A multi-modal generated content compliance checking method

Abstract

The application discloses a kind of multi-modal generated content compliance verification methods, belong to multi-modal information processing technical field.The method includes, obtaining the multi-modal generated content to be verified and corresponding associated information, and obtaining the identification result containing feature and semantic element by multi-modal analysis;Recognition result is mapped to contain content node, subject node, behavior node evidence chain graph, and cross-modal relationship edge that represents the semantic association between different modalities is established;Compliance rules are converted into graph pattern rules containing trigger subgraph and exception subgraph, and graph pattern matching is performed in evidence chain graph to identify risk path;Finally, compliance verification results are output based on risk path.The application can effectively identify cross-modal implicit compliance risks, improve risk identification accuracy and effectiveness of rectification scheme, realize the adaptive iteration of verification system.

Description

Technical Field

[0001] This invention relates to the field of multimodal information processing technology, and in particular to a method for verifying the compliance of multimodal generated content. Background Technology

[0002] With the rapid development of generative artificial intelligence technology, multimodal content generation technology has achieved the fusion and cross-modal creation of multimodal information such as text, images, audio, and video. At the same time, compliance risks associated with multimodal generated content, such as the spread of false information, intellectual property infringement, content misrepresentation, and missing generation identifiers, are becoming increasingly prominent, posing higher demands on content security management.

[0003] For example, CN120688627A discloses a multimodal decision support method for legal compliance review. This scheme achieves the alignment and fusion of text, visual and audio features through a cross-channel attention fusion network, uses a dual-tower matching network to perform similarity matching between business features and regulatory constraint vectors, and generates multi-dimensional interpretable information such as text keyword highlighting, image violation heatmaps and high-risk audio segments based on Grad-CAM. It is suitable for legal compliance risk screening in enterprise operation scenarios.

[0004] For example, CN120751172A discloses a content compliance detection method based on a multimodal large model. It constructs text, visual, and audio features into a heterogeneous graph structure, dynamically adjusts the weights of each modality feature in combination with the scene type, and completes compliance prediction through a graph neural network. It can adapt to the content security detection needs of different scenarios such as live streaming, short videos, and long videos.

[0005] For example, CN121093102A discloses an AI-generated content detection and review method. This method adopts a hierarchical processing path of first verifying text and then visually detecting. It dynamically schedules multimodal model resources based on the text compliance results to avoid redundant calculations and performs structured aggregation output of multimodal detection results. This can effectively reduce the consumption of computing resources in scenarios involving large-scale content processing.

[0006] However, the aforementioned multimodal compliance verification schemes still have some technical limitations. First, existing schemes mostly employ feature vector-level modal fusion or simple result aggregation, lacking a unified structured evidence representation system. They cannot establish semantic association paths between different modal information, making it difficult to identify implicit compliance risks such as cross-modal semantic conflicts, information masking, and content omissions. In particular, they cannot effectively detect malicious circumvention techniques such as audio-visual misalignment, text-image discrepancies, and contradictions between statements and facts. Second, verification rules are mostly in the form of discrete entries or vector constraints, failing to be transformed into a matchable topological graph pattern structure. This makes it impossible to locate a complete risk evidence chain in the global semantic graph, and there is a lack of exemption mechanisms for exceptional scenarios, resulting in insufficient completeness and accuracy of risk identification and difficulty in controlling the false positive rate. Third, existing schemes mostly adopt a single-round full-scale detection mode, lacking a confidence-driven hierarchical verification mechanism. They cannot trigger high-precision secondary identification for low-confidence areas; furthermore, they lack a counterfactual review stage for rectification suggestions, potentially introducing new compliance risks and compromising effectiveness. Fourth, the risk quantification dimension is relatively singular, and it cannot integrate multiple dimensions such as rule weight, scenario sensitivity, and cross-modal conflict factors to achieve comprehensive risk assessment. Furthermore, it lacks the optimization capability of human feedback, and the adaptive iteration capability of the verification system is weak. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0008] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a multimodal generated content compliance verification method to solve the problems mentioned in the background art.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a multimodal generated content compliance verification method, comprising: Obtain the multimodal generated content to be verified; The multimodal generated content is parsed to obtain a recognition result containing multimodal features; The identification result is mapped to an evidence chain graph, which includes nodes used to characterize the identification result; In the evidence chain graph, cross-modal relationship edges are established, which are used to represent the semantic association between nodes of different modalities; The compliance rules are transformed into graph pattern rules, and graph pattern matching is performed in the evidence chain graph to identify risk paths that match the graph pattern rules; Based on the risk path, output the compliance verification result.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention maps heterogeneous multimodal recognition results into a structured evidence chain graph and establishes four types of cross-modal relationship edges: semantic support, semantic conflict, information omission, and content masking. This breaks through the limitations of traditional independent detection of different modalities and can accurately identify hidden compliance risks such as cross-modal semantic misalignment and malicious masking, thereby improving the comprehensiveness of risk coverage.

[0011] 2. This invention transforms the scattered compliance rules into topological graph pattern rules that combine trigger subgraphs and exception subgraphs. It achieves semantic-level subgraph matching through graph attention networks, which can not only locate the complete risk path in the global graph, but also exclude compliance exemption scenarios through exception subgraphs. This can effectively reduce the false alarm rate and ensure the completeness and accuracy of risk identification.

[0012] 3. In addition, this invention also uses confidence-driven triggering of local high-precision secondary recognition to improve the recognition accuracy of complex scenarios while ensuring the efficiency of verification of massive amounts of content; and is equipped with a counterfactual verification mechanism for rectification suggestions to ensure that rectification actions can effectively block risks and not introduce new hidden dangers. It also combines multi-dimensional comprehensive risk quantification and manual feedback iteration to achieve continuous adaptive optimization of the verification system. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the overall process of a multimodal content generation compliance verification method according to an embodiment of the present invention. Detailed Implementation

[0014] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0016] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0017] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for 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. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0018] Example 1

[0019] Reference Figure 1 This is the first embodiment of the present invention, which provides a multimodal generated content compliance verification method, including: S1: Obtain the multimodal generated content to be verified.

[0020] It should be noted that this step aims to build a complete, multi-dimensional data context environment, which includes not only the generated content itself, but also the background and additional attributes of its generation.

