A cross-domain access control method and system

Abstract

The application discloses a cross-domain access control method and system, and relates to the technical field of network security. The method comprises the following steps: obtaining historical interaction data; constructing a time sequence diagram snapshot sequence; constructing an AT-GNN network; the AT-GNN network comprises a cascaded graph attention network layer and a gate recurrent unit layer; calculating the node feature variance of each static graph snapshot in the time sequence diagram snapshot sequence, and selecting a graph snapshot with the maximum node feature variance as an optimal snapshot; using the graph attention network layer to aggregate neighbor node information of the optimal input snapshot through an attention mechanism, so as to obtain a cross-domain subject space embedding containing spatial structure and dynamic time sequence information; the gate recurrent unit layer updates the time state of the cross-domain subject space embedding, so as to obtain a cross-domain subject trust spatiotemporal fusion feature; determining a trust embedding score of the cross-domain subject according to the cross-domain subject trust spatiotemporal fusion feature and a trust score; and realizing the control of the access rule permission of the cross-domain subject to a target resource based on the trust embedding score and a preset trust security threshold.

Description

Technical Field

[0001] This application relates to the field of network security technology, and in particular to a cross-domain access control method and system. Background Technology

[0002] With the rapid development of the Internet of Things, cloud computing, edge computing, and open networks, cross-organizational, cross-domain, and cross-platform data sharing and device collaboration have become the norm. The blurring of network boundaries, dynamic user / device access, and interconnection of heterogeneous systems across multiple domains have led to a sharp increase in security threats such as identity forgery, unauthorized access, data breaches, and abuse of privileges.

[0003] Existing technologies primarily employ Role-Based Access Control (RBAC) and Attribute-Based Access Control (ABAC). However, RBAC, centered on predefined roles, binds permissions to roles rather than the real-time trusted state of cross-domain entities. Roles are statically fixed, and permissions remain constant within the authorization period. Furthermore, it cannot detect real-time changes in trust between cross-domain entities. Therefore, RBAC only supports coarse-grained authorization and lacks trust quantification and reasoning capabilities. ABAC, based on attribute rule matching, lacks trust modeling and reasoning mechanisms. Rules are static logical expressions, and authorization decisions only involve Boolean judgments of attributes. Its rules struggle to adapt to dynamic trust fluctuations in a timely manner, resulting in high latency in attribute collection and rule evaluation. It cannot link trust and permissions in real-time to achieve real-time trust-driven dynamic permission correction. Consequently, neither the RBAC nor ABAC methods can simultaneously achieve fine-grained, interpretable trust assessment and real-time trust-driven dynamic permission correction. Trust assessment results are disconnected from authorization decisions, leading to an unexplainable and untraceable decision-making process, making it difficult to meet the security requirements of open and dynamic networks. Summary of the Invention

[0004] Therefore, it is necessary to provide a cross-domain access control method and system to address the aforementioned technical problems.

[0005] The present invention adopts the following technical solution: This invention provides a cross-domain access control method, comprising: Obtain historical interaction data between the cross-domain entity and the device being accessed when the cross-domain entity initiates a resource access request; and construct a time-series snapshot sequence based on the historical interaction data; Determine the node feature variance of each static graph snapshot in the time series snapshot sequence, and select the static graph snapshot with the largest node feature variance as the optimal input snapshot; An AT-GNN network based on a graph neural network is constructed. The AT-GNN network includes cascaded graph attention network layers and gated recurrent unit layers. The graph attention network layers are used to aggregate the neighbor node information of the optimal input snapshot through an attention mechanism to obtain cross-domain subject spatial embeddings that simultaneously contain spatial structure and dynamic temporal information. The gated recurrent unit layers are used to update the temporal state of the cross-domain subject spatial embeddings to capture the dynamic trust relationship of cross-domain subjects, thereby obtaining cross-domain subject trust spatiotemporal fusion features. Based on the cross-domain subject trust spatiotemporal fusion features and a preset explicit trust score, the trust embedding score of the cross-domain subject is determined. Based on the trust embedding score and the preset trust security threshold, access rule permission control is implemented for cross-domain subjects to the target resources of the device being accessed across domains.

[0006] Preferably, a time-series snapshot sequence is constructed based on historical interaction data of cross-domain entities, specifically including: Historical interaction data is divided into several consecutive time slices according to a preset time step, and a static graph is constructed in each time slice; the static graph consists of an adjacency matrix and a node feature matrix. The adjacency matrix is ​​an n×n matrix generated based on the actual request and interaction records of cross-domain entities within the time slice, where n is the number of nodes in the graph; if the value in the i-th row and j-th column of the matrix exists, it is the weight of the edge, otherwise it is 0, used to express the trust topology relationship between cross-domain entities; The node feature matrix is ​​obtained by generating structural embeddings through random walks on a static graph using the Node2Vec model, and then fusing the behavioral statistical features of cross-domain subject access requests.

