A multi-dimensional scoring mechanism-based data abnormal behavior identification and tracing agent method

CN122817034APending Publication Date: 2026-09-25UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611048108.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-15
Publication Date
2026-09-25

AI Technical Summary

Technical Problem

[0004]本发明提出一种基于多维评分机制的数据异常行为识别及溯源智能体算法,针对数据异常行为识别中样本积累难、异常行为覆盖面不足、不同行业及应用的异常行为差异大,导致传统智能分析算法训练难、识别准确率低、使用范围窄的问题,致力于利用大语言模型(Large Language Model,LLM)推理逻辑能力强的能力,实现数据行为模式断点推断、异常行为分析评分、行为模式沉淀等能力,通过构建智能体,极大的提升数据异常行为识别及溯源的分析能力,简化训练及应用过程

Benefits of technology

[0066]本发明不需要准确训练的样本数据,可以通过对不同行业及应用的日志的持续分析,利用LLM的强大推理能力,自动识别数据流转行为模式,并不断的积累形成针对该行业及应用的特定行为模式库,极大了提升了数据异常行为识别及溯源系统的易用性、准确性和普适性。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122817034A_ABST
    Figure CN122817034A_ABST
Patent Text Reader

Abstract

The application discloses a data abnormal behavior identification and tracing intelligent agent method based on a multi-dimensional scoring mechanism and belongs to the technical field of information security. The method mainly comprises log standardization analysis, file flow diagram generation, LLM-driven abnormal data flow detection and behavior mode library continuous learning. The log standardization analysis automatically extracts key information in combination with the object of data analysis. The file flow diagram generates a data flow diagram for a specific file based on a historical mode library. Multi-dimensional scoring analysis of data subjects, flow paths and abnormal behaviors is realized based on LLM. Continuous learning of the behavior mode library realizes continuous sedimentation of behavior modes and corresponding occurrence probabilities. Through continuous analysis of logs of different industries and applications, the powerful reasoning capability of LLM is used to automatically identify data flow behavior modes, and a specific behavior mode library for the industry and application is continuously accumulated, so that the usability, accuracy and universality of the data abnormal behavior identification and tracing system are greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This paper proposes a data anomaly behavior identification and source tracing intelligent agent method based on a multi-dimensional scoring mechanism, which belongs to the field of information security technology. Background Technology

[0002] Data anomaly identification and tracing is a key research branch of data security. By parsing and analyzing data usage and sharing logs, accumulating data samples, training intelligent models or designing data analysis algorithms, it identifies abnormal data operation behaviors, traces data flow, and pinpoints the causes of data security risks, providing effective support for data security incident response. Currently, it is widely used in data security governance service platforms, data loss prevention systems, and database security auditing systems.

[0003] Existing algorithms for identifying and tracing abnormal data behavior mainly face three problems: First, they must have enough sample data of abnormal data behavior to train intelligent models or build data analysis algorithms; second, they lack reasoning and correlation capabilities and cannot identify abnormal behaviors outside the samples; and finally, due to the significant differences in data operation behaviors across different industries and applications, the generalization ability of the constructed intelligent models and analysis algorithms is weak, and they can only be applied to fixed industries and applications. Summary of the Invention

[0004] This invention proposes a data anomaly behavior identification and source tracing intelligent agent algorithm based on a multi-dimensional scoring mechanism. Addressing the challenges of sample accumulation, insufficient coverage of anomaly behaviors, and significant differences in anomaly behaviors across different industries and applications in data anomaly behavior identification, which lead to difficulties in training traditional intelligent analysis algorithms, low identification accuracy, and narrow application scope, this invention leverages the strong reasoning logic capabilities of Large Language Models (LLMs) to achieve data behavior pattern breakpoint inference, anomaly behavior analysis scoring, and behavior pattern accumulation. By constructing an intelligent agent, it significantly enhances the analytical capabilities for data anomaly behavior identification and source tracing, and simplifies the training and application process.

[0005] The specific technical solution of the present invention is as follows:

[0006] A data anomaly behavior identification and source tracing intelligent agent based on a multi-dimensional scoring mechanism includes the following steps:

[0007] Step 1: Log standardization and parsing; extract data operation log text and associate events based on the probability that two events belong to the same file;

[0008] Step 2: File flow graph generation; For the target file, its flow graph is a directed graph, where nodes correspond to events. Deterministic edges are created based on the preceding and following event nodes of a node. For adjacent edges of the same file entity arranged in chronological order, if the time interval exceeds a threshold, it is marked as a breakpoint.

