Multi-mode AI content security risk traceability and identification detection system

Abstract

The invention belongs to the technical field of digital content security, and discloses a multi-mode AI content security risk traceability and identification detection system. Through obtaining multi-source heterogeneous data of content creation, editing, distribution and presentation of a full link, a content full link feature fingerprint is constructed, and cross-modal consistency analysis and link continuity verification are realized. According to the method, traceability map construction is introduced, potential tampering points are used as map nodes, content risks are quantified through tampering probability weights and risk transfer intensity, and tampering sources and propagation paths are accurately identified. The method has comprehensiveness and accuracy, deep analysis and risk assessment can be carried out on the multi-modal content, and the tampered content is effectively recognized and synthesized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital content security, and more particularly, to a multi-modal AI content security risk tracing and identification detection system. BACKGROUND

[0002] With the rapid development of deep learning and generative AI technology, artificial intelligence generated content (AIGC) has entered a large-scale application stage. Currently, multi-modal generative models such as DALL-E, Midjourney, Stable Diffusion, etc. can create highly realistic images, GPT series can generate fluent and natural text, Sora, Runway, etc. support high-quality video synthesis, and audio synthesis technology has achieved millisecond-level sound cloning. These technologies have shown great value in creative design, content creation, etc., but at the same time, they have brought unprecedented content security challenges.

[0003] However, the existing technology generally lacks an end-to-end content creation, distribution full-link verifiable mechanism, making it difficult to trace and confirm the authenticity of the content. Specifically, traditional detection methods only focus on the surface features of the content presented at the terminal, while ignoring the complete life cycle of the content from the creation source to the user's reception. In actual application, when a video is forwarded through multiple platforms and edited multiple times, the operation traces of the intermediate links are often completely erased, forming a "authenticity black box". For example, after an original news video is replaced with AI tools to replace the faces of the characters, modify the voice content, and spread on social media, the existing technology is difficult to reconstruct the content change timeline, cannot determine the specific link where the tampering occurs, the tools and technical means used, and has no way to trace the original tampering source and responsible subject. This link breakage makes the content tracing system only provide fragmented evidence, rather than a complete evidence chain, seriously restricting the effectiveness of digital forensics. More critically, the existing technology is particularly vulnerable in cross-modal collaborative tampering scenarios - when text, images, audio, and video are collaboratively modified to build a consistent false narrative, the lack of inter-modal consistency verification mechanism makes tampered content easily evade single-modal detection. This lack of traceability integrity not only hinders judicial forensics and news fact checking, but also provides a hidden space for the large-scale dissemination of deepfake content, posing a serious information security threat.

[0004] In view of this, the present application proposes a multi-modal AI content security risk tracing and identification detection system to solve the above problems. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a multi-modal AI content security risk tracing and identification detection system, comprising: The multi-source data acquisition module is configured to acquire multi-source heterogeneous data of the multi-modal content to be detected in the creation and distribution of the whole link, and the multi-source heterogeneous data includes metadata in a content generation stage, operation logs in an editing stage, network transmission features in a distribution stage, and user interaction behavior data in a terminal presentation stage. The feature fingerprint construction module is configured to construct a content whole-link feature fingerprint based on the spatiotemporal distribution characteristics of the data in each stage in the multi-source heterogeneous data, and the content whole-link feature fingerprint includes modality coupling features in the generation stage, operation sequence patterns in the editing stage, topology propagation features in the distribution stage, and behavior response patterns in the terminal presentation stage. The consistency analysis module is configured to extract potential tampering points and risk propagation paths by analyzing the cross-modal consistency and link continuity of the features in each stage in the content whole-link feature fingerprint. The traceability graph construction module is configured to construct a multi-modal content traceability graph according to the distribution characteristics of the potential tampering points and the risk propagation paths, and the multi-modal content traceability graph includes tampering probability weights between nodes and risk transmission intensities between paths. The risk identification and verification module is configured to identify synthetic tampering risks in the content link based on the multi-modal content traceability graph and in combination with a security risk classification model of the multi-modal content, and output an integrity verification result of a traceability link.

[0006] Preferably, the content whole-link feature fingerprint is constructed, including: modal decomposition is performed on the metadata in the generation stage to extract generator parameter distribution features of multi-modal data such as text, image, audio, and video, and modality coupling features are determined through mutual information calculation between parameters among modalities; sequence modeling is performed on the operation logs in the editing stage to identify nonlinear jump patterns in the operation sequence, and operation sequence patterns are extracted through periodic fluctuation analysis of operation timestamps; topology embedding is performed on the network transmission features in the distribution stage to construct a dynamic propagation graph of the distribution network, and topology propagation features are extracted through propagation delay and data packet integrity verification between nodes; behavior clustering is performed on the user interaction behavior data in the terminal presentation stage to identify time sequence abnormal points of the interaction behavior, and behavior response patterns are extracted through context association strength of the abnormal points; The modality coupling features, the operation sequence patterns, the topology propagation features, and the behavior response patterns are aligned across stages to generate the content whole-link feature fingerprint.

[0007] Preferably, the potential tampering points and the risk propagation paths are extracted, including: Perform cross-modal consistency checks on features at each stage of the content's full-link feature fingerprint, calculate the cosine similarity of feature vectors between modalities, and identify cross-modal inconsistency areas through local extreme points of similarity. Perform link continuity analysis on cross-modal inconsistent areas and identify potential tampering points based on the hopping frequency of feature fingerprints on the time axis and the distribution dispersion on the spatial axis. Based on the location distribution of potential tampering points, a risk propagation path map is constructed. The risk propagation path map determines the risk propagation path by weighting the propagation directionality and propagation intensity between potential tampering points; The path stability of the risk propagation path is verified. By introducing the path reconstruction error after random perturbation, unstable paths are eliminated and high-confidence risk propagation paths are retained.

[0008] Preferably, constructing a multimodal content traceability graph includes: Potential tampering points are treated as graph nodes, and directed edges between nodes are constructed based on the feature similarity and time sequence between nodes. Assign tampering probability weights to directed edges. The tampering probability weights are calculated based on the joint probability of the distribution difference of the feature vectors between nodes and the link hopping frequency. Quantify the risk transmission intensity of the risk propagation path. The risk transmission intensity between paths is determined by the normalized ratio of the accumulated tampering probability weight of each node on the path to the path length. Based on the tampering probability weight and risk transmission intensity, a multimodal content traceability map is constructed, and through the connectivity analysis of the map, isolated tampering areas and high-risk transmission areas in the link are identified.

[0009] Preferably, identifying synthetic tampering risks in a content link includes: Construct a security risk classification model for multimodal content. This model uses the full-link feature fingerprint of the content as input, extracts graph features through a multi-layer graph convolutional network, and uses an attention mechanism to weight high-risk nodes. Conduct risk propagation simulation on the multimodal content traceability graph and calculate the tampering diffusion probability of each node in the graph using a random walk algorithm; Based on the output of the tampering diffusion probability and security risk classification model, the synthetic tampering risk in the content chain is identified, and the main modal source of tampering is determined through modal contribution analysis of tampering risk. The synthetic tampering risk is divided into risk levels, and the risk level and the corresponding traceability evidence chain are output through a weighted combination of the tampering diffusion probability and the tampering impact range.

[0010] Preferably, the method for extracting modal coupling features includes: The generator parameter distribution characteristics of each modal data are high-dimensional embedded to generate modal parameter feature vectors; High-coupling parameter pairs between modes are identified through mutual information matrix calculation between modal parameter feature vectors; The high-coupling parameter pairs are dynamically weighted, and the coupling weights of each parameter pair are determined through the fluctuation consistency of the parameter pairs within the generation time window; Based on the coupling weight, the feature vectors of the high-coupling parameter pairs are weighted and fused to generate modal coupling features.

