A method, apparatus and system for trusted mutual recognition of permissions between cross-domain intelligent agents

Abstract

This invention provides a method, apparatus, and system for trusted mutual recognition of permissions between cross-domain intelligent agents, relating to the field of artificial intelligence security technology. The method includes: receiving a call request from a source domain intelligent agent to perform a corresponding operation on a target domain intelligent agent; based on the spatial relationship characteristics between the source domain permission space and the target domain permission space involved in the call request, obtaining a set of affine transformation parameters through a permission transformation model, and using cone volume measurement for evaluation and security verification to finally select affine transformation parameters suitable for permission credential mapping; and using the affine transformation parameters to map the original permission credential of the source domain intelligent agent from the source domain permission space to the target domain permission space, generating a mapped permission credential that can be recognized and verified by the target domain. This invention effectively solves the problem of trusted mutual recognition of permissions in cross-domain intelligent agent collaboration by geometricizing the permission space and utilizing affine transformation, spatial partitioning, and geometric calculation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence security technology, and in particular to a method, apparatus and system for trusted mutual recognition of permissions between cross-domain intelligent agents. Background Technology

[0002] In AI-driven cross-domain collaboration scenarios, the heterogeneity of permission systems has become a major challenge in achieving trusted mutual recognition between intelligent agents. In medical research, the intelligent data analysis agent (a core concept in AI) of the medical research center, Med-Research-AI (Medical Research AI), needs to access the intelligent medical record database management system (Hosp-Record-DB) of the collaborating hospital to perform operations such as reading anonymized medical record data of diabetic patients. The research center's permissions are defined in its two-dimensional geometric space using data sensitivity level and research project level as coordinate axes, with Med-Research-AI's credentials represented by point (4, 8). While the hospital's permission space is also two-dimensional, its coordinate axes are defined as operational risk level and the range of accessible data. The two domains not only have different semantic dimensions of permissions but also inconsistent numerical ranges, resulting in the source domain's permission credentials (4, 8) being unable to be directly understood and used in the target domain.

[0003] Therefore, cross-domain permission verification often relies on pre-established static mapping tables or complex dynamic trust negotiation protocols. Static mapping tables lack flexibility and are difficult to adapt to the dynamic adjustment of permission policies in any domain. On the other hand, complex negotiation protocols introduce significant performance overhead and implementation complexity, which cannot meet the needs of efficient collaboration. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, device and system for trusted mutual recognition of permissions in cross-domain intelligent agents. By geometricizing the permission space and using affine transformation, space partitioning and geometric calculation, the invention effectively solves the problem of trusted mutual recognition of permissions in cross-domain intelligent agent collaboration, and brings significant improvements in accuracy, security, flexibility, precision and efficiency.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a method for trusted mutual recognition of permissions between cross-domain intelligent agents, the method comprising:

[0007] Receive a request from a source domain agent to perform a corresponding operation on a target domain agent;

[0008] Based on the spatial relationship characteristics between the source domain permission space and the target domain permission space involved in the call request, an affine transformation parameter set is obtained through a permission transformation model. Then, the cone volume metric is used for evaluation and security verification. Finally, the affine transformation parameters suitable for permission credential mapping are selected. The original permission credential of the source domain agent is mapped from the source domain permission space to the target domain permission space using the affine transformation parameters, generating a mapped permission credential that can be identified and verified in the target domain.

[0009] Based on the mapped permission credentials, and combined with the set of permission center points of the agents in the target domain permission space, the target domain permission space is divided into multiple permission regions by a clustering algorithm. The target permission region to which the mapped permission credentials belong is determined, and the permission range polygon of the source domain agent in the target domain is generated based on the boundary point set of the target permission region.

[0010] Based on the corresponding operation in the call request, obtain the corresponding operation point in the target domain permission space. Based on the permission range polygon, use the ray-matrix geometric inclusion detection algorithm to determine whether the operation point is located inside the permission range polygon. If the operation point is determined to be inside, allow the execution of the corresponding operation and forward the call request to the target domain agent for execution.

[0011] Furthermore, the invocation request contains the original permission credentials of the source domain agent, which are represented as geometric coordinate points in the source domain permission space and are used for affine transformation and cross-domain mapping.

[0012] Furthermore, based on the spatial relationship characteristics between the source domain permission space and the target domain permission space involved in the call request, an affine transformation parameter set is obtained through a permission transformation model. Then, using a cone volume metric for evaluation and security verification, affine transformation parameters suitable for permission credential mapping are finally selected, including:

[0013] Based on the source domain permission space and target domain permission space involved in the call request, the spatial relationship features between the two are extracted. The spatial relationship features include permission dimension mapping rules, coordinate scaling ratio and spatial rotation angle.

[0014] The extracted spatial relationship features are input into a pre-trained permission conversion model for processing to obtain the output affine transformation parameter set.

[0015] Using the center of the target domain permission space as the vertex and the mapping direction defined by the affine transformation parameters as the cone axis direction, a cone covering the target domain permission space is generated; the actual intersection volume of the cone and the target domain permission space is calculated, and the ratio of this volume value to the total volume of the target domain permission space is used as the cone volume metric value to quantitatively evaluate the degree of coverage of the target domain permission space by the affine transformation parameters;

[0016] The parameters in the affine transformation parameter set are sorted in descending order based on the cone volume metric, and a subset of parameters with metric values ​​higher than a preset threshold are selected. This subset of parameters is then subjected to compatibility and risk verification according to the security policy of the target domain permission space to obtain the parameters that pass the verification. The parameters that pass the verification are used as the final affine transformation parameters for permission credential mapping.

[0017] Furthermore, the original authorization credentials of the source domain agent are mapped from the source domain authorization space to the target domain authorization space using the affine transformation parameters, generating a mapped authorization credential that can be recognized and verified in the target domain, including:

[0018] Read the final affine transformation parameters that have passed verification and parse them into translation vectors, scaling matrices, and rotation matrices for coordinate transformation;

[0019] Based on translation vectors, scaling matrices, and rotation matrices, the geometric coordinates of the original permission credential are sequentially scaled, rotated, and translated to obtain new coordinates in the target domain permission space.

[0020] The obtained new coordinate points are encapsulated according to the format of the target domain permission credential, and a spatial location verification label is attached to generate a mapped permission credential that can be identified and verified by the target domain.

[0021] Furthermore, based on the mapped permission credentials and combined with the set of permission center points of agents in the target domain permission space, a clustering algorithm is used to divide the target domain permission space into multiple permission regions, including:

[0022] Obtain the coordinates of the mapped permission credential in the target domain permission space;

[0023] Based on the coordinate positions, load a predefined set of permission center points for the target domain, where each permission center point corresponds to a permission level;

[0024] Using the loaded set of permission center points as cluster centers, a distance metric algorithm is used to dynamically divide the target domain permission space, generating multiple convex polygon permission regions defined by boundary point sets;

[0025] Based on the divided permission areas, calculate the spatial distance between the coordinates of the mapped permission credential and the center point of each permission;

[0026] Based on the calculated spatial distance, the mapped permission credentials are assigned to the target permission area corresponding to the permission center point with the smallest distance.

[0027] Further, the target permission region to which the mapped permission credential belongs is determined, and a permission range polygon mapped to the source domain agent in the target domain is generated based on the boundary point set of the target permission region, including:

[0028] Obtain the target permission region identifier corresponding to the mapped permission credential;

[0029] Based on the target permission region identifier, retrieve the set of boundary points corresponding to the target permission region;

[0030] Using the retrieved set of boundary points, a closed convex polygon geometry is constructed by connecting them in spatial order.

[0031] A topological verification is performed on the convex polygon geometry, and the polygon that passes the verification is used as the permission scope polygon of the source domain agent in the target domain and bound to the mapped permission credential.

[0032] Furthermore, based on the corresponding operation in the call request, the corresponding operation point in the target domain permission space is obtained. Based on the permission range polygon, a ray-mapping geometric inclusion detection algorithm is used to determine whether the operation point is located inside the permission range polygon. If the operation point is determined to be inside, the corresponding operation is allowed to be executed, and the call request is forwarded to the target domain agent for execution, including:

[0033] Parse the operation description information in the call request and extract the coordinates of the corresponding operation point in the target domain permission space;

[0034] Based on the bounded permission-scoped polygon structure, read its boundary vertex sequence;

[0035] Using the coordinates of the operation point, an infinitely extending detection ray is emitted along a preset direction in the target domain permission space;

[0036] Calculate the coordinates of all intersection points based on the sequence of boundary vertices of the detected ray and the polygon within the permission range;

[0037] Based on the obtained intersection coordinates, count the number of valid intersections. If the number is odd, the operation point is determined to be inside the permission range polygon; if the number is even, it is determined to be outside.

