An AI-generated image dynamic forensics method and system based on hyperbolic space

Abstract

The application discloses a kind of AI generation image dynamic forensics method and system based on hyperbolic space, comprising: extracting the multi-scale feature of image to be forensiced and embedding hyperbolic space, obtaining hyperbolic embedding anchor point;By calculating the distortion degree and penalty value between the anchor point and the real image prototype, a first forensics determination result is generated.Secondly, when the preliminary result is in the fuzzy interval, the system dynamically triggers the cross-modal forensics verification link, and combines physical geometry and semantic logic common sense for deep comparison, to obtain the final determination conclusion.Finally, if it is confirmed as AI generated image, the system will accurately lock the candidate generation engine through the tangent space search, and use the proxy network to trace the training data contribution ranking, and finally output the tamper-proof structured forensics report chain.The application uses the characteristics of hyperbolic space to effectively amplify the hidden artifacts of high-fidelity AI images, while achieving high-precision identification and dynamic balance of system computing power, providing complete engine tracing and data attribution capabilities.

Description

Technical Field

[0001] This invention relates to an AI-generated image dynamic forensics method and system based on hyperbolic space, belonging to the fields of artificial intelligence security and computer vision technology. Background Technology

[0002] With the rapid advancement of generative artificial intelligence (AIGC) technology, deep generative models, represented by stable diffusion models, have achieved breakthroughs in the realism and detail richness of image and video synthesis. While these technologies bring convenience to creative design, film and television production, and other fields, they have also led to the proliferation of deepfake content, seriously threatening the authenticity of news dissemination, the protection of digital copyrights, and the reliability of judicial evidence collection. Therefore, developing efficient, accurate, and highly generalizable AI-generated image detection and tracing technologies is of significant practical importance.

[0003] Currently, AI-generated image detection methods are mainly divided into methods based on traditional handcrafted features and methods based on deep neural networks. Traditional methods mainly rely on the statistical properties of images, such as noise distribution and spectral anomalies, but they often fail when faced with new-generation high-quality generative models. With the development of deep learning, detection methods based on the collaboration of convolutional neural networks (CNN), visual transducers (ViT), and large language models (LLM) have gradually become a research hotspot.

[0004] Regarding multi-domain feature fusion, CN120543486A discloses a collaborative multi-domain AI-generated image detection method and system. This method extracts spatial domain representations through strong and weak texture contrast and introduces a frequency domain channel weighting mechanism. It utilizes Discrete Cosine Transform (DCT) to obtain frequency domain information to guide spatial domain feature fusion, thereby enhancing the model's ability to capture multi-dimensional differences. This spatial-frequency domain collaborative strategy provides an important approach for capturing generated artifacts.

[0005] To leverage the correlations between samples, CN119048843A discloses an AI-generated image detection method based on graph topology learning. This method extracts global visual features using a frozen CLIP model, constructs a graph topology structure to obtain relational features between nodes, and aggregates these features through a graph convolutional network for real / fake classification. This approach improves detection accuracy to some extent by mining the topological relationships between samples.

[0006] Furthermore, with the development of multimodal technology, using large multimodal models for detection has become a new trend. CN121074613A discloses an AI-generated image detection method and system, which obtains open-ended answers by inputting images into a large multimodal language model, extracts and fuses title, backbone, and modification features, and uses prompt word engineering to guide the model in judging authenticity. In addition, CN121033632A discloses an AI-generated image detection method based on multi-agent collaboration, which extracts lightweight visual and non-visual features, uses a large model for adversarial reasoning and result calibration, and displays the results through a visual interactive interface. This indicates that introducing intelligent agents and large model logical reasoning is an effective way to improve the interpretability of detection.

[0007] Regarding the robustness of feature extraction, CN118570599A discloses an AI-generated image detection method based on a frozen ViT feature fusion network. This method utilizes frozen CLIP-ViT to extract shallow and deep features and trains the network through a multi-level feature fusion network, aiming to improve the ability to distinguish between different generative models.

[0008] However, despite the good results achieved by existing technologies on specific datasets, the following major problems still exist when facing unknown generative models and complex application scenarios:

[0009] First, existing methods are mostly based on feature metrics in Euclidean space, which makes it difficult to solve the generalization problem across generators. Whether based on CNNs or Transformers, methods typically assume that image features are distributed in a flat Euclidean space. However, image data is inherently located on curved, high-dimensional manifolds. Existing methods forcibly straighten the curved manifold structure and map it to Euclidean space, causing the feature structure to distort when facing unknown generative models (such as migrating from a diffusion model to a video generation model), leading to detection failure.

[0010] Secondly, existing systems lack in-depth verification of physical consistency and semantic logic. Most existing solutions focus on pixel-level or statistical artifact detection, lacking explicit verification of physical laws such as lighting, shadows, and gravity, as well as biological common sense. Although some methods introduce large models for analysis, they often lack dedicated verification modules for physical laws, making it difficult to discover logical loopholes hidden in high-fidelity images.

[0011] Third, there is a lack of closed-loop governance and traceability capabilities throughout the entire process. Existing multi-agent or large-model methods mainly focus on "authenticity judgment" while neglecting "source tracing" and "training data attribution." In copyright protection and judicial liability determination scenarios, it is not only necessary to determine whether an image was generated by AI, but also to clarify which model generated it and the potential source of training data. Existing technologies are unable to provide a complete chain of evidence.

[0012] Therefore, there is an urgent need to develop a dynamic detection system that can break through the limitations of Euclidean space, integrate physical and semantic logic verification, and has the ability to trace the source of generation and attribute data, so as to improve the detection accuracy, generalization and interpretability in complex adversarial environments. Summary of the Invention

[0013] The purpose of this invention is to overcome the shortcomings of existing AI-generated image detection technologies, such as weak generalization ability when facing unknown generation models, blind spots in single-modal detection, inability to balance detection accuracy and system throughput, and lack of systematic source attribution methods. This invention provides a dynamic evidence collection method and system for AI-generated images based on hyperbolic space.

[0014] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0015] First aspect: A method for dynamic image forensics based on hyperbolic space generated by AI, the method comprising:

[0016] Acquire the image to be used for evidence collection, and extract the multi-scale visual feature representation of the image to be used for evidence collection;

[0017] Based on the topological consistency constraint of hyperbolic divergence, the multi-scale visual feature representation is embedded into a hyperbolic representation space with negative constant curvature to obtain the hyperbolic embedding anchor point of the image to be examined.

[0018] Based on a multi-cluster local reference prototype constructed from real images, the geodesic projection distortion degree and edge dilation penalty value of the hyperbolic embedded anchor point are calculated, and the two are fused into a hyperbolic realism deviation confidence degree.

[0019] A first evidence determination result is generated based on the hyperbolic authenticity deviation confidence level; when the hyperbolic authenticity deviation confidence level falls into the fuzzy decision range based on the hyperbolic curvature parameter adaptive dynamic adjustment, a cross-modal evidence verification link is dynamically triggered, and a second evidence determination result is generated by combining the physical geometric consistency and semantic logic common sense comparison results.

[0020] Based on the first and second evidence collection judgment results, the final evidence collection conclusion of the image to be collected is output. When the final evidence collection conclusion is determined to be an artificial intelligence generated image, the hyperbolic embedding anchor point is input into the source analysis module, and a local tangent space nearest neighbor search with radial penalty is performed in the source tracing feature knowledge base to determine the candidate generation engine. At the same time, the offline influence function approximation proxy network is called to output the contribution ranking of the training data, and the first evidence collection judgment result, the second evidence collection judgment result and the candidate generation engine are fused to generate a structured evidence collection report chain.

[0021] Optionally, the multi-scale visual feature representation includes:

[0022] A frozen visual pre-trained model is used to extract global semantic representations;

[0023] High-frequency filters are used to extract edge artifact features and locally generated texture features from the images to be examined.

[0024] The phase anomaly distribution map of edge artifact features in the frequency domain is calculated, and the global semantic representation is fused with the phase anomaly distribution map across channels. The formula for cross-channel feature fusion is as follows:

[0025] ;

[0026] in, The output features after fusion For global semantic features, Edge features; and These represent the Fourier transform and the inverse Fourier transform, respectively. Frequency domain amplitude information representing edge features; This is a phase anomaly distribution diagram; e represents the base of the natural logarithm; The imaginary unit; This represents element-wise multiplication; Indicates channel splicing operation; by... The amplitude information is preserved and introduced. Phase modulation is performed, via After reconstruction and By fusing these components, an enhanced representation of anomalous regions can be achieved.

[0027] Optionally, obtaining the hyperbolic embedding anchor point of the image to be forensicly examined includes:

[0028] Calculate the hierarchical transitivity probability matrix of multi-scale visual feature representations in Euclidean space to characterize the topological relationships of semantic inclusion and artifact dependence between feature scales;

[0029] By using manifold logarithmic mapping or exponential mapping function, multi-scale visual feature representations are mapped to the Poincaré sphere model or Lorentz hyperboloid model. A topological loss term based on hyperbolic divergence is introduced into the optimization objective of the mapping to force the point integral distribution in hyperbolic space to approximate the hierarchical transfer probability matrix, thereby achieving topologically consistent embedding in non-Euclidean space.

