise multiplication. This represents the splicing along the channel dimension. The final result is the fused spatial features. .
[0023] In step five, the tampered image to be detected is input into the image encoder. The noise encoder captures statistical inconsistencies caused by manipulation, followed by feature downsampling. (14); in, These residuals are processed through channel rearrangement operations, gradually reducing the spatial resolution from Reduce to At the same time, the number of channels is expanded to 128, thus providing a compact and discriminative noise representation for cross-domain fusion. Subsequently, the cross-domain feature fusion module Spatial features through a bidirectional cross-attention mechanism With noise characteristics This module integrates the spatial and noise domains, establishing a collaborative interaction. Based on the CA-BI framework, it implements a complementary reasoning mechanism, enabling the two domains to provide information to each other and refine and correct for each other. (15); in, and These represent the cross-attention operations from the spatial domain to the noise domain and from the noise domain to the spatial domain, respectively, and their specific calculation methods are consistent with Equation (6). This framework establishes two complementary interaction paths between different domains. In the space-to-noise path, with As a query pair Attention computation is performed to enable the semantic context to identify statistically significant tampering patterns in the noise response; conversely, in the noise-to-space path, attention is used to... As a query pair Attention calculations are performed to detect inconsistencies in spatial representations introduced by tampering, using statistical anomalies. This two-way interactive mechanism constructs a closed-loop verification model: spatial priors guide noise analysis to focus on semantically reasonable regions, while noise statistical properties provide pixel-level verification of the integrity of the image's spatial structure. Domain-specific gating weights. By evaluating the relative discriminative power of each domain, the contributions of the two are dynamically balanced. The resulting fusion representation... It also includes spatial consistency and statistical verification information, thereby achieving robust image tampering localization.
[0024] In step six, the Progressive Prediction Head (PPH) refines the cross-domain fusion features through a hierarchical upsampling process, thereby reconstructing the final tampering mask. (16); Among them, decoder It consists of a series of upsampling stages, each containing a Convolution and scale factor are The pixel shuffle operation is used. This structure employs a coarse-to-fine progressive refinement strategy to effectively preserve fine-grained spatial details while smoothly expanding features. The initial stage involves fusing high-level features. and Processing is performed to introduce rich semantic context; subsequent stages gradually restore precise spatial boundaries based on the continuously refined representation. Finally, through a Sigmoid activation function... Convolution generates binary prediction masks .
[0025] In step seven, we introduce a cross-level interaction loss to ensure that the tampering pattern distribution learned by the network is consistent with the pattern distribution derived from the ground truth annotations across multiple scales. We use KL divergence to measure the difference between the two and introduce weight coefficients to balance the contributions of each level, thereby suppressing semantic bias and guiding the model to learn tampering statistical patterns independent of specific object categories, thus improving the generalization ability of cross-scene tampering detection. Furthermore, we combine binary cross-entropy loss and boundary loss to jointly optimize the model.
[0026] The beneficial effects of this invention are as follows: By introducing a cross-scale, cross-domain bidirectional interaction and adaptive gating fusion mechanism, this invention effectively solves the problem of existing methods struggling to balance semantic generalization and localization accuracy. This allows the model to retain semantic context information while avoiding semantic overfitting, enabling it to learn universally applicable tampering patterns. Compared with existing technologies, this invention significantly reduces false detection and false negative rates in tampering detection in complex scenes and on unseen objects, improving the accuracy, robustness, and generalization ability of image tampering localization. Attached Figure Description
[0027] Figure 1 This invention provides t-SNE visualization of the tampered representation learned from baseline high-level features.
[0028] Figure 2 This invention provides a Grad-CAM visualization comparison with methods such as MVSS, Sparse-Vit, and Mesorch.