[0009] Step 3: LLM-driven abnormal data flow detection; First, the flow graph, historical patterns, and external knowledge are converted into structured prompts. For new edges in LLM inference, breakpoint constraints, LLM inference constraints, and LLM output format constraints are constructed. Each node in the flow graph is scored for node anomalies based on its position, event, and file status. Each edge is scored for edge anomalies based on its action transformation, channel, confidence level, and LLM tagging. If the score of any node or edge exceeds the threshold, an alarm is generated.

[0010] Step 4: Continuous learning of the behavior pattern library; extract process sequence patterns, action pair patterns, user baseline patterns, and channel patterns from the flow graph; merge the newly extracted patterns with existing patterns in the historical pattern library based on similarity, and update the confidence of existing patterns and accumulate rules based on the number of observations.

[0011] Furthermore, step 1 is detailed as follows:

[0012] Step 1.1: Given the operation log document data, extract the event number, operation time, behavior, operation file identity, operation principle, operation goal, and preceding operation for event e;

[0013] Step 1.2: Calculate the probability that any two events belong to the same file:

[0014]

[0015] in and This represents two log field data, indirectly representing the event. Indicates the association strategy, The base confidence level for strategy s.

[0016] Furthermore, step 2 is detailed as follows:

[0017] For the target file, its flow graph is a directed graph. , where V is the set of nodes, D is the set of edges, and B is the set of breakpoints;

[0018] Node construction: For each event associated with the target file Create a preceding event node and subsequent event nodes The two nodes represent the file state before and after the event, including the unique number of the event e corresponding to the node, the identity of the file being operated on, the location information, the file attribute status information, and the event operation time;

[0019] Edge construction: for each event Based on preceding event nodes and subsequent event nodes Creating deterministic edges Deterministic edges contain events File attribute status information, time, and operation principles;

[0020] Flow breakpoint detection: For the same file entity, adjacent edges arranged in time sequence and If the time interval exceeds the threshold, it is marked as a breakpoint. ,in This represents the successor node with time sequence k. Represents the preceding node. The timestamp of the edge. The breakpoint detection interval threshold is set internally within the system; each breakpoint It carries context information, which includes interval events, a summary of previous node events, and a summary of subsequent node events.

[0021] Furthermore, step 3 is detailed as follows:

[0022] Step 3.1: Constructing the Context of the Cue Words: Transform the flow graph, historical patterns, and external knowledge into structured cue words, including role positioning, task objectives, evidence chains, breakpoint sets, flow pattern sets, and constraints; where the evidence chain is a natural language summary transformed from deterministic nodes and edges, the breakpoint set is a structured text composed of the evidence before and after each breakpoint, and the flow pattern set is the historical pattern library... Historical patterns;

[0023] Step 3.2: Constructing prompt word constraints;

[0024] Breakpoint constraint: The new edge inferred by LLM satisfies that the predecessor node and the successor node already exist in the flow graph;

[0025] LLM inference constraints: Each inference edge includes the source node, the target node, and the inferred behavior. The LLM itself evaluates the confidence of the inference edge, the natural language inference basis, the inference transmission channel, and the LLM's label for determining whether the inference edge represents abnormal behavior.

[0026] LLM output format constraints: Define the prompts for edge and vertex anomaly detection;

[0027] Step 3.3: The node anomaly score is a weighted composite score based on three dimensions: location, time, and file status.

[0028] The location anomaly score is calculated based on the location similarity between the edge's location information and the flow patterns in the historical pattern library;

[0029] The time anomaly score is calculated based on the empirical probability between the time of using the current node and the occurrence time of the corresponding event in the historical pattern library;

[0030] File status anomaly scoring is based on the grading and scoring of file attribute status associated with the node;

[0031] Step 3.4: The edge anomaly score is a weighted composite of four dimensions: action transformation, channel, confidence, and LLM labeling;

[0032] The action transition anomaly score is calculated based on the action transition statistics in the historical pattern library and the probability of the action on the current edge;

[0033] Channel anomaly scoring is based on the similarity between the edge's transmission channel and its historical transmission pattern.

