Method and apparatus for access control of an agent

Abstract

This invention relates to the field of artificial intelligence technology, providing a method and apparatus for controlling access permissions for intelligent agents. The method includes: based on a graph neural network model, using roles as nodes, calculating the behavioral correlation between nodes based on the role behavior data under each node, obtaining the connection edges between nodes, and constructing a role relationship topology graph; and predicting the permission set of the intelligent agent to be controlled based on the role relationship topology graph. The method provided by this invention, by collecting multi-dimensional behavioral data of the intelligent agent to be controlled in real time, dynamically updating role behavior data using this data, and constructing and analyzing the role relationship topology graph based on a graph neural network, ultimately achieves dynamic and intelligent prediction of the intelligent agent's permission set. It can effectively discover deep connections between roles that are difficult to detect in traditional methods, and adaptively adjust permissions according to the actual changes in the intelligent agent's behavior, greatly improving the flexibility, accuracy, and security of permission management in multi-agent systems.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for controlling access permissions of intelligent agents. Background Technology

[0002] In multi-agent collaborative computing environments, each agent may play multiple roles, and each role enjoys different access permissions. Currently, agent access control is typically implemented using a role-based access control model. This model decouples the complex authorization problem through the mapping relationship between "agent-role" and "role-permission," achieving efficient and clear permission management.

[0003] However, role-based access control models rely on static "role-permission" mapping mechanisms, which are difficult to meet the security requirements of dynamically evolving agent behavior. When agents exhibit abnormal API (Application Programming Interface) call sequences, sudden changes in data access patterns, or other behavioral deviations, fixed permission configurations will lead to lag in permission management. Summary of the Invention

[0004] This invention provides a method and apparatus for controlling access permissions for intelligent agents, which solves the problem that existing permission management methods cannot dynamically adapt to the evolution of intelligent agent behavior, resulting in a lag in permission management.

[0005] This invention provides a method for controlling access permissions for an intelligent agent, comprising: Acquire multi-dimensional behavioral data of the agent to be controlled; The multi-dimensional behavioral data is aggregated according to roles to obtain role-specific behavioral data for each role. Based on the graph neural network model, with roles as nodes, the behavioral correlation degree between nodes is calculated based on the role behavior data under the nodes. Based on the behavioral correlation degree, the connection edges between the nodes are constructed, and the role relationship topology graph is constructed based on the connection edges between the nodes. Based on the aforementioned role relationship topology, the set of permissions for the agent to be controlled is predicted.

[0006] According to the present invention, an access control method for an intelligent agent, wherein calculating the behavioral correlation degree between the nodes based on the role behavior data under the node includes: Based on the role behavior data of any two nodes, the behavior similarity between the two nodes is calculated. Based on the call dependency frequency between the role behavior data under any two nodes, the permission dependency degree between any two nodes is calculated. Extract the execution order and data flow path between the role behavior data under any two nodes, and evaluate the business correlation and data flow correlation between any two nodes based on the execution order and the data flow path. The behavioral correlation degree between any two nodes is calculated based on at least one of the behavioral similarity, permission dependency, business relevance, and data flow relevance between them.

[0007] According to the present invention, an access control method for an intelligent agent, wherein constructing connection edges between the nodes based on the behavioral correlation degree includes: Obtain the security constraint relationship between any two nodes from the pre-configured role security policy; Determine the degree of behavioral correlation between any two nodes; Based on the security constraints between any two nodes and the behavioral correlation, the connection edges between the nodes are obtained.

[0008] According to the present invention, an access control method for an intelligent agent, wherein predicting the access set of the intelligent agent to be controlled based on the role relationship topology graph includes: Based on the role behavior data of each node in the role relationship topology graph, the original node characteristics of each node are determined. Based on the connection edges between the nodes, message propagation is performed on the original node features of each node to obtain the fused node features of each node; Based on the fusion node features of each node, the permission set of the agent to be controlled is predicted.

[0009] According to the present invention, an access control method for an intelligent agent, wherein predicting the access set of the intelligent agent to be controlled based on the fusion node features of each node includes: Based on the fusion node features of each node, the initial permission set of the agent to be controlled is predicted. A sandbox testing environment is constructed, and the initial permission set is applied to the test agent to obtain an updated test agent. Based on the updated test agent being tested in the sandbox test environment, the test results are obtained. If the test results meet the preset test results, the initial permission set is used as the permission set of the agent to be controlled.

[0010] According to the present invention, an access control method for an intelligent agent, wherein the updated test intelligent agent is tested in the sandbox test environment to obtain test results, including: Obtain the actual business request; The updated test agent executes the real business request in the sandbox test environment to obtain the test results.

[0011] According to the access control method for an intelligent agent provided by the present invention, the optimization steps of the graph neural network model include: Obtain historical multi-dimensional behavioral data of the agent to be controlled; Based on the graph neural network model, and by applying the historical multi-dimensional behavioral data, a set of historical permissions is predicted. Based on the historical permission set, determine the predicted permission reward; Based on the permission-based reward prediction, the model parameters of the graph neural network model are optimized to obtain an optimized graph neural network model. The optimized graph neural network model is used to subsequently predict the permission set of the agent to be controlled.

[0012] The present invention also provides an access control device for an intelligent agent, comprising: The acquisition unit acquires multi-dimensional behavioral data of the intelligent agent to be controlled. The aggregation unit aggregates the multi-dimensional behavioral data according to roles to obtain role-specific behavioral data for each role. The graph construction unit, based on a graph neural network model, uses roles as nodes, calculates the behavioral correlation between nodes based on the role behavior data under each node, constructs the connection edges between the nodes based on the behavioral correlation, and constructs a role relationship topology graph based on the connection edges between the nodes. The prediction unit predicts the permission set of the agent to be controlled based on the role relationship topology graph.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the access control method for the intelligent agent as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the access control method for an intelligent agent as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the access control method for an intelligent agent as described above.