[0030] The formula for calculating the topology loss term is as follows: ;

[0031] in, Topology loss term, Indices representing spatial locations or semantic units in multi-scale feature maps, used to identify the first... Each corresponding pair of feature points; and These represent the Euclidean features extracted at the index position under different scale branches; Let f(x) denote a generating function that is strictly convex in hyperbolic space; This represents the Riemann gradient of the generating function on the hyperbolic manifold; Refer to Using as a base point Logarithmic mapping function of manifold onto the tangent plane; Indicates at the base point The Riemann inner product calculated in the local tangent space;

[0032] During the mapping process, a Riemann barrier boundary constraint based on adaptive decay of the feature norm is applied. The formula for the Riemann barrier boundary constraint is as follows: ;in This represents the Riemann barrier boundary constraint term used to prevent numerical anomalies. Regularization weight coefficients used to control the intensity of boundary penalties; The absolute value of the negative constant curvature of the hyperbolic representation space; Let be the feature point vector in hyperbolic space; This represents the Euclidean norm of the feature point vector; A potential barrier triggering boundary threshold is set to prevent features from collapsing toward the boundary at infinity of the manifold. This indicates the truncation penalty function.

[0033] Optionally, the method for constructing the multi-cluster local reference prototype includes:

[0034] Acquire a training set of real images from multiple scenes and map it to the hyperbolic representation space;

[0035] Leveraging the capacity advantage of hyperbolic space for hierarchical tree structures, and based on the hyperbolic geodesic distance between feature points, we perform bottom-up hyperbolic hierarchical clustering of feature points in real images to fit the tree manifold distribution of "major class-subclass-instance" in real scenes.

[0036] Nodes at different levels are extracted into multi-cluster local reference prototypes, and hyperbolic neighborhood bandwidth parameters based on local density are assigned to each local reference prototype according to the hyperbolic geodesic distance.

[0037] Optionally, the acquisition of the geodesic projection distortion degree and edge dilation penalty value includes: acquiring the hyperbolic geodesic distance between the hyperbolic embedding anchor point and the feature points of the multi-cluster local reference prototype, and mapping the distance to the geodesic projection distortion degree, with the greater the distance, the higher the distortion degree;

[0038] Utilizing the geometric property that the volume of hyperbolic space increases exponentially with radius, a nonlinear amplification factor is calculated based on the radial coordinate position of the anchor point in space to adjust the range of influence of distortion.

[0039] When the anchor point approaches the boundary of the hyperbolic representation space, the edge expansion penalty value increases exponentially, providing a quantitative basis for the subsequent dynamic scaling of the fuzzy decision interval based on the Riemann metric tensor, and realizing the geometric consistency constraint of cross-scale, multi-cluster prototypes.

[0040] Optionally, the determination of the fuzzy decision interval includes:

[0041] Obtain the negative curvature constant of the current hyperbolic representation space and the distribution density of the local reference prototype;

[0042] The nonlinear distortion coefficient is calculated based on the negative curvature constant to measure the characteristic.

[0043] Using the nonlinear distortion coefficients and the distribution density, and with the Riemann metric tensor as a constraint, the upper and lower bound thresholds of the fuzzy decision interval are dynamically scaled.

[0044] Optionally, the cross-modal forensics verification link includes:

[0045] Physical and geometric consistency verification specifically involves extracting the scene depth map, surface normal vector, and global ambient lighting direction of the image to be examined using a monocular depth estimation model and an illumination decoupling network.

[0046] The theoretical shadow intensity distribution in the scene is calculated based on Lambert's cosine law;

[0047] By comparing the theoretical shadow intensity distribution with the shadow intensity distribution of the actual pixels in the image to be examined, a local thermal residual map representing the degree of light and shadow inconsistency is generated. The statistical variance of this residual map is extracted as the physical geometric consistency score. The pixel-by-pixel calculation formula for the local thermal residual map is as follows: ;in, Indicates the local thermal residual map at the pixel level. The residual value at the location; (u,v) represents the two-dimensional pixel coordinates of the image, where and These represent the horizontal and vertical indices, respectively; (u,v) represents the actual image captured by the infrared thermal imager or camera at a pixel count. The actual intensity value at that location; N(u,v) represents the thermal emissivity or reflectivity of the material; N(u,v) represents the pixel. The unit normal vector of a point on the surface of the corresponding 3D object; This represents the direction vector of the ambient heat source or main light source; max(..., 0) represents the cutoff function.

[0048] Semantic and logical common sense comparison, specifically:

[0049] Extract multi-dimensional detection templates targeting violations of physical laws and variations in biological structures from a pre-set counterfactual prompt template library;

[0050] The image to be used for evidence collection is combined with the multi-dimensional detection template and input into the multimodal large language model;

[0051] The contrastive normalization algorithm is used to calculate the difference in relative logical response probabilities when the large language model outputs affirmative and negative categorical words, thereby eliminating the prior bias noise of the language model itself and outputting a clean semantic logic verification score.

[0052] Optionally, the second evidence determination result includes:

[0053] The physical geometric consistency score, semantic logic verification score, and metadata tampering risk value of the image to be examined are input into the multilayer perceptron for multi-source evidence alignment.

[0054] The aligned verification confidence and the first evidence determination result are updated using Bayesian posterior probability based on hyperbolic distance decay factor, and the calibrated second evidence determination result is output.

[0055] Optionally, the construction process of the offline influence function approximation proxy network includes:

[0056] During the training phase, a high-precision influence function algorithm is used to calculate the contribution value of the real gradient influence of the pre-constructed training sample set to the generation result of the evidence collection offline; a lightweight ranking neural network is constructed as a proxy network.

[0057] The proxy network is trained using a pairwise sorting hinge loss function, which makes the output training samples' contribution ranking fit the relative magnitude of the actual gradient's influence contribution values, thereby avoiding the computational cost of inverting massive training data matrices. This pairwise sorting hinge loss function is defined as follows: ;

[0058] in, Represents the pairwise sorting hinge loss function. These represent two indices in the sample set, used to form a pairwise comparison relationship; These represent the actual gradient contribution values ​​of the corresponding samples, which are used as supervision signals for ranking. This represents an approximate prediction of the agent network's contribution to the aforementioned impact. This is a sign function used to characterize the actual sorting direction; its output is... , or ; These are preset boundary hyperparameters used to control the relaxation level of the sorting interval; This represents a hinge-like truncation operation, used to incur a loss only when the sorting constraint is violated.

[0059] When ranking the contribution of the output training data, the image to be investigated and the historical training dataset are input into the offline influence function approximation proxy network after training. The proxy network forward propagation directly outputs the approximate contribution score of each historical training sample to the generation of the image to be investigated. The top N training samples with the highest approximate contribution scores are selected as candidate source data for infringement tracing or training attribution.

[0060] The second aspect: A dynamic image forensics system based on hyperbolic space generated by AI, comprising:

[0061] The feature processing module is used to acquire the image to be used for evidence collection and extract the multi-scale visual feature representation of the image to be used for evidence collection.

[0062] Anchor point acquisition module is used to embed the multi-scale visual feature representation into a hyperbolic representation space with negative constant curvature based on the topological consistency constraint of hyperbolic divergence, so as to obtain the hyperbolic embedding anchor points of the image to be examined.

[0063] The confidence calculation module is used to calculate the geodesic projection distortion and edge dilation penalty value of the hyperbolic embedded anchor point based on the multi-cluster local reference prototype constructed from the real image, and to fuse the two into a hyperbolic realism deviation confidence.

[0064] The hyperbolic authenticity assessment module is used to generate a first evidence judgment result based on the hyperbolic authenticity deviation confidence level; when the hyperbolic authenticity deviation confidence level falls into the fuzzy decision range based on the hyperbolic curvature parameter adaptive dynamic adjustment, the cross-modal evidence verification link is dynamically triggered, and a second evidence judgment result is generated by combining the physical geometric consistency and semantic logic common sense comparison results.

[0065] The judgment output module is used to output the final evidence conclusion of the image to be examined based on the first evidence judgment result and the second evidence judgment result. When the final evidence conclusion is determined to be an artificial intelligence generated image, the hyperbolic embedding anchor point is input into the source analysis module, and a local tangent space nearest neighbor search with radial penalty is performed in the source feature knowledge base to determine the candidate generation engine. At the same time, the offline influence function approximate proxy network is called to output the contribution ranking of the training data, and the first evidence judgment result, the second evidence judgment result and the candidate generation engine are fused to generate a structured evidence report chain.

[0066] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0067] 1. Enhance the generalization detection capability for unknown generative models. This invention embeds image features into a hyperbolic representation space with negative constant curvature and introduces topological consistency and Riemann barrier boundary constraints based on hyperbolic divergence. Utilizing the geometric property that the volume of hyperbolic space increases exponentially with radius and the edge dilation penalty, highly realistic generated images with distributions similar to real images in Euclidean space are pushed towards the spatial edges. This scheme amplifies subtle generation artifacts and distortion features through spatial metric transformation, thereby improving the system's accuracy in recognizing generative engines not seen in the training set.

[0068] 2. Balancing Detection Accuracy with System Computational Throughput. This invention constructs an event-driven dynamic forensics scheduling mechanism, dynamically adjusting the upper and lower bounds of the fuzzy decision interval by calculating the nonlinear distortion coefficient and feature distribution density. The system allocates multimodal large-scale model or complex physical verification computing power only to difficult samples falling into the fuzzy interval, replacing the traditional complex verification of all samples. This scheme effectively reduces the ineffective use of computing resources and improves the system's concurrent processing efficiency while maintaining overall detection accuracy.