[0011] Preferably, the method for determining the potential tampering point comprises: Wavelet transform is performed on the time axis jump frequency of the content full link feature fingerprint to extract the high-frequency component and the low-frequency component of the jump frequency; The jump abnormal points on the time axis are determined through the local peak point density of the high-frequency component and the trend smoothness of the low-frequency component; The spatial distribution analysis is performed on the jump abnormal points, and the distribution abnormal area on the spatial axis is determined through the ratio of the spatial dispersion degree of the feature fingerprint to the spatial clustering radius; The intersection of the jump abnormal points and the distribution abnormal area is taken as the potential tampering point, and the false positive points are removed through the context feature consistency of the abnormal points.

[0012] Preferably, the method for assigning the tampering probability weight comprises: The kernel density estimation is performed on the feature vector distribution difference between nodes to calculate the KL divergence of the distribution difference as the feature difference measure; The entropy analysis is performed on the link jump frequency between nodes to determine the link discontinuity measure through the ratio of the distribution entropy of the jump frequency to the average jump interval; The normalized results of the feature difference measure and the link discontinuity measure are weighted and fused to generate the joint probability between nodes; Based on the joint probability, the tampering probability weight between nodes is generated through the softmax function mapping.

[0013] Preferably, the method for simulating risk propagation comprises: The graph embedding is performed on the multi-modal content traceability graph to generate low-dimensional feature vectors of graph nodes; Based on the low-dimensional feature vectors, a plurality of propagation paths are generated through the random walk algorithm, and the access frequencies of nodes on the paths are recorded; The path entropy analysis is performed on the propagation paths to determine the propagation stability of the paths through the ratio of the path entropy to the path length; Based on the propagation stability and the access frequency, the tampering diffusion probability of each node in the graph is calculated, and the high-risk propagation subgraph in the graph is identified through the spatial distribution characteristics of the diffusion probability.

[0014] Preferably, the system further comprises: Constructing a content full-link traceability verification database, the database comprising original feature fingerprints of multi-modal content, tampering point distribution characteristics and statistical characteristics of risk propagation paths; Through the content full-link traceability verification database, quickly traceable matching is performed on newly input multi-modal content, and historical tampering patterns in the content link are identified; Based on the matching results of the historical tampering patterns, the node weights and path intensities of the multi-modal content traceability graph are dynamically adjusted, and the real-time and accuracy of the traceability graph are optimized; The optimization results of the traceability graph and the detection results of the content security risk are jointly output, and a content link integrity verification report is provided.

[0015] The technical effects and advantages of the multi-modal AI content security risk traceability and identification detection system of the present application are as follows: The present application can effectively restore the complete facts of content tampering by establishing an end-to-end content creation and distribution full-link verifiable mechanism. This traceability capability enhances the effectiveness of digital judicial evidence, can accurately lock the source of deep fake content, trace the propagation path, and clarify the responsible subject, thereby providing key evidence support for combating information fraud and network false information. The present application helps to build a content authenticity audit defense line, prevent the generation and spread of fake news and misleading content, and improve the platform content governance capability and public credibility. The cross-modal consistency verification mechanism of the present application can effectively identify carefully designed composite modal fake content, and maintain the health and purity of the information ecosystem. The present application can timely discover and trace improper use. The present application reconstructs the trusted chain of digital content, and provides a technical foundation for building a transparent and reliable digital information environment. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A multi-modal AI content security risk traceability and identification detection system of the present application is shown in the figure. Figure 2 A determination method of potential tampering points of the present application is shown in the figure. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] The application example provides a multi-modal AI content security risk tracing and identification detection system. The execution subject of the system includes but is not limited to the content review platform, the media security monitoring center, the digital forensics system, the multi-modal content analysis platform and the like which can be regarded as the general computing node of the application, and the security detection system includes but is not limited to at least one of the cloud security risk analysis engine, the distributed content tracing system and the intelligent tampering detector.

[0019] The application provides a multi-modal AI content security risk tracing and identification detection system, which realizes the security risk tracing and detection of multi-modal content in the whole process of creation, editing, distribution and presentation through full-link multi-source data acquisition, feature fingerprint construction, cross-modal consistency analysis, tracing graph construction and risk identification verification. It has comprehensiveness and accuracy, can perform deep analysis and risk assessment on multi-modal content, effectively identifies synthetic tampered content, and provides complete trace evidence chain.

[0020] Please refer to Figure 1 In the embodiment of the application, a multi-modal AI content security risk tracing and identification detection system includes: A multi-source data acquisition module is configured to acquire multi-source heterogeneous data of multi-modal content to be detected in a full link in the creation and distribution. The multi-source heterogeneous data includes metadata in the content generation stage, operation logs in the editing stage, network transmission features in the distribution stage, and user interaction behavior data in the terminal presentation stage, and the like. The key information is acquired in real time through a multi-channel data acquisition interface. The metadata in the content generation stage records the creation time of the content, the generation tool used and the parameter setting, and the like. The operation logs in the editing stage record the whole process operation sequence of the content modification in detail. The network transmission features in the distribution stage reflect the path and mode of content propagation. The user interaction behavior data in the terminal presentation stage records the mode of user interaction with the content. These data provide comprehensive raw materials for subsequent analysis, and ensure the integrity and accuracy of the tracing and detection.

[0021] A feature fingerprint construction module is configured to construct a content full-link feature fingerprint based on the space-time distribution characteristics of the data in each stage of the multi-source heterogeneous data. The content full-link feature fingerprint includes the modal coupling feature in the generation stage, the operation sequence mode in the editing stage, the topological propagation feature in the distribution stage, and the behavior response mode in the terminal presentation stage. The modal coupling feature describes the correlation and consistency between different modal contents. The operation sequence mode reflects the behavior characteristics of the editing process. The topological propagation feature embodies the propagation law of the content in the network. The behavior response mode quantifies the interaction characteristics of the user to the content. These multi-dimensional features jointly constitute the "digital DNA" of the content, and provide a basis for subsequent consistency analysis and risk identification.

[0022] A consistency analysis module is configured to extract potential tampering points and risk propagation paths by analyzing the cross-modal consistency and link continuity of the features at each stage in the content full-link feature fingerprint. The module first performs cross-modal consistency testing on the feature fingerprint to identify inconsistent regions between modalities, and then locates abnormal jump points on the time axis and discrete regions in the spatial distribution through link continuity analysis, accurately determines the potential tampering points, and constructs risk propagation paths based on these tampering points, providing key inputs for the construction of the traceability graph.

[0023] A traceability graph construction module is configured to construct a multi-modal content traceability graph according to the distribution characteristics of the potential tampering points and the risk propagation paths. The multi-modal content traceability graph includes tampering probability weights between nodes and risk transmission intensities between paths. The module takes the tampering points as graph nodes, constructs directed edges through feature similarity and time sequence relationship, assigns tampering probability weights to the nodes, and allocates risk transmission intensities to the paths, forming a complete traceability network structure, and intuitively displaying the source, range and influence of content tampering, providing a graph basis for risk identification and verification.