[0034] The confidence anomaly score assigns a score to the inherent anomaly risk of the inferred edge itself due to its uncertainty;

[0035] The LLM-marked anomaly score is obtained directly by using LLM to determine whether the inferred edge represents an anomalous behavior.

[0036] Step 3.5: For any node or edge, if its score exceeds the threshold, an alarm is generated.

[0037] Furthermore, step 4 is detailed as follows:

[0038] Step 4.1: Pattern Extraction: Extract four types of patterns from the flow graph: Process sequence pattern (the set of actions of edges arranged sequentially in the deterministic edge set); Action pair pattern (the set of action pairs of adjacent edges of the same file entity); User baseline pattern (user...). A collection of all operations performed, and records the user. The time window distribution of operations; the channel mode is a set of channel transmissions that are recursively associated across systems by deterministic edges in the edge set;

[0039] Step 4.2: Newly Extracted Pattern With existing patterns in the historical pattern library Merge according to the following rules:

[0040]

[0041] in Represents the update mode. Represents the addition of a pattern. Here is the formula for calculating the similarity between the two patterns;

[0042] Step 4.3: When an existing pattern exists When observed again, its confidence level is updated using Bayesian methods.

[0043] Step 4.4: Rule accumulation: When the accumulated number of a certain pattern exceeds the set accumulation threshold, rule accumulation is triggered. The pattern is solidified into a deterministic association rule, and LLM inference will no longer be required when similar flow patterns appear in the future.

[0044] Furthermore, the association strategies include strong hash association, time window-assisted association, document attribute composite key association, and cross-system recursive association.

[0045] Furthermore, the node anomaly score is calculated as follows:

[0046] Location Anomaly Scoring ,in Represents the location information of the node. For the circulation patterns in the historical pattern library, It is a similarity function;

[0047] Time Anomaly Scoring ,in, For empirical probability, Node operation time; Represents a time window; This refers to the occurrence time of the corresponding event in the historical pattern library; For all Let be the set of users associated with the edges of the target node. ,in, This represents the user operation method for the corresponding edge. The target node of the edge is represented; if there are multiple users, the user with the highest operation frequency is selected as the primary user for calculation.

[0048] File status anomaly score The file status is set into different levels. If the file associated with node v is deleted and there is a corresponding operation and the time meets the requirements, it is rated as the first level. If the file does not exist, it is rated as the second level. If the file is modified but not newly created, it is rated as the third level. All other cases are rated as the fourth level.

[0049] The node anomaly score is a weighted composite score based on three dimensions: location, time, and file status.

[0050]

[0051] Node anomaly constraint weights: Each constraint weight is greater than or equal to 0.

[0052] Furthermore, the edge anomaly score is calculated as follows:

[0053] Action transition anomaly scoring Action transition statistics based on the historical pattern library:

[0054]

[0055] in Represents the probability of transition. For the actions of direct predecessor edges sorted by time under the same file number, the predecessor edge is defined as: ; Represents the probability of conversion in the historical pattern library;

[0056] Channel anomaly scoring :

[0057]

[0058] in The transmission channel representing the edge. Represents the transmission channel and historical transmission mode similarity, If and only if Otherwise, it is 0;

[0059] Confidence anomaly score The inference of edges inherently carries anomaly risk due to its inherent uncertainty:

[0060]

[0061] in This indicates the LLM's own confidence assessment of the inferred edges;

[0062] LLM markers for anomaly scoring LLM directly marks inference edges as anomalous during the inference process:

[0063]

[0064] in, This indicates the marker used by LLM to determine whether this inferred edge represents anomalous behavior.

[0065] The beneficial effects of this invention are as follows:

[0066] This invention does not require accurate training sample data. By continuously analyzing logs from different industries and applications, and leveraging the powerful inference capabilities of LLM, it can automatically identify data flow behavior patterns and continuously accumulate a specific behavior pattern library for each industry and application. This greatly improves the ease of use, accuracy, and universality of the data anomaly behavior identification and tracing system. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0068] To better understand the purpose, structure, and function of this invention, the invention will be described in further detail below with reference to the accompanying drawings.

