A Closed-Loop Optimization Method Based on Behavioral Fingerprint Dynamic Evolution of Environment Decoupling and Zero-Trust Micro-Issuance Strategy
By decomposing the policy execution effect data, quantifying environmental changes, and separating the environmental contribution value, the problem of feedback signal distortion caused by the coupling of environmental factors and policy factors in zero-trust micro-segmentation systems is solved, thereby improving the system's policy adaptation capability and stability in complex network environments.
Patent Information
- Application Number
- CN202611265275.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-17
AI Technical Summary
In existing zero-trust micro-segmentation systems, environmental and policy factors are coupled in the policy execution effect data, leading to distorted feedback signals. This causes the evaluation model to adjust parameters incorrectly based on the distorted feedback, severely limiting the system's protection effectiveness in complex and dynamic network environments.
By collecting multidimensional behavioral data of the target entity, a short-cycle behavioral fingerprint vector is generated. Combined with dynamic trust assessment parameters, a comprehensive trust score is calculated. The policy execution effect data is decomposed into a first effect component associated with the entity's behavior and a second effect component associated with the external environment state. An environmental state encoder is used to quantify environmental changes, extract the environmental contribution value, and generate a feedback correction signal to correct the behavioral fingerprint generation parameters and trust score.
It achieves consistency between the policy optimization direction and the actual effect in complex dynamic network environments, improves the policy adaptation capability and stability of the zero-trust micro-segmentation system, and avoids the interference of environmental fluctuations on feedback signals.
Smart Images

Figure CN123122899A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network security, and in particular to a closed-loop optimization method for zero-trust micro-segmentation strategy based on dynamic evolution of behavioral fingerprints in environmental decoupling. Background Technology
[0002] Zero Trust architecture, as an emerging cybersecurity concept, is based on the core principle of "never trust, always verify," requiring continuous authentication and authorization assessment of all access requests within the network, completely breaking away from traditional security models based on network boundaries. In the practical implementation of Zero Trust architecture, micro-segmentation technology is widely considered a key means of achieving fine-grained access control. By dividing the network into extremely fine logical units and finely controlling the communication traffic between each unit, micro-segmentation effectively curbs lateral movement of attackers within the network. Currently, mainstream zero-trust micro-segmentation systems typically continuously collect user or device behavioral data to build behavioral baselines or profiles, and calculate dynamic trust scores based on the deviation of current behavior from the baseline. The decision engine then dynamically generates corresponding access control policies based on these trust scores, thereby achieving real-time adjustment of access permissions.
[0003] In existing technologies, such as the scheme disclosed in Chinese Patent CN122247725A, "An AI Zero-Trust Adaptive Access Control System Applicable to Sensitive Data Access," a zero-trust adaptive access control system for sensitive data access is proposed. This system constructs behavioral profiles and dynamic behavioral baselines through a user behavior analysis module, and generates real-time trust scores based on multi-dimensional information through a trust score calculation module. The access control decision module then executes hierarchical control accordingly. However, these existing solutions all follow a feedback logic of "collecting effect data → directly feeding the effect data back to each evaluation model → adjusting parameters accordingly." Their fundamental flaw lies in the fact that the effect data generated after policy execution (such as blocking success rate, false blocking rate, etc.) is actually the result of multiple factors working together—including both the effectiveness contribution of the policy itself and the impact of changes in the external environment (such as packet loss due to network congestion, response delay due to high system load, etc.). Because the existing solution adopts an indiscriminate full feedback mechanism, all effect data is used directly for model optimization without distinction. When the external environment fluctuates, the system cannot distinguish whether the deterioration of the effect data is caused by policy failure or environmental changes. This leads to the evaluation model making incorrect adjustments based on data affected by environmental interference, triggering a vicious cycle of "the more the policy is adjusted, the worse it gets." This severely restricts the protection effectiveness of the zero-trust micro-segmentation system in complex and dynamic network environments.
[0004] Furthermore, while existing Chinese patents CN120768583A, "Dynamic Access Control Method for Zero-Trust Networks Based on AI Behavioral Profiling," propose feeding behavioral data back to the AI behavioral profiling engine and adjusting decision thresholds through reinforcement learning, and CN122496297A, "SOAR-Driven Dynamic Orchestration Method for Zero-Trust Architecture," propose iterative optimization of the trust model based on policy effect monitoring, these solutions still use an indiscriminate, uniform feedback mechanism and fail to solve the technical problem of feedback signal distortion caused by the coupling of environmental and policy factors. Summary of the Invention
[0005] The purpose of this invention is to propose a closed-loop optimization method for zero-trust micro-segmentation strategy based on dynamic evolution of behavioral fingerprints with environmental decoupling. This method addresses the technical problem in existing zero-trust micro-segmentation systems where the feedback signal is distorted due to the coupling between environmental and policy factors in the policy execution effect data, thus causing the evaluation model to adjust erroneous parameters based on the disturbed feedback.