[0007] Preferably, the graph attention network layer includes: a cascaded feature transformation layer, an attention computation layer, a normalization layer, and a feature aggregation layer; the graph attention network layer aggregates the neighbor node information of the optimal input snapshot through an attention mechanism to obtain a cross-domain subject spatial embedding that simultaneously contains spatial structure and dynamic temporal information, specifically including: In the feature transformation layer, the features of the neighboring nodes of the optimal input snapshot are subjected to linear projection transformation; In the attention calculation layer, the features of neighboring nodes after linear projection transformation are sequentially concatenated and attention vectors are calculated to obtain non-normalized attention coefficients. In the normalization layer, the non-normalized attention coefficients are normalized within the range of the neighbors of the optimal input snapshot to obtain the attention weights; The neighbor node features of the optimal input snapshot are weighted, aggregated, and activated based on the attention weights to obtain the cross-domain subject space embedding at the current time. The current cross-domain subject space embedding is temporally fused with the cross-domain subject space embedding output by the previous embedding input gating loop unit, and the output is a cross-domain subject space embedding that simultaneously contains spatial structure and dynamic temporal information.

[0008] Preferably, the gated recurrent unit layer includes parallel reset gates and update gates; the method of using the gated recurrent unit layer to update the temporal state of the cross-domain subject spatial embedding to capture the dynamics of cross-domain subject trust relationships and obtain cross-domain subject trust spatiotemporal fusion features specifically includes: The gate filtering is achieved by resetting the gate in the gated loop unit layer, while the effective information in the historical trust state of the previous moment is embedded in the cross-domain subject space. The fusion ratio between the cross-domain subject space embedding at the current moment and the cross-domain subject space embedding at the previous moment is determined by updating the gate; Based on the effective information and fusion ratio, nonlinear transformation and state update are performed on the cross-domain subject spatial embedding that simultaneously contains spatial structure and dynamic temporal information, and the cross-domain subject trust spatiotemporal fusion feature is output.

[0009] Preferably, the process of determining the activation function of the attention mechanism specifically includes: The initial attention score is obtained by concatenating the cross-domain subject trust spatiotemporal fusion feature output by the AT-GNN network at the current time step with the cross-domain subject trust spatiotemporal fusion feature output by the AT-GNN network at the previous time step and then performing a linear transformation. The initial attention score is recalculated using several different activation functions to obtain the attention score corresponding to each activation function; A comprehensive evaluation index for attention scores is constructed, and the attention scores corresponding to each activation function are evaluated to obtain an evaluation score for each activation function. The activation function corresponding to the attention score with the highest evaluation score is determined as the optimal activation function of the attention mechanism.

[0010] Preferably, the comprehensive evaluation index of the attention score is obtained by weighting a distribution evaluation index and a stability evaluation index; The distribution evaluation index is the standard deviation of the attention score corresponding to the activation function; the stability evaluation index is the ratio of the maximum to the minimum value of the attention score corresponding to the activation function.

[0011] Preferably, determining the trust embedding score of a cross-domain subject based on the spatiotemporal fusion features of cross-domain subject trust and a preset trust score specifically includes: The spatiotemporal fusion characteristics of cross-domain subject trust are linearly transformed with the preset explicit trust score, and then the trust embedding score of cross-domain subject is calculated through a variety of different embedding methods. The average of the maximum trust embedding scores under each embedding method is used as the evaluation index. The trust embedding score determined by the embedding method with the highest evaluation index is selected as the final trust embedding score of the cross-domain subject.

[0012] Preferably, before constructing the AT-GNN network based on a graph neural network, a connectivity check is performed on the cross-domain subject and target resource, specifically including: Verify whether the cross-domain entity exists legally, whether the resource is accessible, and whether there are basic connectivity conditions between the two. If the verification fails, the access permission value will be set to zero and access will be denied.

[0013] Preferably, based on the trust embedding score and a preset trust security threshold, access rule-based permission control is implemented for cross-domain subjects to access target resources of the device being accessed across domains, specifically including: Search for existing active access policy rules that cross-domain entities and target resources, including: If a valid rule exists, the preset access permission value and trust embedding score are read from the rule; if no valid rule exists, the trust embedding score is used for further evaluation, including: If the trust embedding score does not exist, the default access rule permissions for cross-domain subjects are zero. If a trust embedding score exists, it is compared with a preset security threshold. If the trust score is less than the security threshold, access permissions are downgraded by decreasing the current permission value by one, ensuring that the value is not negative. If the trust score is greater than or equal to the security threshold, the current access permissions are maintained.

[0014] This invention provides a cross-domain access control system, comprising: The data acquisition module is used to acquire historical interaction data between the cross-domain entity and the accessed device when the cross-domain entity initiates a resource access request; and to construct a time-series snapshot sequence based on the historical interaction data; The snapshot filtering module is used to determine the node feature variance of each static graph snapshot in the time series snapshot sequence, and select the static graph snapshot with the largest node feature variance as the optimal input snapshot. A trust assessment module is used to construct an AT-GNN network based on a graph neural network. The AT-GNN network includes cascaded graph attention network layers and gated recurrent unit layers. The graph attention network layers aggregate neighbor node information of the optimal input snapshot through an attention mechanism to obtain a cross-domain subject spatial embedding that simultaneously contains spatial structure and dynamic temporal information. The gated recurrent unit layers update the temporal state of the cross-domain subject spatial embedding to capture the dynamic trust relationship of cross-domain subjects, thereby obtaining the cross-domain subject trust spatiotemporal fusion feature. Based on the cross-domain subject trust spatiotemporal fusion feature and a preset explicit trust score, the trust embedding score of the cross-domain subject is determined. The access control module is used to implement access control rules for cross-domain subjects to target resources of the cross-domain accessed device based on the trust embedding score and the preset trust security threshold.

