Behavior-aware dynamic access control and permission management methods
By using a kernel-level event-driven architecture and behavioral manifold analysis, access behavior is collected and predicted in real time, solving the response latency and adaptability problems of dynamic access control in existing technologies, and achieving more efficient permission management and security response.
Patent Information
- Application Number
- CN202611364893.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-04
- Publication Date
- 2026-11-17
AI Technical Summary
Existing behavior-aware dynamic access control solutions suffer from problems such as security response delays caused by reactive defense logic, difficulty in distinguishing between legitimate and malicious behavior, and the inability of static behavior baselines to adapt to business changes.
It adopts a kernel-level event-driven architecture to collect access behavior events in real time, generates high-dimensional feature vectors through sliding window serialization, embeds them into a low-dimensional behavior manifold space, calculates the direction and rate of behavior evolution, and realizes predictive hierarchical permission adjustment and online incremental update.
It enhances the proactive defense against progressive attacks, reduces false interception of legitimate business operations, adapts to business changes, balances security and availability, and shortens security response latency.
Smart Images

Figure CN123122943A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network security access control technology, specifically to a behavior-aware dynamic access control and permission management method. Background Technology
[0002] Access control is a core foundational technology in network security and data protection systems. Its core function is to block unauthorized access to resources while ensuring that legitimate entities have the right to access resources, thereby achieving controlled sharing and secure isolation of system resources and sensitive data. Traditional access control systems are mostly based on static models such as discretionary access control, mandatory access control, and role-based access control. They make access decisions by mapping the pre-defined subject identity, role permissions, and object security levels, and are widely used in environments with relatively fixed business architectures and simple access scenarios.
[0003] As enterprises continue to advance digital transformation, cloud-native architectures become widespread, remote work becomes the norm, and the zero-trust security concept is gradually being implemented. Business access scenarios are exhibiting characteristics of high dynamism, multi-terminal access, and blurred resource boundaries. Static permission configuration models are no longer suitable for complex and ever-changing operating environments. Against this backdrop, dynamic access control technology has gradually become a key area of research and application in the industry. Its core idea is to break through the limitations of static permission mapping and adjust access policies in real time by combining various dynamic information during the access process, thereby achieving elastic control of permissions. Existing behavior-aware dynamic access control schemes primarily achieve dynamic permission control by periodically collecting snapshots of behavioral elements and matching them with preset security rule sets to meet the resource access security management needs of different business scenarios. However, these schemes have certain shortcomings. First, they often employ reactive defense logic, tightening permissions only after an abnormal rule is triggered by the behavior state. This results in a significant delay in security response to progressive, covert attacks, limiting proactive defense capabilities. Second, relying solely on single-moment behavior snapshots for rule matching fails to distinguish between legitimate temporary business expansions and malicious progressive attacks, making it difficult to balance access control security and business availability. Third, the commonly used periodic polling approach for behavior collection presents a trade-off between sampling accuracy and system performance, and static behavior baselines are difficult to adapt to dynamic changes in business behavior patterns. Therefore, we propose a behavior-aware dynamic access control and permission management method. Summary of the Invention
[0004] The purpose of this invention is to provide a behavior-aware dynamic access control and permission management method.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a behavior-aware dynamic access control and permission management method, the method comprising the following steps: Step 1: Perform initial configuration and normal behavior baseline construction, collect the subject's historical legitimate access behavior dataset, construct a low-dimensional behavior manifold space and normal behavior region boundary, and initialize the permission level mapping table and three-level risk threshold benchmark; Step 2: Use a kernel-level event-driven architecture to collect access behavior events in real time, abandoning the traditional sampling mode of periodic daemon process polling. Capture each resource access operation in real time and extract multi-dimensional attributes, splicing them into a continuous behavior event stream. Step 3: Serialize and map the continuous behavioral trajectories into feature vectors. After sliding window serialization, a