[0038] If the operation point is located inside the permission scope polygon, an evaluation result is generated that allows the corresponding operation to be executed;

[0039] The evaluation results are appended to the invocation request and forwarded to the target domain agent to trigger the corresponding operation.

[0040] Secondly, a cross-domain intelligent agent trusted mutual recognition system includes:

[0041] The receiving module is used to receive call requests from the source domain agent to perform corresponding operations on the target domain agent;

[0042] The generation module is used to obtain a set of affine transformation parameters through a permission transformation model based on the spatial relationship characteristics between the source domain permission space and the target domain permission space involved in the call request, and to evaluate and verify the security using a cone volume metric. Finally, it selects affine transformation parameters suitable for permission credential mapping, and uses the affine transformation parameters to map the original permission credential of the source domain agent from the source domain permission space to the target domain permission space, generating a mapped permission credential that can be identified and verified in the target domain.

[0043] The partitioning module is used to divide the target domain permission space into multiple permission regions based on the mapped permission credential and the set of permission center points of the agent in the target domain permission space through a clustering algorithm, determine the target permission region to which the mapped permission credential belongs, and generate a polygon of the permission range of the source domain agent in the target domain based on the boundary point set of the target permission region.

[0044] The processing module is used to obtain the corresponding operation point in the target domain permission space according to the corresponding operation in the call request, and use the ray-matrix geometric inclusion detection algorithm to determine whether the operation point is located inside the permission range polygon. If the operation point is determined to be inside, the corresponding operation is allowed to be executed, and the call request is forwarded to the target domain agent for execution.

[0045] Thirdly, a computing device includes:

[0046] One or more processors;

[0047] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0048] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0049] The above-described solution of the present invention has at least the following beneficial effects:

[0050] By introducing permission space topology analysis, spatial relationship features, including permission dimension mapping rules, coordinate scaling ratios, and spatial rotation angles, are dynamically extracted. A pre-trained permission transformation model is used to generate affine transformation parameters, which can accurately map the semantics and numerical range of the source domain permission credentials to the target domain permission space, generating permission credentials that can be identified and verified in the target domain. This overcomes the limitations of traditional static mapping tables in handling semantic differences and numerical range mismatches, and achieves lossless conversion and accurate expression of permission meaning between different domains.

[0051] The affine transformation parameters described in this invention are dynamically determined and need to be verified by the target domain's security policy. This means that when the permission policy, space definition, or security requirements of any domain change, the system can quickly adapt by recalculating and verifying the transformation parameters without having to redefine or frequently update complex static mapping protocols. This significantly improves the adaptability to dynamic environments and solves the problems of poor flexibility and difficulty in coping with dynamic policy adjustments in traditional methods.

[0052] By determining the attribution of mapped permission credentials to convex polygon permission regions divided based on predefined permission center points of the target domain, and binding the credentials to their corresponding permission range polygons, the permission verification problem is transformed into a rigorous geometric spatial inclusion relationship judgment problem. This effectively prevents unauthorized access and ensures the reliability and security of authorization decisions. Furthermore, the security policy verification step added during the mapping process further guarantees that the conversion process and results meet the security requirements of the target domain.

[0053] This invention not only realizes cross-domain mapping of agent permissions, but also further refines the granularity of permission control to specific operations. By mapping the corresponding operation to the operation point in the target domain permission space, and determining whether the point is located inside the permission range polygon of the requesting agent, it determines whether to allow the execution of the operation. This mechanism realizes fine-grained judgment on whether an agent is allowed to execute a corresponding operation in the target domain, improving the precision and flexibility of permission management. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a cross-domain intelligent agent authorization trust mutual recognition method provided by an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of a cross-domain intelligent agent authorization trust mutual recognition system provided by an embodiment of the present invention. Detailed Implementation

[0056] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0057] like Figure 1 As shown, an embodiment of the present invention proposes a method for trusted mutual recognition of permissions between cross-domain intelligent agents, the method comprising the following steps:

[0058] Step 1: Receive a call request from the source domain agent to perform a corresponding operation on the target domain agent;

[0059] Step 2: Based on the spatial relationship characteristics between the source domain permission space and the target domain permission space involved in the call request, obtain the affine transformation parameter set through the permission transformation model, and use the cone volume metric for evaluation and security verification to finally select the affine transformation parameters suitable for permission credential mapping. Then, use the affine transformation parameters to map the original permission credential of the source domain agent from the source domain permission space to the target domain permission space to generate a mapped permission credential that can be identified and verified in the target domain.

[0060] Step 3: Based on the mapped permission credential, and combined with the set of permission center points of the agent in the target domain permission space, the target domain permission space is divided into multiple permission regions by a clustering algorithm. The target permission region to which the mapped permission credential belongs is determined, and the permission range polygon of the source domain agent in the target domain is generated according to the boundary point set of the target permission region.

[0061] Step 4: Based on the corresponding operation in the call request, obtain the corresponding operation point in the target domain permission space. Based on the permission range polygon, use the ray-matrix geometric inclusion detection algorithm to determine whether the operation point is located inside the permission range polygon. If the operation point is determined to be inside, the corresponding operation is allowed to be executed, and the call request is forwarded to the target domain agent for execution.

[0062] In this embodiment of the invention, by standardizing the reception and processing of cross-domain call requests, secure access and initialization of permission requests between heterogeneous systems are achieved. By dynamically analyzing the topological relationship of the permission space and determining the affine transformation parameters, the source domain permission credentials are accurately mapped to a form recognizable by the target domain, effectively solving the heterogeneity problem of different permission systems in semantics and numerical range, and improving the accuracy and cross-domain adaptability of the mapping. Then, based on the predefined permission center point of the target domain, a permission region is generated and the permission range polygon is determined through a spatial partitioning algorithm, realizing refined and geometrical regional management of the mapped permissions, enhancing the granularity and flexibility of permission control. By mapping operations to spatial points and using a geometric inclusion detection algorithm for permission verification, it is ensured that only operations within the permission range are allowed, thereby achieving trusted mutual recognition of cross-domain permissions and automated execution of operations while ensuring security.

[0063] In a preferred embodiment of the present invention, the invocation request includes the original permission credentials of the source domain agent, which are represented as geometric coordinate points in the source domain permission space for affine transformation and cross-domain mapping.

[0064] In embodiments of the present invention, by invoking the request containing the original permission credentials of the source domain agent, the original permission credentials are represented as geometric coordinate points in the source domain permission space, which are used for affine transformation and cross-domain mapping. This design transforms abstract permission information into concrete spatial geometric points, giving the permission credentials a precise, computable, and measurable mathematical representation. At the same time, this approach enhances the standardization and consistency of permission expression, avoids semantic ambiguity caused by differences in permission description languages ​​or structures, and significantly improves the interpretability and processing efficiency in the cross-domain permission mutual recognition process.

[0065] In a preferred embodiment of the present invention, step 1, receiving a call request from the source domain agent to perform a corresponding operation on the target domain agent, specifically involves the following process:

[0066] A standardized secure communication interface module is deployed as the unified access point for the cross-domain permission mutual recognition process. This module is responsible for continuously listening to and receiving cross-domain operation call requests initiated by source domain agents. When a request arrives, the interface module first performs a strict compliance check on the inbound request protocol, verifying whether the message format, encryption method, and transmission protocol conform to the predefined interaction specifications. At the same time, it performs preliminary identity authentication through technologies such as digital signature verification. By parsing the encrypted message, it accurately extracts key operational semantic elements, including the unique identifier of the target domain agent, the specific operation type requested, and, most importantly, the original permission credentials of the source domain agent. The original permission credentials are represented as one or more geometric coordinate points with clear semantics within the source domain permission space.