[0016] The access control method and apparatus for intelligent agents provided by this invention collect multi-dimensional behavioral data of the intelligent agents to be controlled, dynamically update the role behavior data using this data, and construct and analyze the role relationship topology graph based on graph neural networks. Ultimately, it realizes dynamic and intelligent prediction of the set of permissions for intelligent agents, which can effectively discover deep connections between roles that are difficult to detect in traditional methods. It adaptively adjusts permissions according to the actual changes in the behavior of intelligent agents, greatly improving the flexibility, accuracy and security of permission management in multi-agent systems. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts illustrating the access control method for intelligent agents provided by the present invention; Figure 2 This is a schematic diagram of the role relationship topology provided by the present invention; Figure 3 This is the second flowchart of the access control method for intelligent agents provided by the present invention; Figure 4 This is a schematic diagram of the access control system for intelligent agents provided by the present invention; Figure 5 This is a schematic diagram of the access control device for intelligent agents provided by the present invention; Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] To address the aforementioned problems, this invention provides an access control method for intelligent agents, enabling real-time optimization and dynamic binding of agent permissions, effectively improving the security and flexibility of permission management in multi-agent systems or single-agent systems. The following embodiments illustrate its application in a multi-agent system. Figure 1 This is one of the flowcharts illustrating the access control method for intelligent agents provided by the present invention, such as... Figure 1 As shown, the method includes: Step 110: Obtain multi-dimensional behavioral data of the agent to be controlled.

[0021] Here, the intelligent agent to be controlled refers to an autonomous or semi-autonomous software program entity that performs a specific task in a multi-agent system. For example, it could be an AI program responsible for data analysis, a robot executing automated operation and maintenance scripts, or a microservice instance handling user requests. Access control for these intelligent agents is fundamental to ensuring the secure and stable operation of the entire system.

[0022] Furthermore, the multi-dimensional behavioral data here refers to a dataset that can comprehensively and meticulously depict the various activity characteristics of an intelligent agent during operation. For example, multi-dimensional behavioral data may include, but is not limited to: API call level: total number of API calls, API call frequency distribution, percentage of specific API calls, API call success rate, API call response time, etc.; data access level: total data access volume, number of sensitive data accesses, distribution of data access hotspots, data write / read ratio, etc.; resource consumption level: CPU (Central Processing Unit) utilization, memory utilization, network bandwidth usage, etc.; communication interaction level: communication data packet size distribution, communication frequency, communication latency, diversity of communication nodes, etc.; operational behavior level: operation instruction sequence length, operation instruction sequence entropy value, operation time regularity, number of abnormal operation attempts, etc.

[0023] Specifically, a lightweight behavior probe is deployed on the agent to be controlled. This probe is designed to collect real-time, efficient data on the agent's behavior without significantly affecting its normal operation. For example, it can capture multiple behavioral characteristic indicators non-intrusively by hooking operating system calls, monitoring network traffic, and analyzing runtime logs, thus obtaining multi-dimensional behavioral data of the agent to be controlled.

[0024] Step 120: Aggregate the multi-dimensional behavioral data according to roles to obtain role behavioral data for each role.

[0025] Here, role behavior data refers to a dataset formed by categorizing and integrating multi-dimensional behavioral data collected from the controlled intelligence according to the roles assigned to the agents. Roles here include, for example, data analysts, development engineers, and system administrators. It's important to understand that the role behavior data no longer describes the behavior of individual agents, but rather depicts a typical behavioral profile of a group of roles within the system.

[0026] Specifically, firstly, a real-time mapping table between agents and roles can be maintained. Then, upon receiving multi-dimensional behavioral data, the behavioral data of agents belonging to the same role can be aggregated based on this mapping. This aggregation process may include statistical calculations of various behavioral metrics. For example, calculating the average number of API calls, average and variance of CPU utilization, and the union of data access hotspots for an agent under a given role within a unit of time.

[0027] It should be noted that abstracting the individual behavior of the agent to be controlled into role behavior effectively eliminates the behavioral noise caused by the instantaneous task fluctuations of a single agent, making the subsequent role-based correlation analysis more stable and accurate, thereby improving the accuracy of role-based access control.

[0028] Step 130: Based on the graph neural network model, with roles as nodes, calculate the behavioral correlation degree between the nodes based on the role behavior data under the nodes, construct the connection edges between the nodes based on the behavioral correlation degree, and construct the role relationship topology graph based on the connection edges between the nodes.

[0029] Here, behavioral correlation refers to a quantitative metric used to measure the strength of the relationship between different roles in terms of behavioral patterns, such as similarity, dependency, or collaboration. Understandably, high correlation means that two roles have close collaboration or similar behavioral access permission requirements when performing tasks. Furthermore, the role relationship topology graph is a mathematical model represented by a graph structure, its core idea being to network role relationships. In this graph, each node represents a role, the edges connecting nodes represent the behavioral correlation between roles, and the edge weights represent the strength of the correlation. Therefore, the role relationship topology graph can intuitively depict the dynamic and complex relationship network formed by all roles due to their behaviors, and can be considered a core model for permission prediction.

[0030] Specifically, the character behavior data of each role can be input into a graph neural network model. Through the graph neural network model, each role is defined as a node in the graph, and the character behavior data of that role is used as the original node features of that node.

[0031] Then, leveraging the powerful graph structure learning capabilities of graph neural network models, the behavioral correlation between each pair of role nodes can be automatically learned and calculated. For example, the similarity between two role behavior data vectors can be calculated based on their cosine similarity or Euclidean distance; or the correlation strength in the business process can be calculated by analyzing their sequential dependencies in the operational flow. When the calculated behavioral correlation exceeds a preset threshold, a connection edge is established between the two role nodes. Finally, all the role nodes in the graph and the connections between them together constitute a dynamic role relationship topology graph.