[0024] A risk identification and verification module is configured to identify synthetic tampering risks in the content link based on the multi-modal content traceability graph and in combination with a multi-modal content security risk classification model, and output the integrity verification result of the traceability link. The module constructs a special security risk classification model, extracts graph features through a graph convolution network, performs weighted analysis on high-risk nodes in combination with an attention mechanism, simulates the risk propagation process to calculate the tampering diffusion probability, finally identifies the synthetic tampering risks in the content, and generates a detailed risk assessment report and a complete traceability evidence chain according to the risk level and the propagation range.

[0025] The above modules are connected through wired and / or wireless means to realize data transmission between the modules.

[0026] In the embodiment of the application, the detailed implementation steps for constructing the content full-link feature fingerprint include: The metadata generated in the stage is subjected to modal decomposition to extract generator parameter distribution features of multi-modal data such as text, image, audio and video, and modal coupling features are determined through mutual information calculation between parameters in different modalities. Modal decomposition is a basic step of feature extraction, which separates the composite content into independent modal layers for fine analysis. The decomposition process adopts a deep neural network encoder architecture, and special feature extraction modules are designed for different modalities, such as BERT type models for text, CNN variants for images, and spectral analysis networks for audio. The generator parameter distribution features include noise level, color consistency, text semantic features, audio spectral envelope and other key indicators. Mutual information calculation is based on information entropy theory to quantify the statistical dependence between parameters in different modalities, and the formula is: ; wherein is the mutual information of modalities and , is the joint probability distribution, and are the marginal probability distributions, denotes the parameter values or the values of a particular dimension in the feature vector of the first modality (e.g., text), denotes the parameter values or the values of a particular dimension in the feature vector of the second modality (e.g., image).

[0027] The higher the mutual information value, the stronger the correlation between the modalities, and the modal coupling features obtained through principal component analysis of the mutual information matrix provide a benchmark for inter-modal consistency for subsequent analysis.

[0028] The operation log in the editing phase is modeled in sequence, non-linear jump patterns in the operation sequence are identified, and operation sequence patterns are extracted through periodic fluctuation analysis of operation timestamps. Sequential modeling is a key step in capturing editing behavior characteristics, which converts discrete operations into analyzable time sequence patterns. The modeling process uses improved sequence encoding technology to map each operation to a feature vector, including operation type, parameters, duration, and context information. Non-linear jump patterns are identified through Markov chain analysis, calculating the operation transition probability matrix to highlight abnormal state transitions. Periodic fluctuation analysis uses wavelet transform methods to decompose the frequency components of the timestamp sequence, identifying periodic patterns at different scales. These features together constitute the operation sequence pattern, reflecting the behavior characteristics of the content editing process and providing a basis for identifying the differences between artificial editing and automatic generation.

[0029] The network transmission characteristics in the distribution phase are topologically embedded to construct a dynamic propagation graph of the distribution network, and topological propagation features are extracted through the propagation delay between nodes and data packet integrity verification. Topological embedding is an effective method for analyzing network propagation rules, mapping complex network structures to low-dimensional feature space. The embedding process is based on graph neural network technology, representing propagation nodes as high-dimensional vectors that preserve the topological relationships and propagation dynamics between nodes. The dynamic propagation graph is constructed through a time series graph model, capturing the spatio-temporal evolution of content propagation. Propagation delay is calculated by the time difference between nodes, and data packet integrity is evaluated by hash verification. These two indicators together reflect the reliability and consistency of the propagation process. Topological propagation features include key dimensions such as propagation rate, path diversity, node centrality, and community structure, providing a network perspective for tracking content propagation chains.

[0030] The user interaction behavior data in the terminal presentation stage is behavior clustered, time sequence abnormal points of interaction behavior are identified, and behavior response modes are extracted through the context association strength of the abnormal points. Behavior clustering is a basic method for understanding user interaction modes, and similar behaviors are inducted into meaningful categories. The clustering process adopts the density clustering algorithm DBSCAN, and the number and range of categories are automatically determined according to the similarity of behavior characteristics. The time sequence abnormal points are identified by the local outlier factor (LOF) method, the relative density difference of each interaction behavior and its neighborhood is calculated, and the points significantly deviating from the normal mode are identified. The context association strength is calculated by the conditional probability of the previous and subsequent behaviors, and the continuity of the abnormal behavior and the surrounding behavior is quantified. The behavior response mode finally includes user attention distribution, interaction depth, emotional response strength and social communication willingness, etc. dimensions, providing a user perspective for content influence evaluation.

[0031] The modal coupling features, operation sequence modes, topological propagation features and behavior response modes are cross-stage feature aligned to generate content full-link feature fingerprints. Feature alignment is a key step for integrating multi-stage features, ensuring the consistency of features in different stages in the time and semantic dimensions. The alignment process adopts the dynamic time warping (DTW) algorithm to process the time stretching of feature sequences in different stages, realize the time axis alignment, and at the same time adopts the cross-modal attention mechanism to calculate the semantic association between features, realize the alignment in the semantic dimension. The finally generated full-link feature fingerprint is a high-dimensional feature tensor, each dimension corresponds to a time point and a feature dimension in the content life cycle, and the full process digital features of the content from creation to presentation are recorded, providing a comprehensive feature basis for subsequent consistency analysis and risk identification.

[0032] In the embodiment of the application, the detailed implementation steps of extracting potential tampering points and risk propagation paths include: The cross-modal consistency of the features in the content full-link feature fingerprint is tested, the cosine similarity of the feature vectors between the modes is calculated, and the cross-modal inconsistency region is identified through the local extreme points of the similarity. Cross-modal consistency testing is a key step to discover the contradiction between modes, and the potential tampering is revealed by comparing the similarity of features in different modes. The testing process first maps the feature vectors of each mode to a shared feature space through projection transformation, and then calculates the cosine similarity between the mode pairs , the formula is: ; wherein, and are feature vectors of different modes, is the inner product of the vector, and are vector norms.

[0033] The similarity value ranges from -1 to 1, and the closer the value is to 1, the higher the consistency between the modalities. Local extreme points are identified through one-dimensional signal processing techniques, and when the similarity curve shows a significant drop, the corresponding time point is marked as a possible inconsistent region. These regions usually indicate logical or temporal contradictions between different modal content, and are important clues to potential tampering.

[0034] Link continuity analysis is performed on the cross-modal inconsistent regions to determine potential tampering points by the jump frequency of feature fingerprints on the time axis and the distribution dispersion on the spatial axis. Link continuity analysis is the core step of accurately locating tampering positions, and reveals abnormal points by evaluating the continuity of features in time and space dimensions. The analysis process first calculates the inter-frame difference of the feature vector on the time axis to identify time periods with abnormally high jump frequency; then calculates the distribution dispersion in the feature space to quantify the aggregation degree of feature distribution. The jump frequency calculation formula is: ; wherein, is the jump frequency of the time period , is the feature vector of the th frame, is the indicator function, is the jump threshold, is the number of frames.

[0035] The distribution dispersion is calculated by the ratio of the standard deviation to the mean of the feature vector, reflecting the relative dispersion degree of the feature distribution. When the jump frequency and the distribution dispersion both exceed the threshold value, the corresponding region is determined as a potential tampering point, providing a basis node for subsequent risk propagation path analysis.

[0036] Based on the location distribution of potential tampering points, a risk propagation path graph is constructed, which determines the risk propagation path by the propagation directionality and propagation intensity weighting between potential tampering points. The risk propagation path graph is a network model describing the spread of tampering influence, reflecting the causal relationship and influence transmission between tampering points. The construction process first takes the potential tampering points as the nodes of the graph, and then establishes directed edges based on the chronological order and feature correlation to form an initial path graph. The propagation directionality is determined by the timestamp and the gradient direction of feature change, and the propagation intensity is calculated by the feature change amplitude and the influence range. The propagation intensity calculation formula is: ; wherein, is the propagation intensity from node to node , is the feature change amplitude from node to node , is the influence range from node to node is the feature change amplitude from node to node​ The scope of influence, and is the weight coefficient, and .