[0067] In a specific implementation of cross-medical collaboration, the intelligent data analysis agent of the medical research center, namely Med-Research-AI, as the source domain agent, needs to access the intelligent medical record database management agent of the collaborating hospital, namely Hospital-Record-DB, to perform the operation of reading anonymized medical record data of diabetic patients. Med-Research-AI requests and obtains a geometric coordinate point (4, 8) representing its permissions from its source domain permission space, where the horizontal coordinate value 4 represents that it is authorized to process relatively sensitive health data in the data sensitivity level dimension, and the vertical coordinate value 8 represents that it belongs to a research project with high priority and trust in the research project level dimension. The agent initiates a call request to this system through the HTTPS protocol with encryption and digital signature. The message explicitly includes the target domain agent identifier Hospital-Record-DB-01, the operation type of reading anonymized medical record data of diabetic patients, and the aforementioned permission credential point (4, 8). After the receiving module verifies the digital signature to ensure the legitimacy of the request source, it successfully parses these key elements and completes the initialization of the cross-domain mutual recognition process.

[0068] In a preferred embodiment of the present invention, step 2 involves obtaining an affine transformation parameter set through a permission transformation model based on the spatial relationship characteristics between the source domain permission space and the target domain permission space involved in the invocation request. Then, using a cone volume metric for evaluation and security verification, affine transformation parameters suitable for permission credential mapping are finally selected, including:

[0069] Step 21: Based on the source domain permission space and target domain permission space involved in the call request, extract the spatial relationship features between the two. The spatial relationship features include permission dimension mapping rules, coordinate scaling ratio, and spatial rotation angle, specifically including:

[0070] Following step 1, to map the source domain permission coordinates to the target domain permission space, it is necessary to first establish the geometric transformation relationship between the two spaces. By querying a pre-set cross-domain spatial mapping strategy library, feature comparison and semantic relationship analysis are performed on the permission space metadata of the source and target domains, thereby extracting three key feature parameters for defining the spatial affine transformation. Specifically:

[0071] Permission dimension mapping rules: used to establish the semantic correspondence between the coordinate axis dimensions of the source domain permission space and the target domain permission space; by parsing the metadata of the permission spaces of the two domains, based on ontology or predefined semantic mapping tables, a semantic bridge between dimensions is established. For example, the data sensitivity level dimension of the source domain is mapped to the operational risk level dimension of the target domain, and the research project level dimension of the source domain is mapped to the accessible data range dimension of the target domain.

[0072] Coordinate scaling ratio: determined by calculating the ratio of the numerical ranges of the two spaces in the corresponding dimension; for example, if the sensitivity level of the source domain data is in the range of [1, 5] and the operational risk level of the target domain is in the range of [1, 10], then the scaling ratio in this dimension is 2.0.

[0073] Spatial rotation angle: determined by analyzing the angle between the basis vectors of two spatial coordinate systems, used to correct coordinate system orientation deviations caused by differences in permission model design concepts; for example, identifying a 30-degree orientation offset between the target domain coordinate system and the source domain coordinate system.

[0074] In the specific implementation of step 21, the system takes the permission coordinate point (4, 8) of Med-Research-AI in step 1, queries the mapping strategy library, and determines the dimension mapping relationship, including the data sensitivity level corresponding to the target operation risk level and the research project level corresponding to the target accessible data range; after analysis and calculation, the scaling ratio on the operation risk level dimension is 2.0, which comes from the value range [1, 5] to [1, 10], and the spatial rotation angle is determined to be 30°.

[0075] Step 22: Input the extracted spatial relationship features into the pre-trained permission conversion model for processing to obtain the output affine transformation parameter set, specifically including:

[0076] The spatial relationship features extracted in step 21, namely the dimension mapping rules, scaling ratio 2.0, and rotation angle 30°, are used as input vectors and sent to the pre-trained permission conversion model. This model is trained based on a large number of historical mapping examples and performs comprehensive reasoning through the internally integrated mapping rule library and geometric calculation logic to output a complete set of affine transformation parameters, including scaling matrix, rotation matrix, and translation vector.

[0077] The above features are input into the permission conversion model, and the construction process of the permission conversion model is as follows:

[0078] This process needs to cover multiple stages, including data collection, cleaning, standardization, and partitioning. The collected training samples must be successful cross-domain permission mapping cases that have been manually reviewed and approved by various industries within the past five years, with a cumulative number of no less than 150,000 cases, to ensure coverage of various business areas such as healthcare and different permission space configuration scenarios. Each sample must contain complete permission space metadata of the source and target domains, including the definition, numerical range, and semantic description of each dimension, as well as manually confirmed spatial relationship features, such as dimension mapping rules, coordinate scaling ratios, spatial rotation angles, and actual verified affine transformation parameters, such as scaling matrices, rotation matrices, and translation vectors.

[0079] The data cleaning phase requires multiple rounds of verification of the collected samples. First, samples with missing metadata or incomplete parameters are removed to ensure that each sample has complete information. Second, contradictory samples are detected through preset logical rules, such as cases where the source domain and target domain dimension ranges do not match but are labeled with unreasonable scaling ratios. Such samples need to be manually reviewed or directly removed. Finally, statistical methods are used to identify outliers, such as samples with rotation angles exceeding the reasonable range of 0 to 60 degrees. The proportion of outlier samples should be controlled within 3%.

[0080] During the data standardization phase, different types of features need to be processed specifically. For continuous numerical features such as scaling ratios and rotation angles, the min-max method is used to map them to the range of 0 to 1, eliminating the impact of dimensional differences on model training. For semantic features such as dimension mapping rules, one-hot encoding is used to convert them into fixed-length binary vectors, enabling the model to effectively identify the correspondence between different dimensions. For affine transformation parameters used as labels, they need to be converted into a unified vector form, such as converting a two-dimensional scaling matrix into a vector containing two elements, directly representing rotation angles numerically, and converting translation vectors into vectors containing two elements. The dataset should be divided into training, validation, and test sets according to a 7:2:1 ratio. The training set is used for learning model parameters, the validation set is used to adjust hyperparameters during training, and the test set is used to finally evaluate model performance. When dividing, it is necessary to ensure that the proportion of samples from each business domain in the three sets is consistent with the original sample set to avoid affecting the model's generalization ability due to uneven data distribution.

[0081] The model architecture employs a deep neural network, with an overall structure divided into an input layer, a feature fusion layer, and a multi-head output layer. Each layer is optimized for the characteristics of feature space transformation. The number of neurons in the input layer is consistent with the length of the standardized feature vector. For example, when the feature vector contains six elements, including scaling ratio, rotation angle, and dimension mapping rule encoding, the input layer requires six neurons. Its main function is to perform a preliminary linear transformation on the input feature vector, mapping the features to a high-dimensional space to prepare for subsequent feature fusion.

[0082] The feature fusion layer comprises three fully connected sublayers for deep fusion and abstraction of input features, extracting high-level feature representations. The first sublayer has 64 neurons and uses the ReLU activation function to enhance the model's non-linear expressive power. The second sublayer has 32 neurons and also uses the ReLU activation function to further compress the feature dimension and extract more abstract features. The third sublayer has 16 neurons and uses the LeakyReLU activation function to avoid the neuron death problem. At the same time, a residual connection is added between the first and third sublayers, directly superimposing the output of the first sublayer onto the output of the third sublayer, alleviating the gradient decay problem of deep networks and ensuring the integrity of feature propagation.

[0083] The multi-head output layer is designed with three independent output branches based on the different types of affine transformation parameters, generating scaling matrix, rotation matrix, and translation vector respectively. The scaling matrix branch has two neurons and uses a linear activation function. The output is processed by the ReLU function to ensure that the scaling ratio is non-negative. The rotation angle branch has one neuron and uses a linear activation function. The output is constrained to ensure that it is between 0 and 60 degrees. The translation vector branch has two neurons and uses a linear activation function to directly output the two components of the translation vector. The outputs of each branch are combined to form a complete set of affine transformation parameters.

[0084] The model training employs supervised learning, optimizing network parameters through multiple iterations. Before training, network weights are initialized: the weights of the input layer and feature fusion layer use the He initialization method, while the weights of the output layer use the Xavier initialization method. Bias terms are initialized to 0. During training, the Adam optimizer is used to adjust parameters. The initial learning rate is set to 0.001, and it is automatically halved if the validation set loss does not decrease for five consecutive iterations, with a minimum reduction to 0.00001.

[0085] The loss function uses weighted mean squared error, with weights set according to the importance of different parameters. The scaling and rotation matrices are weighted at 1.0, and the translation vector is weighted at 0.8 to emphasize the requirements for spatial scaling and rotation accuracy. During each training round, the training set samples are first input into the network for forward computation to obtain the predicted affine transformation parameters. Then, the predicted and true values ​​are substituted into the loss function to calculate the error. The error is then propagated from the output layer to the input layer via backpropagation. Weights and biases are adjusted based on the error contribution of each neuron, iteratively updating the network parameters.

