CN-121349899-B - Intelligent dynamic behavior safety test method and system based on risk conduction quantization model

CN121349899BCN 121349899 BCN121349899 BCN 121349899BCN-121349899-B

Abstract

The invention relates to the technical field of intelligent body testing, in particular to an intelligent body dynamic behavior safety testing method and system based on a risk conduction quantification model, wherein the method comprises the steps of collecting intelligent body behavior data, including a plurality of risk operation information and risk association relations; the method comprises the steps of taking risk operation as nodes and association relations as edges, constructing a behavior risk association graph, labeling risk attributes and age factors for the nodes, labeling risk conduction intensity for the edges, constructing a risk rule base based on risk labeling results and preset risk judgment logic, inputting behavior data, the behavior risk association graph and the rule base into a trained quantization model, calculating risk conduction probability, influence range and critical nodes, monitoring behavior tracks in real time, suspending tasks, backtracking inference chains and generating intervention records when triggering multidimensional risk thresholds, and fusing multidimensional data to generate a test report. The invention solves the problems of risk omission and response hysteresis associated with the traditional test, and improves the pertinence and reliability of the safety test.

Inventors

LUO JIANFANG
LIU ZIKAI
ZHANG LIN
LI YAN
ZHANG WENBIN
XU JIA
ZHOU LIJING
SONG XUEFENG

Assignees

广州掌动智能科技有限公司
广州掌测信息科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251216

Claims (7)