[0037] The transmission intensity, as the weight of the edge, directly affects the priority of the risk transmission path. The higher the transmission intensity, the more important the corresponding path is, and the more likely it is to be the main risk transmission channel.

[0038] The path stability of the risk propagation path is verified. By introducing the path reconstruction error after random perturbation, unstable paths are eliminated and high-confidence risk propagation paths are retained. Path stability verification is a key step in improving the reliability of risk propagation paths. It screens out truly important paths by evaluating the path's sensitivity to noise. The verification process uses the Monte Carlo simulation method, adding random perturbations to the original feature data multiple times, reconstructing the propagation path, and calculating the path's recurrence rate. The path reconstruction error is calculated using the Jaccard distance to quantify the degree of difference between the original path and the reconstructed path. The stability score calculation formula is: ;in, For path The stability score of The original path With the Secondary reconstruction path Jaccard distance, is the number of simulations.

[0039] The stability score ranges from [0, 1], with higher values ​​indicating more stable paths. Setting an appropriate threshold (typically 0.7-0.8) eliminates low-stability paths and retains high-confidence risk propagation paths to ensure the accuracy and reliability of subsequent risk analysis.

[0040] In an embodiment of the present invention, the detailed implementation steps of constructing a multimodal content traceability graph include: Potential tampering points are treated as graph nodes, and directed edges are constructed between them based on feature similarity and temporal relationships between nodes. Graph node construction is a fundamental step in the traceability graph, organizing discrete tampering points into a network structure. The construction process first maps each potential tampering point into a node in the graph, attaching attributes such as timestamp, location information, feature vector, and impact range to the node. Directed edges are then established based on the relationships between nodes, forming a connected graph structure. Feature similarity is calculated using cosine similarity or Euclidean distance, and temporal relationships are determined by comparing timestamps. Directed edges are established according to the principle of causality, pointing from earlier nodes to later nodes, and the legitimacy of the relationships is verified by the consistency of feature changes. This edge construction method, based on spatiotemporal relationships and feature similarity, ensures the rationality and interpretability of the traceability graph structure.

[0041] Tampering probability weights are assigned to directed edges, and the tampering probability weights are calculated based on the joint probability of the distribution difference of the feature vectors between nodes and the link hopping frequency. Weight assignment is a key step in quantifying the strength of node relationships and reflects the credibility of the causal relationship between tampering points. The assignment process first calculates the feature distribution difference between nodes to quantify the distance of the feature vector distribution; then analyzes the link hopping frequency to evaluate the degree of mutation of the feature change; finally, these two factors are combined through a joint probability model to obtain the final tampering probability weight. The feature distribution difference is calculated by KL divergence, and the link hopping frequency is evaluated by entropy analysis. The KL divergence calculation formula is: ;in, For distribution Relative to The KL divergence of For the event / state in the first probability distribution The probability value of For the same event / state in the second probability distribution The probability value of .

[0042] The entropy calculation of the link hopping frequency reflects the complexity and uncertainty of the hopping pattern. The final tampering probability weight is mapped to the interval [0, 1] through the softmax function, which intuitively represents the credibility of the tampering relationship between nodes.

[0043] The path risk transmission strength of the risk propagation path is quantified. The risk transmission strength between paths is determined by the normalized ratio of the cumulative tampering probability weight of each node on the path to the path length. Path risk transmission strength is a key indicator for assessing the scope of tampering impact and reflects the ability of risk to propagate along a specific path. The quantification process first calculates the sum of the tampering probability weights of all nodes on the path to obtain the original cumulative value; then, the path length factor is considered and the impact of length differences is eliminated through normalization. Finally, the risk transmission strength is obtained and used to rank the path importance. The formula for calculating risk transmission strength is: ;in, For path The risk transmission intensity, For path The tampering probability weight of the edge from node i to node j in, For path length.

[0044] The square root term in the formula acts as a normalization factor, balancing the differences between long and short paths and enabling fair comparison of paths of different lengths. The intensity of risk transmission directly impacts the simulation and prediction of risk propagation; paths with higher intensity are more likely to be the primary risk propagation channel.

[0045] Based on the tampering probability weight and the risk transmission intensity, a multi-modal content traceability graph is constructed, and through connectivity analysis of the graph, isolated tampering areas and high-risk transmission areas in the link are identified. The multi-modal content traceability graph is a comprehensive visualization model for tampering analysis, which intuitively displays the source, range and influence of content tampering. The construction process integrates nodes, directed edges, tampering probability weights and risk transmission intensity to form a complete graph structure. Connectivity analysis is a key step in graph interpretation, which identifies special areas by evaluating the connection mode between nodes. Isolated tampering areas are identified by weakly connected component analysis, representing independent tampering unrelated to the main tampering link; high-risk transmission areas are identified by centrality indicators and community detection algorithms, representing key areas where tampering influence is concentrated or diffused. The identification of these special areas provides a structured perspective for risk assessment, helping analysts quickly locate key risk points and transmission paths, and providing a graph basis for subsequent risk identification and verification.

[0046] In the embodiment of the application, the detailed implementation steps for identifying synthetic tampering risks in content links include: A security risk classification model for multi-modal content is constructed, which takes content full-link feature fingerprints as input, extracts graph features through a multi-layer graph convolution network, and combines attention mechanisms to weight high-risk nodes. The risk classification model is the core component of automatically identifying tampering types, which converts complex graph features into explicit risk classifications. The construction process first designs a multi-layer graph convolution network (GCN) architecture, which captures high-order relationships between nodes through a message passing mechanism; then introduces an attention mechanism to adaptively adjust the importance weights of different nodes and features; finally, the risk category probability is output through a fully connected layer and a softmax classifier. The core calculation formula of GCN is: ; wherein, is the node feature matrix of the th layer, is the adjacency matrix with self-loop added, is the corresponding degree matrix, is the learnable weight matrix, is a nonlinear activation function.

[0047] The attention mechanism highlights the influence of key nodes and edges by calculating feature importance scores, improving the model's sensitivity to high-risk areas. The model training uses a large-scale labeled data set containing various known tampering samples, and optimizes the parameters through a cross-entropy loss function to achieve accurate identification of multiple tampering types.

[0048] Risk propagation simulation is a dynamic method to evaluate the influence range of tampering. Through computational simulation, it predicts the diffusion trend of tampering risk. The simulation process is based on random walk theory, considering tampering risk as a random walk process on the graph, and calculating the probability of each node being visited through multiple simulation. The specific implementation adopts the biased random walk algorithm, and the transition probability of the edge is proportional to the tampering probability weight, reflecting the tendency of actual risk propagation. The tampering diffusion probability calculation formula of node is: ; wherein, is the tampering diffusion probability of node , is the total number of times node is visited in M random walks, is the number of steps in each walk, is the total number of simulations.

[0049] The diffusion probability directly reflects the importance and degree of influence of the node in the risk propagation process, providing a probability basis for the quantitative evaluation of tampering risk.