[0086] Multiple mechanisms are employed during training to prevent overfitting. In addition to using a validation set to monitor model performance, a dropout layer is added after each sub-layer of the feature fusion layer with a dropout rate of 0.2 to randomly discard some neurons and reduce the network's dependence on local features. At the same time, L2 regularization is used with a regularization coefficient of 0.0001 to limit the size of the weight parameters. When the validation set loss does not decrease for 10 consecutive rounds, an early stopping mechanism is triggered to stop training and save the current optimal model parameters.

[0087] After model training, comprehensive performance validation is performed on an independent test set. Evaluation metrics include mean absolute error (MAE) and root mean square error (RMSE), which measure the average deviation and overall dispersion of predicted parameters from actual parameters, respectively. For high-precision fields such as healthcare and finance, the MAE should be controlled within 0.05 and the RMSE within 0.08. Simultaneously, the model's adaptability to different business domains needs to be verified, ensuring that the prediction error in each domain is within the preset threshold. If the error exceeds the threshold in a certain domain, incremental training with additional samples for that domain is required.

[0088] After the model is deployed, a dynamic update mechanism needs to be established to regularly collect new successful cross-domain permission mapping cases. The model should be incrementally trained once every quarter to update the network parameters to adapt to the dynamic changes in the permission space, such as the expansion of the target domain permission dimension or the adjustment of the numerical range. After each update, the performance needs to be re-verified on the test set to ensure that the model always maintains high mapping accuracy.

[0089] Finally, in the model deployment and application phase, the trained model is encapsulated into a callable service module. This module receives the spatial relationship features extracted in step 21 as input, and outputs a complete set of affine transformation parameters after internal forward computation. Simultaneously, this service integrates version management, performance monitoring, and regular update mechanisms, enabling it to adapt to changes in security policies across different target domains and ensuring the continuous provision of accurate and reliable spatial transformation parameter prediction services. The internally calculated affine transformation parameter set is as follows: scaling matrix [2.0, 0; 0, 1.2], rotation matrix [cos30°, -sin30°; sin30°, cos30°], and translation vector [1, 0.5].

[0090] In specific implementation, step 22 inputs the features from step 21 into the permission conversion model. This model is trained using historical cross-domain access data in the medical field. When the input includes features such as data sensitivity to operational risk level mapping rules, scaling ratio 2.0, and rotation angle 30°, the model outputs a set of affine transformation parameters after internal calculation: scaling matrix [2.0, 0; 0, 1.2], rotation matrix [cos30°, -sin30°; sin30°, cos30°], and translation vector [1, 0.5].

[0091] Step 23: Using the center of the target domain permission space as the vertex and the mapping direction defined by the affine transformation parameters as the cone axis direction, generate a cone covering the target domain permission space; calculate the actual intersection volume between the cone and the target domain permission space, and use the ratio of this volume value to the total volume of the target domain permission space as the cone volume metric to quantitatively evaluate the coverage of the target domain permission space by the affine transformation parameters; sort the parameters in the affine transformation parameter set in descending order based on the cone volume metric, and select a subset of parameters with metric values ​​higher than a preset threshold; perform compatibility and risk verification on this parameter subset according to the security policy of the target domain permission space to obtain the verified parameters; use the verified parameters as the final affine transformation parameters for permission credential mapping, specifically including:

[0092] The geometric center of the target domain permission space is taken as the vertex of the cone (this center is calculated from the average value of all permission coordinate points predefined in the target domain, for example, the center coordinates of the permission space of the target domain (cooperating hospital) in the medical scenario are (5, 6)); the mapping direction from the source domain permission point to the target domain permission point defined by a certain set of affine transformation parameters output in step 22 is taken as the axis direction of the cone (that is, the direction of the coordinate transformation vector corresponding to the parameter, for example, the mapping direction of a certain parameter group is the vector direction from the source domain (4, 8) to the target domain (10.24, 9.36).

[0093] Based on the overall boundary of the target domain permission space (determined by the maximum and minimum values ​​of the target domain permission dimension, such as the target domain operation risk level dimension range [1, 10] and the accessible data range dimension range [0, 15] in a medical scenario, forming a rectangular boundary), adjust the base radius and height of the cone to ensure that the generated cone can completely cover the entire target domain permission space, avoiding omissions in subsequent volume calculations due to incomplete coverage.

[0094] Calculate the volume of the actual intersection of the generated cone with the target domain permission space (usually a rectangle or polyhedron structure) (i.e., the volume of the part of the cone that falls inside the target domain permission space); then calculate the total volume of the target domain permission space itself (obtained by multiplying the numerical ranges of each permission dimension, for example, in a medical scenario, the total volume of the target domain is the length of the operation risk level range × the length of the accessible data range, i.e. (10-1)×(15-0)=135).

[0095] The ratio obtained by dividing the volume of the intersection of the cone and the target domain permission space by the total volume of the target domain permission space is the cone volume metric value corresponding to the set of affine transformation parameters. This value is used to quantify the degree of coverage of the parameter to the target domain permission space. The higher the value, the better the integrity of the parameter in covering the scope of legal permissions in the target domain when mapping permissions.

[0096] A preset volume measurement threshold is set (this threshold is set by the target domain according to business needs; for example, in a medical scenario, the threshold is set to 0.7 to ensure that permissions cover the core data area); all affine transformation parameters output in step 22 are sorted from high to low according to their respective conical volume measurement values; parameters with measurement values ​​greater than or equal to the preset threshold are selected to form a high-coverage parameter subset, and parameters with insufficient coverage are excluded (e.g., parameters with a measurement value of 0.6 are removed because they are below the threshold of 0.7).

[0097] Based on all predefined security policy rules in the target domain (the rules are configured by the target domain administrator and are strongly associated with the permission dimension); taking the medical scenario (cooperating hospital) as an example, the security policy rules include: the mapping result of the operation risk level dimension ≤ 8 (avoiding permission mapping with excessively high risk), the mapping result of the accessible data range dimension > 0 (excluding invalid mappings without data access permissions), and specific constraints such as anonymous medical record data access permissions must not be mapped to complete medical record access permissions.

[0098] To comprehensively verify parameter security, representative feature points (covering different permission levels to ensure comprehensive verification) are selected from the source domain permission space, including: extremely high permission points in the source domain, such as the extremely high permission point (5, 9) in the source domain (research center) in a medical scenario, representing the highest data sensitivity and highest research priority; medium permission points, such as (3, 6); and lowest permission points, such as (1, 2). For each set of parameters in the high-coverage parameter subset, these source domain feature points are substituted into the affine transformation logic corresponding to the parameter to calculate the mapping result of each feature point in the target domain permission space, that is, to simulate the permission mapping effect when the parameter is actually used.

[0099] The affine transformation logic specifically includes: based on preset affine transformation parameters (including scaling matrix, rotation matrix, and translation vector), converting the coordinates of feature points in the source domain permission space into recognizable coordinates in the target domain permission space. The specific implementation process is as follows: First, a scaling transformation is performed, representing the source domain feature points as (x1, y1) (x1 corresponds to the first permission dimension of the source domain, and y1 corresponds to the second permission dimension of the source domain). The diagonal elements a (scaling ratio of the first dimension of the source domain to the first dimension of the target domain) and d (scaling ratio of the second dimension of the source domain to the second dimension of the target domain) of the scaling matrix are extracted from the affine transformation parameter set. Through the calculation of x2=x1×a and y2=y1×d, The scaled coordinates (x2, y2) are obtained to eliminate the difference in the numerical range of the corresponding permission dimensions between the source and target domains. Then, a rotation transformation is performed. The rotation matrix (with elements cosθ, -sinθ, sinθ, cosθ, where cosθ is the cosine of θ and sinθ is the sine of θ) is extracted from the affine transformation parameter set based on the angle θ between the source and target domain coordinate systems. The scaled coordinates (x2, y2) are substituted into the calculations x3 = x2 × cosθ - y2 × sinθ and y3 = x2 × sinθ + y2 × cosθ to obtain the rotated coordinates (x3, y3) to correct the directional deviation between the permission space coordinate systems of the source and target domains.