[0069] This invention proposes a data anomaly behavior identification and tracing intelligent agent algorithm based on a multi-dimensional scoring mechanism. The aim is to construct an intelligent data anomaly behavior identification and tracing algorithm based on the powerful reasoning capabilities of LLM (Limited Least Metrics), achieving capabilities such as pattern breakpoint inference, anomaly behavior identification, and behavior pattern accumulation. The method mainly includes log standardization parsing, file flow graph generation, LLM-driven anomaly data flow detection, and continuous learning of the behavior pattern library. Log standardization parsing automatically extracts key information based on the data analysis object; the file flow graph generates a data flow graph for a specific file based on a historical pattern library; LLM-driven anomaly data flow detection uses LLM to achieve multi-dimensional scoring analysis of data subject, flow path, and anomaly behavior; and continuous learning of the behavior pattern library continuously accumulates behavior patterns and their corresponding probabilities. The specific steps are as follows:

[0070] Step 1: Log standardization parsing;

[0071] Step 1.1: Extracting Data Operation Logs (e.g., log data from email systems, cloud storage systems, and data sharing and exchange systems): Given operation log document data and parser configuration After extraction function:

[0072]

[0073] The extraction strategies mainly include various parsing strategies such as regular expression strategies, delimiter strategies, JSON path strategies, and multi-line merging strategies. The transformed log field data includes the event number. ,time ,Behavior File operation identity Operating principles Operational Objectives Preceding operations ; indicates as:

[0074]

[0075] Step 1.2: Text entity association. The core of association is calculating the probability that any two events belong to the same file.

[0076]

[0077] in and This represents two log field data, indirectly representing the event. Indicates the association strategy; association strategies are combined as follows: , Represents a strong hash association. To assist in the correlation of time windows, For document attribute composite key association, Represents cross-system recursive association. For strategy The base confidence level can be calculated by averaging multiple times.

[0078] Step 2: Generate file flow diagram.

[0079] For the target file Its flow graph is a directed graph. , where V is the set of nodes, D is the set of edges, and B is the set of breakpoints.

[0080] Node construction: For each associated node... The incident Create two nodes to represent the file state before and after the event: the preceding event node. and subsequent event nodes This includes the unique ID of event e corresponding to the node, the identity of the file being operated on, its location information, file attribute status information, and the event operation time; it can be represented as:

[0081]

[0082]

[0083] in This represents the unique identifier assigned to the node by the system. Represents location information and Represents file attribute status information, index Represents the preceding event. Represents the subsequent event. This represents the file identity corresponding to the event. Represents the operation time of the event, all nodes Together they form a set of nodes .

[0084] Edge construction: for each event Based on preceding event nodes and subsequent event nodes Creating deterministic edges Deterministic edges contain events File attribute status information, time, operation principles; all deterministic edges Together they form an edge set ;

[0085] Flow breakpoint detection: for adjacent edges of the same file entity and (Arranged in chronological order), if the time interval exceeds the threshold, it is marked as a breakpoint: ,in This represents the successor node with time sequence k. Represents the preceding node. The timestamp of the edge. The breakpoint detection interval threshold is set internally within the system. Each breakpoint... It carries context information, which includes interval events, a summary of previous node events, and a summary of subsequent node events.

[0086] Step 3: Anomaly data flow detection algorithm driven by LLM;

[0087] Step 3.1: Constructing the context of the cue words: Transforming the flow graph, historical patterns, and external knowledge into structured cue words:

[0088]

[0089] in, Representative roles; Represents the mission objective; For the chain of evidence, Transform deterministic nodes and edges into natural language summaries; Represents the set of breakpoints. Structure the evidence before and after each breakpoint into text; Represents a set of circulation patterns. Select historical pattern library Most relevant A historical pattern; Represents the constraints of reasoning. Extract the set of valid nodes from the graph topology as inference constraints.

[0090] Step 3.2: Constructing prompt word constraints:

[0091] Generate breakpoint constraints: for each breakpoint The new edge inferred by LLM must satisfy the previous node. and subsequent nodes It must already exist in the flow graph.

[0092] LLM inference constraints: Mapping prompts to a set of inference edges and an LLM-level anomaly detection set. , Input prompts for the model representing the larger model. Let be the set of inference edges, and each inference edge Not only includes source nodes Target Node and the act of inference It also includes the LLM's own confidence assessment of the inferred edges. Basis for natural language reasoning Inferred transmission channel LLM determines whether this inferred edge represents abnormal behavior. , can be represented as: LLM output format constraints: ,in, Represents the output of the large model. The inference function representing the large model. The descriptive words for inferring anomalies in edges and vertices are defined.