[0006] The method includes the following steps: S1. Collect multi-dimensional behavioral data of the target entity within a preset time window, extract and encode features from the multi-dimensional behavioral data, and generate a short-cycle behavioral fingerprint vector of the target entity. S2. Obtain the dynamic trust assessment parameters of the target entity. The dynamic trust assessment parameters include at least the identity authentication status, device health score and network environment risk index. Calculate the comprehensive trust score of the target entity based on the short-cycle behavioral fingerprint vector and the dynamic trust assessment parameters. S3. Compare the comprehensive trust score with the preset multi-level trust threshold range, determine the current trust level of the target entity based on the comparison result, and adaptively generate the micro-isolation strategy of the target entity based on the current trust level and the anomaly degree of the short-cycle behavioral fingerprint vector. S4. The micro-segmentation strategy is sent to the execution node for execution, and the actual effect data of the target entity and the external environment status parameters of the execution node after the strategy is executed are continuously monitored. S5. Input the external environment state parameters into the environment state encoder to generate the current environment state anchor vector; decompose the actual effect data into a first effect component associated with the target entity behavior and a second effect component associated with the external environment state; S6. Calculate the environmental state offset between the current environmental state anchor vector and the historical environmental state anchor vector; determine the environmental attribution weight of the second effect component based on the environmental state offset; use the environmental attribution weight to extract the environmental contribution value from the second effect component to obtain the net effect value of the strategy. S7. Generate a feedback correction signal based on the net effect value of the strategy and the first effect component; use the feedback correction signal to correct the generation parameters of the short-cycle behavioral fingerprint vector and / or the calculation weight of the comprehensive trust score, so as to achieve environmental adaptive closed-loop optimization of the micro-isolation strategy.
[0007] A closed-loop optimization system for behavior fingerprint dynamic evolution based on environment decoupling and zero-trust micro-segmentation strategy includes: The data acquisition module is used to collect multi-dimensional behavioral data of the target entity within a preset time window; The behavior fingerprint generation module is used to extract and encode features from the multidimensional behavior data to generate a short-period behavior fingerprint vector of the target entity. The trust assessment module is used to obtain the dynamic trust assessment parameters of the target entity and calculate the comprehensive trust score of the target entity based on the short-cycle behavioral fingerprint vector and the dynamic trust assessment parameters. The strategy generation module is used to compare the comprehensive trust score with a preset multi-level trust threshold range, determine the current trust level of the target entity based on the comparison result, and adaptively generate a micro-isolation strategy for the target entity based on the current trust level and the anomaly degree of the short-cycle behavioral fingerprint vector. The strategy execution and monitoring module is used to distribute the micro-segmentation strategy to the execution node for execution, and continuously monitor the actual effect data of the target entity and the external environment status parameters of the execution node after the strategy is executed; The environment decoupling module is used to input the external environment state parameters into the environment state encoder to generate the current environment state anchor vector, and decompose the actual effect data into a first effect component associated with the target entity behavior and a second effect component associated with the external environment state; The strategy effect stripping module is used to calculate the environmental state offset between the current environmental state anchor vector and the historical environmental state anchor vector, determine the environmental attribution weight of the second effect component based on the environmental state offset, and use the environmental attribution weight to strip the environmental contribution value from the second effect component to obtain the net strategy effect value. The closed-loop correction module is used to generate a feedback correction signal based on the net effect value of the strategy and the first effect component, and to use the feedback correction signal to correct the generation parameters of the short-cycle behavioral fingerprint vector and / or the calculation weight of the comprehensive trust score.
[0008] The beneficial effects provided by this invention are as follows: By setting up external environment state sensor probes independent of the policy execution logic, this invention collects environmental parameters such as network bandwidth utilization, latency jitter, and node load in real time, and encodes these parameters into environmental state anchor vectors using an environmental state encoder. Based on this, the actual effect data generated by policy execution is decomposed into a first effect component associated with the target entity's behavior and a second effect component associated with the external environment state. By calculating the offset between the current environmental state anchor vector and the historical baseline anchor vector, the degree of environmental change is quantitatively assessed. Furthermore, the environmental contribution value is extracted from the second effect component to obtain a net policy effect value that truly reflects the effectiveness of the policy itself. The system generates feedback correction signals based solely on this net policy effect value to correct the behavioral fingerprint generation parameters and trust score calculation weights, thereby effectively avoiding interference from environmental fluctuations on the feedback signal. This ensures that the policy optimization direction always remains consistent with the actual policy effect, significantly improving the policy adaptability and stability of the zero-trust micro-segmentation system in complex dynamic network environments. Attached Figure Description
[0009] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0010] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0011] Example 1 Please refer to Figure 1 , Figure 1 This is a schematic diagram of the method flow of the present invention.