[0100] Finally, a translation transformation is performed, and the translation vector (e, f) is extracted from the affine transformation parameter set. e is the translation offset in the first dimension, and f is the translation offset in the second dimension. The rotated coordinates (x3, y3) are substituted into the calculation of x4=x3+e and y4=y3+f to obtain the final coordinates (x4, y4) of the target domain, so as to achieve the alignment of the coordinates with the reference range of the target domain permission space.

[0101] The simulation mapping results of all feature points of each set of parameters are compared and verified one by one with the loaded target domain security policy rules. If all simulation results of a set of parameters fully comply with all security policy rules (no rule is violated), the parameter is determined to pass the security verification. If the simulation result of any feature point of a set of parameters violates any security policy rule (for example, in a medical scenario, a parameter maps the source domain's extremely high privilege point (5, 9) to the target domain (11.0, 11.3), where the operation risk level is 11.0, violating the rule of less than or equal to 8), the parameter is determined to fail the security verification and is removed from the parameter subset. Taking a medical scenario as an example, the initial parameter set was removed because the simulation result (11.0, 11.3) violated the operation risk level ≤ 8. The new parameter set output by the permission conversion model after recalculation (scaling matrix [1.6, 0; 0, 1.0], rotation matrix [cos30°, -sin30°; sin30°, cos30°], translation vector [1, 1]) simulated the mapping of the source domain's extremely high permission point (5, 9) to (7.8, 10.5). The operation risk level 7.8 ≤ 8, the accessible data range 10.5 > 0, and it meets other medical record access permission constraints. Therefore, the parameter was determined to pass the security verification.

[0102] From the high-coverage parameter subset, all parameters that pass the security policy verification are selected; if there is only one set of parameters that passes verification (such as a new parameter set in the medical scenario), then this set of parameters is directly determined as the final affine transformation parameters; if there are multiple sets of parameters that pass verification, then the cone volume metric values ​​of each set of parameters are compared again, and the set with the highest metric value is selected as the final parameter (to ensure optimal permission coverage under the premise of security compliance); the final determined parameters will be used for the subsequent formal mapping from the original permission credentials of the source domain (such as (4,8) in the medical scenario) to the permission space of the target domain.

[0103] In a preferred embodiment of the present invention, step 2, which involves mapping the original permission credentials of the source domain agent from the source domain permission space to the target domain permission space using the affine transformation parameters, and generating a mapped permission credential that can be recognized and verified in the target domain, includes:

[0104] Step 24: Read the verified final affine transformation parameters and parse them into translation vectors, scaling matrices, and rotation matrices for coordinate transformation, specifically including:

[0105] Receive the final affine transformation parameter set output from step 23, which has passed the target domain security policy verification; the implementation process of this step is to call a dedicated parameter parser, read the data structure of the parameter set, and deconstruct and instantiate it into three independently operable mathematical transformation components according to the predefined parameter organization format: a scaling matrix that defines the scaling factors of each dimension, a rotation matrix that defines the rotation orientation of the coordinate system, and a translation vector that defines the offset of the coordinate origin. Through this parsing process, the comprehensive parameter package output from step 23 is converted into a serialized basic geometric transformation operation unit that can be directly called in step 25.

[0106] In specific implementation, step 24 follows the aforementioned medical collaboration case. The system reads the final affine transformation parameter set verified in step 23 and parses its specific composition: the scaling matrix is ​​[1.6, 0; 0, 1.0], indicating a scaling of 1.6 times in the operational risk level dimension, while keeping the original values ​​unchanged in the accessible data range dimension; the rotation matrix is ​​[cos30°, -sin30°; sin30°, cos30°], corresponding to a 30-degree spatial rotation transformation; the translation vector is [1, 1], used to perform the final offset adjustment of the coordinate position after the transformation. At this point, all the basic operation units required for the mapping calculation are ready.

[0107] Step 25: Based on the translation vector, scaling matrix, and rotation matrix, perform scaling, rotation, and translation operations on the geometric coordinates of the original authorization credential in sequence to obtain new coordinates in the target domain authorization space. Specifically, this includes:

[0108] Following a strict geometric transformation sequence, the three basic geometric transformations obtained from step 24 are sequentially applied to the geometric coordinates of the original source domain permission credentials received in step 1. First, a scaling matrix is ​​applied to scale the coordinates in dimensions, multiplying each dimension value of the coordinates by the corresponding scaling factor in the scaling matrix to achieve a linear transformation of the permission value from the source domain range to the target domain range. Next, a rotation matrix is ​​applied to align the scaled coordinates in space, adjusting the orientation of the coordinates in multidimensional space through matrix operations to align it with the coordinate system orientation of the target domain permission space. Finally, a translation vector is applied to offset the coordinates after rotation and alignment, adding each component of the translation vector to the corresponding dimension of the coordinates to achieve precise positioning of the coordinates in the target domain permission space.

[0109] In specific implementation, step 25, according to the embodiment, maps the original permission coordinate point (4, 8) of Med-Research-AI to the target domain space. First, a scaling matrix [1.6, 0; 0, 1.0] is applied to process it, resulting in an intermediate result (6.4, 8.0). Then, a rotation matrix [cos30°, -sin30°; sin30°, cos30°] is applied to the intermediate result for spatial rotation and alignment. Finally, a translation vector [1, 1] is applied to the rotation result for offset, and the new coordinate position of the point in the target domain permission space is calculated to be (7.8, 10.5). Thus, the permission credentials of the source domain have been mathematically and accurately converted to the target domain space.

[0110] Step 26: Encapsulate the obtained new coordinate points according to the format of the target domain permission credential, and attach a spatial location verification label to generate a mapped permission credential that can be identified and verified by the target domain. Specifically, this includes:

[0111] The new coordinates obtained in step 25 are standardized and encapsulated according to the data structure and credential specifications required by the target domain permission system. Specifically, the values ​​of each dimension of the new coordinates are filled into the corresponding fields of the target domain permission credential template. In addition, to ensure the verifiability and tamper-proofness of the mapped credential, the generation process also includes attaching a spatial location verification tag. This tag is generated by a cryptographic hash function based on the value of the new coordinates, the timestamp, and a key that can only be verified by the target domain permission center. Finally, a mapped permission credential with a compliant format, complete content, and verification information is output.

[0112] In step 26, following the previous step, the system obtains the mapped coordinates (7.8, 10.5). Then, a permission credential data structure conforming to the hospital domain specifications is created. 7.8 is filled into the operation risk level field, and 10.5 is filled into the accessible data range field. Simultaneously, using this coordinate data, the current timestamp, and a security key, the system generates a digital tag through cryptographic hashing and attaches it to the credential, forming the final, target-domain-recognizable mapped permission credential. This credential represents the equivalent permissions granted to Med-Research-AI within the hospital's permission space.

[0113] In a preferred embodiment of the present invention, step 3, based on the mapped permission credential and combined with the set of permission center points of agents in the target domain permission space, divides the target domain permission space into multiple permission regions using a clustering algorithm, including:

[0114] Step 31, obtaining the coordinates of the mapped permission credential in the target domain permission space, specifically including:

[0115] Step 31 involves obtaining the coordinates of the mapped permission credential in the target domain permission space, thus completing the data preparation for permission area division. This process involves parsing and extracting information from the standardized permission credential generated in step 26. The system calls a dedicated credential parsing submodule to read the complete data structure of the credential, which follows the permission credential format specification defined by the target domain. The parsing submodule accurately extracts the values ​​representing the permission location by locating predefined coordinate fields. These values ​​are the precise geometric representation of the source domain agent's permissions in the target domain permission space after affine transformation. After extraction, the coordinate values ​​are converted into geometric point data objects that can be processed internally by the system.

[0116] In specific implementation, step 31 follows the aforementioned medical collaboration case. Step 26 generates a target domain permission credential containing the mapped coordinate point (7.8, 10.5). The system parses the credential and extracts its coordinate position in the target domain permission space: the operation risk level dimension value is 7.8, the accessible data range dimension value is 10.5, i.e., the coordinate point (7.8, 10.5).

[0117] Step 32: Based on the coordinate positions, load a predefined set of permission center points for the target domain, where each permission center point corresponds to a permission level, specifically including:

[0118] Step 32: Based on the coordinate positions, load the predefined set of permission center points for the target domain, where each permission center point corresponds to a permission level. This is achieved by the system retrieving and loading the set of permission center points predefined and configured by the domain administrator from the target domain's persistent storage system. Each center point in this set has a fixed coordinate position in the target domain's permission space and a unique correspondence with a specific permission level. These permission levels are defined according to the target domain's security policy and business requirements, forming a complete classification system for all permission states within the target domain. During the loading process, the system verifies the integrity and validity of the center point data. Upon completion, the entire set is loaded into a data structure in memory as a stable reference system.