[0093] Step 3.3: Node Anomaly Scoring: Assign an anomaly score to each node and edge in the flow graph. The quantitative anomaly scoring within the scope upgrades anomaly detection from qualitative judgment to quantitative measurement. For nodes... Its anomaly score is a weighted composite of three dimensions: location, time, and file status.

[0094]

[0095] Node anomaly constraint weights: Each constraint weight is greater than or equal to 0.

[0096] Location Anomaly Scoring ,in This represents the position information of the edge. For the circulation patterns in the historical pattern library, This is the positional similarity function.

[0097] Time Anomaly Scoring ,in, For empirical probability, The time spent using the node; This represents a time window, which can be configured in the system according to actual needs. This refers to the occurrence time of the corresponding event in the historical pattern library; For all Let be the set of users associated with the edges of the target node. ,in, This represents the user operation method for the corresponding edge. The target node represents the edge; if there are multiple users, the user with the highest operation frequency is selected as the master user to calculate the time exception.

[0098] File status anomaly score The scoring is tiered based on file status. If the file associated with node v has a deleted status and a corresponding operation has occurred, and the time meets the requirements, it receives a first-level status score. If the file does not exist, it receives a second-level status score. If the file has been modified but not newly created, it receives a third-level status score. All other cases receive a fourth-level status score. The specific scoring can be set as follows:

[0099]

[0100] Step 3.4: Edge Anomaly Scoring: For edges Its anomaly score is a weighted composite of four dimensions: action transition, channel, confidence level, and LLM labeling.

[0101]

[0102] Node anomaly constraint weights: Each constraint weight is greater than or equal to 0.

[0103] Action transition anomaly scoring Action transition statistics based on the historical pattern library:

[0104]

[0105] in Represents the probability of transition. For the actions of the direct predecessor edges sorted by time under the same FID (file number), This represents the probability of a transition in the historical pattern library. For The preceding edge is defined as: ,in These are the elements of the inferred edge, with the target node being... .

[0106] Channel anomaly scoring :

[0107]

[0108] in The transmission channel representing the edge. Represents the transmission channel and historical transmission mode similarity, If and only if (Channel identifier exact match), otherwise 0.

[0109] Confidence anomaly score The inference of edges inherently carries anomaly risk due to its inherent uncertainty:

[0110]

[0111] LLM markers for anomaly scoring LLM directly marks inference edges as anomalous during the inference process.

[0112]

[0113] Step 3.5: Comprehensive Anomaly Score. The comprehensive anomaly score for the entire flow graph is defined as follows:

[0114]

[0115] Among them, the average outlier of the nodes Average outlier The calculation is as follows:

[0116]

[0117]

[0118] and , .

[0119] Anomaly alarm determination: For any node or edge If its score exceeds the threshold An alarm will be generated:

[0120] ,in This represents an alarm operation, with a default threshold. The threshold can be adjusted according to the security policy; lowering the threshold results in higher recall, while raising the threshold results in higher accuracy.

[0121] Step 4: Continuous learning algorithm for behavior pattern library.

[0122] Step 4.1: Pattern Extraction: From the flow graph completed in Step 2 Extract four types of patterns: ,in It is a set of deterministic edges that have been manually confirmed.

[0123] Process sequence pattern for The set of actions of the edges arranged in chronological order.

[0124] Action to Pattern Adjacent edges of the same file entity and Actions on a set, , ;

[0125] User baseline mode For users A collection of all operations performed, and records the user. The time window distribution of the operation.

[0126] Channel mode It is a set of channel transmissions that are cross-system recursive associations of deterministic edges in the edge set.

[0127] Step 4.2: Pattern merging function: Newly extracted patterns With existing patterns in the historical pattern library Merge the patterns according to the following rules: Calculate the similarity between the two patterns:

[0128] in Represents the update mode. Represents the addition of a pattern. Here is the formula for calculating the similarity between the two patterns, with a pattern similarity threshold of 0.6.

[0129] Step 4.3: Bayesian confidence update: when there is an existing pattern When observed again, its confidence is updated using Bayesian methods:

[0130]

[0131] Among them: the number of observations is ; and All are prior parameters of the Beta distribution; it is recommended to set them as follows: , The upper bound is set to 0.95 to prevent overfitting.