[0015] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: In the cross-domain access control method provided by this invention, the technical problems of insufficient fine-grained trust assessment, disconnect from authorization decision-making, and unexplainable process in existing cross-domain access control are addressed. This invention constructs a temporal graph snapshot sequence and innovatively selects the static graph snapshot with the largest node feature variance as the optimal input snapshot. A graph attention network layer adaptively aggregates the differentiated information of neighboring nodes in the optimal input snapshot through an attention mechanism. Then, a gated recurrent unit layer continuously updates the temporal state of the spatial embedding to obtain spatiotemporal fusion features. These spatiotemporal fusion features are combined with a preset explicit trust score to determine the trust embedding score. This ensures that the trust quantification result integrates both implicit dynamic graph structure information and explicit prior assessment knowledge, forming a complete and interpretable access reasoning chain from dynamic temporal graph construction, optimal snapshot selection, spatiotemporal feature fusion to joint trust computation.

[0016] This invention effectively captures the moments when trust relationships undergo the most significant dynamic changes, avoiding information redundancy and noise interference. It utilizes graph attention network layers and gated recurrent unit layers to accurately extract the spatial trust topology embedding of cross-domain entities and captures the dynamic evolution of trust relationships over time. This yields a spatiotemporal fusion feature that comprehensively characterizes the behavioral patterns and trust evolution trends of cross-domain entities. This spatiotemporal fusion feature, combined with a preset trust score, determines the trust embedding score. This ensures that the trust quantification result for cross-domain access integrates both dynamic graph structure information and prior assessment knowledge. It not only overcomes the limitations of traditional static rules or black-box models, achieving fine-grained, high-precision real-time trust assessment of cross-domain entities, but also, because the trust embedding score calculation process is based on explicit graph structure features and attention weight distribution, possesses natural reasoning path traceability. Ultimately, it provides a transparent, reliable, and deeply coupled quantitative basis for subsequent dynamic permission adjustments based on trust security thresholds, significantly improving the dynamic adaptability, security auditability, and decision interpretability of access control in open cross-domain environments. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0018] Figure 1 A flowchart illustrating a cross-domain access control method provided by the present invention; Figure 2 A mind-map driven access control architecture diagram for a cross-domain access control method provided by the present invention; Figure 3 A mind-map-driven trust assessment model structure diagram for a cross-domain access control method provided by the present invention; Figure 4 The present invention provides a flowchart of a dynamic trust correction access control method for cross-domain access control.

[0019] Figure 5 An AT-GNN neural network architecture diagram for a cross-domain access control method provided by the present invention; Figure 6 This is a schematic diagram of a cross-domain access control system provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in the specification without creative effort are within the scope of protection of this application.

[0021] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating a cross-domain access control method according to the present invention, which specifically includes the following steps: S101: Obtain historical interaction data between the cross-domain entity and the accessed device when the cross-domain entity initiates a resource access request; and construct a time-series snapshot sequence based on the historical interaction data.

[0023] Optionally, a time-series graph snapshot sequence is constructed based on the historical interaction data of cross-domain entities. Specifically, this includes: dividing the historical interaction data into several consecutive time slices according to a preset time step, and constructing a static graph within each time slice; the static graph consists of an adjacency matrix and a node feature matrix; the adjacency matrix is ​​an n×n matrix, generated based on the actual request interaction records of cross-domain entities within the time slice, where n is the number of nodes in the graph; if the value in the i-th row and j-th column of the matrix exists, it is the weight of the edge, otherwise it is 0, used to express the trust topology relationship of cross-domain entities; the node feature matrix generates structural embeddings by random walk on the static graph using the Node2Vec model, and integrates the behavioral statistical features of cross-domain entity access requests.

[0024] Specifically, the historical interaction data of cross-domain entities is constructed into a time-series snapshot sequence, the feature variance of each snapshot node is calculated, and the snapshot with the largest feature variance is selected as the optimal input.

[0025] First, the historical interaction data of cross-domain entities is divided into several consecutive time slices along the time axis. A static graph is constructed within each time slice to represent the current interaction state, ultimately forming a time-series graph snapshot sequence {G¹, G², ..., Gᴷ}. Each snapshot Gᵗ consists of an adjacency matrix Aᵗ and a node feature matrix Xᵗ: the adjacency matrix Aᵗ is generated based on the actual interaction records between cross-domain entities within the time slice and is used to express the trust topology; the node feature matrix Xᵗ is obtained by generating structural embeddings through random walks on the subgraph using Node2Vec and fusing behavioral statistical features. This sequence fully preserves the temporal dynamics and spatial topology of the trust relationship between cross-domain entities, providing a foundational input for subsequent optimal snapshot selection and AT-GNN spatiotemporal feature fusion. Technical details are as follows:

[0026] Initialize the snapshot by setting the initial time step, and generate a static graph snapshot for each time step. .