[0050] Based on the tampering diffusion probability and the output of the security risk classification model, the synthetic tampering risk in the content link is identified, and the main modal source of tampering is determined through modal contribution degree analysis. Tampering risk identification is the core function of the system, which judges the authenticity and security of the content by integrating multiple evidence. The identification process first combines the output category and confidence of the security risk classification model to preliminarily judge the risk type of the content; then it evaluates the influence range of the risk by fusing the tampering diffusion probability; finally, it identifies the main source modal of tampering through modal contribution degree analysis. Modal contribution degree is determined by calculating the influence weight of each modal feature on the final risk judgment, using model explainability techniques such as SHAP value analysis. This multi-dimensional risk identification method not only considers the risk type, but also evaluates the influence range and locates the main risk source, providing a comprehensive basis for security decision-making.

[0051] Risk level classification is performed on the synthetic tampering risk, and the risk level and corresponding traceability evidence chain are output through the weighted combination of tampering diffusion probability and tampering influence range. Risk level classification is the final output of risk assessment, which converts technical analysis results into operational decision-making basis. The classification process first determines the risk score formula, considering the tampering diffusion probability (reflecting the risk propagation ability) and the tampering influence range (reflecting the proportion of affected content), and obtaining the comprehensive risk score through weighted combination; then according to the preset level threshold, the score is mapped to different risk levels, usually divided into low, medium and high three levels or more detailed five levels. The risk level calculation formula is: ​; wherein, is a risk level score of the content , is a maximum tampering diffusion probability, is a tampering influence range, is a weight coefficient.

[0052] It should be noted that the tampering influence range is a quantitative index for calculating the degree of influence of tampering in the content; the specific calculation formula is: ; wherein, is the tampering influence range, the value range is [0, 1], is a set of all nodes in the content provenance graph, is the tampering diffusion probability of node , is a probability threshold for tampering determination (usually set to 0.3-0.5), is an indicator function, which takes a value of 1 when the condition in the parentheses is true, otherwise 0; is the importance weight of node , which is usually based on the content coverage range or visual saliency of the node.

[0053] The provenance evidence chain is then generated by tracing back the source path of the high-risk node, including the time, location, feature change and associated evidence of the key tampering point, forming a complete risk provenance report, providing detailed basis for security decision and forensic analysis.

[0054] In the embodiment of the application, the detailed implementation steps of the modal coupling feature extraction method include: High-dimensional embedding is performed on the generator parameter distribution features of each modal data to generate modal parameter feature vectors. High-dimensional embedding is a key step for mapping complex parameter distribution to a unified feature space, which facilitates subsequent cross-modal analysis. The embedding process uses a deep autoencoder architecture, which inputs the original parameter distribution features into the encoder, compresses them into low-dimensional dense representations through multiple nonlinear transformations, while retaining key information. Different modalities use dedicated encoder structures, such as a Transformer-based encoder for text parameters, a CNN encoder for image parameters, and a waveform analysis encoder for audio parameters. Encoder training uses reconstruction loss and contrastive learning objectives to ensure that the embedding vectors retain the internal structure of the modalities while maximizing the expressiveness of cross-modal information. The final generated feature vectors are usually 128-512 dimensions, providing a unified feature representation for inter-modal information analysis.

[0055] Highly coupled parameter pairs between modalities are identified by mutual information matrix calculation between modal parameter eigenvectors. Mutual information matrix is an effective tool to find the key association between modalities, which quantifies the statistical dependence between parameter pairs. The calculation process first constructs the joint probability distribution and marginal probability distribution between each dimension of the eigenvector, then approximates the probability density by kernel density estimation or histogram method, and finally calculates the mutual information value. For high-dimensional eigenvectors, a block calculation strategy is adopted to segment the vector into sub-blocks with semantic relevance, reducing the computational complexity. By setting a proper threshold (usually the upper quartile of the mutual information distribution), parameter pairs with significantly higher mutual information values than the average level are screened out and marked as highly coupled parameter pairs. These parameter pairs reflect the strong association between different modalities and are the key manifestations of modal coupling, providing important clues for tamper detection.

[0056] The highly coupled parameter pairs are dynamically weighted, and the coupling weights of each parameter pair are determined by the fluctuation consistency of the parameter pair within the generation time window. Dynamic weighting is an important method to evaluate the stability of coupling, which distinguishes stable coupling from accidental association through consistency analysis in the time dimension. The weighting process first tracks the time sequence change trajectory of each highly coupled parameter pair within the generation time window; then calculates the fluctuation consistency to quantify the synchronization of parameter pair changes; finally, according to the consistency level, the weight is assigned, the higher the consistency, the greater the weight. The formula is: ; wherein, is the fluctuation consistency of parameter pairs and , and and are the parameter time sequence fragments within the time window , and is the Pearson correlation coefficient, is the total time length, is the window size.

[0057] The consistency value ranges from 0 to 1, the higher the value, the more synchronized the parameter pair changes, and the more stable the coupling. The coupling weight is normalized to ensure that the sum of all weights is 1, providing a reasonable importance allocation for subsequent feature fusion.

[0058] Based on the coupling weights, the feature vectors of highly coupled parameter pairs are weightedly fused to generate modal coupling features. Weighted fusion is the final step in integrating cross-modal information, combining scattered parameter pairs into a unified coupling feature representation. The fusion process adopts a method that combines attention mechanism and weighted averaging, which not only takes into account global information but also highlights the contribution of key parameter pairs. First, the feature vector of each highly coupled parameter pair is mapped to a shared feature space through a linear transformation; then, the weighted sum is calculated according to the coupling weight; finally, the feature relationship is further extracted through a nonlinear activation function to obtain the final modal coupling feature vector. This feature vector comprehensively reflects the dependency and coupling pattern between different modalities, provides a cross-modal perspective for subsequent consistency analysis, helps to identify unnatural associations or breaks between modalities, and is an important indicator for detecting AI-synthesized content.

[0059] like Figure 2 FIG. 1 is a diagram of a method for determining potential tampering points. In an embodiment of the present invention, the detailed implementation steps of the method for determining potential tampering points include: Perform wavelet transform on the time axis hopping frequency of the full-link feature fingerprint of the content to extract the high-frequency and low-frequency components of the hopping frequency. Wavelet transform is an effective tool for analyzing the multi-scale characteristics of time series, and can capture the local details and global trends of the signal at the same time. The transformation process uses the Daubechies wavelet basis to decompose the original hopping frequency sequence into sub-signals in different frequency bands through multi-level decomposition. The high-frequency component reflects the sudden change and abnormal fluctuation of the feature, which usually corresponds to editing operations or content replacement; the low-frequency component reflects the long-term trend and slow change of the feature, which usually corresponds to the natural evolution of the content. The number of decomposition layers is dynamically adjusted according to the sequence length, usually ,in The key advantage of wavelet transform is that it can accurately locate abnormal points in time series while preserving both time and frequency information, providing a multi-scale perspective for subsequent jump analysis.

[0060] The jump anomaly points on the time axis are determined by the local peak point density of the high-frequency component and the trend smoothness of the low-frequency component. Jump anomaly points are potential signs of tampering in the time dimension, reflecting unnatural changes in content characteristics. The determination process first analyzes the high-frequency component, calculates the density distribution of local peak points, and identifies the time period with concentrated peak anomalies; then evaluates the low-frequency component, calculates the trend smoothness through curve fitting, and identifies areas with abnormal trend changes; finally, combining these two indicators, the jump anomaly points on the time axis are located. The calculation formula for local peak point density is: ;in, For time point The peak density, For the The time position of the peak point, is the indicator function, is the time window radius, is the total number of peak points.

[0061] Trend smoothness is assessed using the variance of the second-order derivative of the fitted curve; larger values ​​indicate more dramatic trend changes. When peak density is significantly higher than the background level and trend smoothness is abnormally low, the corresponding time point is marked as a jump anomaly, providing temporal location for spatial distribution analysis.

