Image tampering localization method based on adversarial evidence debate and reinforcement learning adjudication

Abstract

The application discloses an image tampering positioning method based on adversarial evidence debate and reinforcement learning adjudication, and belongs to the technical field of digital image forensics and content authenticity identification. The scheme constructs a two-stage framework of "evidence confrontation-adjudication decision", including two stages of court debate and adjudication decision. In the court debate stage, multi-layer features are extracted through a shared multi-scale encoder, tampering and real double-hypothesis evidence are generated relying on a double-branch network, and a dynamic debate mechanism and a boundary shaping module are used for refinement, and a probability map and a boundary map are output. In the adjudication decision stage, multi-source evidence in the spatial and frequency domains is integrated to form enhanced features, high conflict areas are identified through controversy modeling, and a segmentation network is driven for repair by means of a reinforcement learning actor-critic structure, and an accurate tampering mask and a pixel-level reliability calibrated by Brier score are output. The application improves the positioning robustness and boundary clarity in complex scenes, solves the pain points of the prior art, and is suitable for judicial evidence and media verification fields.

Description

Technical Field

[0001] This invention relates to the field of digital image forensics and content authenticity identification, specifically to an image tampering location method based on adversarial evidence debate and reinforcement learning adjudication. Background Technology

[0002] In the field of digital image forensics and content authenticity verification, image tampering localization is one of the core research directions. Its core objective is to achieve pixel-level precision segmentation of regions in input images that have been altered through operations such as splicing, copying, moving, erasing, filling, and generative editing. With the popularization of image processing technology and the increasing complexity of tampering methods, higher demands are placed on the accuracy, robustness, and reliability of tampering localization methods. This technology has significant application value in fields such as judicial evidence collection, media verification, and information security.

[0003] Existing image tampering localization techniques mainly revolve around the "trace tracking" paradigm. The core idea is to locate tampering by mining weak, low-level artifacts left in the image. Specific implementation paths fall into two main categories: one is to improve the network structure, such as by introducing multi-scale fusion, attention mechanisms, and correlation modeling, to enhance the model's ability to mine low-level forensic features; the other is to introduce auxiliary clues such as edge maps, frequency domain representations (discrete cosine transform energy, residual signals, etc.), and noise residual maps to expose tampering traces that are not easily perceived in the RGB domain. These methods have achieved some success in simple tampering scenarios, but the overall source of evidence is relatively singular, and there is a lack of systematic testing and adversarial mechanisms against the two competing hypotheses of "authenticity" and "tampering."

[0004] However, existing technologies have significant limitations: First, they are sensitive to post-processing and noise. Operations such as JPEG recompression, scaling, social media transcoding, and noise addition can easily weaken or erase weak low-level artifacts, leading to problems such as missed detections, decreased recall, and blurred boundaries in localization. Second, the evidence is single-source and lacks dual-hypothesis adversarial testing. Most methods only strengthen the "tampered" evidence without simultaneously modeling the "real" evidence, making it difficult to make reliable decisions in areas with complex textures and strong semantic interference. Third, the confidence level is not calibrated, and the output probability is often directly used as the reliability. However, it is prone to overconfidence in cross-distribution or difficult samples, failing to clearly mark disputed areas and trigger targeted review and correction, which seriously affects the application effect of the technology in complex real-world scenarios.

[0005] Therefore, this invention aims to provide a tampering location scheme that is evidence-adversarial, calibrable, and capable of re-reasoning in uncertain regions. It outputs tampering masks and pixel-level reliability (uncertainty) information in complex tampering and cross-distribution scenarios, and performs adaptive refinement on highly controversial regions. Summary of the Invention

[0006] In view of this, the technical problem to be solved by the present invention is to propose an image tampering localization method based on adversarial evidence debate and reinforcement learning adjudication, construct an "evidence adversarial-adjudication" tampering localization framework, explicitly generate and oppose "tampered evidence" and "real evidence", give a conclusion quickly when the evidence is consistent, and perform adaptive re-reasoning and correction in areas of evidence conflict or insufficiency (high uncertainty).