[0027] ; For each calculate and : ; ; Adjacency Matrix It is an n×n matrix, where n is the number of nodes in the graph. If the value in the i-th row and j-th column of the matrix exists, it is the weight of the edge; otherwise, it is 0. The formula for generating the adjacency matrix is: ; The calculation of the node feature matrix is ​​mainly achieved through the Node2Vec model, which aims to map nodes in the graph to a low-dimensional vector space while preserving the structural neighborhood information of the nodes. The formula for generating the node feature matrix is ​​as follows: ; In the formula, v This is the current central node; u These are the neighbor context nodes obtained through random walks; Represents a given center node v Generate its context node u The conditional probability; V It is the set of all nodes in the graph. exp () is an exponential function used to amplify differences in similarity; v⋅u The inner product of node embedding vectors is used to measure the structural similarity between nodes.

[0028] During training, Node2Vec generates node sequences based on random walks. Its optimization objective is to maximize the co-occurrence log probability of a node and its context, as shown in the formula: By maximizing the objective function, the model can maintain the topological proximity relationship between nodes in a low-dimensional space, providing node features with structure awareness for subsequent trust assessment.

[0029] S102: Determine the node feature variance of each static graph snapshot in the time series snapshot sequence, and select the static graph snapshot with the largest node feature variance as the optimal input snapshot.

[0030] Specifically, since snapshots with larger variances imply a more dispersed distribution of node features and better reflect dynamic changes in the data, the graph snapshot selection function S is used: ; In the formula, Let represent the variance of the feature matrix of the t-th snapshot node. A larger variance indicates a more dispersed distribution of node features, containing richer dynamic information, which is more conducive to improving the accuracy and stability of trust assessment.

[0031] S103: Construct an AT-GNN network based on a graph neural network; the AT-GNN network includes cascaded graph attention network layers and gated recurrent unit layers; the graph attention network layers are used to aggregate the neighbor node information of the optimal input snapshot through an attention mechanism to obtain a cross-domain subject spatial embedding that simultaneously contains spatial structure and dynamic temporal information; the gated recurrent unit layers are used to update the temporal state of the cross-domain subject spatial embedding to capture the dynamic trust relationship of cross-domain subjects, thereby obtaining the cross-domain subject trust spatiotemporal fusion feature; based on the cross-domain subject trust spatiotemporal fusion feature and a preset explicit trust score, the trust embedding score of the cross-domain subject is determined; Among them, the preset explicit trust score is the trust score determined by cross-domain subjects for the target resource. It is used to reflect the real trust tendency between nodes and is directly introduced as prior knowledge without the need for model calculation.

[0032] In addition, before constructing the AT-GNN network based on graph neural networks, connectivity verification is performed on cross-domain entities and target resources. Specifically, this includes verifying whether the cross-domain entities exist legally, whether the resources are accessible, and whether there are basic connectivity conditions between them. If the verification fails, the access permission value is set to zero and access is denied.

[0033] Optionally, the graph attention network layer includes: a cascaded feature transformation layer, an attention calculation layer, a normalization layer, and a feature aggregation layer. The graph attention network layer aggregates the neighbor node information of the optimal input snapshot through an attention mechanism to obtain a cross-domain subject space embedding that simultaneously contains spatial structure and dynamic temporal information. Specifically, this includes: in the feature transformation layer, performing a linear projection transformation on the neighbor node features of the optimal input snapshot; in the attention calculation layer, sequentially concatenating and calculating attention vectors on the linearly projected neighbor node features to obtain non-normalized attention coefficients; in the normalization layer, normalizing the non-normalized attention coefficients within the neighbor range of the optimal input snapshot to obtain attention weights; weighted aggregation and activation of the neighbor node features of the optimal input snapshot based on the attention weights to obtain the cross-domain subject space embedding at the current time; and temporally fusing the cross-domain subject space embedding at the current time with the cross-domain subject space embedding output by the previous time-embedded input gated recurrent unit to output a cross-domain subject space embedding that simultaneously contains spatial structure and dynamic temporal information.

[0034] Optionally, the gated recurrent unit layer includes parallel reset gates and update gates; the step of using the gated recurrent unit layer to update the temporal state of the cross-domain subject spatial embedding to capture the dynamics of cross-domain subject trust relationships and obtain cross-domain subject trust spatiotemporal fusion features specifically includes: filtering effective information from the previous time's historical trust state in the cross-domain subject spatial embedding through the reset gate of the gated recurrent unit layer; determining the fusion ratio between the current time's cross-domain subject spatial embedding and the previous time's cross-domain subject spatial embedding through the update gate; and based on the effective information and the fusion ratio, performing nonlinear transformation and state update on the cross-domain subject spatial embedding that simultaneously contains spatial structure and dynamic temporal information to output cross-domain subject trust spatiotemporal fusion features.

[0035] Optionally, the process of determining the activation function of the attention mechanism specifically includes: concatenating the cross-domain subject trust spatiotemporal fusion feature output by the AT-GNN network at the current time step with the cross-domain subject trust spatiotemporal fusion feature output by the AT-GNN network at the previous time step, and then performing a linear transformation to obtain an initial attention score; recalculating the initial attention score using several different activation functions to obtain the attention score corresponding to each activation function; constructing a comprehensive evaluation index for the attention score, evaluating the attention score corresponding to each activation function, and obtaining an evaluation score for the attention score corresponding to each activation function; and determining the activation function corresponding to the attention score with the highest evaluation score as the activation function of the attention mechanism.