[0032] In one embodiment, Figure 2 This is a schematic diagram of the role relationship topology provided by the present invention, such as... Figure 2 As shown in the diagram, the nodes represent different roles in the system, such as system administrators, data analysts, ordinary users, security auditors, and development engineers. The edges connecting the nodes indicate behavioral relationships between these roles, and the weights of the edges quantify the degree of these relationships. For example, the weight of the edge connecting the system administrator and the security auditor is 0.82. The weights of the edges connecting the other two nodes are 0.78, 0.21, and 0.48, respectively. Thus, the role relationship topology diagram clearly reveals the collaborative and interconnected network of different roles in actual work.

[0033] It should be noted that by utilizing the end-to-end learning capabilities of graph neural networks, the implicit relationships between roles are made explicit and structured, constructing a panoramic view that can comprehensively reflect the dynamic collaborative relationships of the system, providing a solid model foundation for accurate permission prediction.

[0034] Step 140: Based on the role relationship topology graph, predict the permission set of the intelligent agent to be controlled.

[0035] Here, the permission set refers to a set of specific access permissions recommended or dynamically assigned to the agent to be controlled. It is understood that the permission set of the agent to be controlled is not statically preset, but dynamically inferred based on the agent's real-time behavior and its position and relationships within the entire role relationship network. This aims to achieve the principle of least necessary permissions, satisfying business needs while avoiding security risks arising from over-granting permissions.

[0036] Specifically, after obtaining the role relationship topology of the agent to be controlled, a graph neural network model can predict the permissions most likely to be needed for each role based on the node and its neighboring nodes and connections in the graph. For example, if the role of the agent to be controlled is "report generator," and there is a strong behavioral association between the "report generator" node and the "data source interface" node in the role topology graph, and the "data source interface" role has permission to access the "core sales database," then the graph neural network model will predict that the "report generator" role is also very likely to need read permission to the "core sales database," and thus add this permission to the set of permissions recommended for the agent to be controlled.

[0037] It should be noted that, based on the role relationship topology graph, the set of permissions for the agent to be controlled is predicted, so that permission allocation is no longer based on rigid rules, but on a deep understanding of the agent's real-time behavior and the dynamic relationship between roles, realizing the transformation from static granting to dynamic prediction.

[0038] The method provided in this invention collects multi-dimensional behavioral data of the agent to be controlled, dynamically updates the role behavior data using this data, and constructs and analyzes the role relationship topology based on graph neural networks. Ultimately, it achieves dynamic and intelligent prediction of the agent's permission set, effectively discovering deep connections between roles that are difficult to detect in traditional methods, and adaptively adjusting permissions according to the actual changes in agent behavior, greatly improving the flexibility, accuracy, and security of permission management in multi-agent systems.

[0039] Based on any of the above embodiments, in step 130, calculating the behavioral correlation between the nodes based on the role behavior data under the node includes: Based on the role behavior data of any two nodes, the behavior similarity between the two nodes is calculated. Based on the call dependency frequency between the role behavior data under any two nodes, the permission dependency degree between any two nodes is calculated. Extract the execution order and data flow path between the role behavior data under any two nodes, and evaluate the business correlation and data flow correlation between any two nodes based on the execution order and the data flow path. The behavioral correlation degree between any two nodes is calculated based on at least one of the behavioral similarity, permission dependency, business relevance, and data flow relevance between them.

[0040] Specifically, the behavioral correlation between nodes can be calculated by considering at least one of the following: behavioral similarity, permission dependency, business relevance, and data flow relevance between any two nodes. When calculating the behavioral correlation using multiple correlation methods, the calculation order of different types of correlation is not important, or they can be calculated in parallel. Furthermore, any type of correlation between nodes can be obtained by calculating the correlation of that type between any two nodes.

[0041] Here, behavioral similarity refers to a quantitative metric used to measure the similarity in the behavioral patterns of controlled agents as two different roles performing their functions. Understandably, high similarity indicates that the two roles may be undertaking similar responsibilities or using similar tools and resources.

[0042] To calculate the behavioral similarity between any two nodes, the following method can be used: First, the role behavior data of any two nodes can be converted into high-dimensional feature vectors. Each dimension corresponds to a behavioral metric, such as the frequency of calls to a specific API or the average CPU utilization. Then, mature vector similarity calculation methods in the field of machine learning can be used, for example, calculating the cosine similarity between the feature vectors of the two role behavior data. The value of cosine similarity ranges from -1 to 1; the closer the value is to 1, the more similar the behavioral patterns of the two roles are. For example, if both the "front-end development" role and the "back-end development" role frequently call the same internal testing API and access the same log database, then calculating their behavioral data vectors will yield a high behavioral similarity score.

[0043] In addition, the degree of permission dependency here refers to the degree to which the execution of a role's behavior depends on the permissions or behavioral results of another role. It usually reflects an asymmetrical relationship between upstream and downstream, front and back.

[0044] To calculate the permission dependency between any two nodes, it can be quantified by analyzing the temporal information and call relationships in the role behavior data. For example, the behavior log sequences of any two roles can be analyzed to count the frequency with which role A's behavior immediately follows role B's behavior, i.e., the call dependency frequency. If a specific operation of role A always occurs or has a high probability of occurring after a specific operation of role B, and A's operation requires a certain state or resource generated by B's operation, then role A can be considered to have a high permission dependency on role B. For example, in the expense reimbursement process, the payment behavior of the "financial approval" role necessarily depends on the approval behavior of the "department manager" role; therefore, the "financial approval" role can be considered to have an extremely high permission dependency on the "department manager" role.

[0045] Furthermore, the business relevance between any two nodes measures the degree of collaboration between the two roles within the same business process, focusing on the task-level collaborative relationship. The data flow relevance between any two nodes, on the other hand, measures how frequently one role's output data is used as input data by another role, focusing on the data-level connection relationship.

