A Method and System for Quantifying the Entire Chain of Liability for Illegal Behavior Based on Dynamic Behavioral Graphs
Patent Information
- Application Number
- CN202611181676.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-05
- Publication Date
- 2026-10-27
AI Technical Summary
[0002]当前针对AI违法行为的治理多集中在前端识别拦截环节，缺少对违法行为全链条参与主体责任划分的量化方案，一方面，AI违法行为涉及多个参与主体，包括AI工具开发提供方、违法行为脚本编写者、引流推广者等多个环节参与者，不同主体参与程度不同，现有责任认定多依赖人工定性，缺乏客观统一的量化标准，容易出现责任认定偏差
[0034]1、本申请通过生成动态行为层级图谱直观展示AI违法行为全链条中各参与主体的行为关联与层级关系，同时通过责任分加权量化出每个主体在整个违法行为链条中所承担的责任大小，既能够帮助办案人员快速梳理违法行为流程、定位核心责任主体，也能为司法裁判中责任认定环节提供可量化的数据参考，解决了当前AI违法行为链条隐蔽、主体责任难以清晰界定的问题，可为AI违法行为案件的办理与防范提供明确的数据支撑。
Smart Images
Figure CN122989137A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of AI-based full-chain responsibility quantification technology, specifically to a method and system for AI-based full-chain responsibility quantification based on dynamic behavior graphs. Background Technology
[0002] Current governance of AI-related illegal activities primarily focuses on the front-end identification and interception stage, lacking a quantitative scheme for delineating the responsibilities of all participants across the entire chain of illegal activities. On one hand, AI-related illegal activities involve multiple participants, including AI tool developers, scriptwriters of illegal activities, and promoters, among others. The degree of participation varies among these participants, and current liability determinations largely rely on manual characterization, lacking objective and unified quantitative standards, which can easily lead to biases in liability assessment. On the other hand, existing anti-fraud data is mostly stored statically and in isolation, failing to fully present the behavioral connections and participation levels of each entity throughout the entire illegal activity process. This makes it difficult to support full-chain liability tracing and quantitative delineation, allowing some responsible parties to escape due punishment and posing significant obstacles to judicial determination of responsibility and sentencing.
[0003] To address this, a method and system for quantifying the full-chain responsibility for AI-related illegal activities based on dynamic behavioral graphs are proposed. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for quantifying the responsibility of AI-related illegal behaviors across the entire chain based on dynamic behavioral graphs. By generating graphs and quantifying data of AI-related illegal behaviors, it can provide data support for subsequent cases.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] A method for quantifying the entire chain of responsibility for illegal behavior based on dynamic behavioral graphs, including:
[0007] Acquire simulated information data of the entire chain of simulated AI illegal behavior, extract subject behavior data based on the key behavioral features of the simulated information data, and generate a subject index directory by associating subject behaviors based on simulated multi-source interaction logs and the subject behavior data;
[0008] Optionally, the subject behavior data is obtained by extracting semantic behavior information of the simulated information data of the initiator and the receiver based on the semantic information extraction model. The subject behavior data includes at least the semantic behavior information of the initiator and the semantic behavior information of the receiver. The semantic behavior information of the initiator includes at least the initiator's behavior subject, the initiator's behavior action, and the initiator's behavior goal. The semantic behavior information of the receiver includes at least the receiver's behavior subject, the receiver's behavior action, and the receiver's behavior goal.
[0009] Based on the simulated multi-source interaction logs and the semantic information extraction model, log semantic information is extracted for the same initiator behavior subject, the same receiver behavior subject, the same initiator behavior target, and the same receiver behavior target. The log semantic information includes at least implementation subject data and implementation behavior data. The implementation behavior data with the same implementation subject data are marked to associate the subject behavior and obtain subject data tags.