high-dimensional feature vector containing static attributes and statistical features is generated, forming a sequence of behavioral trajectory points arranged in time sequence. Step 4: Embed the high-dimensional feature vector into the low-dimensional behavior manifold space, fit and calculate the behavior evolution direction vector and evolution rate, and quantify the evolution trend and the degree of deviation of the behavior. Step 5: Based on the evolution direction vector, evolution rate, and distance from the behavior point to the boundary, calculate the comprehensive risk deviation by weighting, and classify the risk level into three categories: early warning level, pre-tightening level, and emergency blocking level according to the preset three-level threshold. Step 6: Implement predictive hierarchical dynamic adjustment of permissions and implement differentiated control strategies according to risk levels. Among them, the pre-tightening level will proactively restrict the non-essential access permissions of the subject before the behavior exceeds the boundary of the normal behavior area. Step 7: Implement a gradual restoration of permissions and oscillation suppression measures. As the risk continues to decline, restore permissions in stages and set an observation window to avoid frequent fluctuations in permission status. Step 8: Periodically perform online incremental updates to the behavioral baseline and manifold model, and use hot loading to complete the model and threshold calibration without interrupting the normal operation of the real-time access control service.
[0006] As a further aspect of the present invention: In step one, a dataset of historical legal access behaviors of each subject within a preset period is collected, multi-dimensional basic features of access operations are extracted, and an initial low-dimensional behavior manifold space is constructed through offline manifold learning training. The central reference point and closed boundary of the normal behavior area are marked in the manifold space. At the same time, the permission level mapping table is initialized, and three types of permission sets are pre-divided: full permission set, non-essential permission set, and core permission set. The permission configuration templates and risk threshold benchmarks corresponding to the warning level, pre-tightening level, and emergency blocking level are defined respectively, and the loading and verification of the initial parameters of the system are completed.
[0007] As a further aspect of the present invention: In step two, an event-driven architecture using kernel-mode eBPF hooks and / or system call interception is adopted to capture each resource access operation initiated by the subject in real time; for each captured access operation event, six basic attributes corresponding to the event are extracted synchronously, including the subject identity identifier, object resource type, operation execution type, access timestamp, resource sensitivity level, and context environment state, to generate a structured behavioral event with a precise timestamp, which is then spliced together in chronological order to form a continuous behavioral event stream of the subject, thus eliminating the time monitoring blind spot caused by discrete sampling from the data source.
[0008] As a further aspect of the present invention: In step three, the continuous behavioral event stream is subjected to sliding window serialization processing. Using a fixed-duration time window as the unit, all behavioral events within the window are mapped to a high-dimensional feature vector of a unified dimension. The high-dimensional feature vector includes not only the static attribute features of a single event, but also four types of statistical features within the window: access frequency, resource access dispersion, operation sequence entropy value, and sensitive resource ratio. This transforms the continuous access behavior of the subject into a sequence of behavioral trajectory points arranged in chronological order in a high-dimensional feature space.
[0009] As a further aspect of the present invention: In step four, the high-dimensional feature vector corresponding to the current time window is mapped to a low-dimensional behavioral manifold space through a pre-trained manifold embedding algorithm to obtain the coordinates of the current behavioral point in the manifold space; the current behavioral point and N consecutive previous historical behavioral points are selected, and the tangent vector of the behavioral evolution is fitted based on the geodesic distance of the manifold space as the evolution direction vector, and the evolution rate is calculated in combination with the time window length. The core calculation formula for the evolution direction vector is: ; in, for The behavioral evolution direction vector at time step, The current time is calculated for the evolution vector; The total number of historical behavior points involved in the evolution vector calculation. The index variable is used for the summation operation. for The coordinates of the action point in the manifold space at any given moment. Points on the manifold The inverse exponential mapping at a given point is used to calculate the geodesic tangent vector between two points. The time interval between adjacent behavioral points; the evolution direction vector output by this formula represents the evolution trend of the behavior, and the evolution rate represents the severity of the deviation of the behavior.