[0132] Step 4.4: Rule Accumulation: When the accumulated number of a certain pattern exceeds the set accumulation threshold, rule accumulation is triggered. The accumulation threshold is configurable, and it is recommended to set it to 3. After accumulation, the pattern is solidified as a deterministic association rule in Step 2, and LLM inference will no longer be required when similar flow patterns appear in Step 3 in the future.

[0133] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A data anomaly behavior identification and source tracing intelligent agent method based on a multi-dimensional scoring mechanism, characterized in that, Includes the following steps: Step 1: Log standardization and parsing; extract data operation log text and associate events based on the probability that two events belong to the same file; Step 2: File flow graph generation; For the target file, its flow graph is a directed graph, where nodes correspond to events. Deterministic edges are created based on the preceding and following event nodes of a node. For adjacent edges of the same file entity arranged in chronological order, if the time interval exceeds a threshold, it is marked as a breakpoint. Step 3: LLM-driven abnormal data flow detection; First, the flow graph, historical patterns, and external knowledge are converted into structured prompts. For new edges in LLM inference, breakpoint constraints, LLM inference constraints, and LLM output format constraints are constructed. Each node in the flow graph is scored for node anomalies based on its position, event, and file status. Each edge is scored for edge anomalies based on its action transformation, channel, confidence level, and LLM tagging. If the score of any node or edge exceeds the threshold, an alarm is generated. Step 4: Continuous learning of the behavior pattern library; Extract process sequence patterns, action pair patterns, user baseline patterns, and channel patterns from the flow graph; merge the newly extracted patterns with existing patterns in the historical pattern library based on similarity, and update the confidence of existing patterns and accumulate rules based on the number of observations.

2. The data anomaly behavior identification and source tracing intelligent agent method based on a multi-dimensional scoring mechanism according to claim 1, characterized in that, Step 1 is described in detail as follows: Step 1.1: Given the operation log document data, extract the event number, operation time, behavior, operation file identity, operation principle, operation goal, and preceding operation for event e; Step 1.2: Calculate the probability that any two events belong to the same file: in and This represents two log field data, indirectly representing the event. Indicates the association strategy, The base confidence level for strategy s.

3. The data anomaly behavior identification and source tracing intelligent agent method based on a multi-dimensional scoring mechanism according to claim 2, characterized in that, Step 2 is described in detail below: For the target file, its flow graph is a directed graph. , where V is the set of nodes, D is the set of edges, and B is the set of breakpoints; Node construction: For each event associated with the target file Create a preceding event node and subsequent event nodes The two nodes represent the file state before and after the event, including the unique number of the event e corresponding to the node, the identity of the file being operated on, the location information, the file attribute status information, and the event operation time; Edge construction: for each event Based on preceding event nodes and subsequent event nodes Creating deterministic edges Deterministic edges contain events File attribute status information, time, and operation principles; Flow breakpoint detection: For the same file entity, adjacent edges arranged in time sequence and If the time interval exceeds the threshold, it is marked as a breakpoint. ,in This represents the successor node with time sequence k. Represents the previous node. The timestamp of the edge. The breakpoint detection interval threshold is set internally within the system; each breakpoint It carries context information, which includes interval events, a summary of previous node events, and a summary of subsequent node events.

4. The data anomaly behavior identification and source tracing intelligent agent method based on a multi-dimensional scoring mechanism according to claim 3, characterized in that, Step 3 is as follows: Step 3.1: Constructing the Context of the Cue Words: Transform the flow graph, historical patterns, and external knowledge into structured cue words, including role positioning, task objectives, evidence chains, breakpoint sets, flow pattern sets, and constraints; where the evidence chain is a natural language summary transformed from deterministic nodes and edges, the breakpoint set is a structured text composed of the evidence before and after each breakpoint, and the flow pattern set is the historical pattern library... Historical patterns; Step 3.2: Constructing prompt word constraints; Breakpoint constraint: The new edge inferred by LLM satisfies that the predecessor node and the successor node already exist in the flow graph; LLM inference constraints: Each inference edge includes the source node, the target node, and the inferred behavior. The LLM itself evaluates the confidence of the inference edge, the natural language inference basis, the inference transmission channel, and the LLM's label for determining whether the inference edge represents abnormal behavior. LLM output format constraints: Define the prompts for edge and vertex anomaly detection; Step 3.3: The node anomaly score is a weighted composite score based on three dimensions: location, time, and file status. The location anomaly score is calculated based on the location similarity between the edge's location information and the flow patterns in the historical pattern library; The time anomaly score is calculated based on the empirical probability between the time of using the current node and the occurrence time of the corresponding event in the historical pattern library; File status anomaly scoring is based on the grading and scoring of file attribute status associated with the node; Step 3.4: The edge anomaly score is a weighted composite of four dimensions: action transformation, channel, confidence, and LLM labeling; The action transition anomaly score is calculated based on the action transition statistics in the historical pattern library and the probability of the action on the current edge; Channel anomaly scoring is based on the similarity between the edge's transmission channel and its historical transmission pattern. The confidence anomaly score assigns a score to the inherent anomaly risk of the inferred edge itself due to its uncertainty; The LLM-marked anomaly score is obtained directly by using LLM to determine whether the inferred edge represents an anomalous behavior. Step 3.5: For any node or edge, if its score exceeds the threshold, an alarm is generated.