[0069] 3. Reduce judgment distortion and improve objectivity in cross-modal verification. In the cross-modal verification stage, this invention introduces physical geometric consistency verification to quantify the degree of lighting anomalies in the image; simultaneously, a contrastive normalization algorithm is used when calling a multimodal large model for semantic verification. This algorithm suppresses the prior bias of "compliance prompts" commonly found in large models during interaction by calculating the relative difference between the probabilities of affirmative and negative responses. This approach reduces the noise impact of subjective bias and improves the reliability of logical verification results in complex scenarios.

[0070] 4. Reduce the computational complexity of source attribution and construct a complete evidence storage chain. To address the issue of high computational cost in source attribution, this invention employs logarithmic mapping to reduce the dimensionality of features to a local tangent space and applies a radial modulus penalty to reduce the time complexity of large-scale hyperbolic retrieval. Simultaneously, it uses the true gradient influence value as a soft label and trains a lightweight approximation network using pairwise sorting hinge loss, avoiding the computational overhead of matrix inversion operations on massive training data. Finally, the system combines hash encryption and timestamp information binding technologies to achieve information solidification for authenticity verification, platform source attribution, and infringement tracking. Attached Figure Description

[0071] Figure 1 A schematic diagram of the overall architecture of the AI-generated image dynamic evidence collection system based on hyperbolic space provided in an embodiment of the present invention;

[0072] Figure 2 A flowchart of an AI-generated image dynamic forensics method based on hyperbolic space provided in an embodiment of the present invention;

[0073] Figure 3This is a schematic diagram illustrating the geometric principles of multi-scale visual feature mapping to hyperbolic representation space, local reference prototype distribution, and edge dilation penalty calculation in an embodiment of the present invention.

[0074] Figure 4 This is a schematic diagram of the cross-modal forensics verification link (including physical geometry and semantic logic verification) triggered by the dynamic forensics scheduling module in an embodiment of the present invention;

[0075] Figure 5 This is a schematic diagram of the structure of the source attribution engine performing local tangent space retrieval with radial penalty and offline influence function approximation proxy network sorting in an embodiment of the present invention;

[0076] Figure 6 This is a visual diagram illustrating the comprehensive performance evaluation and discrimination distribution of the AI-generated image dynamic forensics method provided in this embodiment of the invention under the conditions of dealing with highly realistic generation models and complex attack scenarios. Detailed Implementation

[0077] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0078] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0079] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0080] Example 1: A Dynamic Evidence Collection Method for AI-Generated Images Based on Hyperbolic Space; This example provides a complete and detailed end-to-end dynamic evidence collection operation process, mainly targeting a highly realistic portrait image widely circulated on social media, suspected to be generated by the latest AI engine, for authenticity determination and source attribution, such as... Figure 1 As shown, the specific steps are as follows:

[0081] Step 1: Obtain the image to be used for evidence collection and extract the multi-scale visual feature representation of the image to be used for evidence collection;

[0082] Step 2: Based on the topological consistency constraint of hyperbolic divergence, the multi-scale visual feature representation is embedded into a hyperbolic representation space with negative constant curvature to obtain the hyperbolic embedding anchor point of the image to be examined.

[0083] Step 3: In the hyperbolic representation space, based on the multi-cluster local reference prototype constructed from the real image, calculate the geodesic projection distortion and edge dilation penalty value of the hyperbolic embedding anchor point, and fuse the two into a hyperbolic realism deviation confidence score; the fusion calculation formula is as follows: in, This represents the final cost of fusion computing. This represents the currently input sample. As a local reference prototype, Indicates the degree of geodesic projection distortion. This represents the penalty for marginal expansion, while and These are the weighting coefficients used to adjust the relative importance of these two indicators;

[0084] Step 4: Generate a first evidence determination result based on the hyperbolic authenticity deviation confidence score; when the hyperbolic authenticity deviation confidence score falls into the fuzzy decision interval based on the adaptive dynamic adjustment of the hyperbolic curvature parameter, dynamically trigger the cross-modal evidence verification link, and generate a second evidence determination result by combining the physical geometric consistency and semantic logical common sense comparison results; based on the first and second evidence determination results, output the final evidence conclusion of the image to be evidenced; when the final evidence conclusion determines that it is an artificial intelligence generated image, input the hyperbolic embedded anchor point into the source analysis module, perform a local tangent space nearest neighbor search with radial penalty in the source feature knowledge base to determine the candidate generation engine; at the same time, call the offline influence function approximation proxy network to output the training data contribution ranking, and fuse the above information to generate a structured evidence report chain.

[0085] In this embodiment, step 1 is specifically as follows: The extraction of multi-scale visual feature representation includes: extracting global semantic representation using a frozen visual pre-trained model; extracting edge artifact features and locally generated texture features in the image to be examined using a high-frequency filter; calculating the phase anomaly distribution map of the edge artifact features in the frequency domain, and fusing the global semantic representation with the phase anomaly distribution map across channels to enhance the grid-like structural defect response left by the generation model during frequency domain downsampling. The cross-channel feature fusion formula is as follows: ;in, The output features after fusion For global semantic features, Edge features; and These represent the Fourier transform and the inverse Fourier transform, respectively. Frequency domain amplitude information representing edge features; This is a phase anomaly distribution diagram; e represents the base of the natural logarithm; The imaginary unit; This represents element-wise multiplication; Indicates channel splicing operation; by... The amplitude information is preserved and introduced. Phase modulation is performed, via After reconstruction and By fusing these components, an enhanced representation of anomalous regions can be achieved.

[0086] In this embodiment, step 2 is specifically as follows: Obtaining hyperbolic embedding anchor points through topological consistency constraints based on hyperbolic divergence includes: calculating the hierarchical transitivity probability matrix of the multi-scale visual feature representation in Euclidean space to characterize the topological relationship between semantic inclusion and artifact dependence between feature scales; mapping the multi-scale visual feature representation to a Poincaré sphere model or a Lorentz hyperboloid model using a manifold logarithmic mapping or exponential mapping function, and introducing a topological loss term based on hyperbolic divergence into the optimization objective of the mapping to force the point integral distribution in hyperbolic space to approximate the hierarchical transitivity probability matrix, thereby achieving topological consistency embedding in non-Euclidean space; the specific calculation formula for this topological loss term is as follows: ;in, and These represent point pairs of multi-scale visual feature representations mapped to hyperbolic space; Let f(x) denote a generating function that is strictly convex in hyperbolic space; This represents the Riemann gradient of the generating function on the hyperbolic manifold; Refer to Using as a base point Logarithmic mapping function of manifold onto the tangent plane; This indicates that at the base point The Riemann inner product is calculated in the local tangent space. During the mapping process, a Riemann barrier boundary constraint based on adaptive decay of the feature norm is applied to prevent the feature vector to collapse to infinity boundary due to extreme high-frequency artifacts in the generated image, thereby avoiding numerical anomalies in the hyperbolic embedding anchor point. The formula for the Riemann barrier boundary constraint is: ;in This represents the Riemann barrier boundary constraint term used to prevent numerical anomalies. Regularization weight coefficients used to control the intensity of boundary penalties; The absolute value of the negative constant curvature of the hyperbolic representation space; Let be the feature point vector in hyperbolic space; This represents the Euclidean norm of the feature point vector; A potential barrier triggering boundary threshold is set to prevent features from collapsing toward the boundary at infinity of the manifold. This means truncating the penalty function, ensuring that the optimization process is penalized only when the feature point approaches the boundary and its function value exceeds the trigger threshold.

[0087] In this embodiment, during the specific implementation process, in step 3, the multi-cluster local reference prototype is constructed in the following way: A training set of real images from multiple scenes is obtained and mapped to the hyperbolic representation space; utilizing the capacity advantage of the hyperbolic space for hierarchical tree structures, feature points of the real images are clustered from bottom to top in a hyperbolic hierarchical manner to fit the tree-like manifold distribution of "major class-subclass-instance" in the real scene; nodes at different levels are extracted as multi-cluster local reference prototypes, and a hyperbolic neighborhood bandwidth parameter based on local density is assigned to each local reference prototype.

[0088] In this embodiment, the geodesic projection distortion and edge dilation penalty values ​​are obtained as follows: The hyperbolic geodesic distance between the hyperbolic embedding anchor point and the nearest local reference prototype is calculated; the greater the distance, the higher the geodesic projection distortion. Utilizing the geometric property that the volume of hyperbolic space increases exponentially with radius, a nonlinear amplification factor is calculated based on the radial coordinate position of the hyperbolic embedding anchor point in space. When the hyperbolic embedding anchor point is closer to the boundary of the hyperbolic representation space, the edge dilation penalty value increases exponentially, thereby using the hyperbolic boundary effect to forcibly peel off and amplify the weak, hidden features of the highly realistic generated image.

[0089] In this embodiment, the fuzzy decision interval is a numerical interval with dynamically floating upper and lower bounds. The determination method includes: obtaining the negative curvature constant of the current hyperbolic representation space and the distribution density of the local reference prototype; calculating the nonlinear distortion coefficient of the feature metric based on the negative curvature constant; and dynamically scaling the upper and lower bound thresholds of the fuzzy decision interval using the nonlinear distortion coefficient and the distribution density, with the Riemann metric tensor as a constraint.