[0010] As a further aspect of the present invention: In step five, the comprehensive risk deviation of the behavior is calculated by weighting the direction of the evolution direction vector, the magnitude of the evolution rate, and the geodesic distance from the current behavior point to the boundary of the normal behavior area; according to the preset three-level threshold range, the current behavior state is divided into corresponding risk levels: when the comprehensive risk deviation is lower than the first threshold, it is determined to be a normal state; when it is between the first threshold and the second threshold, it is determined to be a warning level risk; when it is between the second threshold and the third threshold, it is determined to be a pre-tightening level risk; and when it is higher than the third threshold, it is determined to be an emergency blocking level risk.
[0011] As a further aspect of the present invention: In step six, based on the determined risk level, a differentiated permission adjustment strategy is executed accordingly: Under the warning-level risk, the current permission set remains unchanged, the granularity of operation audit is improved, and all operation context logs and behavior trajectory data are recorded; Under the pre-tightening-level risk, under the premise that the behavior has not exceeded the boundary of the normal behavior area, the set of unnecessary permissions is proactively revoked, batch file access operations and access permissions to highly sensitive resources are restricted, and a defense buffer zone is built in advance; Under the emergency blocking-level risk, all non-core access permissions of the subject are immediately frozen, real-time security alarms are triggered and pushed to the security management platform.
[0012] As a further aspect of the present invention: In step seven, when the comprehensive risk deviation over M consecutive time windows continues to fall below the corresponding threshold and the evolution direction vector points to the center of the normal behavior area, a phased permission restoration mechanism is executed; the permission set is restored step by step in the order of low-sensitivity resource permissions, general operation permissions, non-essential permissions, and full permissions, and a fixed observation window period is set before each level of permission is restored. If the risk deviation remains stable during the observation period, the restoration of this level of permission is completed; if a risk rebound occurs, it is immediately reverted to the previous permission level.
[0013] As a further aspect of the present invention: In step eight, a fixed model update cycle is set, and in each cycle, all legal behavior trajectory data of all subjects confirmed by audit are summarized. The behavior manifold spatial structure and the boundary of the normal behavior area are fine-tuned and updated using an incremental learning method. During the update process, the core topology structure of the historical baseline is retained, and only the newly added legal behavior patterns are adapted. At the same time, the risk thresholds and permission configuration templates at all levels are calibrated simultaneously. The update process adopts a hot loading method so as not to interrupt the normal operation of the real-time access control service.
[0014] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows: 1. This invention uses a control mechanism that combines behavioral evolution trend prediction with hierarchical permission pre-adjustment to advance the response time of access control from after abnormal behavior occurs to the stage when abnormal behavior shows an abnormal evolution trend. When the behavior does not exceed the normal boundary, unnecessary permissions can be proactively reduced to build a front-end defense buffer, effectively shorten the security response time, and improve the system's proactive defense effect against covert attacks such as progressive permission abuse. 2. This invention distinguishes between convergent and divergent evolution patterns of behavior through comprehensive analysis of the direction and rate of behavioral evolution. It breaks through the decision-making limitations of single-moment state matching and can more accurately identify the deep intent of behavior. While ensuring the strength of security protection, it reduces false blocking of legitimate temporary business operations and helps to balance the security of access control with business availability. 3. This invention, through a kernel-level event-driven acquisition architecture and an online incremental update mechanism, abandons the sampling mode of periodic idle polling. While improving the continuity of behavior perception, it controls additional performance overhead. At the same time, it can dynamically adjust the behavior baseline and judgment threshold according to changes in business behavior patterns, thereby improving the adaptability of the solution in different business scenarios and its long-term operational stability. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the method steps in an embodiment of the present invention; Figure 2 This is a schematic diagram of the running system framework process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the enterprise intranet deployment architecture in an embodiment of the present invention. Detailed Implementation
[0016] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.