[0046] To assess the business and data flow correlation between any two nodes, in-depth analysis of role behavior data can extract business process information. By analyzing logs, call chains, and other information, the execution order and data flow path followed by a role when completing a full business process can be reconstructed. Based on the extracted execution order, if two roles frequently appear as adjacent steps in multiple different business process instances, a high degree of business correlation between them can be assessed. For example, the "order receiving" role and the "inventory check" role are always closely linked in the business process. Simultaneously, based on the extracted data flow path, if it is found that report data files generated by role C are always read and analyzed by role D, then a high degree of data flow correlation between role C and role D can be assessed.

[0047] Ultimately, the behavioral correlation between nodes can be calculated using at least one of the following: behavioral similarity, permission dependency, business relevance, and data flow relevance. In this embodiment of the invention, behavioral similarity, permission dependency, business relevance, and data flow relevance are preferably selected to jointly calculate the behavioral correlation between nodes. For example, these dimensions can be weighted and fused. For instance, a set of weights can be set: w1, w2, w3, w4, and the final behavioral correlation can be calculated using the following formula: Behavioral Correlation = w1 * Behavioral Similarity + w2 * Permission Dependency + w3 * Business Relevance + w4 * Data Flow Relevance. These weights can be pre-configured by the system administrator based on experience, or they can be trained and adaptively adjusted using a machine learning model to achieve the optimal correlation evaluation effect.

[0048] The method provided in this invention decomposes a single, vague concept of behavioral association into multiple specific and measurable dimensions, such as behavioral similarity, permission dependency, business relevance, and data flow relevance. It is understood that this multi-dimensional analysis and calculation approach can more profoundly and accurately reveal the complex relationships between roles at multiple levels, including functionality, processes, and data. This results in a role relationship topology diagram with higher fidelity and interpretability, thus providing a solid foundation for the accuracy and rationality of subsequent permission predictions based on this diagram.

[0049] Based on any of the above embodiments, in step 130, constructing the connection edges between the nodes based on the behavioral correlation degree includes: Obtain the security constraint relationship between any two nodes from the pre-configured role security policy; Determine the degree of behavioral correlation between any two nodes; Based on the security constraints between any two nodes and the behavioral correlation, the connection edges between the nodes are obtained.

[0050] Here, pre-configured role security policies refer to a set of rules pre-defined and configured by system security administrators based on enterprise security management regulations, industry compliance requirements, or general security best practices. Additionally, security constraints refer to explicit stipulations extracted from role security policies regarding the relationship between specific roles. It should be noted that security constraints between any two nodes are typically mandatory, such as mutual exclusion, binding, or prohibited association. Understandably, security constraints provide an inviolable set of permission rules for constructing the role relationship topology, preventing the formation of role permissions that violate security principles solely based on behavioral data.

[0051] Specifically, a role security policy library can be maintained first. This library can be a database table or a structured configuration file. The library stores role pairs and their role security policies. For example, the policy library might explicitly define: {"Transaction Initiator" role, "Transaction Auditor" role}: "Mutual Exclusion"; {"Developer" role, "Tester" role}: "Allow Association".

[0052] When it is necessary to determine whether a connection edge has been established between any two nodes, you can first query the pre-configured role security policy to obtain the security constraint relationship defined for this pair of roles.

[0053] In addition, based on the role behavior data of any two roles, one or more dimensions such as behavioral similarity, permission dependency, business relevance and data flow relevance can be comprehensively calculated to obtain a quantitative behavioral relevance that reflects the degree of mutual correlation between them in actual operation.

[0054] Then, the connection edges between nodes can be obtained through the security constraints and behavioral correlation between any two nodes.

[0055] For example, if the security constraints found are "mutually exclusive" or "prohibited association," then regardless of how high the calculated behavioral correlation is—for instance, even if the correlation is as high as 0.99—it will be forcibly determined that no connection can be established between them. This situation itself is a strong security warning signal, indicating potential violations of separation of duties or account abuse, which can trigger an alert to notify the security administrator to conduct an audit. For example, even if the behaviors of the "transaction initiator" role and the "transaction auditor" role are found to be highly similar, they will never be connected in the graph, thus preventing any subsequent cross-recommendation of permissions.

[0056] For example, if the security constraint found is "allow association", or if the pre-configured role security policy does not define a specific relationship between the two roles, then the decision-making power is given to the behavioral association degree. The calculated behavioral association degree will be compared with a preset association degree threshold. Only when the behavioral association degree is greater than the threshold will a connection edge be established between the two nodes.

[0057] The method provided in this invention introduces a role security policy as a pre-check during the construction of the role relationship topology graph, achieving effective supervision and constraint of the behavior analysis results. It perfectly integrates rule-based static security policies and data-based dynamic behavior analysis, ensuring that the constructed role relationship network not only accurately reflects the collaboration patterns in actual business operations, but more importantly, guarantees that this reflection is carried out under the premise of strictly adhering to the organization's security compliance bottom line. This greatly enhances the security and reliability of the entire access control method, effectively avoiding the risk of access recommendations that may arise from purely data-driven approaches and violate core security principles such as separation of duties.

[0058] Based on any of the above embodiments, step 140 includes: Based on the role behavior data of each node in the role relationship topology graph, the original node characteristics of each node are determined. Based on the connection edges between the nodes, message propagation is performed on the original node features of each node to obtain the fused node features of each node; Based on the fusion node features of each node, the permission set of the agent to be controlled is predicted.

[0059] Here, "original node features" refers to the feature vector used to initialize the representation of each node in the graph neural network model. "Fused node features," on the other hand, refer to a new, more informative feature vector formed after a node's original node features have undergone one or more message propagations, absorbing information from its neighboring nodes. In essence, fused node features no longer merely represent the behavior of the role itself, but rather the combined behavioral characteristics of that role and its collaborative environment.