[0029] Figure 3 This is a flowchart of the network structure of the present invention;
[0030] Figure 4 This is the Cross-Attention Bidirectional Interaction (CA-BI) module of the present invention; Detailed Implementation
[0031] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0032] The detailed structure of the multi-scale cross-domain interactive network for image tampering localization described in this invention is as follows: Figure 3 As shown, it includes the following steps:
[0033] Step 1: Selecting the Training Dataset: To comprehensively evaluate the effectiveness and generalization ability of the model, this invention tests the method under two training protocols: the standardized CAT-Protocol and the MVSS-Protocol. The CAT-Protocol constructs a large-scale training set by fusing multiple benchmark datasets, including CASIAv2, IMD2020, tampCOCO, compRAISE, and FantasticReality, and randomly samples 1,800 images from each dataset in each training epoch. In contrast, the MVSS-Protocol is trained only on CASIAv2 and is used to evaluate the model's generalization ability under conditions of limited data diversity.
[0034] Step 2, Measurement Metrics: We use pixel-level F1 score and area under the curve (AUC) as evaluation metrics. Unless otherwise stated, all predicted masks are binarized using a fixed threshold of 0.5.
[0035] Step 3, Implementation Details: This method uses 200 epochs, a batch size of 16, and the AdamW optimizer, with a learning rate starting from... Decay to the cosine annealing strategy A 4-epoch warm-up phase was set up. All experiments were conducted on an NVIDIA GeForce RTX 4090 graphics card.
[0036] Step 4, Training Data Preprocessing: This invention scales the training data to... Then, the images are randomly horizontally flipped. Finally, they are normalized using the mean and standard deviation from ImageNet. ;
[0037] Step 5: Construct the dual-branch network model, including the following steps:
[0038] (5-1) Model loading: Use the pre-trained model trained from the ImageNet dataset to initialize the image domain encoder and the noise domain encoder.
[0039] (5-2) Construct image domain branches;
[0040] (5-2-1) In the image domain, we use Segformer as the backbone network. The tampered image to be detected is input into Segformer to obtain features at four different scales. The feature dimensions follow the design of Segformer. , , , .in These are the low-level detailed features of the first and second phases of Segformer. These are the high-level semantic features for the third and fourth stages of Segformer. Simultaneously, according to formulas (2), (3), and (4)... The GL-CA module enhances the representation, enabling higher-level stages to adaptively synthesize a unified feature representation that is semantically consistent and sensitive to local tampering.
[0041] (5-2-2) In order to achieve high-level semantic features With low-level detail features The bidirectional fusion between them. We input this into the CA-BI module, where the bidirectional interaction constructs two complementary information transmission paths through a gated cross-attention mechanism, thus forming a mutually reinforcing mechanism: the top-down path uses semantic guidance to enhance local details through cross-attention, and the enhanced features are obtained according to formula (5). Symmetrically, the bottom-up path is achieved by exchanging equation (6). and The role is used to calculate semantic features enhanced by details. Enhanced features are obtained using gated attention according to formulas (11) and (12). , Finally, the fused features are obtained by splicing them on the channels according to formula (13). .
[0042] (5-3) Noise Domain: The noise domain is captured by SRM (Steganalysis Rich Model) filtering to capture statistical inconsistencies caused by manipulation, followed by feature downsampling. This is done by progressively reducing the spatial resolution from... Reduce to At the same time, the number of channels is expanded to 128, thus providing a compact and discriminative noise representation for cross-domain fusion.
[0043] (5-4) Cross-domain feature fusion: using the cross-domain feature fusion module Spatial features through a bidirectional cross-attention mechanism With noise characteristics Integration is performed to establish a collaborative interaction relationship between the spatial domain and the noise domain. Based on the CA-BI framework, this module implements a complementary reasoning mechanism, enabling the two domains to provide information to each other and refine and correct each other. The final fusion features are obtained according to formula (15). .
[0044] (5-5) Generate prediction mask: fuse features The input to the Progressive Prediction Head (PPH) refines the cross-domain fusion features through a hierarchical upsampling process. This prediction head consists of a series of upsampling stages, each containing a... Convolution and scale factor are The pixel shuffle operation is used. This structure employs a coarse-to-fine progressive refinement strategy to effectively preserve fine-grained spatial details while smoothly expanding features. The initial stage involves fusing high-level features. and Processing is performed to introduce rich semantic context; subsequent stages gradually restore precise spatial boundaries based on the continuously refined representation. Finally, through a Sigmoid activation function... Convolution generates binary prediction masks .