[0017] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0018] Please see the appendix Figure 1 -Appendix Figure 3 This invention relates to a behavior-aware dynamic access control and permission management method, which includes the following steps: Step 1: Perform initial configuration and normal behavior baseline construction, collect the subject's historical legitimate access behavior dataset, construct a low-dimensional behavior manifold space and normal behavior region boundary, and initialize the permission level mapping table and three-level risk threshold benchmark; Step 2: Use a kernel-level event-driven architecture to collect access behavior events in real time, abandoning the traditional sampling mode of periodic daemon process polling. Capture each resource access operation in real time and extract multi-dimensional attributes, splicing them into a continuous behavior event stream. Step 3: Serialize and map the continuous behavioral trajectories into feature vectors. After sliding window serialization, a high-dimensional feature vector containing static attributes and statistical features is generated, forming a sequence of behavioral trajectory points arranged in time sequence. Step 4: Embed the high-dimensional feature vector into the low-dimensional behavior manifold space, fit and calculate the behavior evolution direction vector and evolution rate, and quantify the evolution trend and the degree of deviation of the behavior. Step 5: Based on the evolution direction vector, evolution rate, and distance from the behavior point to the boundary, calculate the comprehensive risk deviation by weighting, and classify the risk level into three categories: early warning level, pre-tightening level, and emergency blocking level according to the preset three-level threshold. Step 6: Implement predictive hierarchical dynamic adjustment of permissions and implement differentiated control strategies according to risk levels. Among them, the pre-tightening level will proactively restrict the non-essential access permissions of the subject before the behavior exceeds the boundary of the normal behavior area. Step 7: Implement a gradual restoration of permissions and oscillation suppression measures. As the risk continues to decline, restore permissions in stages and set an observation window to avoid frequent fluctuations in permission status. Step 8: Periodically perform online incremental updates to the behavioral baseline and manifold model, and use hot loading to complete the model and threshold calibration without interrupting the normal operation of the real-time access control service.
[0019] Example 1
[0020] This embodiment applies to access control scenarios for enterprise intranet office terminals, implementing dynamic permission management for internal employees' file access, system login, and peripheral device usage. The specific implementation process is as follows: 1. Initial configuration and establishment of normal behavior baseline In the initial stage of system deployment, a dataset of the past 30 days of legal access behavior of all employees within the enterprise was collected, and six basic features were extracted for each access operation: subject identity, object resource type, operation type, timestamp, resource sensitivity level, and environmental status.
[0021] The offline isometric mapping (ISOMAP) manifold learning algorithm is used to embed high-dimensional behavioral features into a 5-dimensional low-dimensional behavioral manifold space, preserving the inherent geometric structure of the behavioral data. Based on historical legal behavioral data, the closed boundary of the normal behavior region is calibrated with a manifold envelope with 95% confidence, and the central reference point of the normal behavior region is calculated.
[0022] Simultaneously, initialize the permission hierarchy mapping table: pre-divide three permission sets: full permission set, non-essential permission set, and core permission set. The full permission set includes five types of permissions: basic office file read / write, internal system access, peripheral device read / write, batch file export, and cross-departmental sensitive file access. The non-essential permission set includes three types of permissions: batch file export, cross-departmental sensitive file access, and peripheral device write. The core permission set includes two types of permissions: basic office file read and internal system login.
[0023] Define a three-tiered risk threshold benchmark: Early warning level threshold Pre-tightening threshold Emergency blocking threshold Complete the loading and verification of all initial parameters.
[0024] 2. Real-time collection of kernel-level event-driven access behavior It adopts an event-driven architecture with kernel-mode eBPF hooks, deployed at the kernel layer of the terminal operating system, to capture every resource access operation initiated by the employee in real time without the need for periodic daemon polling.