[0060] Specifically, the first step involves encoding the role behavior data of each node into feature vectors to obtain the original node features. The second step involves message propagation based on the connections between nodes, resulting in fused node features. More specifically, a graph neural network model can be used to perform aggregation operations along the connections in the role relationship topology graph. For any target node in the graph, the graph neural network model first identifies all its directly connected neighbor nodes. Then, message propagation is performed, collecting the original node features of all these neighbor nodes and merging them into a single neighbor information vector using an aggregation function, such as averaging, summing, or maximizing. Next, the graph neural network model updates the target node's features by combining its own original node features with the newly aggregated neighbor information vector, for example, through concatenation or weighted summation, to obtain the fused node features. It should be noted that this process can iterate multiple times, allowing nodes to acquire information about second-degree, third-degree, and even more distant neighbors, thereby gaining a global awareness of their position in the graph.

[0061] The third step involves predicting the permission set for the agent to be controlled based on the fused node features of each node. Specifically, this can be achieved by determining the permission set corresponding to each role of the agent to be controlled, and then using the permission set corresponding to each role as the permission set for the agent to be controlled. To determine the permission set for any role, first, determine the role to which the agent to be controlled belongs and find the corresponding node in the graph. Then, extract the fused node features obtained after message propagation from that node. This feature vector can be input into a pre-trained prediction module in a graph neural network model, for example, one or more fully connected neural network layers followed by a sigmoid or softmax activation function. This yields a score list of all possible permissions output by the prediction module. Then, a probability threshold can be set, and all permissions with scores higher than this threshold are selected to form the final permission set recommended to the agent to be controlled for that role.

[0062] For example, the original characteristics of the "Report Generator" role may not directly indicate that it needs access to the "Sales Database," but because its neighbor in the graph, the "Data Analyst" role, frequently accesses this database, this information is incorporated into the fused node characteristics of the "Report Generator" role through message propagation. When this information-rich fused node characteristic is fed into the prediction module, the model can predict with high probability, based on this strong correlation, that the "Report Generator" role also needs read permissions to the "Sales Database."

[0063] The method provided in this invention upgrades the permission prediction process from static judgment based on the behavior of a single role to dynamic reasoning based on the entire graph context. By introducing a message propagation mechanism to generate fused node features, permission prediction for a role is no longer isolated but fully considers the behavioral patterns of all its collaborators. This method can uncover potential permission needs hidden behind complex collaborative relationships, achieving more accurate, intelligent, and forward-looking permission recommendations. Thus, while ensuring system security, it greatly improves the automation level and operational efficiency of permission management.

[0064] Based on any of the above embodiments, the step of predicting the permission set of the agent to be controlled based on the fusion node features of each node includes: Based on the fusion node features of each node, the initial permission set of the agent to be controlled is predicted. A sandbox testing environment is constructed, and the initial permission set is applied to the test agent to obtain an updated test agent. Based on the updated test agent being tested in the sandbox test environment, the test results are obtained. If the test results meet the preset test results, the initial permission set is used as the permission set of the agent to be controlled.

[0065] Here, the initial permission set refers to a list of candidate permissions directly predicted by the graph neural network model based on role characteristics that incorporate information from neighboring nodes, without actual verification. It represents the optimal inference made by the model based on multi-dimensional behavioral data and role association analysis, and serves as the input for subsequent verification.

[0066] Specifically, the first step is to predict the initial permission set for the agent to be controlled as a whole, based on the fused node features of each node. More specifically, for the initial permission set of any given role, the role to which the agent belongs is first determined, and the fused node features of that role's nodes are extracted from the role relationship topology graph. This feature vector is then input into the prediction module, which calculates the recommendation score for each permission for that role's node. All permissions with scores exceeding a preset probability threshold are combined to form the initial permission set for that role. Thus, the same method can be used to obtain the initial permission sets for the agent to be controlled as a whole for other roles, i.e., the permission set for the agent to be controlled as a whole for all roles.

[0067] The second step involves constructing a sandbox testing environment. The initial permission set is applied to the test agent to obtain an updated test agent. Here, the sandbox testing environment is a highly simulated virtual execution environment completely isolated from the real production environment. It includes API interfaces, data structures, service dependencies, etc., consistent with the production environment, but all operations are confined within the sandbox and will not affect real data or the system. It should be noted that the core function of the sandbox testing environment is to provide a zero-risk testing environment. Furthermore, the updated test agent here refers to a test instance of the same type as the agent to be controlled, deployed in the sandbox environment, and its access control configuration is set to the initial permission set generated in the first step. Understandably, the behavior of this agent will be used to simulate the actual performance of the agent to be controlled after gaining new permissions.

[0068] In detail, containerization or virtualization technologies can be used to dynamically create a sandbox testing environment. This environment is populated with anonymized test data, a simulated API gateway, and other necessary services based on a preset template. Subsequently, a test agent can be instantiated within this sandbox, and its initial set of permissions can be configured as its unique access credentials, thereby updating the test agent.

[0069] The third step involves testing the updated test agent in a sandbox testing environment to obtain test results. If the test results match the preset test results, the initial permission set is used as the permission set for the agent to be controlled. Here, the test results refer to the data set collected after the updated test agent executes a series of predetermined tasks in the sandbox, reflecting its operational status and permission applicability. Test results may include task success or failure flags, operation logs, exception error messages, and records of access denials. These test results are then compared with the preset test results to make a final decision.

[0070] In detail, a set of predefined automated test cases can be executed in a sandbox testing environment by updating the test agent. These test cases are designed to cover the core business functions of the role. For example, for a "report generator" role, test cases might include generating daily, monthly, and ad-hoc query reports on specific topics. During testing, all behavior logs can be captured in real time by a monitoring system. After testing, the test results are analyzed. If all test cases are executed successfully and no errors due to insufficient permissions are found in the logs, the test results are considered to meet the preset test results. In this case, the initial permission set can be confirmed as valid and sufficient, and it is formally adopted and applied as the final permission set to the agent to be controlled in the production environment.