[0012] This invention provides a closed-loop optimization method for a zero-trust micro-segmentation strategy based on dynamic evolution of behavioral fingerprints with environment decoupling, comprising: Step S1: Collect multi-dimensional behavioral data of the target entity within a preset time window, extract and encode features from the multi-dimensional behavioral data, and generate a short-period behavioral fingerprint vector of the target entity.
[0013] As an optional implementation, in step S1, the target entity refers to an access subject in the network, including but not limited to user accounts, terminal devices, application processes, or service instances. The length of the preset time window is set according to the behavioral characteristics of the target entity and the real-time requirements of the security policy. Preferably, for high-frequency interaction scenarios (such as API call-intensive applications), the length of the preset time window is 60 to 300 seconds; for low-frequency interaction scenarios (such as database management operations), the length of the preset time window is 300 to 600 seconds.
[0014] The multidimensional behavioral data includes multidimensional behavioral characteristics generated by the target entity during network access, specifically including network access behavior sequences, resource call sequences, and operation timing data. The network access behavior sequence records the target IP address, target port, protocol type, and connection frequency of the network connection initiated by the target entity; the resource call sequence records the type of access operation (read, write, execute, delete) and the identifier of the operation object for the target entity to system resources (files, database tables, API interfaces, etc.); the operation timing data records the initiation timestamp, duration, and sequential relationship between each operation.
[0015] The specific process of feature extraction and encoding of multidimensional behavioral data is as follows: First, the collected multidimensional behavioral data is time-series aligned, and behavioral events from different data sources are synchronously arranged according to a unified timestamp; second, the behavioral data is normalized and preprocessed, mapping behavioral features of different dimensions (such as access frequency, data traffic size, operation response time, etc.) to a unified order of magnitude range; finally, an unsupervised feature extraction model based on an autoencoder is used to encode the preprocessed behavioral data sequence, compressing the high-dimensional sparse behavioral data into a dense feature vector of fixed dimensions, which is the short-period behavioral fingerprint vector.
[0016] For example, in one specific embodiment, an application server (target entity) on an enterprise intranet initiated 327 network connection requests, called the database query interface 184 times, and performed 56 file read operations within a 300-second time window. After feature extraction and encoding, a 128-dimensional floating-point feature vector is generated, with the values of each dimension of the vector distributed between -1.0 and 1.0, serving as the behavioral fingerprint of the application server within the current time window.
[0017] Step S2: Obtain the dynamic trust assessment parameters of the target entity. The dynamic trust assessment parameters include at least the identity authentication status, device health score, and network environment risk index. Calculate the comprehensive trust score of the target entity based on the short-cycle behavioral fingerprint vector and the dynamic trust assessment parameters.
[0018] As an optional implementation, in step S2, the dynamic trust assessment parameter is a quantitative representation of the current security status of the target entity. The identity authentication status reflects the trustworthiness of the target entity through identity authentication, including the authentication method (such as multi-factor authentication, single-factor password authentication, certificate authentication, etc.), the time elapsed since the authentication, and whether there are any anomalies during the authentication process (such as multiple authentication failure records). The identity authentication status value ranges from 0 to 100 points, with a value of 90 to 100 points for multi-factor authentication and authentication times within 5 minutes.
[0019] Device health scores reflect the security compliance status of the target entity's terminal device, including operating system patch update status (whether the latest security patches are installed), endpoint detection and response (EDR) or antivirus software effectiveness (whether it is running normally and the virus database is up-to-date), device fingerprint consistency (whether hardware information matches historical records), and the presence of malicious processes. Device health scores range from 0 to 100, with a score of 80 to 100 for a fully compliant device.
[0020] The network environment risk index reflects the security status of the target entity's current network environment, including the geographical credibility of the IP address (whether it belongs to a frequently used login region), network type (internal network / public network / VPN), connection method (wired / wireless), and whether there are known threat activities on the current network. The network environment risk index ranges from 0 to 100 points, with higher values indicating greater network environment risk.
[0021] The comprehensive trust score is calculated as follows: the short-cycle behavioral fingerprint vector is compared with the historical behavioral fingerprint baseline of the target entity to obtain the behavioral anomaly score; at the same time, the identity authentication status, device health score and network environment risk index are weighted and fused according to preset weights to obtain the static trust score; finally, the behavioral anomaly score and the static trust score are weighted and summed according to dynamic weight coefficients to obtain the comprehensive trust score.