[0036] Optionally, the comprehensive evaluation index of attention score is obtained by weighting a distribution evaluation index and a stability evaluation index; the distribution evaluation index is the standard deviation of the attention score corresponding to the activation function; the stability evaluation index is the ratio of the maximum value to the minimum value among the attention scores corresponding to the activation function.

[0037] Optionally, based on the spatiotemporal fusion characteristics of cross-domain subject trust and a preset trust score, the trust embedding score of the cross-domain subject is determined. Specifically, this includes: performing a linear transformation on the spatiotemporal fusion characteristics of cross-domain subject trust and the preset explicit trust score, and then calculating the trust embedding score of the cross-domain subject through multiple different embedding methods; using the average of the maximum values ​​of the trust embedding scores under each embedding method as the evaluation index, and selecting the trust embedding score determined by the embedding method with the highest evaluation index as the final trust embedding score of the cross-domain subject.

[0038] Specifically, the fusion of graph spatial features and time series features is achieved through AT-GNN layers. See the network structure diagram below. Figure 5 The feature with the smallest loss value is selected for output; the attention layer uses Softmax, Tanh and Sigmoid activation functions for comparison and calculation, and selects the optimal attention weight distribution.

[0039] The AT-GNN layer runs the graph attention network and gated recurrent units (GRUs) multiple times. In each trial, the GAT layer and the GRU layer update the node features sequentially. The GAT layer first aggregates neighbor node information through an attention mechanism to generate a spatial embedding; then, the GRU layer receives this embedding and updates the temporal state to capture the dynamic changes in trust relationships. After propagation through L layers, the final node feature representation is obtained. After each trial, calculate the squared loss between the mean of the output features and the target value:

[0040] ; ; In the formula, The loss is calculated by taking the mean of the output features and the squared value of the target value. For node feature representation, For the true value, These are predicted values.

[0041] Due to the non-convex optimization nature of deep neural networks, a single training iteration may get stuck in a local optimum. Therefore, this paper compares the loss values ​​of all independent trials and selects the trial result with the minimum loss as the optimal output of that layer.

[0042] ; In the formula, For optimal output, The loss is calculated by squared the mean of the output features and the target value.

[0043] This multi-trial-based selection mechanism effectively reduces the interference of random weight initialization on model performance, significantly improving the stability and robustness of the AT-GNN layer in the trust evaluation task. Input features and These represent the node features of the current state and the previous state, respectively. and The attention score is obtained by linear transformation after splicing:

[0044] ; In the formula, for and Attention score obtained by linear transformation after concatenation and This represents the node features of the current state and the previous state.

[0045] To determine the most suitable nonlinear mapping function for the attention distribution characteristics of this model, this paper defines three different activation functions in a comparative experiment: Softmax, Tanh, and Sigmoid. For each activation function, the current state is transformed using a linear transformation. and the previous state The attention score is obtained after concatenation, and then an activation function is used to process the score. These activation functions are used for different attention score calculation methods to evaluate which method is more effective:

[0046] ; ; ; In the formula ， for Attention score corresponding to activation function for Attention score corresponding to activation function for Attention score corresponding to activation function.

[0047] To quantify the impact of different activation functions on model performance, this paper calculates a comprehensive evaluation metric for the output of each activation function. This metric consists of two parts: a distribution metric and a stability metric, aiming to balance the discriminative power and robustness of attention.

[0048] First, the distribution evaluation metric measures the dispersion of attention scores, i.e., the model's ability to distinguish between key nodes and noisy nodes. The distribution evaluation metric measures the dispersion of attention scores.

[0049] ; In the formula, As a distribution evaluation index, It is a function of standard deviation. This represents the attention score for each activation function. This metric is calculated by determining the standard deviation of the attention scores; a higher standard deviation indicates a more diverse distribution of attention weights, allowing the model to focus more intently on key interactions with high trust levels.

[0050] The stability evaluation metric measures the ratio of extreme values ​​in the attention distribution, reflecting the model's strength in capturing salient features: ; In the formula, This is a stability evaluation metric. A higher ratio indicates that the attention mechanism can better highlight the maximum value features and suppress the minimum value noise, thereby enhancing the model's sensitivity to abnormal behavior.

[0051] The comprehensive evaluation index is a weighted sum of the distribution evaluation index and the stability evaluation index. This paper sets the weight coefficient for each index to 0.5, to give equal importance to both discrimination and focusing ability.

[0052] ; In the formula, Comprehensive Evaluation Metrics. The comprehensive evaluation metrics for each activation function are compared, and the result with the highest comprehensive evaluation metric is selected as the optimal output. This method systematically evaluates the performance of different activation functions on a specific trust evaluation task, adaptively selecting the most suitable activation function to calculate the attention score, thereby maximizing the feature extraction capability of the attention mechanism while ensuring model stability.

[0053] ; In the formula, This represents the optimal output corresponding to the activation function with the highest comprehensive evaluation index.

[0054] The core task of the final output module is to combine the extracted spatiotemporal features with explicit trust scores to generate the final trust embedding representation. To explore the impact of different nonlinear mappings on the accuracy of trust prediction, this layer defines several different embedding methods and selects the optimal strategy through empirical comparison.