[0021] Furthermore, data input from generative AI applications is received through a pre-defined data interface (e.g., a streaming API interface). In this invention, the content to be verified typically includes multiple data slices of different modalities, such as text, images, audio, and video. Moreover, in real-world generative AI scenarios, malicious users often circumvent single-modal compliance reviews by separating audio and video or misaligning text and images (e.g., a compliant image paired with non-compliant audio). Therefore, this invention, when accessing data, also extracts the timestamps and spatial coordinates of each modal data to establish an initial spatiotemporal synchronization mapping relationship. This ensures that multimodal data within the same generation batch is treated as an indivisible whole for subsequent verification, thereby preventing cross-modal risk escape.

[0022] Furthermore, while acquiring the multimodal generated content, the associated information corresponding to that content is acquired simultaneously. In this embodiment, the associated information includes information in the following three dimensions: Generated instructions: These are the original prompts or control parameters that the user inputs into the AI ​​model. Obtaining generated instructions is crucial because compliance verification not only needs to determine what was generated, but also to trace the user's intention in generating them. This is essential for determining whether there was malicious intent in inducing generation (e.g., a jailbreak attack).

[0023] Application Scenarios: Records the specific business environment in which the content will be distributed (e.g., medical consultation, news broadcasting, children's education, or entertainment pranks). The scenario parameters within these specific business environments will serve as the basis for subsequently dynamically invoking compliance rule sets. For example, the same generated image with exaggerated colors might be compliant in an entertainment prank scenario, but could trigger a violation in a news broadcast scenario due to a breach of the principle of realism.

[0024] User Declarations: Extract explicit declaration text attached by the user when publishing the content (e.g., This video was generated by AI and is purely fictional). These declaration texts will be used as the core comparison benchmark for establishing semantic conflict edges in subsequent steps.

[0025] Furthermore, in order to assign the necessary attributes (modal origin, location information, and confidence level) to the content nodes in the subsequent evidence chain graph, it is also necessary to perform low-level parsing and quality assessment on the incoming data. The processing procedure is as follows: First, extract basic attributes. Parse the modality source (e.g., the specific large model version number or vendor identifier that generated the content) and location information (e.g., the pixel coordinate range of the image, the start and end time frames of the audio and video) from the file header or underlying data stream. Then, calculate the initial confidence score. By calling a lightweight pre-trained evaluation model (e.g., a lightweight convolutional neural network for images, or a signal-to-noise ratio evaluation network for audio), calculate the initial confidence score of each modality data from the following three dimensions: The first dimension is the underlying signal quality, used to assess the clarity and integrity of the raw data. If the generated content has severe mosaic occlusion, blurriness, or audio truncation, the score in this dimension will be significantly reduced, thus indicating a risk of malicious content obscuring in that area during subsequent verification steps.

[0026] The second dimension is metadata integrity, which is used to detect whether the file contains a complete generative AI implicit watermark or a standard EXIF ​​generation identifier. If it is present, the confidence score is improved; if it is missing, it indicates that the compliance of the generation identifier is questionable.

[0027] The third dimension is anti-tampering timeliness, which calculates the time difference between the generation request time and the actual submission verification time of the content. Its purpose in preventing tampering is that if the time difference is too large, it indicates that the generated content has very likely undergone secondary human tampering after generation and before submission for verification (e.g., malicious modification of parts using image editing software). Therefore, the larger the time difference, the more non-linearly the confidence score of the content as the original generated content should decrease.

[0028] Finally, the evaluation results of the above three dimensions are fused using a preset weighting algorithm, and the final confidence score is normalized to a standard range (e.g., between 0 and 1).

[0029] In an feasible process, the weighting coefficients of this fusion can be dynamically optimized and determined by training a machine learning classifier on a historical compliance verification dataset.

[0030] In a feasible process, the initial confidence level is calculated as follows: ; in, Represents any modality from text, images, audio, or video; This indicates the underlying signal quality score for that mode; Indicates metadata integrity score; This represents the time difference between the generation time and the submission verification time. Represents the aging decay constant; The weighting coefficients are and satisfy the following conditions: ; Indicates will Cut off to Interval.

[0031] In a feasible process, if a machine learning classifier is used to determine the above weights, then the training samples are historical verification records. ,in The total number of samples, For the number of samples, , Labels indicating whether manually verified evidence is reliable or unreliable. Training uses cross-entropy. The loss function is defined as follows: weights are updated using the Adam optimizer, with a learning rate of 10. -4 ~10 -3 The batch size is 32 or 64, and training stops when the F1 score on the validation set does not improve for three consecutive rounds.

[0032] S2: Parse the multimodal generated content to obtain a recognition result containing multimodal features.

[0033] Furthermore, for data of different modalities, this invention calls the corresponding deep learning encoding network to perform vectorized representation of the low-level features: Text modality parsing: For the text portions of generated instructions, user declarations, and generated content, a pre-trained language model (e.g., a model based on the Transformer architecture) is used. Its input is a segmented text sequence, and the model captures the contextual dependencies between words through a self-attention mechanism, outputting a text feature representation containing global semantic information.

[0034] Visual modality analysis: For image or video frames, a visual convolutional neural network or a visual Transformer model is employed. Taking video as an example, keyframes can be extracted first at a preset frame rate and used as input to the model. Then, the model extracts global image features and local region features through a multi-layer network, outputting a high-dimensional visual feature representation.

[0035] Audio modality parsing: For audio data, it is converted into an acoustic feature map (e.g., Mel spectrogram), which is then input into an audio convolutional neural network to output an acoustic feature representation.

[0036] The aforementioned multimodal parsing model, in a feasible implementation, includes a text encoder, a visual encoder, an audio encoder, a semantic feature extraction head, and a unified projection layer. The text encoder includes a word embedding layer, a positional encoding layer, and... The Transformer layer consists of a self-attention layer and a text pooling layer; the visual encoder includes an image patch embedding layer or a convolutional feature layer. Layered visual Transformer / convolutional residual layers and region detection heads; audio encoders including short-time Fourier transform or Mel-frequency conversion layers, The system consists of convolutional / Transformer acoustic coding layers and temporal pooling layers. The outputs of the three types of encoders are mapped to the same dimension via linear projection layers. ,form Uniform feature vectors, where 256 or 512 are acceptable.

[0037] The training data for the model includes multimodal samples with compliance labels, subject bounding boxes, text entity labels, action labels, and audio event labels. The training steps include: first, preprocessing and labeling the text, image / video frames, and audio clips separately; then using classification loss... Location loss Cross-modal contrast loss Joint training, joint loss is ,in The weights corresponding to the loss are determined; finally, the entity recognition threshold, action recognition threshold, and audio event threshold are determined on the validation set to ensure that the recall rate and false positive rate meet the preset requirements.