[0010] The subject index directory is obtained by associating the implementation behavior data with initial weights based on the subject data tags. This method can first complete the standardized anchoring of the behavior of each participating subject in the entire chain of AI illegal behavior by deconstructing and sorting out the behavior of different subjects in the simulated scenario from multiple dimensions. This solves the problem of scattered subject information and fragmented behavior segments in interaction logs from different sources. It can quickly map scattered interaction information to specific participating subjects, providing a clear and unified subject basis for the subsequent division of responsibility for the entire chain. It avoids deviations in responsibility determination due to subject information mismatch and behavior omissions, and improves the efficiency and accuracy of the subject sorting link in the responsibility quantification process.
[0011] Based on the main index directory and knowledge base, key behavioral features are extracted, and based on the behavioral association features of the key behavioral features, the main index directory is weighted and matched to generate a dynamic behavioral basic graph.
[0012] Optionally, illegal behavior data can be obtained by extracting AI illegal behavior entries from the knowledge base based on the semantic information extraction model.
[0013] The clustering model uses the number of data from the implementing entities as the number of clusters. Clustering analysis is performed on the data of the implementing behavior and the data of the illegal behavior with the same data label to obtain behavior clusters. The data within the cluster with the smallest distance from the cluster center of the behavior cluster is obtained according to the Euclidean distance algorithm and used as the cluster label of each cluster.
[0014] The behavior similarity between each cluster label and the illegal behavior data is calculated using a cosine similarity algorithm. All cluster data of the clusters with behavior similarity greater than the behavior threshold are marked as suspected illegal behavior data. The dynamic behavior architecture is generated by associating the suspected illegal behavior data with the same subject data label.
[0015] The dynamic behavior architecture is updated by weighting the behavior similarity of the suspected illegal behavior data and the subject data to obtain the dynamic behavior basic map.
[0016] This method constructs a dynamic behavioral foundation graph containing multi-layered relationships between subjects and behaviors. It can connect and integrate the scattered and hidden participating subjects and behaviors at each stage in the entire chain of AI-related illegal activities. This solves the problem of difficulty in locating responsible subjects due to the fragmented division of labor and the concealment of behaviors in AI-related illegal activities. It can intuitively present the degree of participation and behavioral relationships of different subjects in the entire process of illegal activities, providing a clear data foundation for the subsequent quantification of responsibility in the entire chain, and effectively improving the accuracy of responsibility identification and source tracing efficiency for AI-related illegal activities.
[0017] The weights of the dynamic behavior base map are corrected and updated based on the time feature information, and the time feature information and the dynamic behavior base map are matched to generate a dynamic behavior hierarchy map.
[0018] Optionally, the suspected illegal behavior data is extracted from the dynamic behavior baseline graph, and the illegal behavior implementation logic graph is generated based on the number of times and time sequence of the suspected illegal behavior data obtained from the simulated multi-source interaction log to obtain the time feature information;
[0019] Anomaly analysis is performed on the logic diagram of the illegal behavior based on the time anomaly analysis model to obtain anomaly behavior labels. Based on the anomaly behavior labels, weight update values are obtained to correct and update the weights of the dynamic behavior basic graph.
[0020] The dynamic behavior hierarchical map is generated by hierarchically labeling the main data of the dynamic behavior base map based on the abnormal behavior labels.
[0021] This method, by introducing time-dimensional feature correction and hierarchical labeling, can further distinguish the behavioral sequence and degree of abnormality of different participating entities in the entire process of AI-related illegal activities, highlighting the core implementing entities and key behavioral nodes. It solves the problem of difficulty in distinguishing the homogeneity of the main responsibilities of different links, making the graph's presentation of the main participants' participation more in line with the actual characteristics of AI-related illegal activities of "division of labor and cooperation, and sequential advancement," and providing more accurate hierarchical data support for the subsequent precise quantification of the responsibilities of each entity.
[0022] The initial responsibility score of each level is obtained based on the dynamic behavior hierarchy graph and knowledge base. The responsibility score of the dynamic behavior hierarchy graph is obtained by weighting the weights of the dynamic behavior hierarchy graph according to the responsibility score.