[0045] (5-6) The obtained mask image Using multi-level supervision, cross-entropy loss is used to determine whether pixel values have been tampered with, boundary loss is used to improve the accuracy of tampering, and cross-level interaction term loss is used to guide the model to learn semantically independent patterns.
[0046] The above embodiments provide a detailed description of the specific implementation of a multi-scale cross-domain interactive network for image tampering localization proposed in this invention. The description of the above embodiments is only for the purpose of helping to understand the proposed method and core ideas of this invention. Based on the ideas of this invention, there may be some differences in specific implementation methods. Therefore, the content of this specification should not be construed as a limitation of this invention.
Claims
1. A multi-scale cross-domain interaction network for image forgery localization, characterized in that, Including the following steps: Step 1: Construct a dual-branch network structure consisting of an image domain branch and a noise domain branch. Use a cross-domain fusion module to fuse the features from the image and noise domain branches. Input the fused features into the decoder to obtain the final prediction mask. The model is defined as follows: (1); wherein, represents an input image, and represent an image encoder and a noise encoder, respectively, represents a cross-domain fusion module, represents a decoder, represents a predicted tampering mask. Step Two: 1) Based on Step One, the image encoder extracts features at four different scales from the tampered image and introduces multiple Global-Local Context Aggregation (GL-CA) modules into the high-level features to enhance contextual reasoning capabilities, defined as follows: (2); To efficiently capture both global context information and local tamper artifacts simultaneously, the GL-CA module selectively applies only to high-level features. and Each GL-CA module consists of parallel global attention branches and local attention branches. The global branch retains the original self-attention mechanism and is used to model long-range dependencies, thereby obtaining... At the same time, the local branches divide the feature map into several non-overlapping regions. Window, and perform self-attention operations within each window to obtain This is to capture fine-grained local tampering artifacts. Subsequently, the outputs of the two branches are adaptively fused using a gating mechanism. (3); (4); in, This represents a multilayer perceptron. This represents the Sigmoid activation function. The learned gating weights are used to balance the contributions of global context and local details. 2) In order to incorporate high-level semantic features With low-level detail features The bidirectional fusion between the two is achieved through a Cross-Attention Bidirectional Interaction (CA-BI) mechanism. This bidirectional interaction constructs two complementary information transmission paths through a gated cross-attention mechanism, forming a mutually reinforcing mechanism: semantic context is used to guide the precise localization of fine-grained tampering traces. Formally, the top-down path uses semantic guidance to enhance local details through cross-attention. (5); The bottom-up approach, through cross-attention to low-level evidence artifacts, in turn refines and corrects high-level semantic understanding. (6); in, express Query projection, and They represent Key and value projection, Representing feature dimension, These are learnable residual coefficients. , . Then a gating mechanism is used to... and Two-way refinement is performed to obtain fusion features . Step 3: 1) Based on Step 1, the noise encoder captures the statistical inconsistencies caused by manipulation, and then obtains noise features through feature downsampling. Cross-domain integration module Interactive modeling using the cross-domain CA-BI module yields fused features. The calculation method for CA-BI is the same as in step four. 2) Integrate features The input is fed into the Progressive Prediction Head (PPH), which refines the model of the fused features through multi-level progressive decoding, ultimately generating a high-resolution tamper region localization mask. Considering that high-level semantics focuses attention on semantically reasonable regions and that pattern representations are invariant to specific object categories, we add a cross-level interaction loss term to the loss function to guide the model in learning semantically independent patterns: (7); in Indicated in scale The tampering pattern distribution learned in school It was derived from the actual altered image. Finally, the total loss function is: (8); in For binary cross-entropy loss, This is the boundary loss. , These are learnable hyperparameters used to balance the contributions of the three loss functions.