[0090] In this embodiment, the cross-modal evidence verification link includes physical geometric consistency verification, specifically implemented as follows: Using a monocular depth estimation model and an illumination decoupling network, the scene depth map, surface normal vector, and global ambient illumination direction of the image to be verified are extracted; the theoretical shadow intensity distribution in the scene is calculated based on Lambert's cosine law; the theoretical shadow intensity distribution is compared with the shadow intensity distribution of the actual pixels in the image to be verified, generating a local thermal residual map representing the degree of light and shadow inconsistency, and the statistical variance of this residual map is extracted as the physical geometric consistency score. The pixel-by-pixel calculation formula for the local thermal residual map is: ;

[0091] Where R(u,v): represents the local thermal residual map at pixel... The residual value at a given location. It reflects the difference between the actual observed value and the theoretical simulation value. Larger residuals often indicate the presence of anomalies (such as internal defects, surface damage, etc.). (u,v): Represents the two-dimensional pixel coordinates of the image, where... and These represent the horizontal and vertical indices, respectively; (u,v): Represents the actual pixel value captured by the infrared thermal imager or camera. The actual intensity value at that location. : Represents the thermal radiation coefficient or reflectance coefficient of the material. It is a weighting factor that determines the degree to which environmental radiation affects the theoretical intensity at that point. N(u,v): Represents the pixel. The unit normal vector of the corresponding point on the surface of the three-dimensional object. : Represents the direction vector of the ambient heat source or main light source. It is usually also a unit vector pointing towards the radiation source. max(..., 0): This is a cutoff function. Since the dot product can be negative, taking the maximum value ensures that the calculation result is not negative, i.e., shielding the area behind the radiation source from being affected.

[0092] In this embodiment, the cross-modal evidence verification link includes semantic logic common sense comparison, specifically implemented as follows: extracting multi-dimensional detection templates targeting violations of physical laws and variations in biological structures from a preset counterfactual prompt template library; jointly inputting the image to be verified and the multi-dimensional detection templates into a multimodal large language model; and using a contrastive normalization algorithm to calculate the relative logical response probability difference of the large language model when outputting affirmative and negative classification words, thereby eliminating the prior preference noise of the language model itself and outputting a clean semantic logic verification score.

[0093] In this embodiment, after the dynamic triggering of the cross-modal evidence verification link, the generation logic of the second evidence judgment result is as follows: input the physical geometric consistency score, semantic logic verification score, and metadata tampering risk value of the image to be verified into the multilayer perceptron for multi-source evidence alignment; perform Bayesian posterior probability update based on hyperbolic distance decay factor on the aligned verification confidence and the first evidence judgment result, and output the calibrated second evidence judgment result.

[0094] In this embodiment, the step of performing local tangent space nearest neighbor retrieval with radial penalty in the source feature knowledge base includes: constructing a hyperbolic feature knowledge base, which stores historical hyperbolic embedding anchor points and engine labels of samples generated by multiple mainstream AI image generation engines; to reduce the complexity of large-scale hyperbolic distance calculation, the hyperbolic embedding anchor points and historical hyperbolic embedding anchor points of the image to be examined are projected back to the local tangent space where the reference point is located through logarithmic mapping; the penalized cosine similarity with a radial modulus difference penalty term is calculated in the local tangent space, and the candidate generation engine is determined based on the engine label with the highest similarity.

[0095] In this embodiment, the construction process of the offline influence function approximation proxy network includes: during the training phase, using a high-precision influence function algorithm, offline calculation of the true gradient influence contribution value of the pre-constructed training sample set to the generation result of the evidence collection; constructing a lightweight ranking neural network as the proxy network; and training the proxy network using a pairwise ranking hinge loss function, so that the output training sample contribution ranking fits the relative magnitude relationship of the true gradient influence contribution value, thereby avoiding the computational cost of inverting massive training data matrices. The pairwise ranking hinge loss function is defined as: in, The contribution value to the actual gradient impact. For approximate output of the proxy network, For symbolic functions, These are preset boundary hyperparameters;

[0096] When ranking the contribution of the output training data: The image to be investigated and the historical training dataset are input into the offline influence function approximation proxy network after training; the proxy network forward propagation directly outputs the approximate contribution score of each historical training sample to the generation of the image to be investigated; the top N training samples with the highest approximate contribution scores are selected as candidate source data for infringement tracing or training attribution.

[0097] In this embodiment, the structured evidence report chain not only includes evidence conclusions in text format, but also includes: a visual distribution heatmap of the deviation of the image to be examined from the local reference prototype in the hyperbolic representation space; a dynamically labeled area map of physical geometric shadow residuals or local structural anomalies; and the key evidence elements of the structured evidence report chain generate tamper-proof verification codes through a hash encryption mechanism, which are then bound to timestamp information for judicial or copyright evidence preservation.

[0098] The specific implementation steps are as follows:

[0099] Step S1: Extraction of multi-scale visual feature representations and frequency domain phase fusion;

[0100] like Figure 2 As shown, the system first acquires the image to be examined with a resolution of 1024x1024.

[0101] 1. Use a pre-trained visual model with frozen weights to extract the high-dimensional global semantic representation of the image. .

[0102] 2. Simultaneously apply a high-frequency filter to extract edge artifact features and locally generated texture features in details such as hair strands and pores.

[0103] 3. Perform a Fast Fourier Transform (FFT) on the edge features to calculate their phase anomaly distribution map in the frequency domain. .

[0104] 4. The global semantic representation and the phase anomaly distribution map are fused across channels. The fusion formula is defined as follows: .in, This represents the final multi-scale visual feature representation (i.e., the cross-channel fusion result). Represents global semantic representation This is a phase anomaly distribution map. This represents the representation of local features. Represents the learnable weight matrix. This represents the activation function, while and These represent cross-channel fusion and element-wise multiplication operations, respectively. This greatly enhances the grid-like structural defects left by the generative model during the frequency domain downsampling process of the decoder, resulting in the final multi-scale visual feature representation.

[0105] Step S2: Conformal mapping and consistent embedding of non-Euclidean topology;

[0106] The system needs to represent features in Euclidean space. Embedded into curvature with negative constant The hyperbolic representation space.

[0107] 1. Calculation Hierarchical transit probability matrix in Euclidean space This is used to characterize semantic inclusion and artifact dependency relationships between feature scales.

[0108] 2. Using the manifold exponential mapping function The normalized features are projected into the Poincaré sphere to obtain the hyperbolic embedding anchor points of the image to be used for evidence. .

[0109] 3. Introduce a topological loss term based on hyperbolic divergence into the mapping optimization objective. This forces the distribution of points in hyperbolic space to approximate... .

[0110] 4. Apply a Riemann barrier boundary constraint based on adaptive decay of the eigennorm, when the eigennorm approaches the boundary ( A sharp increase is observed to prevent numerical collapse caused by extreme high-frequency noise.

[0111] Step S3: Hyperbolic Authenticity Assessment and First Evidence Determination Result

[0112] like Figure 2 The initial assessment of hyperbolic realism involves retrieving a pre-constructed set of multi-cluster local reference prototypes of real human images within the hyperbolic representation space, using hyperbolic hierarchical clustering. .

[0113] 1. Calculate anchor points Nearest Neighbor Reference Prototype The hyperbolic geodesic distance between them is used to derive the geodesic projection distortion. .

[0114] 2. Read The radial coordinates are used to calculate the edge expansion penalty value using the exponential growth property of hyperbolic volume. The formula is: .in, This represents the edge expansion penalty value. This is the scaling factor (used to control the base strength of the penalty term). This represents the hyperbolic curvature parameter (used to define the geometric curvature properties of hyperbolic space), while and These represent exponential functions and norm operations, respectively. This mechanism forcibly removes and amplifies extremely subtle, hidden distortions in highly realistic portraits.

[0115] 3. The hyperbolic truth deviation confidence level is calculated by combining the two results. (Range 0-1). Based on this, the system generates the first evidence assessment result: suspected AI generation.

[0116] Step S4: Adaptive dynamic adjustment and scheduling of fuzzy decision intervals;

[0117] The dynamic evidence collection scheduling module (event-driven control engine) begins operation. The system acquires the negative curvature constant of the current hyperbolic representation space and the distribution density of the local reference prototype. The nonlinear distortion coefficients were calculated. Using the Riemannian metric tensor as a constraint, the system dynamically scales to obtain the current fuzzy decision interval. .

[0118] because If the feature falls strictly within this fuzzy range, the system determines that the single-modal hyperbolic feature cannot provide an absolutely confident conclusion and immediately triggers the cross-modal forensic verification link dynamically.

[0119] Step S5: Cross-modal forensics verification link and final judgment;

[0120] The system allocates computing power to the time-consuming verification links:

[0121] 1. Physical-Geometric Consistency Verification: A monocular depth estimation and lighting decoupled network is used to extract the depth map of the human face scene and the global ambient lighting direction. Theoretical shadows are calculated based on Lambert's cosine law, and compared with the original image to generate a local thermal residual map (revealing a physically incompatible residual in the shadow below the nose). Variance is extracted to obtain the physical-geometric consistency score. .

[0122] 2. Semantic and Logical Common Sense Comparison: Detection templates are extracted from a counterfactual cue template library and input into a multimodal large language model in conjunction with images. A contrastive normalization algorithm is used to calculate the relative logical response probability difference between affirmative and negative categorical terms (eliminating the prior bias of the large model towards "compliant cue words"), outputting a clean semantic and logical verification score. .