[0025] For each access operation event captured, six basic attributes corresponding to the event are extracted synchronously: subject identity (employee ID), object resource type (office document / system data / peripheral resource), operation execution type (read / write / modify / export), access timestamp (accurate to milliseconds), resource sensitivity level (public / internal / secret / confidential), and context environment status (internal network IP / office hours / device security level). A structured behavioral event with a precise timestamp is generated and concatenated in chronological order to form a continuous behavioral event stream for a single subject, eliminating the time monitoring blind spots of discrete sampling.
[0026] 3. Behavioral Trajectory Serialization and Feature Vector Mapping A sliding window serialization process is used for the continuous behavioral event stream. The sliding window duration is set to 30 seconds and the sliding step size is 5 seconds, that is, a new behavioral feature window is generated every 5 seconds.
[0027] For all behavioral events within each time window, a high-dimensional feature vector of uniform dimension is mapped. The feature vector has 12 dimensions and contains two types of features: Static attribute features: subject role encoding, object resource type proportion, operation type proportion, average resource sensitivity level, and environmental state encoding, totaling 6 dimensions; Statistical characteristics: frequency of access within the window, resource access dispersion (the proportion of accesses to different resources), entropy of operation sequence, proportion of accesses to sensitive resources, and variance of consecutive operation intervals, totaling 6 dimensions.
[0028] Through the above mapping, the subject's continuous access behavior is transformed into a sequence of behavioral trajectory points arranged in time order in a high-dimensional feature space, with the time interval between adjacent trajectory points... .
[0029] 4. Behavioral manifold embedding and evolution vector computation The 12-dimensional high-dimensional feature vector corresponding to the current time window is mapped to a 5-dimensional low-dimensional behavior manifold space using a pre-trained ISOMAP manifold embedding algorithm to obtain the current behavior point. Coordinates in the manifold space.
[0030] Select the current action point and its preceding continuous sequence. For each historical behavior point, the tangent vector of the behavior evolution is fitted based on the geodesic distance in the manifold space, serving as the evolution direction vector. The evolution rate is calculated by combining this vector with the time window length. The core formula for calculating the evolution direction vector is: ; In the formula, For the current moment The corresponding behavioral evolution direction vector; The current time is calculated for the evolution vector; In this embodiment, the total number of historical behavior points participating in the evolution vector calculation is set to 6. The time window interval between two adjacent action points is 5 seconds in this embodiment; The index variable for the summation operation takes values from 0 to... Integers; For a moment The coordinates of the corresponding action point in the low-dimensional action manifold space; For points on the manifold The inverse exponential mapping operator with a base point is used to map adjacent points on a manifold to tangent vectors in the tangent space of that base point, representing the geodesic displacement direction between the two points; For a moment The coordinates of the corresponding behavioral point in the low-dimensional behavioral manifold space.
[0031] Evolution rate The magnitude of the evolution direction vector represents the degree to which the behavior deviates drastically from the normal region.
[0032] 5. Calculation and rating of comprehensive risk deviation The overall risk deviation of the behavior is calculated by weighting the direction of the evolutionary direction vector, the magnitude of the evolution rate, and the geodesic distance from the current behavior point to the boundary of the normal behavior area. The calculation formula is: ; In the formula, The cosine of the angle between the evolution direction and the boundary normal direction, with a value range of [0,1]. The closer the direction is to the boundary, the larger the value becomes. The normalized evolution rate, with a value range of [0,1]; This is the normalized geodesic distance from the current point to the boundary, with a value range of [0,1], where the value is larger as the distance is closer; , , For the weighting coefficients, in this embodiment, they are taken as 0.4, 0.3, and 0.3 respectively, and satisfy the following conditions: .
[0033] Risk levels are determined based on a pre-defined three-tiered threshold range: When the overall risk deviation When this occurs, it is considered a normal state; when At that time, it was determined to be a warning-level risk; when At that time, it was determined to be a pre-tightening level risk; when At that time, it was determined to be an emergency blocking level risk.