[0062] Spatial distribution analysis of outlier points is performed. The ratio of the spatial dispersion of the feature fingerprint to the spatial clustering radius is used to identify areas of abnormal distribution along the spatial axis. Spatial distribution analysis is a key step in identifying tampering in the feature space dimension. It discovers abnormal patterns by evaluating the geometric properties of the feature distribution. The analysis process first projects the feature vectors corresponding to the outlier points into a low-dimensional space, using dimensionality reduction techniques such as t-SNE or Unified Mapping (UMAP) to preserve local structure. Spatial dispersion is then calculated to quantify the dispersion of the feature points. The spatial clustering radius is also estimated to reflect the natural clustering tendency of the features. Finally, the ratio of these two metrics is used to identify areas of abnormal feature distribution. Spatial dispersion is calculated as the average distance from the feature points to the centroid, while the spatial clustering radius is estimated using the optimal parameters of the density clustering algorithm. When the ratio of dispersion to clustering radius is significantly higher than normal, the corresponding area is marked as an abnormal distribution area, indicating that the feature distribution does not conform to the statistical laws of natural content and may indicate tampering or synthesis.

[0063] The intersection of the jump anomaly point and the distribution anomaly area is regarded as a potential tampering point, and the false positive points are eliminated through the consistency of the contextual features of the anomaly point. The final determination of the potential tampering point is the result of dual verification in time and space, which not only takes into account the abnormality of time series changes, but also evaluates the unnatural pattern of feature distribution. The determination process first calculates the intersection of time and space anomalies to preliminarily lock the highly suspicious area; then, through the consistency analysis of context features, the rationality of the anomaly is further verified. The context consistency analysis examines the coherence of the features before and after the anomaly point, and evaluates the naturalness of the change by calculating the similarity gradient of the forward and backward features. The consistency score calculation formula is: ;in, for point The context consistency score of for point The eigenvector of is the similarity function, is the time step.

[0064] The points with abnormally low consistency scores are reserved as the final potential tampering points, and the points with normal consistency are regarded as false positive points and removed. This multi-dimensional tampering point determination method greatly improves the accuracy of detection, reduces false positives, and provides a reliable foundation for subsequent risk assessment positioning.

[0065] In the embodiments of the present application, the detailed implementation steps of the tampering probability weight assignment method include: The feature vector distribution difference between nodes is subjected to kernel density estimation, and the KL divergence of the distribution difference is calculated as the feature difference measure. Kernel density estimation is an effective method for non-parametric distribution fitting, which can accurately capture the complex morphology of feature distribution. The estimation process first selects an appropriate kernel function (usually Gaussian kernel) and bandwidth parameter, then constructs the probability density function based on the feature vector sample, and finally calculates the KL divergence between the distributions. For high-dimensional features, dimensionality reduction preprocessing and piecewise estimation strategy are adopted to reduce the computational complexity. As a classical measure in information theory, KL divergence intuitively reflects the degree of dissimilarity between two distributions, and the larger the value, the more significant the distribution difference. The KL divergence calculation uses the Monte Carlo integration method for approximation, and the feature difference measure directly reflects the degree of change of the content features between nodes, providing an important basis for the distribution perspective of tampering probability evaluation.

[0066] The link jump frequency between nodes is subjected to entropy analysis, and the ratio of the distribution entropy of the jump frequency to the average jump interval is used to determine the link discontinuity measure. Entropy analysis is a classical method for evaluating signal complexity and uncertainty, and when applied to link jump frequency, it can effectively quantify the irregularity of content changes. The analysis process first treats the link jump events as time series, calculates the statistical distribution of the jump interval; then calculates the distribution entropy to reflect the complexity of the jump pattern; at the same time, the average jump interval is calculated to reflect the overall frequency of the jump; finally, the ratio of the two is used to obtain the link discontinuity measure. For discrete distribution, the formula is: ; wherein, is the entropy value of the distribution , the probability of the event , and the specific event in the distribution . The discontinuity measure directly reflects the degree of abnormality of content changes, and the combination of high entropy and low interval usually indicates unnatural frequent changes, which may imply tampering behavior, providing key information for the time series perspective of tampering probability evaluation.

[0067] The discontinuity measure directly reflects the degree of abnormality of content changes, and the combination of high entropy and low interval usually indicates unnatural frequent changes, which may imply tampering behavior, providing key information for the time series perspective of tampering probability evaluation.

[0068] ​The normalized results of the feature difference measure and the link discontinuity measure are weighted and fused to generate the joint probability between nodes. Weighted fusion is a key step in integrating multi-dimensional evidence, and it integrates information from different perspectives by allocating reasonable weights. The fusion process first normalizes the feature difference measure and the discontinuity measure to eliminate dimensional differences and map them to the same [0, 1] interval; then sets the weight coefficients according to the content type and analysis target to reflect the relative importance of different measures in a specific scenario; finally, the joint probability is calculated by weighted average. Normalization usually uses the Min-Max or Z-score method, and weight setting is based on prior knowledge or cross-validation optimization. The joint probability calculation formula is: ; wherein, is the joint probability between nodes, is the normalized KL divergence, is the normalized discontinuity measure, and are weight coefficients, and .

[0069] The joint probability integrates abnormal information in both distribution and timing dimensions, providing a comprehensive basis for the final quantification of tampering probability.

[0070] Based on the joint probability, the tampering probability weight between nodes is generated by mapping through the softmax function. Softmax mapping is the last step to convert the joint probability into a standardized weight, ensuring that the weight distribution is reasonable and has a probability interpretation. The mapping process first applies the softmax function to the joint probability of all outgoing edges of each node, converting the original value to a probability distribution in the [0, 1] interval with a sum of 1; then it performs global scaling to adjust the distribution range of the overall weight, ensuring that high-probability edges are significantly distinguished from low-probability edges. The calculation formula of the softmax function is: ; wherein, is the tampering probability weight from node to node , is the corresponding joint probability, is the temperature parameter, which controls the steepness of the probability distribution, traverses all outgoing edge target nodes of node .

[0071] The temperature parameter is usually set to between 1 and 5, and the larger the value, the more concentrated the weight distribution on high-probability edges, which is more conducive to highlighting the main tampering path. The final tampering probability weight intuitively reflects the credibility of the tampering relationship between nodes, providing a quantitative basis for the construction and analysis of the traceability graph.

[0072] The detailed implementation steps of the method for risk propagation simulation in the embodiment of the application include: The multi-modal content traceability graph is graph embedded to generate low-dimensional feature vectors of the graph nodes. Graph embedding is an effective dimension reduction technique for processing complex graph structures, which compresses high-dimensional graph information into vector representations for easy calculation. The embedding process uses graph neural network (GNN) techniques such as GraphSAGE or GAT to learn the context information and structural features of the nodes through a message passing mechanism. The specific implementation first initializes the node features, using the original attributes of the nodes such as timestamps, location information and feature vectors; then propagates and aggregates the features through multiple layers of GNN, each layer fuses neighborhood information to update the node representation; finally, the embedding quality is optimized through multi-task learning objectives while maintaining node similarity and graph structure integrity. The embedding dimension is usually set to 64-256, which is dynamically adjusted according to the size and complexity of the graph. The low-dimensional feature vectors not only retain the semantic information of the nodes, but also encode the topological relationships of the graph structure, providing an efficient computational basis for subsequent random walks.