[0022] For example, in one specific embodiment, the identity authentication status of a target entity is 95 points (multi-factor authentication passed, authentication time was 2 minutes ago), the device health score is 88 points (patches complete, EDR running normally), the network environment risk index is 25 points (internal network environment, IP address located in a commonly used office area), the behavior anomaly score is 92 points (current behavior highly matches the historical fingerprint baseline), and the comprehensive trust score is = point.
[0023] Step S3: Compare the comprehensive trust score with the preset multi-level trust threshold range, determine the current trust level of the target entity based on the comparison result, and adaptively generate the micro-isolation strategy of the target entity based on the current trust level and the anomaly degree of the short-cycle behavioral fingerprint vector.
[0024] As an optional implementation, in step S3, the multi-level trust threshold range includes at least three levels: a high-trust range, a medium-trust range, and a low-trust range. A high-trust range corresponds to a comprehensive trust score greater than or equal to 80 points, indicating that the target entity's current behavior is trustworthy and the environment is safe; a medium-trust range corresponds to a comprehensive trust score between 60 and 80 points (excluding 80, but including 60), indicating that the target entity has certain risk factors; a low-trust range corresponds to a comprehensive trust score less than 60 points, indicating that the target entity has a high risk.
[0025] Micro-segmentation policies are sets of access control rules dynamically generated based on the trust level and anomaly level of target entities. A micro-segmentation policy includes at least the isolation object identifier, communication permission settings (allow / block), protocol and port range, policy duration, and policy priority. The generation of micro-segmentation policies follows the principles of least privilege and default denial—only granting the target entity the minimum network access permissions necessary to complete its business functions, and blocking any unauthorized access requests.
[0026] Specifically, when the overall trust score is in the high trust range, an allow policy is generated, allowing the target entity to access network resources normally according to its currently assigned business permissions, while only blocking known malicious IPs / ports; when the overall trust score is in the medium trust range, a demotion policy is generated, restricting the target entity's access to sensitive resources (such as core databases and key management systems) and increasing the frequency of traffic audits; when the overall trust score is in the low trust range, an isolation policy is generated, blocking all unnecessary network connections of the target entity, while retaining only necessary management channels.
[0027] For example, in one specific embodiment, a target entity has a comprehensive trust score of 88.4 (high trust range) and an anomaly rate of 12% in its behavioral fingerprint vector (low deviation from the baseline). The system generates an access policy that allows the entity to access the database server (IP: 10.0.1.100, port 3306) and application server (IP: 10.0.2.50, port 8080) required for its business, while blocking all access to other network segments. The policy is effective for 3600 seconds and has a priority of 3 levels (level 1 is the highest and level 5 is the lowest).
[0028] Step S4: Deploy the micro-segmentation strategy to the execution node for execution, and continuously monitor the actual effect data of the target entity and the external environment status parameters of the execution node after the strategy is executed.
[0029] As an optional implementation, in step S4, the execution node refers to a network device with a policy execution agent deployed, including but not limited to security groups in a cloud environment, firewalls in a physical network, network policy controllers (such as the Kubernetes NetworkPolicy controller) in a container environment, and firewall services (such as iptables and nftables) in the host operating system. Policy distribution is completed through standardized interfaces (such as RESTful API, gRPC, or message queues), and the distribution process uses encrypted transmission to ensure that the policy content is not tampered with.
[0030] The actual effect data refers to quantifiable security performance metrics generated after policy execution, including the number of successful blocking attempts (the number of malicious connection requests successfully blocked), the number of false blocking attempts (the number of normal business connection requests incorrectly blocked), access response latency (the additional latency overhead of network access after policy execution), and business continuity metrics (whether core business is affected by policy execution). The actual effect data is collected in real-time and written to the feedback data queue using an event-driven asynchronous monitoring method.
[0031] The external environment state parameters refer to macroscopic environmental variables that are independent of the target entity's behavior and can affect the strategy execution effect. These include the real-time bandwidth utilization of the network where the execution node is located, the packet round-trip latency jitter, the CPU load rate of the execution node, and the memory usage rate. The external environment state parameters are collected in parallel by independent sensor probes at a fixed sampling period (preferably 5 to 15 seconds).
[0032] For example, in one specific implementation, the micro-segmentation policy was issued to a security group in a public cloud environment via a RESTful API, with the policy issuance taking 0.8 seconds. During the monitoring period following policy execution, a total of 47 successful blocking attempts, 2 false blocking attempts, a 15ms increase in access response latency, and a business continuity index of 100% (core business was unaffected) were collected. Simultaneously, sensor probes collected data showing a network bandwidth utilization rate of 62%, a round-trip latency jitter of 8ms, a CPU load rate of 45% on the execution node, and a memory utilization rate of 67%.
[0033] Step S5: Input the external environment state parameters into the environment state encoder to generate the current environment state anchor vector; decompose the actual effect data into a first effect component associated with the target entity behavior and a second effect component associated with the external environment state.