[0038] It should be noted that, through the above steps, the present invention can map the originally heterogeneous pixel, sound wave and character data into a unified high-dimensional continuous feature space.

[0039] Furthermore, to construct subject nodes and behavior nodes, this invention further performs semantic element extraction on top of basic feature extraction. This is because the core of compliance verification often lies in clarifying: who (subject) did what (behavior) in what scenario. Specifically, the semantic element extraction operation is mainly divided into two parts. One part is subject extraction, which primarily uses a NER (Named Entity Recognition) model to extract entity words such as people, organizations, and specific items on the text side; and on the visual side, it calls an object detection model, taking image frames as input and outputting the bounding box coordinates of the target in the image and its category probability. These output core entities are then used as subject nodes in the subsequent graph. The other part is behavior extraction. On the text side, it uses dependency parsing to extract verb-object phrases (e.g., attack - someone); similarly, on the visual side, it calls a spatiotemporal action recognition network, combining motion trajectories between consecutive video frames (e.g., optical flow information) to identify dynamic behaviors in the image (e.g., beating, smoking); and on the audio side, it uses a sound event detection model to identify specific non-speech behaviors (e.g., gunshots, explosions). These extracted or identified actions and events are then used as behavior nodes in the subsequent graph.

[0040] Furthermore, in real-world generative AI violation scenarios, there are often covert circumvention methods such as discrepancies between text and images or misalignment of audio and video (e.g., the video is normal news, but the audio is replaced with inappropriate remarks). Therefore, this invention introduces a cross-modal attention mechanism to perform alignment calculations on the extracted features.

[0041] Specifically, taking visual and text alignment as an example, this invention uses visual features as query conditions to match various words in the text features. Then, an attention weight matrix is ​​calculated, which quantifies the semantic relevance between various local regions in the visual image and various words in the text description. If the generated image content is highly consistent with the user's generation instructions, the relevance score will show a high value.

[0042] Simultaneously, based on the output of the alignment calculation above, the present invention will yield a cross-modal semantic consistency score. By weighted and fused with this consistency score, the updated final confidence score can be calculated.

[0043] In an implementable process, with the first Features of each visual region and the Text lexical features For example, cross-modal attention is calculated as follows: ; ; ; ; ; in, This is the value vector projection matrix in cross-modal attention, used to map text word features into value vectors that can be weighted and converged by visual regions; For the first The visual region and the first Attention weights between text lexical units; The bias term is formed by spatiotemporal alignment or the same subject relationship; It is a non-linear activation function; This represents the query vector projection matrix in cross-modal attention, used to project visual region features. Mapped to query vector ; The key vector projection matrix in cross-modal attention is used to project text lexical features. Mapped to key vector ; The cross-modal attention output projection matrix is ​​used to map the concatenated fused features to the updated cross-modal features. For the first A new feature representation of a visual region after cross-modal attention fusion. For the first The value vector of each text word is composed of text word features. After value projection matrix The mapping yields attention weights. The weighted aggregation object is used to inject textual information into visual region features. It is the first The key vector obtained by mapping the text word features; Represents the query vector With key vector The inner product is used to measure the semantic relevance between the visual region and the text word.

[0044] In an implementable process, cross-modal semantic consistency score It can be represented as: ; in, express Belongs to the visual region index set That is, to traverse all visual regions / local image blocks. This represents the visual semantic projection matrix, used to project visual region features. Mapped into a shared semantic space used for similarity calculation. The text semantic projection matrix is ​​used to project text lexical features. They are mapped to the same shared semantic space.

[0045] In an implementable process, the updated confidence level for: ; in, The initial confidence level obtained from S1 is... This corresponds to the category confidence score output by the recognition model. For normalized weights.

[0046] It should be noted that through the above alignment calculation and update operations, even if the quality of the underlying signal of a single modality is very high, if a semantic disconnect is found between audio and video or text and images (i.e., the consistency score is extremely low), the overall confidence of the multimodal feature will be greatly reduced, and a basis will be provided for subsequent high-precision recognition of low-confidence regions.

[0047] Finally, all the features and elements obtained from the above analysis are structurally encapsulated to form a standardized set of recognition results. This set is divided into three categories according to the needs of subsequent map construction: Content node data: includes extracted multimodal fusion features, and is bound to the acquired modality source, spatial location information (e.g., bounding box coordinates), temporal location information (e.g., audio timestamps), and updated confidence attributes.

[0048] Main node data: contains extracted entity labels and their corresponding feature subsets.

[0049] Behavioral node data: includes extracted actions, event tags, and their spatiotemporal span information.

[0050] S3: Map the identification result to an evidence chain graph, the evidence chain graph including nodes used to characterize the identification result.

[0051] Specifically, combining the multimodal generated content and its associated information (generation instructions, application scenarios, or user statements) obtained in S1, this step divides the nodes of the evidence chain graph into three categories: content nodes, subject nodes, and behavior nodes, in order to comprehensively represent the content itself and its associated context.

[0052] Furthermore, this step does not involve establishing the connecting edges between nodes (edge ​​establishment will be completed in S4), but focuses on the instantiation of the graph vertex set and the vectorized mapping of attribute features. The nodes of the evidence chain graph are divided into three categories: content nodes, main nodes, and behavior nodes.

[0053] It needs to be explained that this content node is used to anchor the specific form of multimodal generated content within the real-world data stream. To ensure the content node possesses complete evidentiary value, the various attributes extracted from S2 also need to be fused. The specific processing method is as follows:

[0054] The system acquires the multimodal semantic features, spatiotemporal location information (e.g., bounding box coordinates of images or timestamps of audio), modality identifiers (e.g., text, generated images), and updated confidence scores from the S2 output. First, the spatiotemporal location information is positionally encoded, transforming it into a relative position vector that the model can understand. Then, the semantic features, position vectors, modality identifiers, and confidence scores are concatenated. The concatenated features are then input into a feature mapping network (e.g., a multilayer perceptron MLP) for nonlinear transformation, outputting a content node vector of uniform dimension.

[0055] In this process, the advantage of directly integrating the confidence score into the node vector is that the content node itself carries the reliability of the evidence and provides native underlying data support for subsequent secondary recognition based on confidence, while also preserving the spatiotemporal topology of the evidence in the original file.