[0055] By inputting feature h t,cu and h t,cvLet u and v represent the features of the node pair respectively. Given the trust score Rv, concatenate the features of the node pair together: ; In the formula, The features are the concatenation of node pairs (u,v); The concatenated features are linearly transformed, and then different activation functions or normalization operations are used to compute the trust embeddings of the node pairs: ; ; ; ; In the formula, Trust embedding for compute node pairs; This is a dynamic spatiotemporal feature weight matrix, used to perform linear transformation on the spatiotemporal fusion features of trust after splicing cross-domain subjects and objects; This is the explicit trust score weight matrix, used to perform weighted mapping on the explicit trust scores provided by the dataset. For bias terms, This refers to explicit trust scores (true trust scores directly labeled in the dataset). σ Sigmoid activation is used, and tanh activation is used for hyperbolic tangent activation. For L2 normalization, softmax is the normalization exponential function.

[0056] To quantify the effectiveness of different embedding methods, for each method, the generated trust embedding h is... t,uv This layer calculates the average of its maximum values ​​as an evaluation metric. This metric reflects the average confidence level of the model when predicting high-confidence relationships.

[0057] ; In the formula, The average confidence level of trust embedding when predicting high-trust relationships; Compare the evaluation metrics of all methods and select the method with the highest evaluation metric as the best embedding method: ； In the formula, The trust embedding score is the one that corresponds to the method with the highest evaluation metric.

[0058] S104: Based on the trust embedding score and the preset trust security threshold, implement access rule permission control for cross-domain subjects to the target resources of the cross-domain accessed device.

[0059] Optionally, based on the trust embedding score and a preset trust security threshold, control is implemented for cross-domain entities' access rules and permissions to target resources. Specifically, this includes: retrieving whether there are any effective access policy rules between the cross-domain entity and the target resource; if a valid rule exists, then reading the preset access permission value and trust embedding score from the rule; if no valid rule exists, then using the trust embedding score for further judgment; if the trust embedding score does not exist, then defaulting the cross-domain entity's access rule permissions to zero; if the trust embedding score exists, then comparing the trust embedding score with the preset security threshold; if the trust score is less than the security threshold, then performing an access permission downgrade, decreasing the current permission value by one and ensuring it is not negative; if the trust score is greater than or equal to the security threshold, then maintaining the current access permission.

[0060] Specifically, see Figure 2 The overall access control process is as follows: Access Request Initiation and Initialization: The system receives resource access requests initiated by cross-domain entities. It records the access start time Tstart. The system performs attribute parsing and anonymization on the cross-domain entity and collects context information. The cross-domain entity identifier, target resource, and access behavior are then passed to the next stage.

[0061] Cross-Origin Subject – Resource Entity Connectivity Verification: Checks whether the cross-origin subject initiating the access is legitimately existing. Checks whether the target resource is accessible. Checks whether the basic connectivity conditions between the cross-origin subject and the resource are met. If the verification fails → permissions are set to 0, the process ends, and a rejection is returned. If the verification passes → proceeds to the rule retrieval stage.

[0062] Valid Access Rule Retrieval: The system searches for existing effective access policy rules for the "Cross-Domain Entity-Resource" pair. If a valid rule exists, it directly reads the following from the rule: the preset permission value (Permission) and the trust factor (Ration, i.e., the trust score output by GMATA). If no valid rule exists, the system calls the GMATA trust assessment module to obtain the real-time trust score (Ration) for the cross-domain entity. If the trust score does not exist, the default permission (Permission) is 0. If the trust score exists, the initial permission is dynamically calculated based on the trust score. Permission Calculation Based on Trust Score: If a rule exists, the rule-based permissions are used. If no rule exists, Permission = f(Ration); the higher the trust level, the higher the permission level.

[0063] Trust threshold determination and dynamic permission downgrading: The trust score Ration output by the AT-GNN network is compared with the system's preset security threshold to achieve dynamic trust correction access control. The process of obtaining the trust score is described in [link to documentation]. Figure 3 For dynamic trust correction access control, see [link to relevant documentation]. Figure 4 Dynamic trust correction access control specifically refers to: If Ration < security threshold: the cross-domain entity is determined to be low-trust and high-risk. Execute permission downgrade: Permission = max(0, original permission - 1); ensure that the permission value will not be negative, and forcibly shrink risky permissions.

[0064] If Ration ≥ security threshold: Keep the currently calculated permissions unchanged.

[0065] Decision log recording and final output: Record the access end time Tend, and calculate the response time RT = Tend − Tstart. Record complete audit logs: cross-domain subject ID, resource ID, access time, trust embedding score, permission calculation process, whether permission downgrade was triggered, and trust inference path (with interpretable basis).

[0066] Example 1: Cross-Domain Device Access Control in the Internet of Things The process consists of: a cross-domain IoT device initiating a resource access request; the access module initializing device attributes; the GMATA module constructing a snapshot of the device interaction sequence diagram and calculating a trust score using AT-GNN and an attention mechanism; the dynamic correction module verifying connectivity, retrieval rules, execution threshold judgment, and permission downgrading; and the policy module outputting the final authorization result and recording logs.