[0034] As an optional implementation, in step S5, the environment state encoder is an encoding network based on a multilayer perceptron (MLP). Its input layer receives a normalized external environment state parameter vector, performs a nonlinear transformation through several hidden layers, and the output layer outputs a fixed-dimensional floating-point vector as the environment state anchor vector. The training objective of the environment state encoder is to minimize the reconstruction error, so that similar anchor vectors can be generated under similar environmental states.
[0035] The first effect component refers to the portion of the actual effect data caused by changes in the target entity's own behavior, reflecting the direct regulatory effect of the strategy on the target entity's behavior. The second effect component refers to the portion of the actual effect data caused by changes in the external environment, reflecting the indirect impact of environmental factors (such as network congestion and system load changes) on the strategy's execution effect.
[0036] The specific method for decomposing the effect data is as follows: obtain the behavioral baseline characteristics of the target entity within the historical reference time period (including the operation frequency baseline, resource access pattern baseline, and communication object baseline), determine the part of the actual effect data that has a statistical correlation with the behavioral baseline characteristics as the first effect component, and determine the part of the actual effect data that has no statistical correlation with the behavioral baseline characteristics but has a statistical correlation with the environmental state parameters as the second effect component.
[0037] For example, in one specific embodiment, the external environment state parameters are [bandwidth utilization 62%, latency jitter 8ms, CPU load 45%, memory utilization 67%]. After encoding by the environment state encoder, a 64-dimensional current environment state anchor vector is generated, with the values of each dimension distributed between -0.5 and 0.5. The total effect value of the actual effect data is a blocking success rate of 95.7%. After decomposition, the first effect component is 88.3% (generated by the regulation effect of the strategy itself), and the second effect component is 7.4% (the additional blocking effect caused by network environment fluctuations).
[0038] Step S6: Calculate the environmental state offset between the current environmental state anchor vector and the historical environmental state anchor vector; determine the environmental attribution weight of the second effect component based on the environmental state offset; use the environmental attribution weight to extract the environmental contribution value from the second effect component to obtain the net effect value of the strategy.
[0039] As an optional implementation, in step S6, the environmental state offset is used to quantify the degree of change of the current environmental state relative to the historical baseline environmental state. The environmental state offset is calculated as follows: the current environmental state anchor vector and the historical environmental state anchor vector are subtracted dimension by dimension to obtain the environmental state difference vector; then, the square root of the sum of the squares of each dimension of the environmental state difference vector is taken to obtain the environmental state offset. :
[0040] in The first anchor point vector of the current environment state dimensional components, The first anchor point vector of historical environment state dimensional components, This represents the total number of dimensions of the anchor vector.
[0041] The environmental attribution weights are used to characterize the proportion of the second effect component that can be attributed to changes in environmental conditions. Calculate using the following formula:
[0042] in The maximum permissible environmental offset is preset (determined based on the standard deviation of historical environmental data). The preset attenuation coefficient and (Preferred, ).when hour, This indicates that environmental changes have reached the maximum tolerable range, and the second effect component is entirely attributed to environmental factors.
[0043] The net effect value of the strategy Calculate using the following formula:
[0044] in This represents the total effectiveness value of the actual results (such as the overall blocking success rate). This is the second effect component.
[0045] For example, in one specific embodiment, the Euclidean distance between the current environment state anchor vector and the historical environment state anchor vector. Preset maximum allowable environmental offset attenuation coefficient Then the environmental attribution weight Second effect component Total effect value Then the net effect value of the strategy .
[0046] Step S7: Generate a feedback correction signal based on the net effect value of the strategy and the first effect component; use the feedback correction signal to correct the generation parameters of the short-cycle behavioral fingerprint vector and / or the calculation weight of the comprehensive trust score, so as to achieve environmental adaptive closed-loop optimization of the micro-isolation strategy.
[0047] As an optional implementation, in step S7, the feedback correction signal is a quantified instruction that includes the correction direction and correction intensity. When the net effect value of the strategy is higher than a preset upper limit threshold of the net effect, the feedback correction signal indicates positive correction (enhancing the current strategy direction); when the net effect value of the strategy is lower than a preset lower limit threshold of the net effect, the feedback correction signal indicates negative correction (adjusting the current strategy direction).
[0048] The specific steps for correcting the generation parameters of short-cycle behavioral fingerprint vectors include: when the net effect value of the policy is higher than the preset upper limit threshold of net effect (e.g., 95%), it indicates that the current behavioral fingerprint can effectively reflect the behavioral characteristics of the target entity. In this case, the length of the preset time window is shortened (e.g., from 300 seconds to 180 seconds) to increase the update frequency of the behavioral fingerprint vector and achieve more refined behavioral monitoring; when the net effect value of the policy is lower than the preset lower limit threshold of net effect (e.g., 80%), it indicates that the current behavioral fingerprint is updated too frequently, resulting in fingerprint instability. In this case, the length of the preset time window is extended (e.g., from 300 seconds to 480 seconds) to reduce the update frequency of the behavioral fingerprint vector and obtain a more stable behavioral baseline.