[0007] To achieve the above objectives, the present invention provides the following technical solution: an image tampering localization method based on adversarial evidence debate and reinforcement learning for adjudication, including a court debate stage and a adjudication stage;

[0008] The court debate stage includes constructing a dual-branch segmentation network for the prosecution (falsification hypothesis) and the defense on a shared multi-scale encoder, forming falsified evidence and true evidence respectively, and refining adversarial evidence through a dynamic debate mechanism and a boundary shaping module; The court debate phase specifically includes the following steps: S1: Feature extraction, acquiring the image to be detected and extracting multi-layer features through a shared multi-scale encoder; S2: Dual Hypothesis Evidence Generation, which generates tampered evidence and real evidence based on encoded features, and simultaneously outputs the corresponding probability map tP and boundary map tE. S3: Evidence Refinement, which uses a dynamic debate mechanism and boundary shaping module to adversarially refine evidence with dual hypotheses; The adjudication stage includes introducing an adjudication model, integrating multi-source evidence (spatial domain and frequency domain), strategically re-reasoning and condition repairing highly controversial areas, and outputting the final tampering mask and pixel-level reliability. The adjudication stage specifically includes the following steps: S4: Multi-source evidence aggregation, integrating spatial and frequency domain evidence to form enhanced evidence features; S5: Controversy modeling, generating pixel-level controversy maps and constructing block-level state vectors based on enhanced evidence features; S6: Reinforcement learning drives repair by outputting motion graphs through an actor-critic structure, which in turn drives the conditional segmentation network to generate the final tamper mask PM. S7: Reliability Output, calculates and outputs pixel-level reliability information Rel.

[0009] Preferably, the multi-layer features extracted by the shared multi-scale encoder in S1 include features from layer 1 to layer 5, covering low-level detail features and high-level semantic features; the image to be detected includes tampered images that have been JPEG recompressed, scaled, platform transcoded, or noise-added.

[0010] As a preferred embodiment, the dual-hypothesis evidence described in S2 is generated by constructing a prosecution branch and a defense branch in parallel. The prosecution branch outputs a tamper probability graph as tP, a tamper boundary graph as tE, and tamper semantic evidence features as tF. The defense branch uses (1-G) as supervision and G as a tamper truth mask to output a true probability graph rP, a true boundary graph rE, and true semantic evidence features rF.

[0011] Preferably, the dynamic debate mechanism described in S3 includes: S3.1, Bidirectional cross-attention with divergence suppression, based on the divergence graph D between the high-level features of the two branches, introduces a suppression term -λD into the attention weight to achieve complementary fusion of consistent regions and information isolation of disputed regions; S3.2, the cross-current coupling module, constructs a bounded difference based on feature difference, which is divided into Δ=tanh(MF-AF) and gated as α=sigmoid(conv([MF,AF])). It achieves evidence redistribution by updating MF'=MF+α·Δ and AF'=AF-α·Δ through symmetrical push-pull. Here, MF is the tampered evidence feature, AF is the real evidence feature, MF' is the enhanced tampered evidence feature, and AF' is the enhanced real evidence feature.

[0012] S3.3, a multi-level cascaded interaction, gradually suppresses noise and amplifies differences in disputed areas.

[0013] Preferably, the boundary shaping module in S3 extracts the edge prior of the input image through the Laplacian operator and fuses it with the multi-scale features of the encoder to generate a boundary map; it adopts symmetric boundary consistency constraints to align tE and rE spaces, and injects boundary information into the semantic stream to improve the sharpness of the segmentation boundary.

[0014] Preferably, the frequency domain evidence in S4 includes Laplace high-frequency information, SRM residual filter response, and block discrete cosine transform energy; the tampered semantic evidence feature tF and the real semantic evidence feature rF are projected onto the evidence space through a two-dimensional multilayer perceptron adapter, and fused with the spatial domain evidence and frequency domain evidence to form the enhanced evidence feature EV. EV is the final evidence feature obtained after fusing the tampered, real, and original image information.

[0015] Preferably, the block-level state vector in S5 includes: evidence statistics (mean, standard deviation, maximum value, Shannon entropy), conflict or uncertainty indicators, and the uncertainty indicators include the average degree of dispute within the block, the complementary consistency gap |tP-(1-rP)|, and the comprehensive prediction uncertainty H(tP)+H(1-rP); the value of the dispute graph dM is positively correlated with the intensity of the evidence conflict between the prosecution and the defense.