[0034] 6. Predictive hierarchical permission dynamic adjustment Based on the determined risk level, corresponding differentiated permission adjustment strategies will be implemented: Warning-level risk: Keep the current permission set unchanged, improve the granularity of operation audit, record all operation context logs and behavior trajectory data, and extend the log storage period from the default 7 days to 30 days; Pre-tightening risk level: Under the premise that the behavior has not exceeded the boundary of the normal behavior area, proactively withdraw the set of unnecessary permissions, that is, close the three types of permissions: batch file export, cross-departmental access to sensitive files, and external device write, and retain only core permissions and basic office permissions, thus building a defense buffer in advance; Emergency blocking level risk: Immediately freeze all non-core access permissions of the entity, retain only basic system login permissions, and trigger real-time security alerts, which will be pushed to the enterprise security management platform for security operations personnel to investigate.
[0035] 7. Gradual recovery of permissions and oscillation suppression When continuous When the overall risk deviation of a time window continues to fall below the corresponding threshold and the evolution direction vector points to the center of the normal behavior area, the phased permission restoration mechanism is triggered.
[0036] The restoration process is carried out in four stages: "low-sensitivity resource permissions → general operation permissions → non-essential permissions → full permissions". A two-minute observation window is set before each level of permissions is restored: if the risk deviation remains stable and does not rebound during the observation period, the restoration of this level of permissions is completed; if the risk deviation rises above the current level threshold during the observation period, the permissions are immediately reverted to the previous level to avoid frequent fluctuations in permission status that may affect business continuity.
[0037] 8. Online incremental updates of behavioral baselines The model update cycle is set to 24 hours. During the off-peak business hours of each cycle (such as 2:00-3:00 am), the legal behavior trajectory data of all entities confirmed by audit are aggregated, and the behavior manifold spatial structure and the boundary of the normal behavior area are fine-tuned and updated using the incremental ISOMAP learning method.
[0038] The update process retains the core topology of the historical baseline, adapts only the newly added legal behavior patterns, and simultaneously calibrates risk thresholds and permission configuration templates at all levels. The update adopts a hot-loading method, so that the normal operation of the real-time access control service is not interrupted during the loading of the new model, ensuring that the business is unaware of the changes.
[0039] Example 2
[0040] This embodiment is applied to the access control scenario of cloud platform container clusters. It performs dynamic permission control for operations such as system calls, cross-service API requests, and sensitive configuration reading of container instances. The overall method and process are consistent with Embodiment 1, but are adapted and adjusted for the high dynamism of cloud-native scenarios, as follows: In this embodiment, the behavior event collection adopts a kernel-mode system call interception mechanism to capture events for the process of each container. The collected behavior attributes include six categories: container ID, namespace, system call type, target service requested, resource sensitivity level, and cluster node status.
[0041] The sliding window duration is set to 10 seconds, the sliding step size is 2 seconds, and the number of historical behavior points involved in the evolution vector calculation is [not specified]. It adapts to the characteristics of container access behavior that is frequent and changes rapidly.
[0042] The permission set is divided into: the full permission set includes 5 categories: basic system calls, cross-service API calls, sensitive configuration reading, resource quota expansion, and host machine resource access; the non-essential permission set includes 3 categories: resource quota expansion, host machine resource access, and cross-namespace calls; and the core permission set includes 2 categories: basic system calls and API calls within this service.
[0043] The three risk thresholds are set as follows: It meets the high security requirements of cloud environments.
[0044] The model update cycle is set to 6 hours, and incremental manifold learning is used to quickly adapt to the dynamic scaling of container instances and business changes, ensuring the accuracy of access control in dynamic environments.