[0049] The specific adjustments to the calculation weights of the comprehensive trust score include: when the net effect value of the strategy is positive and the first effect component is higher than the preset component threshold, it indicates that the behavior of the target entity is the dominant factor affecting the effect of the strategy, and in this case, the calculation weight of the device health score in the comprehensive trust score is increased; when the net effect value of the strategy is negative and the first effect component is lower than the preset component threshold, it indicates that environmental factors are the dominant factor affecting the effect of the strategy, and in this case, the calculation weight of the network environment risk index in the comprehensive trust score is decreased.
[0050] For example, in one specific embodiment, the net effect value of the strategy First effect component The preset upper limit of net effect threshold is 92%, the lower limit of net effect threshold is 80%, and the component threshold is 85%. Because... The feedback correction signal indicates positive correction. Further analysis of the first effect component is needed. Therefore, the weight of the device health score was increased (from 25% to 30%), and the weights of other dimensions were adjusted accordingly to make the calculation of the comprehensive trust score more accurate.
[0051] Example 2 As an optional implementation, the external environment status parameters mentioned in step S5 include the following four items: real-time bandwidth utilization of the network where the execution node is located, data packet round-trip time jitter value, CPU load rate of the execution node, and memory usage rate.
[0052] The real-time bandwidth utilization rate refers to the percentage of the actual bandwidth currently used by the network link where the execution node is located relative to the total bandwidth of the link. It is obtained in real time through Simple Network Management Protocol (SNMP) or network traffic monitoring tools (such as sFlow, NetFlow), with a preferred collection period of 5 seconds.
[0053] The round-trip time jitter value refers to the standard deviation of the round-trip time (RTT) of multiple consecutive data packets, reflecting the stability of network transmission. The RTT jitter value is obtained through active probing (such as sending ICMP Echo requests) or passive monitoring (such as analyzing the round-trip time of a TCP session), and is measured in milliseconds (ms). Preferably, a RTT jitter value less than 5ms indicates network stability, while a value greater than 20ms indicates significant network fluctuations.
[0054] The CPU load rate refers to the average utilization rate of the central processing unit of the execution node during the sampling period, obtained by reading from the system call provided by the operating system (such as / proc / stat in Linux), in percentage (%). Preferably, a CPU load rate below 70% is considered to indicate sufficient system resources, and a CPU load rate above 90% is considered to indicate that the system is under high load.
[0055] The memory utilization rate refers to the ratio of the physical memory used by the execution node to the total physical memory, obtained through system calls provided by the operating system (such as / proc / meminfo in Linux), and is expressed as a percentage (%). Preferably, a memory utilization rate below 80% is considered sufficient, and a rate above 95% is considered strained.
[0056] The environmental state encoder is an encoding network based on a multilayer perceptron (MLP). Its structure includes an input layer (4 neurons, corresponding to four external environmental state parameters), two hidden layers (32 neurons and 16 neurons respectively, with ReLU activation function), and an output layer (8 neurons, with no activation function). The input of the encoding network is the normalized external environmental state parameters (each parameter is normalized to the interval between 0 and 1), and the output is an 8-dimensional floating-point vector as the anchor vector of the current environmental state.
[0057] For example, in one specific embodiment, the network bandwidth utilization of an execution node is 62%, the round-trip time jitter is 8ms, the CPU load is 45%, and the memory usage is 67%, which are normalized to [0.62, 0.40, 0.45, 0.67]. After being input into the multilayer perceptron coding network, the output 8-dimensional anchor vector is [0.23, -0.15, 0.47, -0.08, 0.31, 0.12, -0.26, 0.19].
[0058] Example 3 As an optional implementation, step S5 decomposes the actual effect data into a first effect component and a second effect component, specifically including the following sub-steps: S51. Obtain the behavioral baseline characteristics of the target entity within a historical reference time period. The behavioral baseline characteristics include the operation frequency baseline, resource access pattern baseline, and communication object baseline.
[0059] The historical reference time period refers to the historical time period of the target entity under normal behavior, preferably the past 7 to 30 days, excluding abnormal events (such as security alarms and failures). The operation frequency baseline is the average frequency and standard deviation of various operations (such as network connection, database query, and file access) of the target entity within the historical reference time period; the resource access pattern baseline is the distribution of resource types accessed by the target entity within the historical reference time period (such as the proportion of read operations, the proportion of write operations, and the frequency of access to sensitive resources); the communication object baseline is the set of peer IP addresses, port sets, and communication traffic distribution of the target entity within the historical reference time period.