[0023] Optionally, the responsibility score of the dynamic behavior hierarchy graph is calculated using a responsibility score weighting function, including:
[0024] ;
[0025] in, The weights are the weights of the main body of the dynamic behavior hierarchy graph at the i-th level. The weights of the dynamic behavior hierarchy graph at the i-th level are given by [the weights]. The number of illegal behavior data of the subject in the i-th dynamic behavior hierarchy graph. The maximum value among the initial responsibility scores of the t dynamic behavior hierarchy graphs. For cosine similarity, This refers to the j-th illegal act of the subject in the i-th dynamic behavior hierarchy graph. For the t-th historical illegal behavior standard behavior in the knowledge base, The initial responsibility score is given for the t-th historical illegal behavior standard behavior in the knowledge base. This method combines the responsibility score of historical illegal behavior standard behavior with behavior similarity to weight and correct the weights of each level of the graph. This makes the final responsibility score more closely reflect the severity of the actual illegal behavior, avoiding the shortcomings of fixed weights that cannot adapt to different illegal behavior patterns. At the same time, relying on the hierarchical structure of the dynamic behavior hierarchy graph, the position and degree of participation of each illegal behavior participant in the whole chain can be clearly located, realizing the responsibility division of the entire process from the initiation of illegal behavior, traffic diversion, transfer to the distribution of spoils. This solves the problem of difficulty in accurately quantifying responsibility in AI illegal behavior scenarios with multiple participants and concealed division of labor, and provides a quantifiable basis for responsibility determination for subsequent anti-fraud handling and judicial discretion.
[0026] Output the responsibility score and the dynamic behavior hierarchy graph.
[0027] This application also provides an AI-based system for quantifying the entire chain of responsibility for illegal behavior based on dynamic behavioral graphs, including:
[0028] The main index directory module is used to acquire simulated information data of the entire chain of simulated AI illegal behavior, extract main behavior data based on the key behavioral features of the simulated information data, and generate a main index directory by associating main behaviors with simulated multi-source interaction logs and the main behavior data.
[0029] The dynamic behavior basic graph module is used to extract key behavior features based on the main index directory and knowledge base, and to perform weighted association matching on the main index directory based on the behavior association features of the key behavior features to generate a dynamic behavior basic graph.
[0030] The dynamic behavior hierarchy graph module is used to correct and update the weights of the dynamic behavior base graph based on time feature information, and to generate a dynamic behavior hierarchy graph by matching the time feature information and the dynamic behavior base graph.
[0031] The responsibility score generation module is used to obtain the initial responsibility score of each level based on the dynamic behavior hierarchy graph and the knowledge base, and to perform data weighting on the weights of the dynamic behavior hierarchy graph according to the responsibility score to obtain the responsibility score of the dynamic behavior hierarchy graph.
[0032] The output module is used to output the responsibility score and the dynamic behavior hierarchy map.
[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0034] 1. This application visually displays the behavioral relationships and hierarchical connections among various participants in the entire chain of AI-related illegal activities by generating a dynamic behavioral hierarchy map. At the same time, it quantifies the extent of responsibility borne by each entity in the entire chain of illegal activities through a weighted responsibility score. This not only helps investigators quickly sort out the illegal activity process and identify the core responsible parties, but also provides quantifiable data references for the determination of responsibility in judicial decisions. It solves the current problems of the concealment of AI-related illegal activity chains and the difficulty in clearly defining the responsibilities of the main parties, and can provide clear data support for the handling and prevention of AI-related illegal activity cases. Attached Figure Description
[0035] Figure 1 This is a flowchart of the method of the present invention.
[0036] Figure 2 This is a system flowchart of the present invention. Detailed Implementation
[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0038] This invention provides a method and system for quantifying the entire chain of responsibility for illegal behavior based on dynamic behavioral graphs. The technical solution is as follows:
[0039] Example 1:
[0040] A method for quantifying the entire chain of responsibility for illegal behavior based on dynamic behavioral graphs, referencing Figure 1 As shown, it includes:
[0041] The simulation information data of the entire chain of simulated AI illegal behavior is obtained. Based on the key behavioral features of the simulation information data, the subject behavior data is extracted. Based on the simulated multi-source interaction logs and the subject behavior data, the subject behavior is associated to generate a subject index directory. The entire chain of simulated AI illegal behavior is compiled and provided by relevant technical personnel.