[0045] Summary of Implementation Results Compared to traditional periodic snapshot-based dynamic access control schemes (sampling period of 30 seconds), this solution, through event-driven data collection and evolution trend prediction, increases the lead time for security responses to progressive privilege abuse attacks by over 70%, significantly compressing the attack window. By differentiating behavioral intent through evolution patterns, it reduces the false interception rate of legitimate temporary business operations by approximately 45%, effectively balancing security and availability. At the same time, the kernel-level event-driven architecture keeps the system's additional CPU overhead to within 3%, and combined with the incremental update mechanism, it can adapt to business changes over the long term, demonstrating significant advantages in three dimensions: security protection effectiveness, business availability, and system performance.
[0046] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for dynamic access control and permission management based on behavior perception, characterized in that, The method includes the following steps: Step 1: Perform initial configuration and normal behavior baseline construction, collect the subject's historical legitimate access behavior dataset, construct a low-dimensional behavior manifold space and normal behavior region boundary, and initialize the permission level mapping table and three-level risk threshold benchmark; Step 2: Use a kernel-level event-driven architecture to collect access behavior events in real time, abandoning the traditional sampling mode of periodic daemon process polling. Capture each resource access operation in real time and extract multi-dimensional attributes, splicing them into a continuous behavior event stream. Step 3: Serialize and map the continuous behavioral trajectories into feature vectors. After sliding window serialization, a high-dimensional feature vector containing static attributes and statistical features is generated, forming a sequence of behavioral trajectory points arranged in time sequence. Step 4: Embed the high-dimensional feature vector into the low-dimensional behavior manifold space, fit and calculate the behavior evolution direction vector and evolution rate, and quantify the evolution trend and the degree of deviation of the behavior. Step 5: Based on the evolution direction vector, evolution rate, and distance from the behavior point to the boundary, calculate the comprehensive risk deviation by weighting, and classify the risk level into three categories: early warning level, pre-tightening level, and emergency blocking level according to the preset three-level threshold. Step 6: Implement predictive hierarchical dynamic adjustment of permissions and implement differentiated control strategies according to risk levels. Among them, the pre-tightening level will proactively restrict the non-essential access permissions of the subject before the behavior exceeds the boundary of the normal behavior area. Step 7: Implement a gradual restoration of permissions and oscillation suppression measures. As the risk continues to decline, restore permissions in stages and set an observation window to avoid frequent fluctuations in permission status. Step 8: Periodically perform online incremental updates to the behavioral baseline and manifold model, and use hot loading to complete the model and threshold calibration without interrupting the normal operation of the real-time access control service.
2. The method for dynamic access control and permission management based on behavior perception according to claim 1, characterized in that, In step one, a dataset of historical legal access behaviors of each subject within a preset period is collected, multidimensional basic features of access operations are extracted, and an initial low-dimensional behavior manifold space is constructed through offline manifold learning training. The central reference point and closed boundary of the normal behavior region are marked in the manifold space. Simultaneously, initialize the permission hierarchy mapping table, pre-divide three permission sets: full permission set, non-essential permission set, and core permission set, define the permission configuration templates and risk threshold benchmarks corresponding to the warning level, pre-tightening level, and emergency blocking level, and complete the loading and verification of the system's initial parameters.
3. The method for dynamic access control and privilege management based on behavior perception according to claim 1, characterized in that, In step two, an event-driven architecture using kernel-mode eBPF hooks and / or system call interception is employed to capture every resource access operation initiated by the subject in real time. For each captured access operation event, six basic attributes are extracted synchronously: subject identity, object resource type, operation execution type, access timestamp, resource sensitivity level, and context environment state. This generates a structured behavioral event with a precise timestamp, which is then concatenated in chronological order to form a continuous behavioral event stream for the subject, eliminating the time monitoring blind spots caused by discrete sampling from the data source.