[0060] The method for obtaining behavioral baseline characteristics is as follows: statistical analysis of behavioral data within a historical reference time period, and calculation of the mean of each behavioral indicator. and standard deviation ,by This is considered a normal range of behavior.
[0061] For example, in one specific embodiment, the baseline operation frequency of an application server within a historical reference time period is as follows: database queries average 15.3 times per minute (standard deviation 3.2 times), file reads average 4.7 times per minute (standard deviation 1.8 times), and network connections average 8.2 times per minute (standard deviation 2.1 times). The baseline resource access pattern is: read operations account for 72%, write operations account for 18%, and sensitive resource access accounts for 10%. The baseline communication objects are: the main communication peers include the database server (10.0.1.100:3306), the cache server (10.0.1.200:6379), and the log server (10.0.2.100:514).
[0062] S52. The portion of the actual effect data that has a statistical correlation with the behavioral baseline features is determined as the first effect component, and the portion of the actual effect data that has no statistical correlation with the behavioral baseline features but has a statistical correlation with the external environment state parameters is determined as the second effect component.
[0063] The term "statistical correlation" refers to a statistically significant correlation between changes in actual effect data and changes in behavioral baseline characteristics. This is determined by calculating the Pearson correlation coefficient: when the absolute value of the correlation coefficient is greater than 0.5, a statistical correlation is considered to exist.
[0064] For example, in one specific embodiment, the overall blocking success rate of the actual effect data was 95.7%. Among them, the portion that was statistically correlated with the operating frequency baseline (deviation from the normal value + 2.3 standard deviations) was 88.3%, which was identified as the first effect component; the portion that was not statistically correlated with the operating frequency baseline but was statistically correlated with the network bandwidth utilization rate (increased from 45% to 62%) was 7.4%, which was identified as the second effect component.
[0065] Example 4 As an optional implementation, step S6, calculating the environmental state offset, specifically includes the following sub-steps: S61. Calculate the difference between the current environmental state anchor vector and the historical environmental state anchor vector in each dimension to obtain the environmental state difference vector.
[0066] The historical environmental state anchor vector refers to the statistical center vector (median or mean of each dimension) of the environmental state anchor vectors at each sampling time within a historical reference time period (preferably the past 24 to 72 hours). Dimensional difference calculation refers to subtracting the corresponding dimension component of the historical environmental state anchor vector from each component of the current environmental state anchor vector.
[0067] For example, in one specific embodiment, the current environmental state anchor vector is [0.23, -0.15, 0.47, -0.08, 0.31, 0.12, -0.26, 0.19], the historical environmental state anchor vector is [0.18, -0.12, 0.35, -0.05, 0.28, 0.10, -0.20, 0.15], and the environmental state difference vector is calculated by dimension-wise difference as [0.05, -0.03, 0.12, -0.03, 0.03, 0.02, -0.06, 0.04].
[0068] S62. Take the square root of the sum of the squares of each dimension of the environmental state difference vector to obtain the environmental state offset.
[0069] Environment state offset The calculation formula is:
[0070] in The first anchor point vector of the current environment state dimensional components, The first anchor point vector of historical environment state dimensional components, The total number of dimensions of the anchor vector (in this embodiment) ).
[0071] Continuing with the example above, the sum of the squares of each dimension of the environmental state difference vector is: Take the square root to get .
[0072] Example 5 As an optional implementation, the environmental attribution weights Calculate using the following formula:
[0073] in This is the preset maximum allowable environmental offset. The preset attenuation coefficient and .when hour, .
[0074] The value is determined based on the statistical distribution of historical environmental state data. Specifically, environmental state anchor vectors are collected for all sampling times over the past 30 days, and the offset of each sampling time relative to the historical baseline anchor vector is calculated. The 95th percentile of these offsets is then taken as the threshold value. For example, if 95% of the environmental offsets over the past 30 days are less than 0.8, then set... .
[0075] This is the attenuation coefficient, used to adjust the sensitivity of environmental attribution. The value ranges from 0.5 to 0.95, with 0.85 being preferred. A larger value is used when a more aggressive approach to stripping away environmental contributions is needed, and a smaller value is used when a more conservative approach to preserving environmental contributions is needed. The value can be preset by the system administrator according to the security policy requirements of the deployment scenario, or determined through cross-validation and optimization of historical data.
[0076] The net effect value of the strategy Calculate using the following formula:
[0077] in The total effect value of the actual effect data. This is the second effect component.
[0078] Continuing with the previous example, , , Then the environmental attribution weight Overall effect value Second effect component Then the net effect value of the strategy .