[0067] Effects: Enables fine-grained trust assessment for cross-domain devices; automatically downgrades permissions for abnormal / low-trust devices; ensures the entire process is explainable, traceable, and auditable; effectively prevents device forgery, unauthorized access, and unauthorized operations.

[0068] Example 2: Dynamic Authorization for Multi-Tenant Cloud Computing Composition: Multi-tenant cross-domain access to cloud resources; real-time collection of tenant behavior and context by the system; output of tenant trust scores; dynamic modules assign corresponding permission levels according to trust scores; automatic demotion of permissions if trust falls below the threshold.

[0069] Effects: Enables dynamic adjustment of tenant permissions based on risk; avoids permission overflow caused by static authorization; and improves the security and compliance of cross-domain access on the cloud platform.

[0070] The above describes a cross-domain access control method provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding cross-domain access control system, such as... Figure 6 As shown.

[0071] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

[0072] The data acquisition module 601 is used to acquire historical interaction data between the cross-domain entity and the accessed device when the cross-domain entity initiates a resource access request; and to construct a time-series snapshot sequence based on the historical interaction data; The snapshot filtering module 602 is used to determine the node feature variance of each static graph snapshot in the time series snapshot sequence, and select the static graph snapshot with the largest node feature variance as the optimal input snapshot. The trust assessment module 603 is used to construct an AT-GNN network based on a graph neural network. The AT-GNN network includes cascaded graph attention network layers and gated recurrent unit layers. The graph attention network layers aggregate the neighbor node information of the optimal input snapshot through an attention mechanism to obtain a cross-domain subject spatial embedding that simultaneously contains spatial structure and dynamic temporal information. The gated recurrent unit layers update the temporal state of the cross-domain subject spatial embedding to capture the dynamic trust relationship of cross-domain subjects, thereby obtaining the cross-domain subject trust spatiotemporal fusion feature. Based on the cross-domain subject trust spatiotemporal fusion feature and a preset explicit trust score, the trust embedding score of the cross-domain subject is determined. The access control module 604 is used to implement access rule access control for cross-domain subjects to target resources of the cross-domain accessed device based on the trust embedding score and the preset trust security threshold.

[0073] For specific limitations regarding a cross-domain access control system, please refer to the limitations of a cross-domain access control method described above, which will not be repeated here. Each module in the aforementioned cross-domain access control system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

Claims

1. A cross-domain access control method, characterized in that, include: Who performs the following steps: Obtain historical interaction data between the cross-domain entity and the device being accessed when the cross-domain entity initiates a resource access request; and construct a time-series snapshot sequence based on the historical interaction data; Determine the node feature variance of each static graph snapshot in the time series snapshot sequence, and select the static graph snapshot with the largest node feature variance as the optimal input snapshot; An AT-GNN network based on a graph neural network is constructed. The AT-GNN network includes cascaded graph attention network layers and gated recurrent unit layers. The graph attention network layers are used to aggregate the neighbor node information of the optimal input snapshot through an attention mechanism to obtain cross-domain subject spatial embeddings that simultaneously contain spatial structure and dynamic temporal information. The gated recurrent unit layers are used to update the temporal state of the cross-domain subject spatial embeddings to capture the dynamic trust relationship of cross-domain subjects, thereby obtaining cross-domain subject trust spatiotemporal fusion features. Based on the cross-domain subject trust spatiotemporal fusion features and a preset explicit trust score, the trust embedding score of the cross-domain subject is determined. Based on the trust embedding score and the preset trust security threshold, access rule permission control is implemented for cross-domain subjects to the target resources of the device being accessed across domains.

2. The cross-domain access control method as described in claim 1, characterized in that, The construction of a time-series snapshot sequence based on historical interaction data of cross-domain entities specifically includes: Historical interaction data is divided into several consecutive time slices according to a preset time step, and a static graph is constructed in each time slice; the static graph consists of an adjacency matrix and a node feature matrix. The adjacency matrix is ​​an n×n matrix generated based on the actual request and interaction records of cross-domain entities within the time slice, where n is the number of nodes in the graph; if the value in the i-th row and j-th column of the matrix exists, it is the weight of the edge, otherwise it is 0, used to express the trust topology relationship between cross-domain entities; The node feature matrix is ​​obtained by generating structural embeddings through random walks on a static graph using the Node2Vec model, and then fusing the behavioral statistical features of cross-domain subject access requests.

3. The cross-domain access control method as described in claim 1, characterized in that, The graph attention network layer comprises: a cascaded feature transformation layer, an attention computation layer, a normalization layer, and a feature aggregation layer. Utilizing this graph attention network layer, the neighbor node information of the optimal input snapshot is aggregated through an attention mechanism to obtain a cross-domain subject spatial embedding that simultaneously contains spatial structure and dynamic temporal information. Specifically, this includes: In the feature transformation layer, the features of the neighboring nodes of the optimal input snapshot are subjected to linear projection transformation; In the attention calculation layer, the features of neighboring nodes after linear projection transformation are sequentially concatenated and attention vectors are calculated to obtain non-normalized attention coefficients. In the normalization layer, the non-normalized attention coefficients are normalized within the range of the neighbors of the optimal input snapshot to obtain the attention weights; The neighbor node features of the optimal input snapshot are weighted, aggregated, and activated based on the attention weights to obtain the cross-domain subject space embedding at the current time. The current cross-domain subject space embedding is temporally fused with the cross-domain subject space embedding output by the previous embedding input gating loop unit, and the output is a cross-domain subject space embedding that simultaneously contains spatial structure and dynamic temporal information.