[0056] It's important to explain that the subject node represents the core entities involved in the generated content (e.g., specific political figures, protected minors, registered trademarks, etc.). In multimodal scenarios, the same subject may appear in different forms (e.g., a text instruction might mention a celebrity, but the generated image might show that celebrity's face). To bridge this semantic gap between modalities, a unified spatial mapping of subject features is needed. The specific processing is as follows: The original entity features extracted from S2 are obtained, mapped to a shared semantic space through a pre-trained projection matrix, and L2 normalization is performed on the mapped vectors to obtain the final main node vectors.

[0057] It should be noted that the normalization operation here mainly forces all subject nodes to be bound to a unit hypersphere. In this way, regardless of whether the subject comes from text or vision, as long as they semantically refer to the same real entity, their cosine distance on the hypersphere will approach infinitely.

[0058] It needs to be explained that behavior nodes are used to represent actions performed by the subject or events that occur (e.g., spreading rumors, violent attacks, infringement). Since behaviors often have spatiotemporal continuity (especially video and audio modalities), it is necessary to aggregate the dynamic feature sequences extracted from S2. The specific processing method is as follows: For a given action, multiple instantaneous action feature sequences extracted along the time axis are processed using temporal average pooling. Specifically, all feature vectors within the duration of the action are summed and averaged, and then a single action node vector is generated through linear transformation and an activation function.

[0059] It should be noted that time-series aggregation can effectively filter out redundant frames and noise fluctuations during the execution of actions (such as irrelevant background changes in the video), thereby extracting the core semantic meaning of the behavior over the entire time span and ensuring that the behavior nodes can accurately reflect the essence of the actions required for compliance verification.

[0060] Furthermore, after the above steps, the recognition result of S2 can be transformed into three types of node sets in the graph space: content node set, subject node set, and behavior node set.

[0061] At this point, the present invention initializes the data structure of the evidence chain graph in memory. At the end of the current S3, this graph will only contain the aforementioned instantiated set of nodes, while the set of relationship edges between nodes will be empty. Furthermore, all nodes are independent entities, but each carries rich semantic, spatiotemporal, and confidence attributes.

[0062] In an feasible process, the chain of evidence diagram In memory, it can be formally represented as: ; in, A collection of content nodes. For the main node set, A set of behavior nodes; This is the set of relation edges, initialized to an empty set at the end of S3; For a collection of node attributes, , representing the node semantic vector, spatial location, time interval, modality source, and confidence level, respectively; This is a set of compliance rule indexes used to invoke the corresponding graph pattern rules in S5. This formalization clarifies how the output of S2 enters the S3 graph construction process.

[0063] S4: Establish cross-modal relationship edges in the evidence chain graph, which are used to characterize the semantic association between nodes of different modalities.

[0064] Furthermore, in multimodal scenarios, the faces of someone in text and someone in an image are fundamentally different in their underlying data structures. To establish a connection, subject normalization is the first step.

[0065] Specifically, the features of subject nodes belonging to different modalities are mapped to a shared semantic space, and the semantic similarity between these node features is calculated. When the semantic similarity between a text entity node and a visual entity node is detected to be greater than a preset subject identity threshold, it is determined that the two nodes refer to the same physical entity in the real world, and a basic identity edge is established between them.

[0066] Furthermore, for modalities that include a time dimension (such as video and audio), compliance risks are often instantaneous. Therefore, it is necessary to assess the degree of spatiotemporal synchronization between nodes.

[0067] Specifically, the time interval attributes carried by dynamic nodes (such as video frame nodes and audio sound nodes) are extracted, and the overlap ratio of these two time intervals on the time axis is calculated. Only when the overlap ratio is greater than a preset time alignment threshold is the spatiotemporal synchronization of the two nodes determined, and subsequent semantic relationship calculations between the two nodes are allowed.

[0068] In a feasible process, the association score of candidate cross-modal node pairs is: ; in, The shared semantic vector after subject normalization. For time intervals, For spatial regions, For node categories or declare attributes, This represents the weight of the corresponding item. If ( To establish a cross-modal relationship edge, if the subject identity score, time overlap ratio, or spatial overlap ratio meets the corresponding threshold, then the node pair is sent into the cross-modal relationship inference network; otherwise, no relationship edge is established. Intersection over Union (IoU) is used to indicate the degree of overlap between two spatial frames or two time intervals.

[0069] It should be noted that assessing the spatiotemporal synchronization between nodes can effectively prevent the incorrect association of unrelated events from different time periods (e.g., normal footage from one minute and illegal audio from the next minute), thereby reducing the false alarm rate of compliance verification.

[0070] Furthermore, between node pairs that have completed subject normalization and spatiotemporal alignment, a pre-trained cross-modal relation inference network is invoked to determine the specific edge type.

[0071] Specifically, the processing of this cross-modal relational reasoning network is as follows: First, the feature vectors of the source node and the target node are concatenated to obtain a joint feature vector. Then, this joint feature vector is input into a feature fusion layer containing multiple layers of cross-attention mechanisms, enabling the model to fully learn fine-grained semantic alignment and conflict relationships between features of different modalities. Finally, through a fully connected layer and a Softmax activation function, a probability distribution vector with four dimensions is output, corresponding to the predicted probabilities of four types of relation edges: semantic support, semantic conflict, information omission, and content occlusion. The category with the highest probability and greater than a preset confidence threshold is selected as the final relation edge type.

[0072] It is important to emphasize that this cross-modal relationship reasoning network requires supervised training in a specific compliance scenario before deployment. During this supervised training, a sample set containing a large number of multimodal text / image / audio / video pairs is constructed, and their relationship labels are manually annotated (e.g., labeling real news footage paired with fabricated illegal audio as semantic conflict). The model iterates through parameters by minimizing the cross-entropy loss function, thereby gaining the ability to accurately identify complex semantic interactions in multimodal generated content.

[0073] In a feasible implementation, the network structure and training process of the cross-modal relational reasoning network are as follows: The input layer receives the feature vectors of the source node and the target node. First, both are unified in dimension through a linear mapping layer. Then, they enter a multi-layer cross-attention layer, using the source node features as the query vector and the target node features as the key and value vectors, to calculate the cross-modal attention weights, thereby extracting the directional semantic association features between the source modality and the target modality. Next, the output of the cross-attention layer is residually connected to the original source / target features and layer normalized, and then concatenated to obtain a joint feature vector. Finally, a fully connected layer composed of an MLP (Multilayer Perceptron) is passed through the Softmax activation function to output a probability distribution vector with a dimension of 4 (corresponding to semantic support, semantic conflict, information omission, and content masking, respectively). Model training: In the offline training phase, a multimodal compliance verification sample set is constructed, containing a massive number of image-text and audio-video node pairs, and four types of relationship labels are manually labeled (e.g., node pairs of real news images paired with fake illegal audio are labeled as semantic conflict, and the label is one-hot encoded). During training, samples are input into the network to obtain the predicted probability distribution. The cross-entropy loss function is used to calculate the loss between the predicted distribution and the true labels, and the network parameters are updated through backpropagation using the Adam optimizer. To address the imbalanced sample problem, a Focal Loss mechanism is introduced into the loss function, increasing the penalty weight for difficult samples such as those with content occlusion, until the model's accuracy on the validation set converges.