[0071] If any test case fails, and log analysis indicates the failure is due to a missing necessary permission, or monitoring reveals the agent attempted to access a non-essential sensitive resource, even if the initial permission set incorrectly included that permission, then the test result is considered inconsistent with the preset test results. In this case, the initial permission set can be rejected, and the failure details can be recorded. It should be noted that the failure details can serve as feedback data for future optimization or retraining of the graph neural network model, thereby achieving continuous iteration and self-improvement of the model.

[0072] The method provided in this invention adds permission verification in a sandbox testing environment between permission prediction and actual application. This allows for a dual verification of the validity and security of the predicted permission set without affecting the stability and security of online business. This greatly reduces the risk of business interruption or security vulnerabilities caused by model prediction errors, and provides a decisive security guarantee for the implementation of intelligent permission management systems.

[0073] Based on any of the above embodiments, the updated test agent is tested in the sandbox test environment to obtain test results, including: Obtain the actual business request; The updated test agent executes the real business request in the sandbox test environment to obtain the test results.

[0074] Specifically, firstly, network traffic mirroring tools can be used to capture API requests from the production environment, obtaining real business requests. This allows the sandbox testing environment to receive requests similar to those in the production environment, accurately simulating real user behavior and business operation patterns. Then, the test agent can be updated to execute real business requests in the sandbox testing environment to obtain test results. While executing real business requests in the sandbox testing environment, load generation tools can simulate a large number of concurrent user accesses, generating a business load similar to that of the production environment. For example, JMeter can be configured to simulate 1000 concurrent users, following common business operation processes in the production environment, to test the permissions of new roles and evaluate their performance and stability under high load conditions.

[0075] It should be noted that during the construction of the sandbox testing environment, VPN (Virtual Private Network) technology can be used to create an independent network space for the sandbox environment. This allows network traffic within the sandbox environment to be transmitted through a VPN tunnel, isolating it from the external network. Simultaneously, firewall rules can be configured to strictly control communication between the sandbox environment and external systems, allowing only specific ports and protocols for data exchange. For example, only allowing the sandbox environment to access the testing API server can prohibit access to the production database or other sensitive resources.

[0076] In addition, resource quotas are set at the operating system level to strictly limit the use of resources such as CPU, memory, and disk I / O in the sandbox testing environment. For example, the maximum memory usage in the sandbox testing environment is limited to 1GB. When memory usage exceeds this limit during testing, the system will automatically terminate the relevant processes to ensure system stability.

[0077] Based on any of the above embodiments, the optimization steps of the graph neural network model include: Obtain historical multi-dimensional behavioral data of the agent to be controlled; Based on the graph neural network model, and by applying the historical multi-dimensional behavioral data, a set of historical permissions is predicted. Based on the historical permission set, determine the predicted permission reward; Based on the permission-based reward prediction, the model parameters of the graph neural network model are optimized to obtain an optimized graph neural network model. The optimized graph neural network model is used to subsequently predict the permission set of the agent to be controlled.

[0078] Specifically, multi-dimensional behavioral data acquired within a preset time period can be used as historical multi-dimensional behavioral data. Then, a graph neural network model can be used to apply this historical multi-dimensional behavioral data to predict the historical permission set. It should be noted that the method for predicting the historical permission set based on the graph neural network model can be implemented using the access control method for the agent described in any of the above embodiments. Next, the predicted historical permission set and the standard permission set can be evaluated to calculate a reward score. For example, a positive score can be assigned to each historical permission correctly predicted by the model, and a negative score can be assigned to historical permissions incorrectly predicted by the model.

[0079] Furthermore, algorithms such as policy gradient from reinforcement learning can be employed. The reward for permission prediction calculated in the previous step is used as a feedback signal to guide the update of the graph neural network model parameters. Specifically, a loss function can be calculated using the permission prediction reward. The goal of this loss function is to increase the probability of prediction actions that yield high rewards, while decreasing the probability of prediction actions that result in low rewards. Subsequently, optimization algorithms such as gradient descent are used to fine-tune all model parameters in the direction that reduces the loss function. This process is iterated repeatedly on the entire historical dataset until the overall performance of the model reaches a satisfactory level or no longer improves significantly.

[0080] It should be noted that the optimized graph neural network model obtained after the above optimization steps will replace the old model in the production environment and will be used for real-time permission prediction of new agents to be controlled.

[0081] The method provided in this embodiment introduces a self-learning and iterative optimization capability into the entire access control system. This makes the graph neural network model no longer a static entity that remains unchanged after deployment, but rather able to continuously learn from historical experience and dynamically adapt to business changes and new behavioral patterns, thereby continuously improving its prediction accuracy. This significantly reduces the access control risks caused by model aging, concept drift, and other issues, ensuring the long-term effectiveness and reliability of the entire intelligent access control method.

[0082] Based on any of the above embodiments Figure 3 This is the second flowchart illustrating the access control method for intelligent agents provided by this invention, as shown below. Figure 3 As shown, the method includes: The first step is to collect behavioral data of the intelligent agent. This means acquiring multi-dimensional behavioral data of the intelligent agent to be controlled.

[0083] The second step is to construct the topology graph of the intelligent agent role relationship. In detail, the multi-dimensional behavioral data is aggregated according to the role to obtain the role behavior data under each role; based on the graph neural network model, with the role as the node, the behavioral correlation degree between the nodes is calculated based on the role behavior data under the node, and the connection edges between the nodes are constructed based on the behavioral correlation degree, and the role relationship topology graph is constructed based on the connection edges between the nodes.