1. The intelligent agent dynamic behavior safety test method based on the risk conduction quantification model is characterized by comprising the following steps of: acquiring behavior data in the execution process of the intelligent agent task, wherein the behavior data comprises a plurality of risk operation information and risk association relations among the risk operation information; Taking a plurality of risk operation information as map nodes, taking a risk association relationship as a map edge, constructing a behavior risk association map, and carrying out risk labeling on the map nodes, wherein the risk labeling comprises a risk attribute labeling, a risk aging factor labeling and a risk conduction intensity labeling, and the method specifically comprises the following steps: taking a plurality of risk operation information as map nodes, taking risk association relations as map edges, and preliminarily constructing a topological structure of a behavior risk association map; Labeling risk attributes of the map nodes, wherein the risk attributes comprise the association degree of the risk level and the sensitive resource; Labeling the risk aging factors of the map nodes, wherein the risk aging factors are calculated by multiplying the risk initial level quantized values by the time attenuation coefficients; labeling the risk conduction intensity of the sides of the atlas, wherein the risk conduction intensity is obtained by quantitatively calculating the linkage occurrence probability of the associated risk operation corresponding to each atlas side to be labeled by multiplying the linkage hazard amplification coefficient; integrating the topological structure of the behavior risk association graph and the risk labeling result to form a complete behavior risk association graph; Constructing a risk rule base comprising risk event triggering conditions and risk grade judging criteria based on a risk labeling result and preset risk judging logic, wherein the risk event triggering conditions refer to conditions for judging to trigger a risk event when an intelligent agent has an operation combination in a certain high-frequency association path; the construction of the preset risk judging logic specifically comprises the following steps: Extracting a high-frequency risk conduction path from a behavior risk association graph, screening association risk operation combinations with the linkage occurrence probability being greater than or equal to a preset threshold value, and sorting the association risk operation combinations into a graph core feature data set, wherein the high-frequency risk conduction path refers to a risk association path with the occurrence frequency being higher than the preset frequency threshold value in the behavior risk association graph; converting the high-frequency risk conduction path into a risk event triggering condition and defining linkage judgment logic between risk operations, wherein the linkage judgment logic refers to a rule for judging whether the associated operation forms a risk event or not; Setting a risk level adjustment rule of an associated risk operation combination based on a risk conduction intensity label of a behavior risk association map, wherein the risk level adjustment rule refers to a rule for adjusting a risk event level corresponding to an associated path according to the risk conduction intensity of the path; integrating the atlas core feature data set, the risk event triggering condition and the risk level adjustment rule to obtain a preset risk judgment logic; Inputting the behavior data, the behavior risk association graph and the risk rule base into a trained risk conduction quantization model, quantitatively calculating risk conduction probability, risk influence range and critical risk nodes, and outputting a risk conduction quantization result, wherein the trained risk conduction quantization model adopts a fusion framework of a dynamic Bayesian network and a graph neural network, and is established with a fusion interaction mechanism of the dynamic Bayesian network and the graph neural network, the dynamic Bayesian network is used for training risk time sequence conduction logic, and the graph neural network is used for training risk structure association analysis logic, and the method specifically comprises the following steps: performing data preprocessing on the behavior data, the behavior risk association graph and the risk rule base according to preset rules to obtain a current characteristic parameter vector; inputting the risk conduction intensity labels in the current characteristic parameter vector and the behavior risk associated map into a trained risk conduction quantization model, and calculating risk conduction probability through dynamic Bayesian network quantization; Defining a risk influence range through a graph neural network based on risk conduction probability and risk grade judgment standards in a risk rule base; Combining the risk conduction probability and the risk influence range, and identifying critical risk nodes through a fusion interaction mechanism of a dynamic Bayesian network and a graph neural network; Integrating risk conduction probability, risk influence range and critical risk nodes, obtaining a risk conduction quantification result and outputting the risk conduction quantification result; According to the risk conduction quantification result and the behavior risk association map, monitoring the behavior track of the intelligent agent in real time, and when the behavior track of the intelligent agent triggers a preset multidimensional risk threshold, suspending task execution of the intelligent agent and tracing an inference chain to generate a risk intervention record; And merging and analyzing the behavior data, the behavior risk association map, the risk conduction quantification result and the risk intervention record, and generating a test report based on the analysis result.
2. The method for testing the dynamic behavior safety of an agent based on a risk conduction quantification model according to claim 1, wherein the construction of the trained risk conduction quantification model specifically comprises the following steps: Adopting a fusion framework of a dynamic Bayesian network and a graph neural network, and constructing a risk conduction quantization model by taking risk conduction probability, a risk influence range and critical risk nodes as core quantization indexes; Collecting an agent behavior log, a risk event history record and system topological structure data, and preprocessing the data according to a preset rule to obtain a risk operation characteristic parameter vector which is suitable for the input requirement of a fusion framework; Inputting the risk operation characteristic parameter vector into a risk conduction quantization model, training a risk time sequence conduction logic based on a dynamic Bayesian network, and training a risk structure association analysis logic based on a graph neural network to enable the risk conduction quantization model to be respectively adapted to the calculation requirements of different core quantization indexes; Establishing a fusion interaction mechanism of a dynamic Bayesian network and a graph neural network, and optimizing parameters of a risk conduction quantization model by combining risk grade judgment standards in a risk rule base; and obtaining a trained risk conduction quantization model based on the optimized risk conduction quantization model, wherein the trained risk conduction quantization model is used for quantitatively calculating risk conduction probability, risk influence range and critical risk nodes and outputting a risk conduction quantization result.
3. The method for testing the dynamic behavior safety of an intelligent agent based on a risk conduction quantification model according to claim 2, wherein the steps of monitoring the behavior track of the intelligent agent in real time according to the risk conduction quantification result and the behavior risk association map, suspending the task execution of the intelligent agent and tracing back the inference chain when the behavior track of the intelligent agent triggers a preset multidimensional risk threshold, and generating a risk intervention record specifically comprise the following steps: Constructing a three-dimensional risk threshold matrix based on risk grade judging standards and system safety requirements in a risk rule base, wherein the three-dimensional risk threshold matrix comprises a risk conduction probability threshold, a risk influence range threshold and a critical risk node association density threshold, and the critical risk node association density is calculated through a coupling coefficient of risk conduction intensity and a risk influence range; comparing the risk conduction quantization result with a three-dimensional risk threshold matrix through a fusion interaction mechanism of the trained risk conduction quantization model, and respectively calculating risk deviation degree of each dimension; monitoring the behavior track of the intelligent agent in real time, and judging a triggering risk early-warning condition when the risk deviation degree of any dimension of the behavior track exceeds a preset critical value or more than two dimensions reach a preset early-warning threshold value at the same time; suspending task execution of the intelligent agent, tracing back an intelligent agent behavior reasoning chain based on risk timeliness factor marking and risk conduction intensity marking of the behavior risk correlation map, and positioning a risk triggering source; recording risk triggering dimension, risk triggering source, intervention time and intervention measures, and generating a risk intervention record; And according to the risk intervention effect and the historical risk data, adaptively adjusting parameters of the three-dimensional risk threshold matrix through the trained risk conduction quantization model.
4. The method for testing the dynamic behavior safety of an agent based on a risk conduction quantification model according to claim 3, wherein the step of suspending the task execution of the agent, labeling a risk aging factor and a risk conduction intensity based on a behavior risk correlation map, backtracking an agent behavior inference chain, and positioning a risk triggering source specifically comprises the following steps: Suspending task execution of the intelligent agent, splitting time sequence nodes of an intelligent agent behavior inference chain based on risk aging factor labeling of a behavior risk association map, and combing behavior execution sequences and association relations of the time sequence nodes according to time sequence; calculating the risk conduction confidence coefficient of the association path between each time sequence node by combining the risk conduction intensity label of the behavior risk association map, wherein the risk conduction confidence coefficient is obtained by weighted summation of the risk conduction intensity quantized value and the risk aging factor quantized value; screening associated paths with risk conduction confidence coefficient higher than a preset confidence threshold value, forming a key risk conduction link, and tracing an initial trigger node of the key risk conduction link; Verifying the matching degree of the behavior characteristics of the initial trigger node and the risk event trigger conditions in the risk rule base, and judging the initial trigger node as a risk trigger source when the matching degree reaches a preset matching threshold value.
5. A risk conduction quantization model-based agent dynamic behavior safety test system for implementing the steps of the risk conduction quantization model-based agent dynamic behavior safety test method according to any one of claims 1 to 4, comprising: the data acquisition module is used for acquiring behavior data in the execution process of the intelligent agent task, wherein the behavior data comprises a plurality of risk operation information and risk association relations among the risk operation information; The risk labeling module is used for taking a plurality of risk operation information as a map node, taking a risk association relationship as a map edge, constructing a behavior risk association map, and labeling the map node with risks, wherein the risk labeling comprises a risk attribute labeling, a risk aging factor labeling and a risk conduction intensity labeling; the rule base construction module is used for constructing a risk rule base containing risk event triggering conditions and risk grade judging standards based on the risk labeling result and preset risk judging logic; The risk calculation module is used for inputting the behavior data, the behavior risk association patterns and the risk rule base into a trained risk conduction quantification model, quantitatively calculating risk conduction probability, risk influence range and critical risk nodes, and outputting risk conduction quantification results; The backtracking reasoning module is used for monitoring the behavior track of the intelligent body in real time according to the risk conduction quantification result and the behavior risk association map, and suspending the task execution of the intelligent body and backtracking the reasoning chain when the behavior track of the intelligent body triggers a preset multidimensional risk threshold value to generate a risk intervention record; And the report generation module is used for fusing and analyzing the behavior data, the behavior risk association map, the risk conduction quantification result and the risk intervention record, and generating a test report based on the analysis result.
6. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the steps of a method for testing the dynamic behaviour safety of an agent based on a risk conduction quantification model as defined in any one of claims 1-4.
7. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the steps of a method for testing the dynamic behaviour safety of an agent based on a risk conduction quantization model according to any one of claims 1 to 4.