[0074] More specifically, the cross-modal relational reasoning network includes two linear mapping layers. The system consists of a multi-layered cross-attention layer, a residual connection layer, a multi-layered normalization layer, a splicing layer, and two MLP classification heads. The number of attention heads can be 2-4, the number of attention heads can be 4-8, and the hidden layer dimension can be 256 or 512. The input for each training sample is... ,in and These are the features of the source node and the target node, respectively. Includes modality type, temporal overlap ratio, spatial overlap ratio, and metadata consistency flags; the output is... ,in These are the features resulting from the fusion of the source and target nodes. It is a classification weight matrix. It is a bias term. The four components correspond to semantic support, semantic conflict, information omission, and content masking, respectively. The network is trained using Focal Loss with class weights. ,in One-hot encoding for four types of relationship labels, To predict probabilities, For category weights, This is the focusing parameter and can be 2. The training, validation, and test sets are divided into three categories in an 8:1:1 ratio. The AdamW optimizer is used, with a learning rate of 10. -4 The weight decay is 10. -5 The test will stop when the average F1 value of the validation set reaches the preset value or when it does not improve for 5 consecutive rounds.

[0075] Specifically, the four types of relation edges are established as follows: Semantic conflict edges are established to characterize semantic contradictions, conflicts, or deceptiveness in content of different modalities. Specifically, this invention employs explicit rule calculation for scenarios involving file metadata or implicit watermarking. Extract the declaration nodes parsed from the user input text (e.g., text description: this image is a real photograph), and simultaneously extract the AI-generated content identifier nodes parsed from file metadata or implicit watermarks (e.g., underlying identifier: AI generated). Logically compare the authenticity declaration attributes of these two nodes. When it is determined that the attributes of the two nodes are inconsistent (i.e., there is a contradiction), a semantic conflict edge is forcibly established between the two nodes.

[0076] It should be noted that establishing semantic conflict edges can accurately combat deepfake fraud, a common phenomenon in generative AI applications (i.e., using AI to generate false content but claiming it to be real).

[0077] An information omission edge is established to characterize a mandatory constraint imposed by one modality but failed to be enforced by another. Key entity features required to be present in the source modality (e.g., constraint text nodes parsed from preset security instructions or user statements) are extracted and compared with the similarity of all content node features in the target modality (e.g., generated images or text). If the maximum similarity between all content nodes in the target modality and the key entity feature is still lower than a preset omission threshold, it indicates that the target modality is missing this key information, and an information omission edge is established from the source modality to the target modality. For example, compliance rules require that generated text involving medical advice must include a disclaimer. If no disclaimer content is matched in the generated text nodes, an information omission edge is established to effectively identify the risk of safety barriers failing.

[0078] Content occlusion edges are established to characterize the adversarial behavior of malicious users attempting to conceal inappropriate content by reducing local quality. The establishment of these edges relies on the confidence and spatial coordinate attributes resolved in previous steps. This primarily involves detecting subject nodes (e.g., sensitive figures) and visual content nodes (e.g., a region in an image) within the graph. When it is determined that these two nodes highly overlap in spatial coordinates, and the confidence score of the visual content node is extremely low (indicating severe blurring, pixelation, or artificial smearing in the region), a malicious occlusion risk is identified, and a content occlusion edge is established between the subject node and the visual content node.

[0079] In a feasible process, content occlusion edges The conditions for its establishment can be written as: and ; in, This refers to the spatial region corresponding to the main node. The spatial area for visual content nodes. The spatial overlap threshold, This is the low confidence threshold. For nodes The confidence level, and the aforementioned The meanings are the same, only the node numbers are different.

[0080] In a feasible process, the conditions for establishing an information omission edge can be written as: ; in, To enforce the shared semantic vectors of entities, The threshold for determining omissions. For nodes Normalized shared semantic vector. express It belongs to the target modal node set, such as the set of generated text nodes, image nodes, or video nodes to be checked for missing necessary content.

[0081] It should be noted that establishing content masking edges can combine spatial topology with signal quality, enabling the actual deployed system to resist malice and evade censorship. It can keenly detect compliance risks such as attempts to cover up violations, thereby preventing illegal content from evading verification through localized damage.

[0082] S5: Transform compliance rules into graph pattern rules, and perform graph pattern matching in the evidence chain graph to identify risk paths that match the graph pattern rules.

[0083] It should be noted that in traditional compliance verification systems, rules are typically represented by IF-THEN codes. This approach struggles to address cross-modal semantic misalignments in multimodal scenarios. In this invention, however, after processing by S4, the multimodal content is structured into an evidence chain graph containing nodes and relational edges. Therefore, the purpose of this step is to transform abstract compliance clauses into a concrete topological graph structure and use a graph neural network algorithm to perform fuzzy semantic subgraph search within the global graph, thereby accurately locating the evidence chain of violations.

[0084] It is important to emphasize that the graph pattern rules in this invention are not simple matching templates, but rather binary logic pairs composed of trigger subgraphs and exception subgraphs. This is significant because, in generative AI applications, some seemingly illegal content may be legal in a specific context (e.g., generating fake news footage is generally illegal, but if the user explicitly states that it is a movie special effects demonstration, it is a compliant exception). Therefore, this invention pre-constructs a graph pattern rule library: Triggering subgraphs are used to represent the topological structure of core violations. For example, for a fraud rule, the triggering subgraph is defined as follows: [Declaration node: Real shot] points to [Content node: AI generated image] through a semantic conflict edge.

[0085] An exception subgraph is used to represent the contextual topology that can exempt the violation. In the example above, the structure of the exception subgraph is defined as follows: [Main Node: Visual Effects] points to [Content Node: AI-Generated Image] through a semantic support edge.