[0119] In specific implementation, step 32 continues the aforementioned embodiment, whereby the system loads a predefined set of permission centers from the configuration library of the cooperating hospital. This set contains three core points: permission center A (2, 3), corresponding to permission to access only public health statistics data; permission center B (6, 7), corresponding to permission to access anonymized medical record data; and permission center C (8, 10), corresponding to permission to access detailed medical record data, but requiring additional approval permissions.

[0120] Step 33: Using the loaded set of permission center points as cluster centers, a distance metric algorithm is used to dynamically divide the target domain permission space, generating multiple convex polygon permission regions defined by boundary point sets, specifically including:

[0121] Step 33: Using the loaded set of permission center points as cluster centers, a distance metric algorithm is used to dynamically partition the target domain permission space, generating multiple convex polygon permission regions defined by boundary point sets. The implementation process involves using the set of permission center points loaded in step 32 as fixed cluster centers, and employing a spatial partitioning algorithm, such as Voronoi diagram partitioning, using Euclidean distance as the metric, to divide the entire target domain permission space into several non-overlapping convex polygon sub-regions. Each sub-region uniquely corresponds to a permission center point and contains all points in the space closest to that center point. The output of the partitioning algorithm is that each permission region is precisely defined by its set of boundary points, thus discretizing the continuous space into permission regions with clear geometric boundaries.

[0122] In step 33, the system uses the three loaded permission center points A(2,3), B(6,7), and C(8,10) as the core to perform spatial partitioning calculations. After the calculations are completed, the target domain permission space is divided into three convex polygon regions: region A (around center point A), region B (around center point B), and region C (around center point C). The shape and extent of each region are defined by the set of coordinates of its boundary vertices. For example, region B is bounded by its boundary point set {(4,5), (5,6), (7,8), (6,9)}.

[0123] Step 34: Based on the divided permission regions, calculate the spatial distance between the coordinates of the mapped permission credential and the center point of each permission region, specifically including:

[0124] Step 34: Based on the divided permission areas, calculate the spatial distance between the coordinates of the mapped permission credential and each permission center point. The implementation process involves obtaining the coordinates of the mapped permission credential obtained in step 31, and calculating the geometric spatial distance between this point and each permission center point loaded in step 32. The distance metric is based on the principle of multidimensional Euclidean distance. The square root of the sum of the squares of the differences in coordinate values ​​of two points in each dimension is calculated to obtain a scalarized distance value, providing a comparable quantitative indicator for the proximity between the permission credential coordinates and each permission level.

[0125] In step 34, the system calculates the distances from the mapping point (7.8, 10.5) to the three permission center points. It calculates the distance d1 to point A (2, 3), the distance d2 to point B (6, 7), and the distance d3 to point C (8, 10). The results show that the value of distance d2 is less than d1 and d3, meaning that point (7.8, 10.5) is geometrically closest to the center point B representing the permission to access de-identified medical record data.

[0126] Step 35: Based on the calculated spatial distance, assign the mapped permission credential to the target permission region corresponding to the permission center point with the smallest distance, specifically including:

[0127] Step 35: Based on the calculated spatial distance, the mapped permission credential is assigned to the target permission area corresponding to the permission center point with the smallest distance. The implementation process involves receiving and comparing all spatial distance results calculated in step 34, filtering out the smallest distance value through comparison operations, and recording its corresponding permission center point. Based on the correspondence between the permission center point and the permission area established in step 33, the mapped permission credential is finally assigned to the target permission area under the jurisdiction of the permission center point corresponding to the smallest distance, thus completing the positioning and classification of the source domain agent's permissions in the target domain space.

[0128] In the specific implementation of step 35, following the previous step, since the distance d2 from the mapping point (7.8, 10.5) to the center point B (6, 7) is the smallest, the system determines that the credential should belong to the permission area corresponding to the center point B, that is, area B, which can access the de-identified medical record data area; this attribution result means that the permission level of the source domain agent Med-Research-AI in the target domain hospital is to access the de-identified medical record data.

[0129] In a preferred embodiment of the present invention, determining the target permission region to which the mapped permission credential belongs, and generating a permission range polygon of the source domain agent in the target domain based on the boundary point set of the target permission region, includes:

[0130] Step 36, obtaining the target permission region identifier corresponding to the mapped permission credential, specifically including:

[0131] Obtaining the target permission region identifier corresponding to the mapped permission credential is an intermediate step connecting permission region attribution determination and geometric structure construction. The implementation process is that the system receives the permission region attribution determination result generated in step 35 through the internal data interface. The result data object explicitly contains the unique identifier of the target permission region to which the mapped permission credential belongs. This identifier is a symbol code that corresponds one-to-one with the permission center point loaded in step 32.

[0132] In specific implementation, step 36 follows the aforementioned medical collaboration case. Step 35 determines that the mapped credentials should belong to region B. The system obtains the determination result, namely the target permission region identifier Region_B. This identifier uniquely corresponds to the permission center point B(6,7) and the access permission level of the anonymized medical record data it represents.

[0133] Step 37: Based on the target permission region identifier, retrieve the set of boundary points corresponding to the target permission region, specifically including:

[0134] Based on the target permission area identifier, the corresponding boundary point set is retrieved to provide the geometric data basis for the construction of the permission range polygon. The implementation process is as follows: using the target permission area identifier obtained in step 36 as the query key, the permission area metadata database generated and stored after spatial division in step 33 is retrieved. The system quickly locates the geometric definition data of the corresponding permission area according to the identifier, and reads the boundary point set that defines the spatial range of the area. The boundary point set is an ordered set containing multiple coordinate points. Connecting these points in sequence can outline the geometric contour of the target permission area.

[0135] In specific implementation, step 37 involves the system using the region identifier Region_B as an index to query the permission region geometric information database generated and stored in step 33. After successful retrieval, the system reads the boundary point set of region B: {(4,5), (5,6), (7,8), (6,9)}. This point set precisely defines the geometric range of region B in the target domain permission space.

[0136] Step 38: Using the retrieved set of boundary points, construct a closed convex polygon geometry by connecting them in spatial order, specifically including:

[0137] Step 38: Using the retrieved set of boundary points, a closed convex polygon geometry is constructed by connecting them in spatial order. This is achieved by the system receiving the set of boundary points retrieved in step 37 and reconstructing its geometry. First, based on the spatial coordinate relationships of the boundary points, spatial sorting algorithms such as polar angle sorting are used to arrange the point set into an ordered sequence that can form closed loops in space. Then, following this order, adjacent boundary points are connected sequentially with straight line segments, and the first and last points are connected, ultimately constructing a closed, simply connected convex polygon geometry. Mathematically, this convex polygon represents the geometric range within which the source domain agent is allowed to exercise its authority in the target domain.

[0138] In specific implementation, step 38 involves the system acquiring the boundary point set {(4,5), (5,6), (7,8), (6,9)} of region B, performing spatial sorting to determine the connection order as (4,5), (5,6), (7,8), (6,9), (4,5), and connecting the points in this order to construct a quadrilateral convex polygon. This polygon represents the specific geometric range of the accessible anonymized medical record data permissions within the target domain space.

[0139] Step 39 involves performing topological verification on the convex polygon geometry and binding the verified polygons, as the permission scope polygons of the source domain agent in the target domain, to the mapped permission credentials. Specifically, this includes:

[0140] Step 39 involves performing topological verification on the convex polygon geometry and binding the verified polygons to the mapped permission credentials as the permission range polygons. The implementation process involves performing strict topological consistency verification on the convex polygons constructed in Step 38, including verifying whether the polygon is a closed structure, whether all edges do not intersect (e.g., simple polygons), whether it is a convex polygon, and the directionality of the vertex sequence. Only when a valid geometric region is confirmed through all topological verifications will the system bind and associate it with the mapped permission credentials generated in Step 26. This is specifically achieved by adding geometric attribute fields to the credential data structure and storing them in the polygon structure.

[0141] In specific implementation of step 39, the system performs topological verification on the constructed quadrilateral convex polygon. After the verification is passed, this polygon structure is attached as a new attribute to the mapped permission credential of Med-Research-AI, and formally defined as the permission range polygon of the credential in the target domain. Thereafter, any operation permission verification of the agent in the hospital domain will be judged based on this quadrilateral area as the geometric boundary.