[0042] Optionally, the subject behavior data is obtained by extracting semantic behavior information of the simulated information data of the initiator and the receiver based on a semantic information extraction model (such as the BERT model). The subject behavior data includes at least the semantic behavior information of the initiator and the semantic behavior information of the receiver. The semantic behavior information of the initiator includes at least the initiator's behavior subject, the initiator's behavior action, and the initiator's behavior goal. The semantic behavior information of the receiver includes at least the receiver's behavior subject, the receiver's behavior action, and the receiver's behavior goal. The data content includes, for example, illegal behavior scripts, transfer guidance information, victim transfer records, fund diversion details, etc.
[0043] Based on the simulated multi-source interaction logs and the semantic information extraction model (e.g., a pre-trained BERT model), log semantic information is extracted for the same initiator, receiver, initiator, and receiver behavior targets. The log semantic information includes at least implementation subject data and implementation behavior data. The implementation behavior data with the same implementation subject data are labeled to associate subject behavior, resulting in subject data labels. Specifically, association is performed using a default association weight, which is 1.0 by default, but can also be set based on expert experience. Simulated multi-source interaction logs include, for example, simulated AI illegal behavior communication logs, simulated AI illegal behavior cash transfer record logs, and simulated AI-generated fake contract signing logs.
[0044] The subject index directory is obtained by associating the implementation behavior data with initial weights based on the subject data tags. This method can first complete the standardized anchoring of the behavior of each participating subject in the entire chain of AI illegal behavior by deconstructing and sorting out the behavior of different subjects in the simulated scenario from multiple dimensions. This solves the problem of scattered subject information and fragmented behavior segments in interaction logs from different sources. It can quickly map scattered interaction information to specific participating subjects, providing a clear and unified subject basis for the subsequent division of responsibility for the entire chain. It avoids deviations in responsibility determination due to subject information mismatch and behavior omissions, and improves the efficiency and accuracy of the subject sorting link in the responsibility quantification process.
[0045] Key behavioral features are extracted from the main index directory and knowledge base. Based on the behavioral association features of these key behavioral features, a weighted association matching process is performed on the main index directory to generate a dynamic behavioral baseline graph. The knowledge base is provided by relevant experts and includes illegal behaviors, legal basis, etc.
[0046] Optionally, illegal behavior data can be obtained by extracting AI illegal behavior entries from the knowledge base based on the semantic information extraction model (e.g., a pre-trained BERT model).
[0047] The clustering model uses the number of data from the implementing entities as the number of clusters. Clustering analysis is performed on the data of the implementing behavior and the data of the illegal behavior with the same data label to obtain behavior clusters. The data within each cluster with the smallest distance from the cluster center is obtained using the Euclidean distance algorithm as the cluster label of each cluster. The clustering model selected is the K-means model, and the size of K is consistent with the number of data from the implementing entities.
[0048] The behavior similarity between each cluster label and the illegal behavior data is calculated using a cosine similarity algorithm. All cluster data of the clusters with behavior similarity greater than a behavior threshold are marked as suspected illegal behavior data. The dynamic behavior architecture is generated by associating the suspected illegal behavior data with the same subject data labels. The behavior threshold can be set to 0.7, 0.75, 0.78, 0.8, 0.81, etc., and can be specifically set by experts.
[0049] The dynamic behavior architecture is updated by weighting the behavior similarity of the suspected illegal behavior data and the subject data to obtain the dynamic behavior basic map.
[0050] This method constructs a dynamic behavioral foundation graph containing multi-layered relationships between subjects and behaviors. It can connect and integrate the scattered and hidden participating subjects and behaviors at each stage in the entire chain of AI-related illegal activities. This solves the problem of difficulty in locating responsible subjects due to the fragmented division of labor and the concealment of behaviors in AI-related illegal activities. It can intuitively present the degree of participation and behavioral relationships of different subjects in the entire process of illegal activities, providing a clear data foundation for the subsequent quantification of responsibility in the entire chain, and effectively improving the accuracy of responsibility identification and source tracing efficiency for AI-related illegal activities.