4. The cross-domain access control method as described in claim 1, characterized in that, The gated recurrent unit layer includes parallel reset gates and update gates; the method of using the gated recurrent unit layer to update the temporal state of cross-domain subject spatial embedding to capture the dynamics of cross-domain subject trust relationships and obtain cross-domain subject trust spatiotemporal fusion features specifically includes: The gate filtering is achieved by resetting the gate in the gated loop unit layer, while the effective information in the historical trust state of the previous moment is embedded in the cross-domain subject space. The fusion ratio between the cross-domain subject space embedding at the current moment and the cross-domain subject space embedding at the previous moment is determined by updating the gate; Based on the effective information and fusion ratio, nonlinear transformation and state update are performed on the cross-domain subject spatial embedding that simultaneously contains spatial structure and dynamic temporal information, and the cross-domain subject trust spatiotemporal fusion feature is output.

5. The cross-domain access control method as described in claim 1, characterized in that, The process of determining the activation function of the attention mechanism specifically includes: The initial attention score is obtained by concatenating the cross-domain subject trust spatiotemporal fusion feature output by the AT-GNN network at the current time step with the cross-domain subject trust spatiotemporal fusion feature output by the AT-GNN network at the previous time step and then performing a linear transformation. The initial attention score is recalculated using several different activation functions to obtain the attention score corresponding to each activation function; A comprehensive evaluation index for attention scores is constructed, and the attention scores corresponding to each activation function are evaluated to obtain an evaluation score for each activation function. The activation function corresponding to the attention score with the highest evaluation score is determined as the optimal activation function of the attention mechanism.

6. The cross-domain access control method as described in claim 5, characterized in that, The comprehensive evaluation index of the attention score is obtained by weighting the distribution evaluation index and the stability evaluation index; The distribution evaluation index is the standard deviation of the attention score corresponding to the activation function; the stability evaluation index is the ratio of the maximum to the minimum value of the attention score corresponding to the activation function.

7. The cross-domain access control method as described in claim 1, characterized in that, The method of determining the trust embedding score of cross-domain entities based on cross-domain entity trust spatiotemporal fusion features and a preset trust score specifically includes: The spatiotemporal fusion characteristics of cross-domain subject trust are linearly transformed with the preset explicit trust score, and then the trust embedding score of cross-domain subject is calculated through a variety of different embedding methods. The average of the maximum trust embedding scores under each embedding method is used as the evaluation index. The trust embedding score determined by the embedding method with the highest evaluation index is selected as the final trust embedding score of the cross-domain subject.

8. The cross-domain access control method as described in claim 1, characterized in that, Before constructing the AT-GNN network based on a graph neural network, connectivity verification is performed on cross-domain entities and target resources, specifically including: Verify whether the cross-domain entity exists legally, whether the resource is accessible, and whether there are basic connectivity conditions between the two. If the verification fails, the access permission value will be set to zero and access will be denied.

9. The cross-domain access control method according to claim 1, characterized in that, Based on the trust embedding score and a preset trust security threshold, access rule-based permission control is implemented for cross-domain subjects to access target resources of the device being accessed across domains, specifically including: Search for existing active access policy rules that cross-domain entities and target resources, including: If a valid rule exists, the preset access permission value and trust embedding score are read from the rule; if no valid rule exists, the trust embedding score is used for further evaluation, including: If the trust embedding score does not exist, the default access rule permissions for cross-domain subjects are zero. If a trust embedding score exists, it is compared with a preset security threshold. If the trust score is less than the security threshold, access permissions are downgraded by decreasing the current permission value by one, ensuring that the value is not negative. If the trust score is greater than or equal to the security threshold, the current access permissions are maintained.

10. A cross-domain access control system, characterized in that, include: The data acquisition module is used to acquire historical interaction data between the cross-domain entity and the accessed device when the cross-domain entity initiates a resource access request; And a time-series snapshot sequence is constructed based on historical interaction data; The snapshot filtering module is used to determine the node feature variance of each static graph snapshot in the time series snapshot sequence, and select the static graph snapshot with the largest node feature variance as the optimal input snapshot. A trust assessment module is used to construct an AT-GNN network based on a graph neural network. The AT-GNN network includes cascaded graph attention network layers and gated recurrent unit layers. The graph attention network layers aggregate neighbor node information of the optimal input snapshot through an attention mechanism to obtain a cross-domain subject spatial embedding that simultaneously contains spatial structure and dynamic temporal information. The gated recurrent unit layers update the temporal state of the cross-domain subject spatial embedding to capture the dynamic trust relationship of cross-domain subjects, thereby obtaining the cross-domain subject trust spatiotemporal fusion feature. Based on the cross-domain subject trust spatiotemporal fusion feature and a preset explicit trust score, the trust embedding score of the cross-domain subject is determined. The access control module is used to implement access control rules for cross-domain subjects to target resources of the cross-domain accessed device based on the trust embedding score and the preset trust security threshold.

Images