[0084] The third step is to perform agent sandbox testing. Specifically, based on the role behavior data of each node in the role relationship topology graph, the original node features of each node are determined. Then, based on the connection edges between nodes, message propagation is performed on the original node features of each node to obtain the fused node features. Next, based on the fused node features, the initial permission set of the agent to be controlled is predicted. A sandbox testing environment is constructed, and the initial permission set is applied to the test agent to obtain an updated test agent. Testing is conducted on the updated test agent in the sandbox testing environment to obtain test results. If the test results meet the preset test results, the initial permission set is used as the permission set of the agent to be controlled.

[0085] The fourth step is to dynamically bind the roles of the intelligent agents. Specifically, the set of permissions for each intelligent agent to be controlled is applied to its corresponding role, thus achieving dynamic binding between roles and permissions.

[0086] Based on any of the above embodiments Figure 4 This is a schematic diagram of the access control system for intelligent agents provided by the present invention, as shown below. Figure 4 As shown, the system comprises a base layer, a support layer, and an application layer. The base layer includes a sandbox environment, reinforcement learning algorithms, and behavior probes.

[0087] The sandbox environment serves as a testing environment where an initial set of permissions is applied to the test agent, resulting in an updated test agent. This updated test agent then executes real business requests within the sandbox environment to obtain test results. If the test results match the preset test results, the initial set of permissions is used as the permission set for the agent to be controlled.

[0088] Reinforcement learning algorithms are used to periodically update the model parameters of the graph neural network model, ensuring the long-term effectiveness and reliability of the entire intelligent access control method. Additionally, behavior probes are used to acquire multi-dimensional behavioral data of the agent to be controlled, enabling access control of the agent.

[0089] The supporting layer includes a role relationship topology graph and a role relationship update algorithm. It should be noted that the role relationship topology graph and the role relationship update algorithm are the two core components of the graph neural network model. In constructing the role relationship topology graph, multi-dimensional behavioral data is aggregated according to roles to obtain the role behavior data under each role. Based on the graph neural network model, with roles as nodes, the behavioral correlation between nodes is calculated based on the role behavior data under each node. Based on the behavioral correlation, connecting edges between nodes are constructed, and the role relationship topology graph is built based on these connecting edges. The role relationship update algorithm is mainly used to predict the permission set of the agent to be controlled based on the role relationship topology graph.

[0090] The application layer includes dynamic role binding, which means that by applying the set of permissions of the intelligent agent to be controlled to the corresponding role of the intelligent agent, dynamic binding between roles and permissions is achieved.

[0091] Based on any of the above embodiments Figure 5 This is a schematic diagram of the access control device for intelligent agents provided by the present invention, as shown below. Figure 5 As shown, the device includes: Acquisition unit 510 acquires multi-dimensional behavioral data of the intelligent agent to be controlled; Aggregation unit 520 aggregates the multi-dimensional behavioral data according to roles to obtain role behavioral data for each role. Graph construction unit 530, based on a graph neural network model, uses roles as nodes, calculates the behavioral correlation between nodes based on the role behavior data under the nodes, constructs the connection edges between the nodes based on the behavioral correlation, and constructs a role relationship topology graph based on the connection edges between the nodes. The prediction unit 540 predicts the permission set of the intelligent agent to be controlled based on the role relationship topology graph.

[0092] The device provided in this invention collects multi-dimensional behavioral data of the intelligent agent to be controlled, dynamically updates the role behavior data using this data, and constructs and analyzes the role relationship topology based on graph neural networks. Ultimately, it achieves dynamic and intelligent prediction of the set of permissions for the intelligent agent, effectively discovering deep connections between roles that are difficult to detect in traditional methods, and adaptively adjusting permissions according to the actual changes in the intelligent agent's behavior, greatly improving the flexibility, accuracy, and security of permission management in multi-agent systems.

[0093] Based on any of the above embodiments, the graph construction unit is specifically used for: Based on the role behavior data of any two nodes, the behavior similarity between the two nodes is calculated. Based on the call dependency frequency between the role behavior data under any two nodes, the permission dependency degree between any two nodes is calculated. Extract the execution order and data flow path between the role behavior data under any two nodes, and evaluate the business correlation and data flow correlation between any two nodes based on the execution order and the data flow path. The behavioral correlation degree between any two nodes is calculated based on at least one of the behavioral similarity, permission dependency, business relevance, and data flow relevance between them.

[0094] Based on any of the above embodiments, the graph construction unit is further specifically used for: Obtain the security constraint relationship between any two nodes from the pre-configured role security policy; Determine the degree of behavioral correlation between any two nodes; Based on the security constraints between any two nodes and the behavioral correlation, the connection edges between the nodes are obtained.

[0095] Based on any of the above embodiments, the prediction unit is specifically used for: Based on the role behavior data of each node in the role relationship topology graph, the original node characteristics of each node are determined. Based on the connection edges between the nodes, message propagation is performed on the original node features of each node to obtain the fused node features of each node; Based on the fusion node features of each node, the permission set of the agent to be controlled is predicted.

[0096] Based on any of the above embodiments, the prediction unit is further specifically used for: Based on the fusion node features of each node, the initial permission set of the agent to be controlled is predicted. A sandbox testing environment is constructed, and the initial permission set is applied to the test agent to obtain an updated test agent. Based on the updated test agent being tested in the sandbox test environment, the test results are obtained. If the test results meet the preset test results, the initial permission set is used as the permission set of the agent to be controlled.

[0097] Based on any of the above embodiments, the prediction unit is further specifically used for: Obtain the actual business request; The updated test agent executes the real business request in the sandbox test environment to obtain the test results.

[0098] Based on any of the above embodiments, the device further includes an optimization unit, which is specifically used for: Obtain historical multi-dimensional behavioral data of the agent to be controlled; Based on the graph neural network model, and by applying the historical multi-dimensional behavioral data, a set of historical permissions is predicted. Based on the historical permission set, determine the predicted permission reward; Based on the permission-based reward prediction, the model parameters of the graph neural network model are optimized to obtain an optimized graph neural network model. The optimized graph neural network model is used to subsequently predict the permission set of the agent to be controlled.