[0051] The weights of the dynamic behavior base map are corrected and updated based on the time feature information, and the time feature information and the dynamic behavior base map are matched to generate a dynamic behavior hierarchy map.
[0052] Optionally, the suspected illegal behavior data is extracted from the dynamic behavior baseline graph. The implementation number and time sequence of the suspected illegal behavior data are obtained based on the simulated multi-source interaction log to generate an illegal behavior implementation logic graph to obtain the time feature information. A two-dimensional coordinate axis is generated based on the implementation number and time sequence, with the vertical axis representing the implementation number and the horizontal axis representing time. The time can be set in hours, days, 5 days, a week, half a month, a month, etc., and can be freely set with reference to expert opinions.
[0053] Anomaly analysis is performed on the logic graph of the illegal behavior based on a time anomaly analysis model (such as a pre-trained LSTM model, Transformer model, etc.) to obtain anomaly behavior labels. The weights of the dynamic behavior base graph are then corrected and updated based on the weight update values obtained from the anomaly behavior labels. The anomaly behavior labels can include multiple levels of labels, such as minor impact, moderate impact, major impact, and extremely significant impact. Data labels can be set as different types of labels such as numeric labels and encrypted labels. The specific number of levels and label types are set based on expert experience.
[0054] The dynamic behavior hierarchical map is generated by hierarchically labeling the main data of the dynamic behavior base map based on the abnormal behavior labels.
[0055] This method, by introducing time-dimensional feature correction and hierarchical labeling, can further distinguish the behavioral sequence and degree of abnormality of different participating entities in the entire process of AI-related illegal activities, highlighting the core implementing entities and key behavioral nodes. It solves the problem of difficulty in distinguishing the homogeneity of the main responsibilities of different links, making the graph's presentation of the main participants' participation more in line with the actual characteristics of AI-related illegal activities of "division of labor and cooperation, and sequential advancement," and providing more accurate hierarchical data support for the subsequent precise quantification of the responsibilities of each entity.
[0056] The initial responsibility score for each level is obtained based on the dynamic behavior hierarchy graph and knowledge base. Specifically, the historical responsibility score data of the historical hierarchy graph is obtained from the knowledge base and assigned accordingly. The historical responsibility score data is set based on expert experience. The responsibility score of the dynamic behavior hierarchy graph is obtained by weighting the weights of the dynamic behavior hierarchy graph according to the responsibility score.
[0057] Optionally, the responsibility score of the dynamic behavior hierarchy graph is calculated using a responsibility score weighting function, including:
[0058] ;
[0059] in, The weights are the weights of the main body of the dynamic behavior hierarchy graph at the i-th level. The weights of the dynamic behavior hierarchy graph at the i-th level are given by [the weights]. The number of illegal behavior data of the subject in the i-th dynamic behavior hierarchy graph. The maximum value among the initial responsibility scores of the t dynamic behavior hierarchy graphs. For cosine similarity, This refers to the j-th illegal act of the subject in the i-th dynamic behavior hierarchy graph. For the t-th historical illegal behavior standard behavior in the knowledge base, Let t be the initial liability score for the t-th historical illegal behavior standard behavior in the knowledge base. Based on expert experience, various scores are set, with variable values such as 0.2, 0.5, 0.7, or even 10, 20, 50. This method combines the responsibility score of historical illegal behavior with behavioral similarity to weight and adjust the weights of each level of the graph. This ensures that the final responsibility score more closely reflects the severity of the actual illegal behavior, avoiding the shortcomings of fixed weights that cannot adapt to different illegal behavior patterns. Furthermore, relying on the hierarchical structure of the dynamic behavioral graph, the position and degree of participation of each participant in the illegal behavior can be clearly identified throughout the entire chain. This enables the delineation of responsibility for the entire process from initiation, traffic generation, and transfer to the distribution of spoils. It solves the problem of difficulty in accurately quantifying responsibility in AI-driven illegal behavior scenarios where multiple parties participate and their roles are concealed, providing a quantifiable basis for responsibility determination in subsequent anti-fraud handling and judicial discretion.