[0123] 3. The risk value of metadata tampering is input into a multilayer perceptron for multi-source evidence alignment, and then Bayesian posterior probability update based on hyperbolic distance decay factor is performed with the first judgment result. The confidence level of the final calibration output of the second evidence judgment result jumps to [value missing]. .

[0124] Based on this, the system outputs the final evidence conclusion for the image to be examined: it is highly confirmed that the image was generated by artificial intelligence.

[0125] Step S6: Source tracing and attribution, and tracing infringement through proxy networks;

[0126] 1. Local tangent space nearest neighbor search with radial penalty: Through logarithmic mapping Projecting back to the local tangent space (Euclidean space) where the reference point is located significantly reduces retrieval complexity. In a hyperbolic feature knowledge base containing historical anchor points of mainstream engines, a penalized cosine similarity with a radial modulus difference penalty term is calculated to lock the candidate generation engine as "Midjourney V6".

[0127] 2. Offline Influence Function Approximation Proxy Network Invocation: The system inputs images into a lightweight proxy network trained with a pairwise sorting hinge loss function. The network avoids the massive inverse operations of the data matrix, instantly forward propagating and outputting approximate contribution scores of historical training data. The top-3 high-contribution real images are extracted as the training data contribution ranking results (for copyright infringement tracing).

[0128] Step S7: Generate solid evidence from the structured forensic report chain;

[0129] The report assembly module integrates the above processes into a structured forensic report chain. The report includes:

[0130] A visualization heatmap of the anchor point deviation distribution in a hyperbolic Poincaré disk.

[0131] Dynamically labeled area map of physical geometric shadow residual anomalies.

[0132] Finally, the system uses the SHA-256 algorithm to generate an immutable verification code by hashing the feature anchors, engine tags, traceability sorting, and timestamp information, and outputs a judicial-grade test report with an anti-counterfeiting traceability electronic seal.

[0133] Example 2: Adaptive Dynamic Adjustment and Cross-Modal Scheduling Method for Fuzzy Decision Intervals Based on Riemannian Metric Tensor and Negative Curvature; This example focuses on the dynamic forensics scheduling module of the present invention, detailing how the system calculates nonlinear distortion coefficients using the unique geometric properties of hyperbolic space when facing massive concurrent detection requests, and dynamically generates fuzzy decision intervals with Riemannian metric tensors as constraints. This mechanism can accurately identify high-difficulty samples on the "edge of true / false decision" and automatically schedule high-computing-power cross-modal verification links for them, thereby perfectly achieving a dynamic balance between forensics detection accuracy and system concurrent computing throughput.

[0134] The specific implementation steps are as follows:

[0135] Step S1: Quantization of local feature distribution density and nonlinear distortion coefficient;

[0136] like Figure 3 As shown, after the system obtains the visual features of the image to be used for evidence collection, it first retrieves the surrounding multi-cluster local reference prototypes in the hyperbolic representation space. The system uses a density estimation method with hyperbolic distance attenuation characteristics to calculate the anchor points. Distribution density of the neighborhood of the manifold :

[0137] ;

[0138] Where N represents the total number of local reference prototypes retrieved in the hyperbolic representation space. The distance is the hyperbolic geodesic distance. The hyperbolic neighborhood bandwidth parameter is based on local density.

[0139] Subsequently, the system extracts the negative curvature constant of the current hyperbolic space. (This system is set up) To quantify the extreme distortion effect at the hyperbolic space boundary, the system innovatively constructs a nonlinear distortion coefficient as a characteristic metric. :

[0140] ;

[0141] in, This is the global scaling factor. Let Riemann norm be the Riemann norm on the tangent space of the manifold. To control for the curvature-adaptive exponent that controls distortion as it increases with radial distance, this formula ensures that the distortion coefficient increases significantly in regions where the reference prototype is sparse and close to the manifold boundary.

[0142] Step S2: Dynamically generate fuzzy decision intervals constrained by Riemannian metric tensors;

[0143] Obtaining nonlinear distortion coefficients Then, the system uses this as a benchmark and dynamically generates a decision threshold by combining the Riemannian metric tensor under conformal mapping. In the Poincaré sphere model, the point... The Riemannian metric tensor matrix at that location is Among them, conformal factor , It is a unit array.

[0144] The system sets a benchmark Euclidean criterion center point. Furthermore, by introducing the trace and determinant of the metric tensor, the fuzzy decision interval is dynamically scaled to generate the fuzzy decision interval. Its unique formula is defined as:

[0145] ;

[0146] ;

[0147] in, Represents the sample points to be collected; This represents the nonlinear distortion coefficient corresponding to that point; The midpoint of the Poincaré sphere model The Riemannian metric tensor matrix at that location; It is the conformal factor that enables metric transformation; Represents the identity matrix; Refers to the curvature parameter of hyperbolic space; Represents the Euclidean norm of the sample points; It is the system's preset benchmark Euclidean determination center point; The order of the tensor matrix represents the dimension of the feature space. and Let represent the trace and determinant of the metric tensor matrix, respectively; the final... and These represent the lower and upper bounds of the fuzzy decision interval that the system dynamically stretches or shrinks based on the metric tensor, respectively.

[0148] The core physical meaning of this dynamic interval is: when the sample to be examined is located in the central region of dense features ( When the sample falls into the sparse region at the edge of hyperbolic space (i.e., the region where highly realistic generated images are often pushed), the metric tensor expands rapidly, causing the interval to be widened significantly, forcing the system to be more cautious about such "difficult cases".

[0149] Step S3: Triggering of event-driven cross-modal verification link;

[0150] The system calculates the hyperbolic realism deviation confidence level of the current image to be used for evidence collection. And execute the condition judgment mechanism:

[0151] 1. If The system directly outputs the first evidence determination result as "confirmed AI generation", thus blocking subsequent calculations.

[0152] 2. If It directly outputs the first evidence judgment result as "real image" and blocks subsequent calculations.

[0153] 3. If and only if When the system identifies the current sample as a high-fidelity adversarial sample, it triggers an event interrupt and routes the task to the dynamic forensics scheduling module. The scheduling module then dynamically allocates GPU cluster computing power to the task and initiates physical geometric consistency verification and semantic logical common sense comparison in parallel.

[0154] Step S4: Bayesian posterior probability update based on hyperbolic distance decay factor;

[0155] The calculation is completed in the cross-modal verification link, and the physical-geometric consistency score is extracted. Semantic logic verification score Subsequently, the system does not simply perform numerical averaging, but instead uses a Bayesian posterior probability update based on the hyperbolic distance decay factor to generate a second evidence determination result.

[0156] The system first defines a confidence attenuation factor based on hyperbolic geodesic distance. Subsequently, based on the first determination probability... For the prior probability, perform a posterior update using multi-source evidence:

[0157] ;

[0158] in, Represents the confidence decay factor based on hyperbolic geodesic distance; This parameter represents the decay rate of control reliability as it decreases with distance. The hyperbolic geodesic distance between the sample point and the nearest cluster center; Represents the sample points to be collected; Refers to the cluster center in hyperbolic space that is closest to the sample point; Indicates the system decision hypothesis The prior probability of its validity; The classification confidence level is derived from the initial determination based on hyperbolic space. This represents the posterior probability made by the system after updating through the fusion of multi-source evidence; This represents cross-modal evidence that incorporates multidimensional information; The modality index refers to the multi-source evidence that participates in the posterior update; and These represent the physical layer modality and the semantic layer modality in multi-source evidence, respectively. Representing the The support score of each feature mode for the judgment result.

[0159] Through the above mathematical deduction, when the system integrates physical geometric consistency and semantic logical common sense evidence, it can dynamically suppress unreliable feature weights caused by excessive hyperbolic distance, and finally output a highly calibrated final evidence conclusion, greatly reducing the blind spot of single-modal detection.

[0160] Example 3: Cross-modal forensic verification method based on physical geometric consistency and contrast normalization semantic comparison. This example focuses on the cross-modal forensic verification link of the present invention. It discloses in detail how the system accurately identifies the microscopic inconsistencies in physical common sense and biological structure of high-fidelity AI-generated images when the single-modal hyperbolic features of the image to be verified fall into the "fuzzy decision interval" through in-depth physical light and shadow calculation and large model semantic logic debiasing algorithm, which greatly reduces the judgment distortion rate in complex scenes.

[0161] The specific implementation steps are as follows:

[0162] Step S1: Physical-geometric consistency verification and light and shadow residual quantization based on Lambert's cosine law;

[0163] When the dynamic scheduling module triggers cross-modal verification, the system first performs physical and geometric consistency verification on the image. This applies to the image to be used for evidence collection. The system calls a lightweight monocular depth estimation model and an illumination decoupled network to output scene depth maps respectively. Pixel-level surface normal vector and the estimated global ambient lighting direction vector .

[0164] Based on Lambert's cosine law, the system calculates the theoretical diffuse shadow intensity distribution for each pixel in the scene. :

[0165] ;

[0166] in, This represents the decoupled surface albedo. Subsequently, the system extracts the grayscale shadow intensity of the actual pixels in the image. The absolute deviation between the two is calculated to generate a local thermal residual map representing the degree of light and shadow inconsistency. :

[0167] ;

[0168] in, For salient foreground masks (such as a person's face or the main object). The smoothing constant is used. The system extracts the statistical variance and mean peak value of the high-frequency outlier regions in the residual plot, normalizes them, and outputs the physical-geometric consistency score. (The higher the score, the more serious the physical violation).