2. A multi-scale cross-domain interactive network for image tampering localization according to claim 1. Its characteristics are: 1) In step two, the tampered image to be detected is input into the image encoder to extract tampering features at four different scales. .in Enhanced by the GL-CA module, it enables the discovery of fine-grained local tampering artifacts in high-level semantic features and the modeling of global long-range dependencies. 2) GL-CA uses two parallel attention mechanisms, and their computation methods are as follows: (9); (10); in, , and These represent the query, key, and value projection of the input feature map within each local window, respectively. , and These represent global feature queries, key-value projections, and value projections, respectively. The dimension of the key is represented. Subsequently, the outputs of the two branches are adaptively fused using gating mechanism formulas (3) and (4) to obtain the features. . 3) A cross-attention bidirectional interaction (CA-BI) module was introduced to achieve high-level semantic features. With low-level detail features Two-way integration. Through and Alternating between formulas (6) and (7) as queries for cross-attention, the key and value projection calculation yields enhanced features. , To better integrate the enhanced features, we employ multi-scale adaptive fusion to integrate bidirectional refined features: (11); (12); (13); in, Indicates passage The channel-wise gated weights are obtained through convolution and learning via the Sigmoid activation function. ⊙ represents element-wise multiplication. This represents the splicing along the channel dimension. The final result is the fused spatial features. .
3. A multi-scale cross-domain interactive network for image tampering localization according to claim 1. Its characteristics are: 1) In step three, the tampered image to be detected is input into the image encoder. The noise encoder captures statistical inconsistencies caused by manipulation, followed by feature downsampling: (14); in, These residuals are processed through channel rearrangement operations, gradually reducing the spatial resolution from Reduce to At the same time, the number of channels is expanded to 128, thus providing a compact and discriminative noise representation for cross-domain fusion. Subsequently, the cross-domain feature fusion module Spatial features through a bidirectional cross-attention mechanism With noise characteristics This module integrates the spatial and noise domains, establishing a collaborative interaction. Based on the CA-BI framework, it implements a complementary reasoning mechanism, enabling the two domains to provide information to each other and refine and correct for each other. (15); in, and These represent the cross-attention operations from the spatial domain to the noise domain and from the noise domain to the spatial domain, respectively, and their specific calculation methods are consistent with Equation (6). This framework establishes two complementary interaction paths between different domains. In the space-to-noise path, with As a query pair Attention computation is performed to enable the semantic context to identify statistically significant tampering patterns in the noise response; conversely, in the noise-to-space path, attention is used to... As a query pair Attention calculations are performed to detect inconsistencies in spatial representations introduced by tampering through statistical anomalies. This two-way interactive mechanism constructs a closed-loop verification model: spatial priors guide noise analysis to focus on semantically reasonable regions, while noise statistical properties provide pixel-level verification of the integrity of the image's spatial structure. Domain-specific gating weights. By evaluating the relative discriminative power of each domain, the contributions of the two are dynamically balanced. The resulting fusion representation... It also includes spatial consistency and statistical verification information, thereby achieving robust image tampering localization. 2) The Progressive Prediction Head (PPH) refines cross-domain fusion features through a hierarchical upsampling process, thereby reconstructing the final tampering mask: (16); Among them, decoder It consists of a series of upsampling stages, each stage containing a Convolution and scale factor are The pixel shuffle operation is used. This structure employs a coarse-to-fine progressive refinement strategy to effectively preserve fine-grained spatial details while smoothly expanding features. The initial stage involves fusing high-level features. and Processing is performed to introduce rich semantic context; subsequent stages gradually restore precise spatial boundaries based on the continuously refined representation. Finally, through a Sigmoid activation function... Convolution generates binary prediction masks . 3) We introduce a cross-level interaction loss to ensure that the tampering pattern distribution learned by the network is consistent with the pattern distribution derived from the ground truth annotations across multiple scales. We use KL divergence to measure the difference between the two and introduce weight coefficients to balance the contributions of each level, thereby suppressing semantic bias and guiding the model to learn tampering statistical patterns independent of specific object categories, thus improving the generalization ability of tampering detection across different scenarios. Furthermore, we combine binary cross-entropy loss and boundary loss to jointly optimize the model.