[0142] In a preferred embodiment of the present invention, step 4 involves obtaining the corresponding operation point in the target domain permission space based on the corresponding operation in the call request, and using a ray-matrix geometric inclusion detection algorithm to determine whether the operation point is located inside the permission range polygon. If the operation point is determined to be inside, the corresponding operation is allowed to be executed, and the call request is forwarded to the target domain agent for execution. This includes:

[0143] Step 41: Parse the operation description information in the call request and extract the coordinates of the corresponding operation point in the target domain permission space, specifically including:

[0144] Parsing the operation description information in the call request and extracting the operation point coordinates corresponding to the operation in the target domain permission space is the data preparation step for operation permission verification. The implementation process involves re-parsed the original call request received and stored in step 1, focusing on processing the operation description information. The system calls the operation semantic parser to identify and understand the specific operation type described in the request, and converts the abstract operation type into a geometric point with a clear coordinate value in the target domain permission space according to the operation-permission coordinate mapping rules preset in the target domain system. The operation point coordinates quantitatively represent the minimum permission level required to perform this operation or its specific location in the permission space.

[0145] In specific implementation, step 41 follows the medical collaboration example in steps 1-3. The system parses the call request issued by Med-Research-AI and identifies the requested operation as reading anonymized medical record data of diabetic patients. According to the hospital domain, i.e. the target domain, and the predefined mapping rules, the operation is converted into a fixed operation point in the permission space, with coordinates defined as (5.5, 6.5). Here, the operation risk level dimension value of 5.5 indicates that the operation has a medium to high risk level, and the accessible data range dimension value of 6.5 indicates that a large range of anonymized data needs to be accessed.

[0146] Step 42, based on the bounded permission-limited polygon structure, read its boundary vertex sequence, specifically including:

[0147] Based on the bound permission range polygon structure, reading its boundary vertex sequence is the key to constructing the geometry detection environment. The implementation process is as follows: from the mapped permission credential processed in step 39, the system obtains the bound geometric attribute information, accesses the data structure of the credential, locates and reads the convex polygon geometry representing its permission range, and then extracts the ordered coordinate sequence of the boundary vertices from the polygon structure. This vertex sequence precisely defines a closed convex polygon region, which geometrically represents the effective permission range granted to the source domain agent in the target domain.

[0148] In specific implementation, step 42 involves the system retrieving the permission scope polygon bound in step 39 from the mapped permission credentials of Med-Research-AI, i.e., the boundary of region B. The system reads the boundary vertex sequence of this polygon, for example, the coordinates of four points stored in sequence: [(4, 5), (5, 6), (7, 8), (6, 9)]. This sequence defines the geometric boundary of permission verification.

[0149] Step 43: Using the coordinates of the operation point, emit an infinitely extending detection ray along a preset direction in the target domain permission space, specifically including:

[0150] The initial action of the geometric inclusion detection algorithm is to emit an infinitely long detection ray along a preset direction in the target domain permission space using the coordinates of the operation point. This is achieved by simulating an infinitely long ray originating from the operation point to determine its positional relationship with the permission polygon. The direction of this ray is preset to be parallel to the positive direction of a certain coordinate axis in the permission space, such as horizontally to the right, to simplify the subsequent intersection calculation logic.

[0151] In specific implementation, step 43 involves the system emitting an infinitely extending detection ray horizontally to the right along the positive X-axis direction of the target domain permission space, starting from the operation point (5.5, 6.5).

[0152] Step 44: Calculate the coordinates of all intersection points based on the sequence of boundary vertices of the detected ray and the polygon within the permission range. Specifically, this includes:

[0153] The core calculation step of the ray method is to calculate the coordinates of all intersection points based on the sequence of boundary vertices of the detection ray and the polygon within the permission range. The implementation process is as follows: based on the principles of computational geometry, the polygon boundary vertex sequence obtained in step 42 is traversed, and each edge of the polygon is processed in turn. For the line segment connecting two adjacent vertices, the system determines whether the detection ray intersects with the currently processed edge. If they intersect, the precise coordinates of the intersection point are solved according to the geometric relationship. After the traversal is completed, a set of coordinates containing all valid intersection points is obtained.

[0154] In specific implementation of step 44, the system calculates the intersection point between the horizontal ray to the right and the four sides defined by the vertex sequence [(4,5), (5,6), (7,8), (6,9)]. After calculation, the ray intersects two sides of the polygon, and the coordinates of the two intersection points are obtained as (6.2, 6.5) and (7.5, 6.5).

[0155] Step 45: Based on the obtained intersection coordinates, count the number of valid intersections. If the number is odd, determine that the operation point is inside the authorized polygon; if the number is even, determine that it is outside. Specifically, this includes:

[0156] Based on the obtained intersection coordinates, the number of valid intersections is counted. If the number is odd, the operation point is determined to be inside the authorized polygon; if the number is even, it is determined to be outside. This is the decision-making step for determining the positional relationship. The implementation process is as follows: First, all intersections calculated in step 44 are validated to exclude duplicate intersections that coincide with polygon vertices or tangent cases. Then, the total number of intersections after validation is counted. Finally, the core parity determination rule of the ray method is applied: if the total number of intersections is odd, the operation point is determined to be inside the authorized polygon; if the number is even, including zero, it is determined to be outside.

[0157] In specific implementation, step 45 verifies that both intersection points obtained in step 44 are valid intersection points. The total number of intersection points is 2, which is an even number. According to the odd-even determination rule, the operation point (5.5, 6.5) is determined to be outside the permission range polygon.

[0158] Step 46: If the operation point is located inside the permission range polygon, generate an evaluation result allowing the corresponding operation to be executed, specifically including:

[0159] If the operation point is located inside the permission range polygon, generating an evaluation result that allows the corresponding operation to be executed is a key step in generating business logic based on the geometric determination result. The implementation process is that the system receives the position determination conclusion output in step 45, and only generates a positive evaluation result that allows execution when the determination result is that the operation point is located inside the permission range polygon. This evaluation result is a structured data object that includes the final decision, the basis for the decision, information such as geometric inclusion detection and timestamp.

[0160] In practice, step 46 will not generate an evaluation result that allows execution because step 45 determines that the operation point is outside the scope of permissions.

[0161] Step 47: Attach the evaluation result to the invocation request and forward it to the target domain agent to trigger the corresponding operation execution, specifically including:

[0162] The final output of the entire mutual authorization process involves attaching the evaluation result to the invocation request and forwarding it to the target domain agent to trigger the corresponding operation. This is achieved by the system encapsulating and integrating the execution permission evaluation result generated in step 46 with the original invocation request received in step 1, forming an enhanced, authorized new request message. This new request is then forwarded to the target domain agent via a secure communication channel. Based on the attached authoritative evaluation result, the target domain agent directly triggers the execution logic of the corresponding operation without requiring further complex permission verification.

[0163] In specific implementation of step 47, assuming that an operation point is determined to be located internally, the system generates an evaluation result allowing execution. The system appends this result to the original request to read medical record data and forwards it to the target domain agent Hosp-Record-DB. After verifying the validity of the evaluation result, Hosp-Record-DB executes the read operation and returns the anonymized diabetes medical record data to the source domain agent Med-Research-AI.

[0164] like Figure 2 As shown, embodiments of the present invention also provide a cross-domain intelligent agent trusted mutual recognition system, comprising:

[0165] The receiving module is used to receive call requests from the source domain agent to perform corresponding operations on the target domain agent;

[0166] The generation module is used to obtain a set of affine transformation parameters through a permission transformation model based on the spatial relationship characteristics between the source domain permission space and the target domain permission space involved in the call request, and to evaluate and verify the security using a cone volume metric. Finally, it selects affine transformation parameters suitable for permission credential mapping, and uses the affine transformation parameters to map the original permission credential of the source domain agent from the source domain permission space to the target domain permission space, generating a mapped permission credential that can be identified and verified in the target domain.

[0167] The partitioning module is used to divide the target domain permission space into multiple permission regions based on the mapped permission credential and the set of permission center points of the agent in the target domain permission space through a clustering algorithm, determine the target permission region to which the mapped permission credential belongs, and generate a polygon of the permission range of the source domain agent in the target domain based on the boundary point set of the target permission region.