[0086] Furthermore, since the nodes and edges generated in S3 and S4 contain high-dimensional continuous feature vectors (rather than simple scattered labels), traditional exact subgraph isomorphism algorithms cannot be directly applied. Therefore, this invention introduces the idea of ​​GAT (Graph Attention Network) to perform fuzzy semantic matching.

[0087] Specifically, the evidence chain graph constructed by S4 is input into the improved GAT model. To fully utilize the four types of cross-modal relation edges (semantic support, conflict, omission, and occlusion) established in S4, the following processing operations are performed when calculating the information transfer weights (i.e., attention coefficients) between nodes: For any target node in the graph, extract the node's own feature vector, the feature vectors of its neighboring nodes, and the feature vector of the edge connecting these two nodes (this edge feature vector is obtained by one-hot encoding and linear mapping of the edge's type label). Concatenate these three vectors, then use a single-layer feedforward neural network and the LeakyReLU nonlinear activation function to calculate the unnormalized attention score; finally, normalize it using the Softmax function to obtain the final attention weight.

[0088] It should be noted that by introducing edge features into the computation, the strength of information transmission between nodes depends not only on their own semantic content, but also on the type of cross-modal interaction that occurs between them. For example, if there is a semantically conflicting edge between two nodes, the edge feature vector will prompt the attention mechanism to calculate a higher weight, causing the risk signal of conflict to be amplified and spread rapidly during the feature update process of the graph.

[0089] Subsequently, the calculated attention weights are used to perform a weighted summation of the features of neighboring nodes, thereby updating the features of the target node.

[0090] In a feasible process, in the improved GAT model, the target node with neighboring nodes Unnormalized attention scores between for: ; in, and They are nodes and nodes eigenvectors, The feature vector of the edge connecting the two. This represents vector concatenation. and These are learnable parameters.

[0091] Normalized attention weights for: ; The node update formula is: ; in, Representing neighboring nodes For the target node Information contribution weight; This represents the projection matrix of the value vectors in the GAT layer; This represents a node after aggregating neighbor information. The new hidden representation.

[0092] It should be emphasized that, in order to avoid confusion, this invention will... This represents the value vector projection matrix in cross-modal attention. This represents the value vector projection matrix in the graph attention network; both are value projection parameters of the attention mechanism, but they serve different model layers and do not share parameters by default.

[0093] After updating node features, graph pooling is performed. It's important to note that, to ensure the effectiveness of similarity calculations, the pre-defined trigger and exception subgraphs are pre-mapped into standardized trigger and exception graph feature vectors offline using the GAT model and graph pooling layer, which share weights with the aforementioned model. During online matching, the extracted candidate subgraphs generate candidate graph feature vectors using the same graph pooling operation (e.g., global average pooling). Subsequently, cosine similarity is used to calculate the first similarity between the candidate graph feature vector and the trigger graph feature vector, and the second similarity with the exception graph feature vector.

[0094] In a feasible implementation, the training samples for improving the GAT model are: ,in These are candidate subgraphs extracted from historical content. To trigger a subgraph, As an exception subgraph, This indicates a risk label that has been manually confirmed. The model includes an input linear layer, Edge enhancement GAT layer, graph pooling layer, and similarity output layer; The value can be 2 or 3. The training objective is to make the risk subgraph similar to the corresponding trigger subgraph and dissimilar to the exception subgraph. The loss function can be expressed as: ,in The similarity between the candidate subgraph and the triggering subgraph. The similarity between the candidate subgraph and the exception subgraph. For the exception matching threshold, This is the coefficient used in the GAT training loss to adjust the weight of the exception subgraph penalty term.

[0095] Furthermore, after updating the graph features, a K-hop neighbor sampling algorithm based on core nodes is used to extract multiple local candidate subgraphs from the evidence chain graph. Specifically, taking the core node in the graph (the core node is specifically the main node or content node with a confidence level below a preset threshold) as the center, all nodes and edges within the K-layer topological relationship are extracted outward to form candidate subgraphs. Subsequently, graph pooling operations are used to map the candidate subgraphs, preset trigger subgraphs, and exception subgraphs to global graph feature vectors. The first similarity (e.g., cosine similarity) between the candidate subgraph and the trigger subgraph, and the second similarity between the candidate subgraph and the exception subgraph are calculated. Simultaneously, the following composite judgment is performed to confirm the risk:

[0096] The following conditions must be met simultaneously for a compliance risk to be identified: Triggering condition is met: The first similarity is greater than the preset trigger matching threshold (i.e., the candidate subgraph is highly suspected of violating the rules in terms of semantics and structure).

[0097] The exception condition is not met: the second similarity is less than the preset exception matching threshold (i.e., the candidate subgraph does not have the contextual conditions to exempt the violation).

[0098] In other words, the GAT model is mainly used to output the semantic similarity between candidate subgraphs, triggering subgraphs, and exception subgraphs. The final risk determination is still completed by the graph pattern rule that the triggering subgraph is valid and the exception subgraph is invalid. For highly deterministic rules such as contradictions between metadata and user claims, explicit logical comparison results are preferred and used as the basis for forcibly establishing semantic conflict edges. This avoids directly replacing the logical judgment of compliance rules with the fuzzy similarity of the neural network.

[0099] It should be noted that, among the above judgment conditions, only when a certain local structure in the evidence chain graph highly matches the violation characteristics and simultaneously does not meet the exemption conditions, is the local structure ultimately confirmed as a real compliance risk. Once the risk is confirmed, all nodes contained in the candidate subgraph (e.g., specific violation image regions, corresponding malicious prompt words) and the cross-modal relationship edges connecting them (e.g., semantic conflict edges) are completely extracted to form a structured risk path.

[0100] S6: Based on the risk path, output the compliance verification result.

[0101] It should be noted that after graph pattern matching in S5, the present invention has initially identified the chain of evidence of violations (i.e., risk paths) in the multimodal generated content. However, in order to ensure that the final compliance verification result is both highly accurate (avoiding false positives) and highly available (providing actionable remediation solutions), this step does not directly output a simple binary conclusion of violation / compliance, but instead constructs a comprehensive decision-making mechanism.

[0102] Furthermore, in actual engineering deployments, to ensure the real-time performance of verification of massive amounts of multimodal content, traditional solutions typically employ lightweight analytical models as the initial steps. This can lead to insufficient confidence in feature extraction for some complex scenarios (e.g., minor local forgeries, blurred background text). Therefore, this invention introduces an on-demand triggered secondary verification mechanism.