[0099] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute an access control method for an intelligent agent. This method includes: acquiring multi-dimensional behavioral data of the intelligent agent to be controlled; aggregating the multi-dimensional behavioral data according to roles to obtain role behavioral data under each role; calculating the behavioral correlation degree between nodes based on a graph neural network model, with roles as nodes and the role behavioral data under each node; constructing connection edges between nodes based on the behavioral correlation degree; constructing a role relationship topology graph based on the connection edges between nodes; and predicting the permission set of the intelligent agent to be controlled based on the role relationship topology graph.

[0100] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the access control method for intelligent agents provided by the above methods. The method includes: acquiring multi-dimensional behavioral data of the intelligent agent to be controlled; aggregating the multi-dimensional behavioral data according to roles to obtain role behavioral data under each role; calculating the behavioral correlation degree between the nodes based on the role behavioral data under the nodes, using roles as nodes, and constructing connection edges between the nodes based on the behavioral correlation degree; constructing a role relationship topology graph based on the connection edges between the nodes; and predicting the permission set of the intelligent agent to be controlled based on the role relationship topology graph.

[0102] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements an access control method for an intelligent agent provided by the methods described above. The method includes: acquiring multi-dimensional behavioral data of the intelligent agent to be controlled; aggregating the multi-dimensional behavioral data according to roles to obtain role behavioral data under each role; calculating the behavioral correlation degree between the nodes based on the role behavioral data under the nodes, using roles as nodes, and constructing connection edges between the nodes based on the behavioral correlation degree; constructing a role relationship topology graph based on the connection edges between the nodes; and predicting the permission set of the intelligent agent to be controlled based on the role relationship topology graph.

[0103] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0104] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for controlling access permissions for an intelligent agent, characterized in that, include: Acquire multi-dimensional behavioral data of the agent to be controlled; The multi-dimensional behavioral data is aggregated according to roles to obtain role-specific behavioral data for each role. Based on the graph neural network model, with roles as nodes, the behavioral correlation degree between nodes is calculated based on the role behavior data under the nodes. Based on the behavioral correlation degree, the connection edges between the nodes are constructed, and the role relationship topology graph is constructed based on the connection edges between the nodes. Based on the aforementioned role relationship topology, the set of permissions for the agent to be controlled is predicted.

2. The access control method for an intelligent agent according to claim 1, characterized in that, The calculation of behavioral correlation between nodes based on the role behavior data under the nodes includes: Based on the role behavior data of any two nodes, the behavior similarity between the two nodes is calculated. Based on the call dependency frequency between the role behavior data under any two nodes, the permission dependency degree between any two nodes is calculated. Extract the execution order and data flow path between the role behavior data under any two nodes, and evaluate the business correlation and data flow correlation between any two nodes based on the execution order and the data flow path. The behavioral correlation degree between any two nodes is calculated based on at least one of the behavioral similarity, permission dependency, business relevance, and data flow relevance between them.

3. The access control method for an intelligent agent according to claim 1, characterized in that, The process of constructing connection edges between nodes based on the behavioral correlation includes: Obtain the security constraint relationship between any two nodes from the pre-configured role security policy; Determine the degree of behavioral correlation between any two nodes; Based on the security constraints between any two nodes and the behavioral correlation, the connection edges between the nodes are obtained.

4. The access control method for an intelligent agent according to any one of claims 1 to 3, characterized in that, The permission set of the agent to be controlled, predicted based on the role relationship topology graph, includes: Based on the role behavior data of each node in the role relationship topology graph, the original node characteristics of each node are determined. Based on the connection edges between the nodes, message propagation is performed on the original node features of each node to obtain the fused node features of each node; Based on the fusion node features of each node, the permission set of the agent to be controlled is predicted.

5. The access control method for an intelligent agent according to claim 4, characterized in that, The set of permissions for the agent to be controlled, predicted based on the fused node features of each node, includes: Based on the fusion node features of each node, the initial permission set of the agent to be controlled is predicted. A sandbox testing environment is constructed, and the initial permission set is applied to the test agent to obtain an updated test agent. Based on the updated test agent being tested in the sandbox test environment, the test results are obtained. If the test results meet the preset test results, the initial permission set is used as the permission set of the agent to be controlled.

6. The access control method for an intelligent agent according to claim 5, characterized in that, The updated test agent is tested in the sandbox test environment to obtain test results, including: Obtain the actual business request; The updated test agent executes the real business request in the sandbox test environment to obtain the test results.

7. The access control method for an intelligent agent according to any one of claims 1 to 3, characterized in that, The optimization steps of the graph neural network model include: Obtain historical multi-dimensional behavioral data of the agent to be controlled; Based on the graph neural network model, and by applying the historical multi-dimensional behavioral data, a set of historical permissions is predicted. Based on the historical permission set, determine the predicted permission reward; Based on the permission-based reward prediction, the model parameters of the graph neural network model are optimized to obtain an optimized graph neural network model. The optimized graph neural network model is used to subsequently predict the permission set of the agent to be controlled.

8. An access control device for an intelligent agent, characterized in that, include: The acquisition unit acquires multi-dimensional behavioral data of the intelligent agent to be controlled. The aggregation unit aggregates the multi-dimensional behavioral data according to roles to obtain role-specific behavioral data for each role. The graph construction unit, based on a graph neural network model, uses roles as nodes, calculates the behavioral correlation between nodes based on the role behavior data under each node, constructs the connection edges between the nodes based on the behavioral correlation, and constructs a role relationship topology graph based on the connection edges between the nodes. The prediction unit predicts the permission set of the agent to be controlled based on the role relationship topology graph.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the access control method for the intelligent agent as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the access control method for the intelligent agent as described in any one of claims 1 to 7.

Images