[0169] Step S2: Multi-dimensional counterfactual template extraction and comparison with large language models using normalization;

[0170] While performing physical verification, the system performs semantic logic common sense comparison in parallel. The system has a pre-built counterfactual prompt template library. For the subject recognition result of the current image (such as "person"), it extracts multi-dimensional detection templates for violations of physical laws (such as "whether the number of fingers is abnormal" or "whether the water reflection is symmetrical") and biological structural variations (such as "whether the pupil shape is regular").

[0171] To eliminate the prior preference noise of "conforming to user prompts" that is prevalent in multimodal large language models (MLLM) during response, the system employs a unique contrastive normalization algorithm. For the same detection point, the system constructs a positive prompt. and negative prompts By combining the image input to the large model, the original log probabilities of the classification term "contrary to common sense" in the model output are obtained separately. .

[0172] The system calculates the clean semantic logic check score using the following comparison normalization formula. :

[0173] ;

[0174] in, This refers to the inherent bias compensation term statistically derived from massive benchmark tests of the language model. The algorithm successfully eliminates the "illusionary" bias inherent in the language model itself, outputting a highly objective logical judgment score.

[0175] Step S3: Alignment of Multi-Source Evidence and Calibration of Final Evidence Conclusion

[0176] The system extracts the EXIF ​​information and underlying coding features of the image to be used for evidence collection, and calculates the metadata tampering risk value. .

[0177] Subsequently, as Figure 4 As shown, the system will Physical consistency score Semantic logic score and Metadata risk scores are concatenated into feature vectors, which are then input into a pre-trained multilayer perceptron (MLP) for multi-source evidence alignment, mapping them to a unified confidence space to obtain the aligned verification confidence. .

[0178] Finally, the system retrieves the unresolved first forensic determination result (i.e., the unimodal hyperbolic confidence level) from "Example 2" and compares it with... A Bayesian posterior probability update based on the hyperbolic distance decay factor is performed, ultimately outputting a highly robust second forensic determination result. This determination result, as the final forensic conclusion for the image to be examined, is sent to the next step, the source tracing and report assembly module.

[0179] Example 4: Source attribution and judicial evidence consolidation method based on dimensionality reduction retrieval and approximate proxy network. This example focuses on the source attribution engine and report assembly module of the present invention. It discloses in detail how to overcome the computing power bottleneck of large-scale non-Euclidean space retrieval and massive data matrix inverse operation after the system determines that the image is generated by AI, so as to achieve extremely fast generation engine locking and infringement training data tracing, and finally generate a structured evidence report chain with judicial validity.

[0180] The specific implementation steps are as follows:

[0181] Step S1: Logarithmic mapping dimensionality reduction and local tangent space nearest neighbor search with radial penalty;

[0182] like Figure 5 As shown, once the image to be used for evidence is confirmed to be AI-generated, the system obtains its hyperbolic embedding anchor point in the Poincaré sphere model. To quickly retrieve candidate generation engines from a hyperbolic feature knowledge base containing tens of millions of samples, the system reduces the complexity of large-scale hyperbolic distance calculations by using the origin... Using the manifold logarithmic mapping as the reference point anchor point Project back to the local tangent space of the reference point (i.e., standard Euclidean space):

[0183] ;

[0184] Historical sample anchors stored in the knowledge base It has also been pre-projected as a tangent space vector In the local tangent space, the system not only calculates the conventional cosine similarity but also innovatively introduces a radial modulus difference penalty term to preserve the geometric properties of hyperbolic space edge dilation. The system calculates the penalized cosine similarity between the vector to be verified and the historical vectors. :

[0185] ;

[0186] In the formula, This is the radial penalty weight. The system performs nearest neighbor retrieval based on this similarity, selecting the engine label corresponding to the highest similarity, thereby quickly locking in candidate generation engines.

[0187] Step S2: Approximate surrogate network attribution based on offline influence function of pairwise ordering hinge loss;

[0188] like Figure 5 As shown, to trace whether the generative model "remembered" or infringed on copyrighted real images during training, the system invoked a pre-trained offline influence function approximation proxy network. Traditional methods of calculating the true gradient contribution value require inverse operations on massive training data matrices, consuming enormous computational resources. This system, in the offline phase, utilizes a high-precision influence function algorithm to pre-calculate the true gradient contribution value for a small batch of samples. Subsequently, a lightweight sorting neural network was constructed. As a proxy network, and employing a pairwise ordered hinge loss function Conduct training:

[0189] ;

[0190] in, For the training sample pair set, For sorting margin, The ranking margin is a hyperparameter that forces the proxy network to not only rank the samples correctly in terms of relative influence, but also that the difference between two predicted scores must be greater than or equal to the set boundary value. It refers to a lightweight ranking neural network that is responsible for quickly predicting the influence scores between samples; Represents the sample to be tested, equivalent to This is the benchmark for calculating influence; and These represent two different training samples in the set that are being compared. This indicates that the agent network compares two training samples with the test sample. The difference in predicted influence; The function represents the sign of influence. It returns 1 when the internal value is positive, -1 when it is negative, and 0 when it is 0. It is used to extract the relative level (direction) of true influence. This represents the actual gradient impact value pre-calculated using a high-precision algorithm during the offline phase; This indicates that two training samples are paired with a test sample. The difference in the true influence.

[0191] In actual evidence collection and source tracing, the system inputs the image to be investigated and a suspected copyright image library into the proxy network. This network, through simple forward propagation, directly outputs an approximate score of the contribution of each historical training sample to the generation of the image to be investigated, completely avoiding computational bottlenecks. The system selects the top N training samples with the highest scores as candidate source data for infringement tracing.

[0192] Step S3: Multidimensional evidence fusion and generation of structured evidence report chain;

[0193] Once all the above detection and tracing processes are completed, the report assembly module begins to integrate the information and generate a structured forensic report chain. The report not only includes plain text forensic conclusions such as "determined to be AI-generated" and "candidate engine is the Stable Diffusion model," but also rigidly embeds the following multimodal visual evidence:

[0194] 1. A heatmap showing the distribution of the image to be examined deviating from the local reference prototype in the hyperbolic representation space.

[0195] 2. Dynamically labeled area map of physical geometric shadow residuals or local structural anomalies.

[0196] 3. Comparison of thumbnails of the top-ranked source data in infringement tracing.

[0197] Finally, the system packages the key evidence elements of the structured evidence report chain (including feature anchor vector, hyperbolic authenticity deviation confidence level, cross-modal verification score, and proxy network output results), generates an immutable verification code through the SHA-256 secure hash algorithm and hash encryption mechanism, and strongly binds it with the timestamp information provided by the National Time Service Center, automatically synchronizing it to the judicial evidence storage blockchain node to form a complete evidence closed loop.

[0198] Example 5: System training process, experimental verification, and ablation experiment analysis

[0199] This embodiment details the training process of the system proposed in this invention, as well as experimental verification, performance analysis, and ablation experiments on a large-scale hybrid generated dataset. Through rigorous quantitative data, this embodiment demonstrates the significant advantages of this invention in cross-domain generalization, robustness, source tracing efficiency, and system throughput.

[0200] The system training process is divided into two stages:

[0201] Phase 1: Pre-training of the hyperbolic representation space

[0202] Only the manifold map and classification head were trained. A "known domain" dataset containing various generative models based on adversarial generative networks was used. Data augmentation strategies included random cropping, horizontal flipping, Gaussian blur (σ∈[0.1,2.0]), and JPEG compression (compression strength Q∈[30,100]). This augmentation forced the model to learn robust geometric features rather than fragile high-frequency artifacts. The optimizer used was Riemannian Adam (RADAM) (a Riemannian optimizer) with a learning rate of 1e-4.

[0203] Phase Two: Fine-tuning of the dynamic evidence collection scheduling strategy

[0204] With a fixed backbone network and classification head, the decision gating parameters (such as the upper and lower bounds of the fuzzy decision interval) of the event-driven dynamic forensics scheduling module (i.e., the control engine within the system) are trained. and A fine-tuned set containing "hard samples" (i.e., samples falling near the fuzzy decision boundary) is constructed. A reinforcement learning strategy is employed, using detection accuracy and the computational cost of cross-modal forensics verification links as the reward function. Optimize scheduling strategies to minimize the use of expensive external tools while ensuring accuracy.

[0205] Experiment setup and dataset construction:

[0206] To fully verify the effectiveness, robustness, and operational efficiency of the AI-generated image dynamic forensics method and system based on hyperbolic space proposed in this invention, systematic testing was conducted.

[0207] Hardware environment: All experiments were conducted on a high-performance server equipped with four NVIDIA A100 (80GB) GPUs, and the deep learning framework used was PyTorch 2.0.

[0208] Hybrid Dataset: To simulate realistic open-world scenarios, a hybrid dataset containing 200,000 images was constructed. The ground truth domain was sampled from ImageNet, FFHQ, and COCO datasets. The known generative domain includes images generated by multiple adversarial generative network models such as ProGAN, StyleGAN2, and BigGAN, used in the training phase. The unknown generative domain includes images generated by current mainstream diffusion-based models such as LatentDiffusion, Adversarial Diffusion, DALL-E 2, and Midjourney v5, used only in the testing phase to evaluate generalization ability.