[0103] Specifically, for each node in the extracted risk path, obtain its associated preliminary risk assessment value and feature extraction confidence level. Then execute the following triggering process: When the initial risk assessment value associated with a node exceeds the preset risk threshold (i.e., there is a suspected high risk), or the confidence level of the underlying feature extraction is lower than the preset confidence level threshold (i.e., the feature extraction is extremely uncertain), the secondary identification process is triggered.

[0104] In an implementable process, the triggering condition for secondary identification can be expressed as: .in, The logical OR condition means that either of the two conditions must be met. For nodes The preliminary risk assessment value obtained in the initial graph pattern matching. Node confidence, As a risk threshold, The confidence threshold. This is an indicator function. When At that time, based on the spatial location of the node or time interval Extracting local slices from the original multimodal content ,in It is a local slice extraction function, used to extract slices based on spatial location. or time interval Extract the corresponding data from the original content and input the slice into the corresponding expert-level model. It is the original multimodal generated content data to be verified, which may include text, images, audio, video and their metadata.

[0105] Furthermore, once triggered, based on the node's location information, the corresponding data slice in the original file is extracted (e.g., extracting pixel regions of specific bounding boxes in an image, specific frames in a video, or specific segments in an audio file), and a secondary recognition model with significantly higher accuracy than the initial parsing model is invoked. For example, if the node is a suspected forged face image region, a heavyweight forgery detection network based on frequency domain analysis and micro-expression capture is invoked; if the node is blurry text, a fine-grained optical character recognition model is invoked.

[0106] Furthermore, after completing the secondary identification, the high-precision feature vectors output by the expert model and the new confidence scores are used to replace the corresponding attributes of the original nodes in the evidence chain graph, thereby completing a local update of the graph. Subsequently, graph pattern matching is re-executed in the updated evidence chain graph to confirm whether the risk path still holds true.

[0107] Specifically, in an implementable process, the output of this expert-level model includes new class probabilities. Local feature vectors and confidence level The map has been updated to... ,in Represents the spatial location information of the nodes. Indicates the time interval of the node. This indicates the origin of the node modality. The original node number and the historical version number of existing edges are retained to trace changes in risk paths before and after secondary identification. The training data for the expert-level model comes from manually reviewed difficult samples, low-confidence samples, and false positive / false negative samples. During training, the corresponding task's classification loss, detection localization loss, or sequence transcription loss is used, and the results are determined on the validation set. and .

[0108] It should be noted that through the secondary identification operation, the risk path can accurately locate the local area that needs to be reviewed, avoiding blindly calling high computing power models for the whole content, thereby achieving intensive use of computing power resources.

[0109] Furthermore, to provide control and management support to the business side, it is necessary to transform the risk paths of the topology into intuitive quantitative indicators. Therefore, this invention constructs a multi-dimensional risk quantification model to comprehensively extract feature data from the following dimensions for weighted calculation, thereby obtaining the final comprehensive risk score:

[0110] Basic rule weight: Extract the specific compliance rules violated by the risk path and assign a basic weight based on the importance of the rule in laws, regulations or platform specifications.

[0111] Dynamic adjustment of the severity of the main body: Analyze the attributes of the main body nodes involved in the risk path. If the main body node involves special protection groups (e.g., minors) or highly sensitive entities, the basic risk value is amplified exponentially.

[0112] Cross-modal conflict penalty factor: This is determined by statistically analyzing the number and intensity of semantic conflict edges and content masking edges contained in the risk path. If such cross-modal interaction edges exist, it indicates that the generated content not only violates regulations but also involves subjective deception and malicious concealment (e.g., using cheerful audio to mask violent visuals). Therefore, this invention introduces a penalty factor to significantly amplify the risk value.

[0113] Scenario sensitivity adaptation: Adjustments are made based on the acquired application scenario-related information. For example, the sensitivity coefficient for medical consultation scenarios is much higher than that for entertainment and prank scenarios, thus reflecting the scenario adaptability of compliance verification.

[0114] Evidence confidence decay: Aggregates the confidence levels of all nodes along the risk path. If the overall reliability of the underlying evidence is low, it is used as a decay coefficient to reduce the final risk alarm level, thereby effectively controlling the false alarm rate.

[0115] In an feasible process, the calculation logic for the comprehensive risk score adopts a multi-factor weighted and penalty model, as shown in the following formula: ; in, This is a risky path; This is the basic weight of the rule, with a value ranging from 0 to 100; This represents the severity coefficient of the rule. This is a dynamic adjustment coefficient for the severity of the subject matter; for example, 1.5 is used for minors and 1.0 is used for ordinary subjects. The scene sensitivity adaptation coefficient; This represents the cumulative number or intensity of semantic conflict edges and content occlusion edges in the risk path; edge confidence; Representing an edge Belongs to the risk path The set of edges; This represents an edge type function used to return edges. Types such as semantic conflict edges, content occlusion edges, etc.; This is the cross-modal conflict penalty factor. For example, a value of 0.2 means that for each additional high-confidence conflict edge, the risk value is increased by 20%. This is the confidence decay coefficient of evidence. Indicates risk path The set of nodes contained therein. If Exceeding alarm threshold If so, it is determined to be a substantive violation; if If the risk level is below the review threshold, a manual review prompt will be displayed instead of a low-risk conclusion being drawn directly.

[0116] Furthermore, after calculating the comprehensive risk score, based on the topology of the risk path, corresponding preliminary rectification suggestions are mapped out through a preset knowledge graph (e.g., for the information omission edge, the suggestion is to add AI-generated identifiers in specific time periods of the video; for the semantic conflict edge, the suggestion is to modify the prompt words and remove false statements).

[0117] Furthermore, to prevent rectification actions from creating new compliance loopholes, this invention also introduces a counterfactual verification mechanism.

[0118] Specifically, a complete copy of the current evidence chain graph is first created in memory. Then, within the copy, graph editing operations simulate the actions corresponding to the initial rectification suggestions. To achieve automatic conversion of natural language rectification suggestions into graph structure operations, this invention pre-configures a rectification action mapping dictionary. This dictionary defines the mapping relationship between different types of rectification suggestions and graph editing instructions (e.g., InsertNode adds a node, DeleteEdge deletes an edge, UpdateAttr modifies an attribute). For example, if the initial rectification suggestion is to add an AI-generated identifier, the mapping dictionary is queried to convert it into a graph editing instruction: an InsertNode operation is executed in the copy to inject a compliance declaration node (with the attribute being AI-generated), and an AddEdge operation is executed to establish a semantic support edge between this node and the corresponding visual content node; if the suggestion is to delete a non-compliant image, it is converted into an instruction: a DeleteNode operation is executed in the copy to remove the corresponding visual content node, and all its associated edges are cascaded and deleted. Finally, graph pattern matching is re-executed on the updated copy. Only when the original risk path is successfully blocked in the replica after the simulated rectification action, and no new risk path is generated in the replica, is the rectification recommendation confirmed to have passed the counterfactual review and marked as a safe and executable final rectification recommendation.