Description

Intelligent dynamic behavior safety test method and system based on risk conduction quantization model Technical Field The invention relates to the technical field of intelligent body testing, in particular to an intelligent body dynamic behavior safety testing method and system based on a risk conduction quantization model. Background With the rapid development of artificial intelligence technology, agents (agents) have been widely used in various fields such as financial services, industrial control, intelligent interaction, etc. The intelligent agent has the capabilities of autonomous decision making and dynamic task execution, the behavior process of the intelligent agent relates to a plurality of links such as perception input, logic reasoning, resource calling, result output and the like, and the behavior mode has the characteristics of dynamic property, relevance and complexity. Under the background, the safety test of the dynamic behavior of the intelligent agent becomes a key link for guaranteeing the stable operation of the system. The existing intelligent agent behavior safety test method mainly surrounds risk operation identification, fixed rule matching and single-dimension risk assessment expansion, and the core aim is to find out explicit security holes (such as authority override, data leakage, flow violations and the like) in the intelligent agent execution process. However, the prior art has the following significant drawbacks in practical applications: On one hand, quantitative analysis on risk conduction is lacking, isolation detection on single risk operation is often focused, association relations and risk conduction effects among a plurality of risk operations are ignored, risk propagation intensity, influence range and time efficiency characteristics cannot be quantified, so that risk assessment is unilateral and accuracy is insufficient, on the other hand, a risk tracing mechanism is fuzzy, manual tracing behavior logs are relied on after risks are detected, systematic carding on a risk conduction link is lacking, initial risk trigger nodes and key conduction paths are difficult to rapidly locate, intervention response is lagged, repair pertinence is not strong, in addition, a risk judgment threshold value is a manually preset fixed value, and cannot be dynamically adjusted according to actual risk intervention effects and historical risk data, misjudgment and missed judgment problems are easy to occur after long-term use, and intelligent adaptation capability is weak. Therefore, a safety test method capable of adapting to dynamic behavior characteristics of an intelligent agent, quantifying a risk conduction process, precisely tracing a source of risk and realizing self-adaptive optimization is needed, so that the problems of static state, one-sided state, difficult tracing, insufficient intellectualization and the like in the prior art are solved, and the accuracy, the high efficiency and the adaptability of the safety test of the intelligent agent behavior are improved. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides an intelligent agent dynamic behavior safety test method and system based on a risk conduction quantization model, which are used for solving the problems in the prior art. One embodiment of the invention provides an intelligent agent dynamic behavior safety test method based on a risk conduction quantization model, which comprises the following steps: acquiring behavior data in the execution process of the intelligent agent task, wherein the behavior data comprises a plurality of risk operation information and risk association relations among the risk operation information; Taking a plurality of risk operation information as a map node, taking a risk association relationship as a map edge, constructing a behavior risk association map, and carrying out risk labeling on the map node, wherein the risk labeling comprises a risk attribute labeling, a risk aging factor labeling and a risk conduction intensity labeling; constructing a risk rule base containing risk event triggering conditions and risk grade judging standards based on a risk labeling result and preset risk judging logic; Inputting the behavior data, the behavior risk association map and the risk rule base into a trained risk conduction quantification model, quantitatively calculating risk conduction probability, risk influence range and critical risk nodes, and outputting risk conduction quantification results; According to the risk conduction quantification result and the behavior risk association map, monitoring the behavior track of the intelligent agent in real time, and when the behavior track of the intelligent agent triggers a preset multidimensional risk threshold, suspending task execution of the intelligent agent and tracing an inference chain to generate a risk intervention record; And merging and analyzing the behavior data, the behavior risk ass