[0209] Cross-domain generalization capability analysis:

[0210] This invention compares the proposed method with current mainstream baseline models.

[0211] Performance in known domains: In known domains such as StyleGAN, this method performs on par with other state-of-the-art (SOTA) methods, maintaining an accuracy above 98%. Breakthrough in unknown domains: For unseen diffusion models, traditional Euclidean space-based methods often fail. Thanks to the hyperbolic embedding anchor mapping and the topological loss term based on hyperbolic divergence described in Example 2, this method can better capture the hierarchical structural differences of the generated images in the latent space. Experimental results show that on datasets generated by unknown diffusion models, such as... Figure 6 As shown, the detection accuracy (ACC) of this method reaches 89.2%, which is a performance improvement of 14.7 percentage points compared to the 74.5% of the Euclidean space-based ResNet-50 baseline model. Figure 6 As shown, under hyperbolic space projection, the real image and the generated images from different sources exhibit clear cluster separation, while Euclidean space projection shows severe aliasing.

[0212] Anti-interference capability under complex attacks:

[0213] To verify the robustness of the model in social media dissemination scenarios, various post-processing attacks were applied to the test images.

[0214] Performance Maintenance: Thanks to the cross-modal forensic verification (comparison of physical geometric consistency and semantic logical common sense) in Example 3, semantic consistency features compensate for the defects caused by the destruction of pixel-level features. For example... Figure 6 As shown, under the combined attack of JPEG compression (intensity Q=50) and Gaussian blur (radius 0.5), the detection AUC of our method decreases by less than 3% (from 99.1% to 96.2%), while the performance of methods relying solely on frequency domain artifacts decreases by more than 15%. This demonstrates that our method has extremely strong anti-interference capabilities and practical application potential.

[0215] Efficiency and accuracy of attribution:

[0216] For tasks requiring the identification of specific image sources, the performance of the offline influence function approximation proxy network module in Example 4 was verified. Computation speed: Traditional attribution methods based on data matrix inversion take approximately 90 seconds on a single image. This method utilizes a lightweight proxy network trained with a pairwise sorted hinge loss function, such as... Figure 6As shown, locating the training source of a single image on a single NVIDIA A100 takes only 45ms, representing a speedup of approximately 2000 times, meeting the demands of large-scale real-time stream processing. In a source tracing task involving 20 different generators, this method achieves a Top-1 accuracy of 76.8% and a Top-10 hit rate exceeding 85%, effectively solving the challenge of distinguishing subtle differences in the generated source fingerprints.

[0217] System gains from dynamic forensics scheduling:

[0218] The advantages of event-driven dynamic forensics scheduling over single-model inference were verified. Throughput testing and overall performance: Under simulated high-concurrency requests, the dynamic forensics scheduling module dynamically allocated computing resources based on image confidence. For 60% of simple samples, scheduling was performed to a lightweight single-modal hyperbolic geometry discrimination link; for 40% of "difficult samples," scheduling was performed to a cloud-based cross-modal forensics verification link. Compared to using a heavy-duty verification model across the board, the system's overall average inference latency was reduced by 65%, GPU memory usage was reduced by 40%, and the overall detection accuracy did not show a statistically significant decrease.

[0219] Ablation experiment:

[0220] To verify the independent contribution of each core module to the overall performance of this invention, this invention uses data generated against diffusion as the test set and conducts a comparative experiment on the basis of the ResNet-50 backbone network, progressively stacking modules:

[0221] 1. Experimental group A serves as the baseline: Only the ResNet-50 model based on Euclidean space metrics is used, without introducing cross-modal verification and dynamic scheduling. Experimental results show that its detection accuracy (ACC) is only 74.5%, and the source tracing Top-1 accuracy is 42.1%, indicating that traditional methods have limited performance when facing attacks from unknown domains.

[0222] 2. Experimental Group B: Hyperbolic representation mapping and a topological loss term based on hyperbolic divergence were introduced on top of the baseline. Results showed a significant jump in detection accuracy to 86.3%, and Top-1 accuracy for source tracing improved to 68.4%. This substantial improvement confirms the crucial role of non-Euclidean geometry in capturing the hierarchical structure of generated features.

[0223] 3. Experimental Group C: Based on Experimental Group B, a cross-modal evidence verification link was further added. Thanks to the introduction of multimodal semantic features, the model's detection accuracy was further improved to 87.2%, and the Top-1 accuracy for source tracing reached 71.2%.

[0224] 4. Experimental Group D: Based on Experimental Group C, a dynamic forensic scheduling and approximate proxy network source tracing module were introduced. Experiments show that although the detection accuracy remained at 89.2%, it demonstrated a significant advantage in source tracing tasks, achieving the highest Top-1 accuracy of 76.8% across the group. This indicates that the scheduling strategy, while ensuring high detection accuracy, greatly optimized the attribution capability of specific sources and the system's operational efficiency.

[0225] The ablation experiments show that hyperbolic space metric is key to improving generalization performance in unknown domains, while multimodal verification further enhances robustness.

[0226] Example 6 discloses an AI-generated image dynamic forensics system based on hyperbolic space, comprising:

[0227] The feature processing module is used to acquire the image to be used for evidence collection and extract the multi-scale visual feature representation of the image to be used for evidence collection.

[0228] Anchor point acquisition module is used to embed the multi-scale visual feature representation into a hyperbolic representation space with negative constant curvature based on the topological consistency constraint of hyperbolic divergence, so as to obtain the hyperbolic embedding anchor points of the image to be examined.

[0229] The confidence calculation module is used to calculate the geodesic projection distortion and edge dilation penalty value of the hyperbolic embedded anchor point based on the multi-cluster local reference prototype constructed from the real image, and to fuse the two into a hyperbolic realism deviation confidence.

[0230] The hyperbolic authenticity assessment module is used to generate a first evidence judgment result based on the hyperbolic authenticity deviation confidence level; when the hyperbolic authenticity deviation confidence level falls into the fuzzy decision range based on the hyperbolic curvature parameter adaptive dynamic adjustment, the cross-modal evidence verification link is dynamically triggered, and a second evidence judgment result is generated by combining the physical geometric consistency and semantic logic common sense comparison results.

[0231] The judgment output module is used to output the final evidence conclusion of the image to be examined based on the first evidence judgment result and the second evidence judgment result. When the final evidence conclusion is determined to be an artificial intelligence generated image, the hyperbolic embedding anchor point is input into the source analysis module, and a local tangent space nearest neighbor search with radial penalty is performed in the source feature knowledge base to determine the candidate generation engine. At the same time, the offline influence function approximate proxy network is called to output the contribution ranking of the training data, and the first evidence judgment result, the second evidence judgment result and the candidate generation engine are fused to generate a structured evidence report chain.

[0232] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic image forensics based on hyperbolic space generated by AI, characterized in that, The method includes: Acquire the image to be used for evidence collection, and extract the multi-scale visual feature representation of the image to be used for evidence collection; Based on the topological consistency constraint of hyperbolic divergence, the multi-scale visual feature representation is embedded into a hyperbolic representation space with negative constant curvature to obtain the hyperbolic embedding anchor point of the image to be examined. Based on a multi-cluster local reference prototype constructed from real images, the geodesic projection distortion degree and edge dilation penalty value of the hyperbolic embedded anchor point are calculated, and the two are fused into a hyperbolic realism deviation confidence degree. A first evidence determination result is generated based on the hyperbolic authenticity deviation confidence level; when the hyperbolic authenticity deviation confidence level falls into the fuzzy decision range based on the hyperbolic curvature parameter adaptive dynamic adjustment, a cross-modal evidence verification link is dynamically triggered, and a second evidence determination result is generated by combining the physical geometric consistency and semantic logic common sense comparison results. Based on the first and second evidence collection judgment results, the final evidence collection conclusion of the image to be collected is output. When the final evidence collection conclusion is determined to be an artificial intelligence generated image, the hyperbolic embedding anchor point is input into the source analysis module, and a local tangent space nearest neighbor search with radial penalty is performed in the source tracing feature knowledge base to determine the candidate generation engine. At the same time, the offline influence function approximation proxy network is called to output the contribution ranking of the training data, and the first evidence collection judgment result, the second evidence collection judgment result and the candidate generation engine are fused to generate a structured evidence collection report chain.

2. The AI-generated image dynamic forensics method based on hyperbolic space according to claim 1, characterized in that, The multi-scale visual feature representation includes: A frozen visual pre-trained model is used to extract global semantic representations; High-frequency filters are used to extract edge artifact features and locally generated texture features from the images to be examined. The phase anomaly distribution map of edge artifact features in the frequency domain is calculated, and the global semantic representation is fused with the phase anomaly distribution map across channels. The formula for cross-channel feature fusion is as follows: ； in, The output features after fusion For global semantic features, Edge features; and These represent the Fourier transform and the inverse Fourier transform, respectively. Frequency domain amplitude information representing edge features; This is a phase anomaly distribution diagram; e represents the base of the natural logarithm; The imaginary unit; This represents element-wise multiplication; Indicates channel splicing operation; by... The amplitude information is preserved and introduced. Phase modulation is performed, via After reconstruction and By fusing these components, an enhanced representation of anomalous regions can be achieved.