[0016] Preferably, the actor network in S6 outputs a discrete action set through Gumbel-Softmax and a direct-pass estimator. The discrete action set includes three types of actions: conservative, error-correcting, and reconstruction. The conditional segmentation network adopts a lightweight U-shaped structure, uses the action graph as a conditional encoding, and combines multi-source evidence to form evidence features EV and state vectors to generate the final tampering mask PM.

[0017] Preferably, the pixel-level reliability Rel described in S7 is generated by constructing pseudo-label supervision based on evidence entropy and cross-branch consistency, and calibrated by Brier score.

[0018] Compared with existing technologies, the image tampering localization method based on adversarial evidence debate and reinforcement learning adjudication provided by this invention has the following beneficial effects: (1) Robustness improvement: By using dual evidence adversarial methods of "tampering / real" and multi-source evidence aggregation (including frequency domain priors), the dependence on a single fragile artifact is reduced, making the model more stable under degradations such as compression, transcoding, blurring, and noise.

[0019] (2) Clearer boundaries: The consistency constraints of edge prior and symmetric boundary make the altered boundary geometrically consistent with the real boundary. Combined with edge-guided feature injection, it can effectively improve the problems of boundary adhesion, voids and blurred contours.

[0020] (3) Uncertainty can be quantified and corrected: the pixel-level reliability Rel is output and calibrated, which can directly locate the "disputed area"; the referee concentrates the computing resources on the high-entropy / high-conflict area to perform re-inference through the reinforcement learning strategy, thereby reducing overconfidence and missed detection on difficult samples.

[0021] (4) High scalability: The prosecution / defense branches, evidence construction methods, action space and judge segmentation network can all be replaced with different backbones and modules to adapt to different types of tampering and computing power constraints. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the image tampering localization method based on adversarial evidence debate and reinforcement learning judge of the present invention; Figure 2 This is a schematic diagram of the court debate stage in an embodiment of the present invention; Figure 3 This is a schematic diagram of the adjudication stage in an embodiment of the present invention; Figure 4 This table compares the performance of the present invention with that of the most advanced existing models. Figure 5 This is a diagram showing a comparison of the visualization results of the present invention with those of the most advanced existing models; Figure 6This table compares the performance of the present invention with that of the most advanced existing models for locating compressed images on social networking platforms. Figure 7 This table compares the performance of the present invention with that of the most advanced existing models under traditional image attacks. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0025] Example Please refer to Figures 1 to 7 As shown: To address the problems mentioned in the technical solutions, embodiments of this application provide an image tampering localization method based on adversarial evidence debate and reinforcement learning adjudication, specifically including: (A) Courtroom Debate Stage: Based on the shared multi-scale encoder, a dual-branch segmentation network is constructed for the prosecution (alleged tampering) and the defense, generating tampered evidence and genuine evidence respectively. A dynamic debate mechanism is used to achieve adversarial interactive refinement of the dual-branch evidence, and a boundary shaping module is combined to optimize the segmentation boundaries. The specific process is as follows: Figure 1 As shown.

[0026] (B) Adjudication Stage: An adjudication model is introduced, integrating multi-source evidence from the spatial and frequency domains. Strategic re-inference and condition repair are performed on highly contentious areas, ultimately outputting a tampering mask and pixel-level reliability. The specific process is as follows: Figure 2 As shown.

[0027] (A) During the court debate stage, such as Figure 2 As shown; Step 1: Shared Encoding and Dual Hypothesis Output; Input image I is processed by a shared multi-scale encoder to extract multi-level features (such as features from layers 1 to 5). Based on these shared features, branches for the prosecution and defense are constructed respectively, achieving parallel output of dual hypotheses: The prosecution branch focuses on the segmentation of "tampered regions," outputting a tampered probability map tP and a tampered boundary map tE, and generating tampered semantic evidence features tF. The defense branch focuses on the segmentation of "real regions," outputting a real probability map rP and a real boundary map rE, and generating real semantic evidence features rF. rP is trained using (1-G) supervised training (G is a tampered ground truth mask) to ensure that the model accurately predicts the real regions.