[0168] The processing module is used to obtain the corresponding operation point in the target domain permission space according to the corresponding operation in the call request, and use the ray-matrix geometric inclusion detection algorithm to determine whether the operation point is located inside the permission range polygon. If the operation point is determined to be inside, the corresponding operation is allowed to be executed, and the call request is forwarded to the target domain agent for execution.

[0169] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0170] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0171] This technology (including methods, devices, and systems) only provides a technical framework for cross-domain intelligent agent permission mutual recognition and does not involve the active collection, storage, or additional processing of user privacy data. Any privacy data, such as medical records, involved in the implementation of this technology must be handled by the specific application party in strict accordance with laws, regulations, and industry privacy protection standards to ensure the legality of data sources and compliance with regulations. This technology assumes no legal responsibility for any consequences arising from the illegal acquisition or use of data.

[0172] The effectiveness of permission mapping and verification in this technology depends on the accuracy of the source / target domain permission space metadata provided by the application, the rationality of the security policy configuration, and the quality of the basic data. This technology assumes no responsibility for any permission recognition errors or operational risks caused by errors in the application's metadata, improper policy configuration, or data distortion.

[0173] This technology does not provide additional guarantees for the compliance of business processes in specific industries (such as healthcare and finance). Users need to adapt and adjust the technical solution according to their own industry regulatory requirements to ensure that the overall business process complies with industry standards. This technology shall not be liable for any problems caused by failure to adapt to industry compliance requirements.

[0174] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for trusted mutual recognition of permissions among cross-domain intelligent agents, characterized in that, The method includes: Receive a request from a source domain agent to perform a corresponding operation on a target domain agent; Based on the spatial relationship characteristics between the source domain permission space and the target domain permission space involved in the invocation request, an affine transformation parameter set is obtained through a permission transformation model. Then, using a cone volume metric for evaluation and security verification, affine transformation parameters suitable for permission credential mapping are finally selected. These affine transformation parameters are then used to map the original permission credentials of the source domain agent from the source domain permission space to the target domain permission space, generating a mapped permission credential that can be recognized and verified in the target domain, including: Based on the source and target permission spaces involved in the call request, spatial relationship features between them are extracted. These features include permission dimension mapping rules, coordinate scaling ratios, and spatial rotation angles. The extracted spatial relationship features are input into a pre-trained permission transformation model for processing, resulting in an output set of affine transformation parameters. Using the center of the target permission space as the vertex and the mapping direction defined by the affine transformation parameters as the cone axis direction, a cone covering the target permission space is generated. The actual intersection volume between this cone and the target permission space is calculated, and the ratio of this volume to the total volume of the target permission space is used as a cone volume metric to quantitatively evaluate the coverage of the target permission space by the affine transformation parameters. Based on the cone volume metric, the affine transformation parameter set is... The parameters in the set are sorted in descending order, and a subset of parameters with metric values ​​higher than a preset threshold are selected. This subset of parameters is then subjected to compatibility and risk verification according to the security policy of the target domain's permission space to obtain the verified parameters. The verified parameters are used as the final affine transformation parameters for permission credential mapping. The verified final affine transformation parameters are read and parsed into translation vectors, scaling matrices, and rotation matrices for coordinate transformation. Based on the translation vectors, scaling matrices, and rotation matrices, the geometric coordinates of the original permission credential are sequentially scaled, rotated, and translated to obtain new coordinates in the target domain's permission space. The obtained new coordinates are encapsulated according to the format of the target domain's permission credential and a spatial location verification label is attached to generate a mapped permission credential that can be recognized and verified by the target domain. Based on the mapped permission credentials, and combined with the set of permission center points of the agents in the target domain permission space, the target domain permission space is divided into multiple permission regions by a clustering algorithm. The target permission region to which the mapped permission credentials belong is determined, and the permission range polygon of the source domain agent in the target domain is generated based on the boundary point set of the target permission region. Based on the corresponding operation in the call request, obtain the corresponding operation point in the target domain permission space. Based on the permission range polygon, use the ray-matrix geometric inclusion detection algorithm to determine whether the operation point is located inside the permission range polygon. If the operation point is determined to be inside, allow the execution of the corresponding operation and forward the call request to the target domain agent for execution.

2. The method for trusted mutual recognition of permissions among cross-domain intelligent agents according to claim 1, characterized in that, The invocation request contains the original permission credentials of the source domain agent. The original permission credentials are represented as geometric coordinate points in the source domain permission space and are used for affine transformation and cross-domain mapping.

3. The method for trusted mutual recognition of permissions among cross-domain intelligent agents according to claim 2, characterized in that, Based on the mapped permission credentials, and combined with the set of permission center points of agents in the target domain permission space, the target domain permission space is divided into multiple permission regions using a clustering algorithm, including: Obtain the coordinates of the mapped permission credential in the target domain permission space; Based on the coordinate positions, load a predefined set of permission center points for the target domain, where each permission center point corresponds to a permission level; Using the loaded set of permission center points as cluster centers, a distance metric algorithm is used to dynamically divide the target domain permission space, generating multiple convex polygon permission regions defined by boundary point sets; Based on the divided permission areas, calculate the spatial distance between the coordinates of the mapped permission credential and the center point of each permission; Based on the calculated spatial distance, the mapped permission credentials are assigned to the target permission area corresponding to the permission center point with the smallest distance.

4. The method for trusted mutual recognition of permissions among cross-domain intelligent agents according to claim 3, characterized in that, Determine the target permission region to which the mapped permission credential belongs, and generate a polygon representing the permission range of the source domain agent in the target domain based on the boundary point set of the target permission region, including: Obtain the target permission region identifier corresponding to the mapped permission credential; Based on the target permission region identifier, retrieve the set of boundary points corresponding to the target permission region; Using the retrieved set of boundary points, a closed convex polygon geometry is constructed by connecting them in spatial order. A topological verification is performed on the convex polygon geometry, and the polygon that passes the verification is used as the permission scope polygon of the source domain agent in the target domain and bound to the mapped permission credential.

5. The method for trusted mutual recognition of permissions among cross-domain intelligent agents according to claim 4, characterized in that, Based on the corresponding operation in the call request, obtain its corresponding operation point in the target domain permission space. Based on the permission range polygon, use a ray-mapping geometric inclusion detection algorithm to determine whether the operation point is located inside the permission range polygon. If the operation point is determined to be inside, allow the execution of the corresponding operation, and forward the call request to the target domain agent for execution, including: Parse the operation description information in the call request and extract the coordinates of the corresponding operation point in the target domain permission space; Based on the bounded permission-scoped polygon structure, read its boundary vertex sequence; Using the coordinates of the operation point, an infinitely extending detection ray is emitted along a preset direction in the target domain permission space; Calculate the coordinates of all intersection points based on the sequence of boundary vertices of the detected ray and the polygon within the permission range; Based on the obtained intersection coordinates, count the number of valid intersections. If the number is odd, the operation point is determined to be inside the permission range polygon; if the number is even, it is determined to be outside. If the operation point is located inside the permission scope polygon, an evaluation result is generated that allows the corresponding operation to be executed; The evaluation results are appended to the invocation request and forwarded to the target domain agent to trigger the corresponding operation.

6. A cross-domain intelligent agent authorization trust mutual recognition system, wherein the system implements the method as described in any one of claims 1 to 5, characterized in that, include: The receiving module is used to receive call requests from the source domain agent to perform corresponding operations on the target domain agent; The generation module is used to obtain a set of affine transformation parameters through a permission transformation model based on the spatial relationship characteristics between the source domain permission space and the target domain permission space involved in the call request, and to evaluate and verify the security using a cone volume metric. Finally, it selects affine transformation parameters suitable for permission credential mapping, and uses the affine transformation parameters to map the original permission credential of the source domain agent from the source domain permission space to the target domain permission space, generating a mapped permission credential that can be identified and verified in the target domain. The partitioning module is used to divide the target domain permission space into multiple permission regions based on the mapped permission credential and the set of permission center points of the agent in the target domain permission space through a clustering algorithm, determine the target permission region to which the mapped permission credential belongs, and generate a polygon of the permission range of the source domain agent in the target domain based on the boundary point set of the target permission region. The processing module is used to obtain the corresponding operation point in the target domain permission space according to the corresponding operation in the call request, and use the ray-matrix geometric inclusion detection algorithm to determine whether the operation point is located inside the permission range polygon. If the operation point is determined to be inside, the corresponding operation is allowed to be executed, and the call request is forwarded to the target domain agent for execution.

7. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images