5. The data anomaly behavior identification and source tracing intelligent agent method based on a multi-dimensional scoring mechanism according to claim 4, characterized in that, Step 4 is as follows: Step 4.1: Pattern Extraction: Extract four types of patterns from the flow graph: Process sequence pattern (the set of actions of edges arranged sequentially in the deterministic edge set); Action pair pattern (the set of action pairs of adjacent edges of the same file entity); User baseline pattern (user...). A collection of all operations performed, and records the user. The time window distribution of operations; the channel mode is a set of channel transmissions that are recursively associated across systems by deterministic edges in the edge set; Step 4.2: Newly Extracted Pattern With existing patterns in the historical pattern library Merge according to the following rules: in Represents the update mode. Represents the addition of a pattern. Here is the formula for calculating the similarity between the two patterns; Step 4.3: When an existing pattern exists When observed again, its confidence level is updated using Bayesian methods. Step 4.4: Rule accumulation: When the accumulated number of a certain pattern exceeds the set accumulation threshold, rule accumulation is triggered. The pattern is solidified into a deterministic association rule, and LLM inference will no longer be required when similar flow patterns appear in the future.

6. The data anomaly behavior identification and source tracing intelligent agent method based on a multi-dimensional scoring mechanism according to claim 5, characterized in that, The association strategies include strong hash association, time window-assisted association, document attribute composite key association, and cross-system recursive association.

7. The data anomaly behavior identification and source tracing intelligent agent method based on a multi-dimensional scoring mechanism according to claim 6, characterized in that, The node anomaly score is calculated in the following specific way: Location Anomaly Scoring ,in Represents the location information of the node. For the circulation patterns in the historical pattern library, It is a similarity function; Time Anomaly Scoring ,in, For empirical probability, Node operation time; Represents a time window; This refers to the occurrence time of the corresponding event in the historical pattern library; For all Let be the set of users associated with the edges of the target node. ,in, This represents the user operation method for the corresponding edge. The target node of the edge is represented; if there are multiple users, the user with the highest operation frequency is selected as the primary user for calculation. File status anomaly score The file status is set into different levels. If the file associated with node v is deleted and there is a corresponding operation and the time meets the requirements, it is rated as the first level. If the file does not exist, it is rated as the second level. If the file is modified but not newly created, it is rated as the third level. All other cases are rated as the fourth level. The node anomaly score is a weighted composite score based on three dimensions: location, time, and file status. Node anomaly constraint weights: Each constraint weight is greater than or equal to 0.

8. The data anomaly behavior identification and source tracing intelligent agent method based on a multi-dimensional scoring mechanism according to claim 7, characterized in that, The edge anomaly score is calculated in detail as follows: Action transition anomaly scoring Action transition statistics based on the historical pattern library: in Represents the probability of transition. For the actions of direct predecessor edges sorted by time under the same file number, the predecessor edge is defined as: ; Represents the probability of conversion in the historical pattern library; Channel anomaly scoring : in The transmission channel representing the edge. Represents the transmission channel and historical transmission mode similarity, If and only if Otherwise, it is 0; Confidence anomaly score The inference of edges inherently carries anomaly risk due to its inherent uncertainty: in This indicates the LLM's own confidence assessment of the inferred edges; LLM markers for anomaly scoring LLM directly marks inference edges as anomalous during the inference process: in, This indicates the marker used by LLM to determine whether this inferred edge represents anomalous behavior.