[0073] Based on the low-dimensional feature vectors, multiple propagation paths are generated through the random walk algorithm, and the access frequency of each node on the path is recorded. Random walk is a classic method for simulating risk propagation, which reflects the randomness and tendency of risk diffusion through probabilistic path generation. The walk process starts from a high-risk source node, determines the next step transition probability according to the tampering probability weight of the outgoing edge, and simulates the path of risk propagation along the network. To improve the accuracy of the simulation, a second-order random walk strategy is used, which considers the joint influence of the current position and the previous position, better capturing the inertia and directionality of the propagation. The number of walk steps is usually set to 3-5 times the diameter of the graph to ensure coverage of the main propagation range; the number of repetitions is set to 10-20 times the number of nodes to ensure the reliability of the statistical results. After each walk, the access of each node on the path is recorded, and the access frequency statistics are accumulated to intuitively reflect the participation and importance of the nodes in the risk propagation process.

[0074] The propagation paths are analyzed for path entropy, and the propagation stability of the path is determined by the ratio of path entropy to path length. Path entropy analysis is an important method for evaluating the regularity of propagation, which quantifies the structural characteristics and change patterns of the path through information theory metrics. The analysis process first considers each propagation path as a node sequence, calculates the conditional entropy of the node transition probability, and reflects the uncertainty and complexity of the path; then considers the path length factor, eliminates the influence of length difference through the ratio of entropy to length; finally, the propagation stability index of the path is obtained, which evaluates the predictability and consistency of risk propagation. The path entropy calculation formula is: ; wherein, is the entropy value of the path , is the path Middle node Follow-up node Conditional probability, Path length.

[0075] The propagation stability is inversely proportional to the path entropy, and the path with high stability usually shows low entropy and high regularity, which is the main channel of risk propagation, and has higher reference value for the calculation of tampering diffusion probability.

[0076] Based on the propagation stability and the access frequency, the tampering diffusion probability of each node in the graph is calculated, and through the spatial distribution characteristics of the diffusion probability, the high-risk propagation subgraph in the graph is identified. The tampering diffusion probability is the core index of risk assessment, which directly reflects the possibility and degree of the influence of the node by tampering. The calculation process first comprehensively evaluates the importance of the node in the risk propagation according to the access frequency and the stability of the path participated by the node; then through normalization processing, the original score is converted into a standardized probability value; finally, the diffusion probability distribution of all nodes in the graph is formed. The calculation formula is: ; wherein, The tampering diffusion probability of node The access frequency, The set of all paths containing node The stability score of path The stability score of path All nodes in the graph. Based on the diffusion probability distribution, through community detection algorithm and density threshold analysis, the connected region with probability value significantly higher than the background level is identified and marked as a high-risk propagation subgraph. These subgraphs are the key areas of tampering influence, which usually reflect the influence range and propagation mode of a specific type of tampering, and provide a structured perspective for subsequent risk classification and intervention strategies.

[0077] In the embodiment of the application, the system further comprises:

[0078] In the embodiment of the application, the system further comprises: ​A content full-link traceability verification database is constructed, which includes the original feature fingerprints of multi-modal content, the tampering point distribution characteristics, and the statistical characteristics of risk propagation paths. The traceability verification database is the core component of system knowledge accumulation, and improves the detection capability through continuous learning and updating. The construction process first designs a multi-level data structure, including the bottom layer of original feature storage layer, the middle layer of tampering feature index layer, and the top layer of propagation mode summary layer; then through the online learning mechanism, the new results of system analysis are continuously integrated into the database to form a dynamically growing knowledge base; at the same time, regular cleaning and optimization are implemented to ensure the accuracy and efficiency of the database. The feature fingerprints are stored by compression encoding, the tampering point distribution is organized by spatial indexing technology, and the propagation path is managed by graph database. The database not only stores the detailed features of known tampering cases, but also contains typical patterns and evolution trends of various tampering techniques, providing a rich reference basis for fast traceability matching.

[0079] Through the content full-link traceability verification database, fast traceability matching is performed on the newly input multi-modal content to identify the historical tampering patterns in the content link. Fast traceability matching is a key technology to improve detection efficiency, which quickly locates the possible tampering type and source through pattern matching. The matching process first extracts the feature fingerprints of the new content, then performs multi-level search in the database, from coarse-grained type matching to fine-grained pattern comparison, gradually narrowing down the candidate range; finally, through similarity calculation, the matching degree with historical cases is identified to determine the possible tampering pattern. The search adopts hierarchical indexing and approximate nearest neighbor algorithm, which significantly improves the query efficiency of large-scale database. The matching result includes the most similar historical case, the matching degree score and the possible tampering type suggestion, providing valuable prior information for subsequent analysis, accelerating the traceability process and improving the accuracy and efficiency of detection.

[0080] Based on the matching results of historical tampering patterns, the node weights and path intensities of the multi-modal content traceability graph are dynamically adjusted to optimize the real-time and accuracy of the traceability graph. Graph dynamic adjustment is an adaptive mechanism to improve traceability accuracy, guided by historical experience for current analysis. The adjustment process first identifies the feature regions similar to historical patterns in the current content according to the matching results; then based on the credibility of matching, the weights of related nodes and the intensities of paths in the traceability graph are adjusted, strengthening the importance of high credibility regions and weakening the influence of uncertain regions; finally, the key indicators of the graph are recalculated to ensure that the adjusted graph structure is more consistent with the actual tampering situation. The adjustment strategy adopts a Bayesian update framework, which combines historical matching as prior information with current observation evidence to obtain more accurate posterior estimation. This dynamic optimization mechanism enables the system to continuously learn and adapt to new tampering techniques and patterns, maintaining the advancedness and effectiveness of detection capability.

[0081] The optimization results of the traceability map and the detection results of content security risks are jointly output to provide an integrity verification report for the content chain. The integrity verification report is the final output of the system, which comprehensively displays the security status and traceability evidence of the content. The report generation process first integrates the optimized traceability map and risk detection results to form a unified security assessment view; then, based on the needs of different users, multi-level report content is generated, ranging from a concise risk level overview to detailed technical analysis evidence; finally, through visualization technology, the complex map structure and risk distribution are intuitively displayed for easy understanding and decision-making. The report content includes key information such as content authenticity score, tampering area marking, risk propagation path analysis, tampering technology determination and risk level recommendation, providing comprehensive technical support for content review, digital forensics and security decision-making. The report output supports multiple formats and interfaces, which is convenient for integration with other security systems to form a complete content security protection system.

[0082] This paper implements full-link security risk tracing and detection for multimodal AI content through multi-source data acquisition, feature fingerprint construction, consistency analysis, traceability graph construction, and risk identification and verification. The present invention's graphical traceability method can accurately locate the source of tampering and propagation paths, effectively identifying security risks in synthetic content, and providing a systematic solution for authenticity verification of digital content.