[0119] In an implementation process, the rectification suggestion can be formalized as a diagram editing instruction. ,in It is one of InsertNode, DeleteNode, AddEdge, DeleteEdge, or UpdateAttr. For the node or edge being edited, For newly added or modified attribute values, These are compliance constraints that must not be violated. The conditions for passing a counterfactual review are: ,in A copy of the map after the rectification order has been executed. This is the graph pattern corresponding to the original risk path. For the rule base, This refers to potential new risk patterns or risk paths that may emerge after counterfactual verification. Indicates a mismatch. The logical AND condition means that both conditions must be true simultaneously.

[0120] Finally, the above processing results are structurally assembled to output the final compliance verification result. This result not only includes a comprehensive risk score and a conclusion on whether a violation has occurred, but also displays the complete risk path in the form of a visual graph, along with final rectification suggestions that have passed counterfactual review. Simultaneously, this invention also provides a manual review interface for receiving feedback from auditors on the compliance verification results (e.g., confirmation of violations, false alarms, or omissions). This feedback is used as a monitoring signal to dynamically update the matching threshold of the graph pattern rules, or to fine-tune the various weight parameters used to calculate the comprehensive risk score.

[0121] In an implementable process, the parameter updates from manual review feedback can be represented as: ; in, This includes graph pattern matching thresholds, rule weights, and risk score weights; This indicates that the violation was manually confirmed. This indicates a false alarm that has been manually confirmed. To update the step size; This indicates the overall risk score relative to the parameter. The gradient or approximate incremental direction. For non-differentiable rule thresholds, grid search or Bayesian optimization can be used to redetermine them on a manually verified sample set. This is the set of parameters updated based on manual feedback. This is the set of parameters before the manual feedback update; These are the lower and upper limits allowed during parameter updates, used to prevent thresholds or weights from going out of bounds.

[0122] It should be noted that by introducing human feedback, the compliance verification scheme of this invention has the ability to continuously learn and adaptively iterate, and can continuously improve its sensitivity and accuracy in multimodal risk identification as laws and regulations are updated and new violation methods emerge.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for compliance verification of multimodal generated content, characterized in that, include: Obtain the multimodal generated content to be verified; The multimodal generated content is parsed to obtain a recognition result containing multimodal features; The identification result is mapped to an evidence chain graph, which includes nodes used to characterize the identification result; In the evidence chain graph, cross-modal relationship edges are established, which are used to represent the semantic association between nodes of different modalities; The compliance rules are transformed into graph pattern rules, and graph pattern matching is performed in the evidence chain graph to identify risk paths that match the graph pattern rules; Based on the risk path, output the compliance verification result.

2. The multimodal generated content compliance verification method as described in claim 1, characterized in that, The process of obtaining the multimodal generated content to be verified also includes: Obtain associated information, including generation instructions, application scenarios, or user declarations; The nodes in the evidence chain graph include content nodes, subject nodes, and behavior nodes. The content node records include modality source, location information, and confidence attribute.

3. The multimodal generated content compliance verification method as described in claim 1, characterized in that, Establishing cross-modal relationship edges in the evidence chain graph includes: Based on subject normalization, spatiotemporal alignment, and semantic relation calculation, semantic support edges, semantic conflict edges, information omission edges, or content occlusion edges are established between nodes belonging to different modalities.

4. The multimodal generated content compliance verification method as described in claim 3, characterized in that, The establishment of the semantic conflict edge includes: When the claim node extracted from the user claim is inconsistent with the AI-generated content identifier node extracted from the file metadata or implicit watermark, the semantic conflict edge is established between the claim node and the AI-generated content identifier node.

5. The multimodal generated content compliance verification method as described in claim 1, characterized in that, The graph pattern rules include triggering subgraphs and exception subgraphs: Performing graph pattern matching in the evidence chain graph includes: Candidate subgraphs are obtained by sampling neighbors in the evidence chain graph based on the core nodes; Determine whether there exists a subgraph structure in the candidate subgraph that matches the trigger subgraph but does not match the exception subgraph.

6. The multimodal generated content compliance verification method as described in claim 1, characterized in that, The method further includes: Based on the risk value or confidence level of the nodes in the risk path, a secondary identification is triggered on the corresponding local content in the multimodal generated content; The evidence chain graph is updated based on the results of the secondary identification, and graph pattern matching is re-executed in the updated evidence chain graph.

7. The multimodal generated content compliance verification method as described in claim 6, characterized in that, Triggering secondary recognition of the corresponding local content in the multimodal generated content includes: When the risk value of a node in the risk path exceeds a preset threshold, or the confidence level of a node is lower than a preset threshold, the image region, video frame, or audio segment corresponding to the node is extracted, and a recognition model with higher accuracy than the initial analysis model is called for secondary recognition.

8. The multimodal generated content compliance verification method as described in claim 1, characterized in that, The method further includes: Based on the aforementioned risk path, rectification suggestions are generated; A counterfactual review will be conducted on the aforementioned rectification recommendations, and the counterfactual review includes: Based on a preset rectification action mapping dictionary, the rectification suggestions are converted into graph editing instructions; In a copy of the evidence chain graph, the corresponding rectification actions are simulated according to the graph editing instructions, and graph pattern matching is re-executed to confirm the change in risk status; The rectification recommendations obtained through the counterfactual review will be output as part of the compliance verification results.

9. The multimodal generated content compliance verification method as described in claim 8, characterized in that, The confirmation of the change in risk status includes: When the original risk path is blocked and no new risk path is generated in the copy of the evidence chain graph after the rectification action is simulated, the rectification suggestion is determined to pass the review.

10. The multimodal generated content compliance verification method as described in claim 1, characterized in that, The method further includes: Based on the aforementioned risk path, and combined with rule weights, risk severity, evidence confidence, scenario sensitivity, and cross-modal conflict factors, a comprehensive risk score is calculated. Receive manual review feedback on the compliance verification results, and update the graph pattern rules or the weights used to calculate the comprehensive risk score based on the feedback.

Images