[0060] Output the responsibility score and the dynamic behavior hierarchy graph. The output can be based directly on the data score and graph, or it can be combined with the Agent and other parameters for a comprehensive judgment to generate a solution before outputting.
[0061] Example 2:
[0062] This application also provides an AI-based system for quantifying the entire chain of responsibility for illegal behavior based on dynamic behavioral graphs, referring to... Figure 2 As shown, it includes:
[0063] The main index directory module is used to acquire simulated information data of the entire chain of simulated AI illegal behavior, extract main behavior data based on the key behavioral features of the simulated information data, and generate a main index directory by associating main behaviors with simulated multi-source interaction logs and the main behavior data.
[0064] The dynamic behavior basic graph module is used to extract key behavior features based on the main index directory and knowledge base, and to perform weighted association matching on the main index directory based on the behavior association features of the key behavior features to generate a dynamic behavior basic graph.
[0065] The dynamic behavior hierarchy graph module is used to correct and update the weights of the dynamic behavior base graph based on time feature information, and to generate a dynamic behavior hierarchy graph by matching the time feature information and the dynamic behavior base graph.
[0066] The responsibility score generation module is used to obtain the initial responsibility score of each level based on the dynamic behavior hierarchy graph and the knowledge base, and to perform data weighting on the weights of the dynamic behavior hierarchy graph according to the responsibility score to obtain the responsibility score of the dynamic behavior hierarchy graph.
[0067] The output module is used to output the responsibility score and the dynamic behavior hierarchy map.
[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for quantifying the entire chain of responsibility for illegal behavior based on dynamic behavioral graphs, characterized in that: include: Acquire simulated information data of the entire chain of simulated AI illegal behavior, extract subject behavior data based on the key behavioral features of the simulated information data, and generate a subject index directory by associating subject behaviors based on simulated multi-source interaction logs and the subject behavior data; Based on the main index directory and knowledge base, key behavioral features are extracted, and based on the behavioral association features of the key behavioral features, the main index directory is weighted and matched to generate a dynamic behavioral basic graph. The weights of the dynamic behavior base map are corrected and updated based on the time feature information, and the time feature information and the dynamic behavior base map are matched to generate a dynamic behavior hierarchy map. The initial responsibility score of each level is obtained based on the dynamic behavior hierarchy graph and knowledge base. The responsibility score of the dynamic behavior hierarchy graph is obtained by weighting the weights of the dynamic behavior hierarchy graph according to the responsibility score. Output the responsibility score and the dynamic behavior hierarchy graph.
2. The method for quantifying the full-chain responsibility of AI-driven illegal behavior based on dynamic behavioral graphs according to claim 1, characterized in that, Based on the key behavioral features of the simulated information data, subject behavior data is extracted. Based on the simulated multi-source interaction logs and the subject behavior data, a subject index directory is generated by associating subject behaviors, including: The subject behavior data is obtained by extracting semantic behavior information of the simulated information data of the initiator and the receiver based on the semantic information extraction model. The subject behavior data includes at least the semantic behavior information of the initiator and the semantic behavior information of the receiver. The semantic behavior information of the initiator includes at least the initiator's behavior subject, the initiator's behavior action, and the initiator's behavior goal. The semantic behavior information of the receiver includes at least the receiver's behavior subject, the receiver's behavior action, and the receiver's behavior goal. Based on the simulated multi-source interaction logs and the semantic information extraction model, log semantic information is extracted for the same initiator behavior subject, the same receiver behavior subject, the same initiator behavior target, and the same receiver behavior target. The log semantic information includes at least implementation subject data and implementation behavior data. The implementation behavior data with the same implementation subject data are marked to associate the subject behavior and obtain subject data tags. The subject index directory is obtained by associating the implementation behavior data with initial weights based on the subject data tags.