[0083] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art will be able to modify the technical solutions described in the foregoing embodiments or to substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0084] It should be noted that the formulas in this manual are all dimensionless and calculated numerically. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0085] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A multi-modal AI content security risk tracing and identification detection system, characterized in that, The method comprises the following steps: a multi-source data acquisition module for acquiring multi-source heterogeneous data of the multi-modal content to be detected in the creation and distribution of the whole link, the multi-source heterogeneous data including metadata of the content generation stage, operation logs of the editing stage, network transmission characteristics of the distribution stage, and user interaction behavior data of the terminal presentation stage; a feature fingerprint construction module for constructing a content whole link feature fingerprint based on the spatiotemporal distribution characteristics of the data of each stage in the multi-source heterogeneous data, the content whole link feature fingerprint including modal coupling characteristics of the generation stage, operation sequence patterns of the editing stage, topological propagation characteristics of the distribution stage, and behavior response patterns of the terminal presentation stage; a consistency analysis module for extracting potential tampering points and risk propagation paths by analyzing the cross-modal consistency and link continuity of the features of each stage in the content whole link feature fingerprint; a traceability graph construction module for constructing a multi-modal content traceability graph according to the distribution characteristics of the potential tampering points and risk propagation paths, the multi-modal content traceability graph including tampering probability weights between nodes and risk transmission intensities between paths; a risk identification and verification module for identifying synthetic tampering risks in the content link based on the multi-modal content traceability graph and combining a multi-modal content security risk classification model, and outputting a completeness verification result of the traceability link.

2. The multi-modal AI content security risk tracing and identification detection system of claim 1, wherein, The construction of the content whole link feature fingerprint comprises: modal decomposition of the metadata of the generation stage, extraction of the generator parameter distribution characteristics of the multi-modal data, and determination of the modal coupling characteristics through inter-modal parameter mutual information calculation; sequence modeling of the operation logs of the editing stage, identification of the nonlinear jump patterns in the operation sequence, and extraction of the operation sequence patterns through periodic fluctuation analysis of the operation timestamps; topological embedding of the network transmission characteristics of the distribution stage, construction of a dynamic propagation graph of the distribution network, and extraction of the topological propagation characteristics through propagation delay and data packet integrity verification between nodes; behavior clustering of the user interaction behavior data of the terminal presentation stage, identification of the time series abnormal points of the interaction behavior, and extraction of the behavior response patterns through the context association strength of the abnormal points; cross-stage feature alignment of the modal coupling characteristics, operation sequence patterns, topological propagation characteristics, and behavior response patterns to generate the content whole link feature fingerprint.

3. The multi-modal AI content security risk tracing and identification detection system of claim 1, wherein, The extraction of the potential tampering points and risk propagation paths comprises: cross-modal consistency testing of the features of each stage in the content whole link feature fingerprint, calculation of the cosine similarity of the feature vectors between modes, and identification of the cross-modal inconsistency regions through local extreme points of the similarity; link continuity analysis of the cross-modal inconsistency regions, determination of the potential tampering points through the jump frequency of the feature fingerprints on the time axis and the distribution dispersion degree on the spatial axis; construction of a risk propagation path graph based on the position distribution of the potential tampering points, the risk propagation path graph determining the risk propagation paths through the propagation directionality and intensity weighting between the potential tampering points; The path stability of the risk propagation path is verified, the path reconstruction error after introducing random disturbance is removed, the unstable path is removed, and the risk propagation path with high confidence is reserved.

4. The multi-modal AI content security risk tracing and identification detection system of claim 1, wherein, The multi-modal content traceability graph is constructed, including: The potential tampering point is taken as a graph node, and a directed edge between nodes is constructed through the feature similarity and time sequence relationship between nodes; The tampering probability weight of the directed edge is valued, and the tampering probability weight is calculated based on the joint probability of the distribution difference of the feature vector between nodes and the link jump frequency; The risk transmission strength of the risk propagation path is quantified, and the risk transmission strength between paths is determined by the normalized ratio of the cumulative tampering probability weight of each node on the path and the path length; Based on the tampering probability weight and the risk transmission strength, a multi-modal content traceability graph is constructed, and isolated tampering areas and high-risk propagation areas in the link are identified through connectivity analysis of the graph.

5. The multi-modal AI content security risk tracing and identification detection system of claim 1, wherein, The synthetic tampering risk in the content link is identified, including: A security risk classification model of multi-modal content is constructed, the security risk classification model takes the content full-link feature fingerprint as input, extracts graph features through a multi-layer graph convolution network, and combines an attention mechanism to weight high-risk nodes; Risk propagation simulation is performed on the multi-modal content traceability graph, and the tampering diffusion probability of each node in the graph is calculated through a random walk algorithm; Based on the tampering diffusion probability and the output of the security risk classification model, the synthetic tampering risk in the content link is identified, and the main modal source of tampering is determined through modal contribution degree analysis of the tampering risk; The synthetic tampering risk is divided into risk levels, and the risk level and the corresponding traceability evidence chain are output by the weighted combination of the tampering diffusion probability and the tampering influence range.

6. The multi-modal AI content security risk tracing and identification detection system of claim 2, wherein, The extraction method of the modal coupling feature includes: The generator parameter distribution feature of each modal data is high-dimensional embedded to generate a modal parameter feature vector; High-coupling parameter pairs between modal parameter feature vectors are identified through mutual information matrix calculation; The high-coupling parameter pairs are dynamically weighted to determine the coupling weight of each parameter pair through the fluctuation consistency of the parameter pairs within the generation time window; Based on the coupling weight, the feature vectors of the high-coupling parameter pairs are weighted and fused to generate the modal coupling feature.

7. The multi-modal AI content security risk tracing and identification detection system of claim 3, wherein, The determination method of the potential tampering point includes: The time axis jump frequency of the content full-link feature fingerprint is wavelet transformed to extract the high-frequency component and the low-frequency component of the jump frequency; The jump abnormal points on the time axis are determined through the local peak point density of the high-frequency component and the trend smoothness of the low-frequency component; The spatial distribution of the jump abnormal points is analyzed, and the distribution abnormal area on the spatial axis is determined through the ratio of the spatial dispersion degree of the feature fingerprint to the spatial clustering radius; The intersection of the jump abnormal points and the distribution abnormal area is taken as the potential tampering point, and the false positive points are removed through the context feature consistency of the abnormal points.

8. The multi-modal AI content security risk tracing and identification detection system of claim 4, wherein, The valuation method of the tampering probability weight includes: The feature vector distribution difference between nodes is kernel density estimated, and the KL divergence of the distribution difference is calculated as a feature difference measure; The link hop frequency between nodes is analyzed by entropy value, and the link discontinuity measure is determined by the ratio of the distribution entropy of the hop frequency to the average hop interval; The feature difference measure and the normalized result of the link discontinuity measure are weighted and fused to generate a joint probability between nodes; Based on the joint probability, the tampering probability weight between nodes is generated by mapping through a softmax function.

9. The multi-modal AI content security risk tracing and identification detection system of claim 5, wherein, The method of risk propagation simulation comprises: Performing graph embedding on the multi-modal content traceability graph to generate low-dimensional feature vectors of graph nodes; Based on the low-dimensional feature vectors, generate multiple propagation paths through random walk algorithm, and record the access frequency of each node on the path; Performing path entropy analysis on the propagation path, and determining the propagation stability of the path by the ratio of the path entropy to the path length; Based on the propagation stability and the access frequency, calculate the tampering diffusion probability of each node in the graph, and identify the high-risk propagation subgraph in the graph through the spatial distribution characteristics of the diffusion probability.

10. The multi-modal AI content security risk tracing and identification detection system of claim 1, wherein, The system further comprises: Construct a content full-link traceability verification database, which includes the original feature fingerprint of multi-modal content, the tampering point distribution characteristics and the statistical characteristics of risk propagation path; Through the content full-link traceability verification database, the new input multi-modal content is quickly traced and matched to identify the historical tampering mode in the content link; Based on the matching result of the historical tampering mode, dynamically adjust the node weight and path intensity of the multi-modal content traceability graph to optimize the real-time and accuracy of the traceability graph; The optimization result of the traceability graph and the detection result of the content security risk are jointly output to provide a completeness verification report of the content link.

Images