[0028] Step 2: Dynamic Debate Mechanism (DDM) To balance information sharing and misleading avoidance between the two branches, a dynamic debate mechanism is introduced. This mechanism achieves precise screening and strengthening of evidence through multi-dimensional interaction, specifically comprising three parts: (1) Bidirectional cross-attention with divergence suppression; First, calculate the divergence map D between the high-level features of the two branches (which can be solved by means of channel mean squared error, mean squared error, etc.); embed the suppression term -λD into the bidirectional cross-attention weights to reduce the absorption of cross-branch information in feature conflict regions to avoid misleading, and enhance the fusion of complementary information in feature consistency regions to improve the reliability of evidence.

[0029] (2) Cross-current coupling module; Based on the bi-branch eigenvalue difference, a bounded difference Δ=tanh(MF-AF) and a gating coefficient α=sigmoid(conv([MF,AF])) are constructed. Local evidence redistribution is achieved through a symmetric push-pull update formula (MF'=MF+α·Δ, AF'=AF-α·Δ). This mechanism strengthens strong evidence, yields weak evidence, and ensures the conservation of the total bi-branch response, thereby improving the discriminative power of the evidence.

[0030] (3) Detailed dynamic debate; The above-mentioned interaction process can be cascaded at multiple feature levels to gradually suppress noise interference in consistent areas, amplify feature differences in disputed areas, and continuously improve the separability of evidence.

[0031] Step 3: Boundary Shaping Module and Edge Guidance; In binary classification scenarios, "tampered regions" and "real regions" share the same geometric boundary. To address this, edge priors are extracted from the input image (e.g., high-frequency edge maps are obtained through the Laplacian operator), and these priors are fused with the low-level detail features and high-level semantic features of the encoder to accurately generate boundary prediction maps tE and rE. Simultaneously, an attention-based noise reduction mechanism is used to highlight key boundary information.

[0032] Furthermore, a symmetric boundary consistency constraint is introduced to ensure that tE and rE are precisely aligned in space, and boundary information is injected into the bi-branch semantic flow (such as through edge-guided feature modules) to finally obtain a segmentation result with clear contours and sharp boundaries.

[0033] (B) During the adjudication stage, such as Figure 3 As shown; Step four: Multi-source evidence aggregation; The adjudication model integrates multi-dimensional evidence to construct a complete evidence system: on the one hand, it summarizes the output results (tP, rP, tE, rE) of the prosecution and the defense; on the other hand, it incorporates frequency domain features and residual information (including but not limited to Laplace high-frequency features, SRM residual filter response, block discrete cosine transform (block-DCT) energy, etc.). The bi-branch semantic evidence features tF and rF are projected onto the evidence space through a two-dimensional multilayer perceptron adapter and fused with the aforementioned multi-source information to form the enhanced evidence feature EV.

[0034] Step 5: Controversy modeling and state construction; A pixel-level dispute map dM is generated based on the enhanced evidence feature EV. A larger dM value indicates a stronger conflict between the prosecution's and defense's evidence. The image / features are divided into N non-overlapping patches, and a state vector is constructed for each patch, containing two core types of information: (1) Evidence statistics: including mean, standard deviation, maximum value, Shannon entropy, etc., used to characterize the distribution characteristics and uncertainty of evidence; (2) Conflict / uncertainty indicators: including the average degree of dispute within the block, the complementary consistency gap |tP-(1-rP)|, the comprehensive prediction uncertainty H(tP)+H(1-rP), etc., to accurately quantify the degree of conflict between the two branches of evidence.

[0035] Step six, reinforcement learning strategy-driven condition repair; The referee model employs an actor-critic structure: the actor network maps the state vector of each block to a set of discrete actions (such as "conservative retention," "error correction adjustment," and "reconstruction optimization"). To achieve end-to-end training, Gumbel-Softmax and a direct-pass estimator are used to generate differentiable action maps. These action maps are used as conditional encodings, and together with enhanced evidence features (EV) and evidence statistics, are input into a lightweight U-shaped segmentation network, ultimately outputting an accurate tamper mask (PM).

[0036] Step 7: Training Objectives and Reasoning Process; (1) Segmentation and boundary supervision; For the tampered probability map tP, the true probability map rP, and the final tampered mask PM, a structure-aware segmentation loss (such as a combination of weighted binary cross-entropy and weighted cross-union loss) is used; for the boundary prediction maps tE and rE, a boundary loss (such as a combination of binary cross-entropy and Dice loss) is used to ensure accurate alignment of the bi-branch boundaries.