3. The AI-generated image dynamic forensics method based on hyperbolic space according to claim 1, characterized in that, The process of obtaining hyperbolic embedding anchor points for the image to be used for evidence includes: Calculate the hierarchical transitivity probability matrix of multi-scale visual feature representations in Euclidean space to characterize the topological relationships of semantic inclusion and artifact dependence between feature scales; By using manifold logarithmic mapping or exponential mapping function, multi-scale visual feature representations are mapped to the Poincaré sphere model or Lorentz hyperboloid model. A topological loss term based on hyperbolic divergence is introduced into the optimization objective of the mapping to force the point integral distribution in hyperbolic space to approximate the hierarchical transfer probability matrix, thereby achieving topologically consistent embedding in non-Euclidean space. The formula for calculating the topology loss term is as follows: ; in, Topology loss term, Indices representing spatial locations or semantic units in multi-scale feature maps, used to identify the first... Each corresponding pair of feature points; and These represent the Euclidean features extracted at the index position under different scale branches; Let f(x) denote a generating function that is strictly convex in hyperbolic space; This represents the Riemann gradient of the generating function on the hyperbolic manifold; Refer to Using as a base point Logarithmic mapping function of manifold onto the tangent plane; Indicates at the base point The Riemann inner product calculated in the local tangent space; During the mapping process, a Riemann barrier boundary constraint based on adaptive decay of the feature norm is applied. The formula for the Riemann barrier boundary constraint is as follows: ;in This represents the Riemann barrier boundary constraint term used to prevent numerical anomalies. Regularization weight coefficients used to control the intensity of boundary penalties; The absolute value of the negative constant curvature of the hyperbolic representation space; represents the feature point vector in hyperbolic space; This represents the Euclidean norm of the feature point vector; A potential barrier triggering boundary threshold is set to prevent features from collapsing toward the boundary at infinity of the manifold. This indicates the truncation penalty function.

4. The AI-generated image dynamic forensics method based on hyperbolic space according to claim 1, characterized in that, The method for constructing the multi-cluster local reference prototype includes: Acquire a training set of real images from multiple scenes and map it to the hyperbolic representation space; Leveraging the capacity advantage of hyperbolic space for hierarchical tree structures, and based on the hyperbolic geodesic distance between feature points, we perform bottom-up hyperbolic hierarchical clustering of feature points in real images to fit the tree manifold distribution of "major class-subclass-instance" in real scenes; Nodes at different levels are extracted into multi-cluster local reference prototypes, and hyperbolic neighborhood bandwidth parameters based on local density are assigned to each local reference prototype according to the hyperbolic geodesic distance.

5. The AI-generated image dynamic forensics method based on hyperbolic space according to claim 4, characterized in that, The acquisition of the geodesic projection distortion degree and edge dilation penalty value includes: obtaining the hyperbolic geodesic distance between the hyperbolic embedding anchor point and the feature points of the multi-cluster local reference prototype, and mapping the distance to the geodesic projection distortion degree, with the greater the distance, the higher the distortion degree; Utilizing the geometric property that the volume of hyperbolic space increases exponentially with radius, a nonlinear amplification factor is calculated based on the radial coordinate position of the anchor point in space to adjust the range of influence of distortion. When the anchor point approaches the boundary of the hyperbolic representation space, the edge expansion penalty value increases exponentially, providing a quantitative basis for the subsequent dynamic scaling of the fuzzy decision interval based on the Riemann metric tensor, and realizing the geometric consistency constraint of cross-scale, multi-cluster prototypes.

6. The AI-generated image dynamic forensics method based on hyperbolic space according to claim 1, characterized in that, The determination of the fuzzy decision interval includes: Obtain the negative curvature constant of the current hyperbolic representation space and the distribution density of the local reference prototype; The nonlinear distortion coefficient is calculated based on the negative curvature constant to measure the characteristic. Using the nonlinear distortion coefficients and the distribution density, and with the Riemann metric tensor as a constraint, the upper and lower bound thresholds of the fuzzy decision interval are dynamically scaled.

7. The AI-generated image dynamic forensics method based on hyperbolic space according to claim 1, characterized in that, The cross-modal forensics verification link includes: Physical and geometric consistency verification specifically involves extracting the scene depth map, surface normal vector, and global ambient lighting direction of the image to be examined using a monocular depth estimation model and an illumination decoupling network. Theoretical shadow intensity distribution in the scene is calculated based on Lambert's cosine law: By comparing the theoretical shadow intensity distribution with the shadow intensity distribution of the actual pixels in the image to be examined, a local thermal residual map representing the degree of light and shadow inconsistency is generated. The statistical variance of this residual map is extracted as the physical geometric consistency score. The pixel-by-pixel calculation formula for the local thermal residual map is as follows: ;in, Indicates the local thermal residual map at the pixel level. The residual value at the location; (u,v) represents the two-dimensional pixel coordinates of the image, where and These represent the horizontal and vertical indices, respectively; (u,v) represents the actual image captured by the infrared thermal imager or camera at a pixel count. The actual intensity value at that location; N(u,v) represents the thermal emissivity or reflectivity of the material; N(u,v) represents the pixel. The unit normal vector of a point on the surface of the corresponding 3D object; This represents the direction vector of the ambient heat source or main light source; max(..., 0) represents the cutoff function. Semantic and logical common sense comparison, specifically: Extract multi-dimensional detection templates targeting violations of physical laws and variations in biological structures from a pre-set counterfactual prompt template library; The image to be used for evidence collection is combined with the multi-dimensional detection template and input into the multimodal large language model; The contrastive normalization algorithm is used to calculate the difference in relative logical response probabilities when the large language model outputs affirmative and negative categorical words, thereby eliminating the prior bias noise of the language model itself and outputting a clean semantic logic verification score.

8. The AI-generated image dynamic forensics method based on hyperbolic space according to claim 7, characterized in that, The second evidence collection and determination results include: The physical geometric consistency score, semantic logic verification score, and metadata tampering risk value of the image to be examined are input into the multilayer perceptron for multi-source evidence alignment. The aligned verification confidence and the first evidence determination result are updated using Bayesian posterior probability based on hyperbolic distance decay factor, and the calibrated second evidence determination result is output.

9. The AI-generated image dynamic forensics method based on hyperbolic space according to claim 1, characterized in that, The construction process of the offline influence function approximation proxy network includes: During the training phase, a high-precision influence function algorithm is used to calculate the contribution value of the real gradient influence of the pre-constructed training sample set to the generation result of the evidence collection offline; a lightweight ranking neural network is constructed as a proxy network. The proxy network is trained using a pairwise sorting hinge loss function, which makes the output training samples' contribution ranking fit the relative magnitude of the actual gradient's influence contribution values, thereby avoiding the computational cost of inverting massive training data matrices. This pairwise sorting hinge loss function is defined as follows: ; in, Represents the pairwise sorting hinge loss function. These represent two indices in the sample set, used to form a pairwise comparison relationship; These represent the actual gradient contribution values ​​of the corresponding samples, which are used as supervision signals for ranking. This represents an approximate prediction of the agent network's contribution to the aforementioned impact. This is a sign function used to characterize the actual sorting direction; its output is... , or ; These are preset boundary hyperparameters used to control the relaxation level of the sorting interval; This represents a hinge-like truncation operation, used to incur a loss only when the sorting constraint is violated. When ranking the contribution of the output training data, the image to be investigated and the historical training dataset are input into the offline influence function approximation proxy network after training. The proxy network forward propagation directly outputs the approximate contribution score of each historical training sample to the generation of the image to be investigated. The top N training samples with the highest approximate contribution scores are selected as candidate source data for infringement tracing or training attribution.

10. A dynamic image forensics system based on hyperbolic space generated by AI, characterized in that, include: The feature processing module is used to acquire the image to be used for evidence collection and extract the multi-scale visual feature representation of the image to be used for evidence collection. Anchor point acquisition module is used to embed the multi-scale visual feature representation into a hyperbolic representation space with negative constant curvature based on the topological consistency constraint of hyperbolic divergence, so as to obtain the hyperbolic embedding anchor points of the image to be examined. The confidence calculation module is used to calculate the geodesic projection distortion and edge dilation penalty value of the hyperbolic embedded anchor point based on the multi-cluster local reference prototype constructed from the real image, and to fuse the two into a hyperbolic realism deviation confidence. The hyperbolic authenticity assessment module is used to generate a first evidence determination result based on the hyperbolic authenticity deviation confidence level. When the hyperbolic authenticity deviates from the confidence level and falls into the fuzzy decision range based on the adaptive dynamic adjustment of the hyperbolic curvature parameter, the cross-modal evidence collection and verification link is dynamically triggered, and a second evidence collection and judgment result is generated by combining the physical geometric consistency and semantic logical common sense comparison results. The determination output module is used to output the final evidence conclusion of the image to be evidenced based on the first evidence determination result and the second evidence determination result. When the final evidence conclusion is determined to be an artificial intelligence generated image, the hyperbolic embedding anchor point is input into the source analysis module, and a local tangent space nearest neighbor search with radial penalty is performed in the source feature knowledge base to determine the candidate generation engine. Simultaneously, the offline influence function is invoked to approximate the contribution ranking of the training data output by the proxy network, and the first evidence judgment result, the second evidence judgment result, and the candidate generation engine are integrated to generate a structured evidence report chain.

Images