4. The method for dynamic access control and permission management based on behavior perception according to claim 1, characterized in that, In step three, the continuous stream of behavioral events is processed by sliding window serialization. Using a fixed-duration time window as the unit, all behavioral events within the window are mapped to a high-dimensional feature vector of the same dimension. The high-dimensional feature vector includes not only the static attribute features of a single event, but also four types of statistical features: access frequency within the window, resource access dispersion, operation sequence entropy, and sensitive resource ratio. This transforms the subject's continuous access behavior into a sequence of behavioral trajectory points arranged in chronological order in the high-dimensional feature space.
5. The method for dynamic access control and privilege management based on behavior perception according to claim 1, characterized in that, In step four, the high-dimensional feature vector corresponding to the current time window is mapped to the low-dimensional behavior manifold space using a pre-trained manifold embedding algorithm to obtain the coordinates of the current behavior point in the manifold space. The current behavior point and N consecutive previous historical behavior points are selected, and the tangent vector of the behavior evolution is fitted based on the geodesic distance in the manifold space as the evolution direction vector. The evolution rate is then calculated in conjunction with the time window length. The core formula for calculating the evolution direction vector is: ; in, for The behavioral evolution direction vector at time step, The current time is calculated for the evolution vector; The total number of historical behavior points involved in the evolution vector calculation. The index variable is used for the summation operation. for The coordinates of the action point in the manifold space at any given moment. Points on the manifold The inverse exponential mapping at a given point is used to calculate the geodesic tangent vector between two points. The time interval between adjacent rows.
6. The behavior-aware dynamic access control and permission management method according to claim 1, characterized in that, In step five, the comprehensive risk deviation of the behavior is calculated by weighting the direction of the evolution direction vector, the magnitude of the evolution rate, and the geodesic distance from the current behavior point to the boundary of the normal behavior area. Based on the preset three-level threshold range, the current behavior status is divided into corresponding risk levels: when the comprehensive risk deviation is lower than the first threshold, it is judged as a normal state; when it is between the first and second thresholds, it is judged as a warning level risk; when it is between the second and third thresholds, it is judged as a pre-tightening level risk; and when it is higher than the third threshold, it is judged as an emergency blocking level risk.
7. The behavior-aware dynamic access control and permission management method according to claim 1, characterized in that, In step six, a differentiated permission adjustment strategy is executed according to the determined risk level: under the warning level risk, the current permission set remains unchanged, the granularity of operation audit is improved, and all operation context logs and behavior trajectory data are recorded; Under the pre-tightening risk level, before the behavior exceeds the boundary of the normal behavior area, the set of unnecessary permissions is proactively revoked, batch file access operations and access permissions to highly sensitive resources are restricted, and a defense buffer zone is built in advance. In the event of an emergency blocking level risk, immediately freeze all non-core access permissions of the entity, trigger a real-time security alert, and push it to the security management platform.
8. The behavior-aware dynamic access control and permission management method according to claim 1, characterized in that, In step seven, when the overall risk deviation over M consecutive time windows continues to fall below the corresponding threshold and the evolution direction vector points to the center of the normal behavior area, a phased permission restoration mechanism is executed. The permission set is restored step by step in the order of low-sensitivity resource permissions, general operation permissions, non-essential permissions, and full permissions. A fixed observation window period is set before each level of permission is restored. If the risk deviation remains stable during the observation period, the restoration of the current level of permission is completed. If a risk rebound occurs, the permission level is immediately reverted to the previous level.
9. The behavior-aware dynamic access control and permission management method according to claim 1, characterized in that, In step eight, a fixed model update cycle is set, and in each cycle, all legal behavioral trajectory data of all subjects that have been audited and confirmed are summarized. Incremental learning is used to fine-tune and update the behavioral manifold spatial structure and the boundary of the normal behavior area. The update process retains the core topology of the historical baseline, adapts only the newly added legal behavior patterns, and simultaneously calibrates risk thresholds and permission configuration templates at all levels. Furthermore, the update process adopts a hot-reload method to ensure that the normal operation of the real-time access control service is not interrupted.