[0037] (2) Consistency and reliability calibration; Under reliability gating constraints, the symmetric KL divergence between tP and (1-rP) is calculated to force complementary consistency between the two branches. Pixel-level reliability Rel is supervised by constructing pseudo-labels through evidence entropy and cross-branch consistency. At the same time, calibration terms such as Brier score are combined to suppress model overconfidence and improve prediction reliability.

[0038] (3) Enhance learning optimization; A strong baseline without referee intervention is defined as B = max(tP, 1 - rP), and the reward function is defined as the cross-union gain (CUI) of the referee's output relative to the baseline: r = IoU(PM, G) - IoU(B, G). The actor network updates its parameters through policy gradients, and the critic network stabilizes its training process by regressing reward values.

[0039] (4) Reasoning stage; the reasoning process is divided into two steps: first, the court debate stage is executed, and tP, rP, tE, rE and bi-branch semantic evidence are output; then the judge model identifies high uncertainty blocks based on the dispute graph dM, performs strategic repair, and finally outputs the tampered mask PM and pixel-level reliability Rel.

[0040] In summary, the present solution has the following advantages through the above embodiments: (1) Dual Hypothesis Evidence Generation: Two segmented evidence streams, one for the prosecution (tampering) and one for the defense (true), are constructed in parallel on a shared multi-scale encoder, and the tampering probability map tP and the true probability map rP (and their complementary relationship 1-rP) are output respectively.

[0041] (2) Dynamic debate mechanism: Introducing bidirectional cross-attention with “disagreement suppression” in the two-stream interaction, and cross-stream coupling module based on gating push-pull update, to achieve complementary fusion of consistent areas, information isolation of disputed areas and redistribution of evidence.

[0042] (3) Boundary shaping and symmetric boundary consistency: extract edge priors from the input image and fuse them with multi-scale features to constrain the boundary prediction alignment of the two hypotheses in order to improve boundary clarity.

[0043] (4) Judge model: Based on multi-source evidence aggregation and pixel / block level dispute modeling, a reinforcement learning strategy network is used to select regional processing actions, drive the conditional segmentation network to re-infer and repair high uncertainty regions, and output the final mask.

[0044] (5) Reliability calibration: Pixel-level reliability Rel is output based on prediction entropy and cross-hypothesis consistency, and is used as training gating and final output.

[0045] In summary, compared with existing single-stream or direct fusion schemes, the difference of this invention is that it not only generates "tampered evidence" but also generates "genuine evidence" and resists verification within a unified framework; it avoids misleading in areas of evidence conflict through a divergence suppression mechanism, and the judge performs strategic review in areas of high uncertainty, rather than providing an uncalibrable confidence map through one-time forward reasoning.

[0046] Please refer to the above work process. Figures 1 to 3 .

[0047] It should be noted that the term "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for locating image tampering based on adversarial evidence debate and reinforcement learning-based adjudication, characterized in that, This includes the court debate stage and the judgment stage; The court debate stage includes constructing a dual-branch segmentation network for the prosecution (falsification hypothesis) and the defense on a shared multi-scale encoder, forming falsified evidence and true evidence respectively, and refining adversarial evidence through a dynamic debate mechanism and a boundary shaping module; The court debate phase specifically includes the following steps: S1: Feature extraction, acquiring the image to be detected and extracting multi-layer features through a shared multi-scale encoder; S2: Dual Hypothesis Evidence Generation, which generates tampered evidence and real evidence based on encoded features, and simultaneously outputs the corresponding probability map tP and boundary map tE. S3: Evidence Refinement, which uses a dynamic debate mechanism and boundary shaping module to adversarially refine evidence with dual hypotheses; The adjudication stage includes introducing an adjudication model, integrating multi-source evidence (spatial domain and frequency domain), strategically re-reasoning and condition repairing highly controversial areas, and outputting the final tampering mask and pixel-level reliability. The adjudication stage specifically includes the following steps: S4: Multi-source evidence aggregation, integrating spatial and frequency domain evidence to form enhanced evidence features; S5: Controversy modeling, generating pixel-level controversy maps and constructing block-level state vectors based on enhanced evidence features; S6: Reinforcement learning drives repair by outputting motion graphs through an actor-critic structure, which in turn drives the conditional segmentation network to generate the final tamper mask PM. S7: Reliability Output, calculates and outputs pixel-level reliability information Rel.