3. The method for quantifying the full-chain responsibility for AI-related illegal behavior based on dynamic behavioral graphs according to claim 2, characterized in that, Based on the main index directory and knowledge base, key behavioral features are extracted. Then, based on the behavioral association features of these key behavioral features, weighted association matching is performed on the main index directory to generate a dynamic behavioral baseline graph, including: Based on the semantic information extraction model, the AI illegal behavior entries in the knowledge base are extracted to obtain illegal behavior data; The clustering model uses the number of data from the implementing entities as the number of clusters. Clustering analysis is performed on the data of the implementing behavior and the data of the illegal behavior with the same data label to obtain behavior clusters. The data within the cluster with the smallest distance from the cluster center of the behavior cluster is obtained according to the Euclidean distance algorithm and used as the cluster label of each cluster. The behavior similarity between each cluster label and the illegal behavior data is calculated using a cosine similarity algorithm. All cluster data of the clusters with behavior similarity greater than the behavior threshold are marked as suspected illegal behavior data. The dynamic behavior architecture is generated by associating the suspected illegal behavior data with the same subject data label. The dynamic behavior architecture is updated by weighting the behavior similarity of the suspected illegal behavior data and the subject data to obtain the dynamic behavior basic map.
4. The method for quantifying the full-chain responsibility of AI-driven illegal behavior based on dynamic behavioral graphs according to claim 3, characterized in that, The weights of the dynamic behavior baseline graph are corrected and updated based on time feature information. Data matching between the time feature information and the dynamic behavior baseline graph is then performed to generate a dynamic behavior hierarchy graph, including: The suspected illegal behavior data is extracted by obtaining the dynamic behavior basic graph, and the implementation number and time sequence of the suspected illegal behavior data are obtained based on the simulated multi-source interaction log to generate an illegal behavior implementation logic graph and obtain the time feature information. Anomaly analysis is performed on the logic diagram of the illegal behavior based on the time anomaly analysis model to obtain anomaly behavior labels. Based on the anomaly behavior labels, weight update values are obtained to correct and update the weights of the dynamic behavior basic graph. The dynamic behavior hierarchical map is generated by hierarchically labeling the main data of the dynamic behavior base map based on the abnormal behavior labels.
5. The method for quantifying the full-chain responsibility for AI-related illegal behavior based on dynamic behavioral graphs according to claim 1, characterized in that, Based on the dynamic behavior hierarchy graph and knowledge base, initial responsibility scores are obtained for each level. The dynamic behavior hierarchy graph is then weighted according to these responsibility scores to obtain the final dynamic behavior hierarchy graph. The responsibility scores of the dynamic behavior hierarchy graph are calculated using a responsibility score weighting function, including: ； in, The weights are the weights of the main body of the dynamic behavior hierarchy graph at the i-th level. The weights of the dynamic behavior hierarchy graph at the i-th level are given by [the weights]. The number of illegal behavior data of the subject in the i-th dynamic behavior hierarchy graph. The maximum value among the initial responsibility scores of the t dynamic behavior hierarchy graphs. For cosine similarity, This refers to the j-th illegal act of the subject in the i-th dynamic behavior hierarchy graph. For the t-th historical illegal behavior standard behavior in the knowledge base, Let t be the initial liability score for the t-th historical illegal behavior standard behavior in the knowledge base.
6. An AI-based system for quantifying the entire chain of responsibility for illegal activities based on dynamic behavioral graphs, characterized in that: include: The main index directory module is used to acquire simulated information data of the entire chain of simulated AI illegal behavior, extract main behavior data based on the key behavioral features of the simulated information data, and generate a main index directory by associating main behaviors with simulated multi-source interaction logs and the main behavior data. The dynamic behavior basic graph module is used to extract key behavior features based on the main index directory and knowledge base, and to perform weighted association matching on the main index directory based on the behavior association features of the key behavior features to generate a dynamic behavior basic graph. The dynamic behavior hierarchy graph module is used to correct and update the weights of the dynamic behavior base graph based on time feature information, and to generate a dynamic behavior hierarchy graph by matching the time feature information and the dynamic behavior base graph. The responsibility score generation module is used to obtain the initial responsibility score of each level based on the dynamic behavior hierarchy graph and the knowledge base, and to perform data weighting on the weights of the dynamic behavior hierarchy graph according to the responsibility score to obtain the responsibility score of the dynamic behavior hierarchy graph. The output module is used to output the responsibility score and the dynamic behavior hierarchy map.