2. The image tampering localization method based on adversarial evidence debate and reinforcement learning judge as described in claim 1, characterized in that, The multi-layer features extracted by the shared multi-scale encoder described in S1 include features from layer 1 to layer 5, covering low-level detail features and high-level semantic features; the image to be detected includes tampered images that have been recompressed, scaled, transcoded by the platform, or subjected to noise.

3. The image tampering localization method based on adversarial evidence debate and reinforcement learning judges as described in claim 1, characterized in that, The dual-hypothesis evidence described in S2 is generated through the parallel construction of the prosecution branch and the defense branch. The prosecution branch outputs a tP tamper probability graph, a tE tamper boundary graph, and a tF tamper semantic evidence feature. The defense branch uses (1-G) as supervision and G as a tamper truth mask to output a true probability graph rP, a true boundary graph rE, and a true semantic evidence feature rF.

4. The image tampering localization method based on adversarial evidence debate and reinforcement learning judges as described in claim 1, characterized in that, The dynamic debate mechanism described in S3 includes: S3.1, Bidirectional cross-attention with divergence suppression, based on the divergence graph D between the high-level features of the two branches, introduces a suppression term -λD into the attention weight to achieve complementary fusion of consistent regions and information isolation of disputed regions; S3.2, the cross-current coupling module, constructs a bounded difference based on feature difference, which is divided into Δ=tanh(MF-AF) and gated as α=sigmoid(conv([MF,AF])). It achieves evidence redistribution by updating MF'=MF+α·Δ and AF'=AF-α·Δ through symmetrical push-pull. Here, MF is the tampered evidence feature, AF is the real evidence feature, MF' is the enhanced tampered evidence feature, and AF' is the enhanced real evidence feature. S3.3, a multi-level cascaded interaction, gradually suppresses noise and amplifies differences in disputed areas.

5. The image tampering localization method based on adversarial evidence debate and reinforcement learning judges as described in claim 1, characterized in that, The boundary shaping module described in S3 extracts the edge prior of the input image through the Laplacian operator and fuses it with the multi-scale features of the encoder to generate a boundary map; it adopts symmetric boundary consistency constraints to align tE and rE spaces, and injects boundary information into the semantic stream to improve the sharpness of the segmentation boundary.

6. The image tampering localization method based on adversarial evidence debate and reinforcement learning judges as described in claim 1, characterized in that, The frequency domain evidence described in S4 includes Laplace high-frequency information, SRM residual filter response, and block discrete cosine transform energy; the tampered semantic evidence feature tF and the true semantic evidence feature rF are projected onto the evidence space through a two-dimensional multilayer perceptron adapter, and fused with the spatial domain evidence and frequency domain evidence to form the enhanced evidence feature EV.

7. The image tampering localization method based on adversarial evidence debate and reinforcement learning judges according to claim 1, characterized in that, The block-level state vector in S5 includes: evidence statistics (mean, standard deviation, maximum value, Shannon entropy), conflict or certainty indicators, and uncertainty indicators including the average degree of dispute within the block, complementary consistency gap |tP-(1-rP)|, and comprehensive prediction uncertainty H(tP)+H(1-rP); the magnitude of the dispute graph dM is positively correlated with the intensity of the evidence conflict between the prosecution and the defense.

8. The image tampering localization method based on adversarial evidence debate and reinforcement learning judges according to claim 1, characterized in that, The actor network described in S6 outputs a discrete action set through Gumbel-Softmax and a direct-pass estimator. The discrete action set includes three types of actions: conservative, error-correcting, and reconstruction. The conditional segmentation network adopts a lightweight U-shaped structure, uses the action graph as a conditional encoding, and combines multi-source evidence to form evidence features EV and state vectors to generate the final tampering mask PM.

9. The image tampering localization method based on adversarial evidence debate and reinforcement learning judges according to claim 1, characterized in that, The pixel-level reliability Rel described in S7 is generated by constructing pseudo-label supervision based on evidence entropy and cross-branch consistency, and calibrated by Brier score.

Images