[0079] Example 6 As an optional implementation, step S3, which adaptively generates the micro-segmentation strategy for the target entity, specifically includes the following sub-steps: S31. The comprehensive trust score is compared with a first trust threshold and a second trust threshold, wherein the first trust threshold is greater than the second trust threshold.
[0080] The first trust threshold and the second trust threshold are trust score boundary values preset according to the system security policy. Preferably, the first trust threshold is set to 80 points and the second trust threshold is set to 60 points. The above thresholds can be adjusted according to the security level requirements of the actual deployment environment. In high-security scenarios, the thresholds can be increased (e.g., the first threshold is 85 points and the second threshold is 65 points), and in low-security scenarios, the thresholds can be decreased (e.g., the first threshold is 75 points and the second threshold is 55 points).
[0081] S32. When the comprehensive trust score is greater than or equal to the first trust threshold, a release policy is generated to allow the target entity to access normally according to the current permissions.
[0082] The allowance policy permits target entities to access network resources normally according to their currently assigned business permissions. Specific rules include: allowing target entities to access authorized servers and ports required for their business operations; allowing target entities to perform the types of operations required for their business operations; blocking known malicious IP addresses and ports; and restricting access that is not essential for their business operations. The allowance policy priority is set to medium (e.g., level 3), and the effective duration is set according to the business cycle (e.g., 8 hours or 24 hours).
[0083] S33. When the comprehensive trust score is between the second trust threshold and the first trust threshold, a demotion strategy is generated to restrict the target entity's access to sensitive resources.
[0084] The privilege reduction policy adds access restrictions to the existing allowance policy. Specific rules include: restricting target entities' access to sensitive resources (such as core databases, key management services, and configuration file servers), allowing only read operations and prohibiting write or delete operations; increasing the audit frequency of target entity traffic (from the usual once every 10 minutes to once every 2 minutes); and performing secondary authentication on cross-security domain access requests initiated by the target entity. The priority of the privilege reduction policy is set to high (e.g., level 2), and the effective duration is set based on the risk assessment results (e.g., 1 hour to 4 hours).
[0085] S34. When the comprehensive trust score is less than or equal to the second trust threshold, an isolation strategy is generated to block all unnecessary network connections of the target entity.
[0086] Isolation policies are the strictest access control measures. Specific rules include: blocking all non-essential network connections to the target entity, retaining only management channels (such as SSH port 22 or RDP port 3389) for security administrators to troubleshoot and recover from; prohibiting the target entity from initiating any new network connection requests; and forcibly terminating established sessions. The isolation policy is set to the highest priority (e.g., level 1), and its duration is determined based on the severity of the security incident (e.g., 30 minutes to 2 hours). Isolation can only be lifted after the security administrator confirms the risk has been eliminated.
[0087] For example, in one specific embodiment, the overall trust score of a target entity is 43.1 points (below the second trust threshold of 60 points). The system generates an isolation policy: block the entity from accessing all internal network IP addresses, and only allow it to connect to the management jump server (IP: 10.0.0.100) via SSH (port 22). The policy priority is level 1, and the effective duration is 3600 seconds.
[0088] Example 7 As an optional implementation, step S7 uses the feedback correction signal to correct the generation parameters of the short-period behavioral fingerprint vector, specifically including the following operations: When the net effect value of the strategy exceeds a preset upper limit threshold for net effect, the length of the preset time window is shortened, thereby increasing the update frequency of the short-cycle behavioral fingerprint vector. The upper limit threshold for net effect is preferably set to 92%. For example, if the current preset time window length is 300 seconds, when the net effect value of the strategy reaches 94.5% (above 92%), the preset time window length is shortened to 180 seconds, increasing the update frequency of the behavioral fingerprint vector from once every 300 seconds to once every 180 seconds, thus achieving more refined behavioral monitoring.
[0089] When the net effect value of the strategy falls below a preset lower limit threshold, the length of the preset time window is extended, reducing the update frequency of the short-cycle behavioral fingerprint vector. The lower limit threshold is preferably set to 80%. For example, if the current preset time window length is 300 seconds, when the net effect value drops to 75% (below 80%), the preset time window length is extended to 480 seconds, reducing the update frequency of the behavioral fingerprint vector from once every 300 seconds to once every 480 seconds. This results in a more stable behavioral baseline and avoids misjudgments caused by frequent fingerprint updates.
[0090] The length of the preset time window is adjusted in steps, with each adjustment preferably ranging from 30 to 60 seconds. The single adjustment range does not exceed 30% of the current window length to avoid sudden changes in behavioral fingerprints due to excessive adjustment range.
[0091] Example 8 As an optional implementation, step S7 uses the feedback correction signal to correct the calculation weights of the comprehensive trust score, specifically including the following operations: When the net effect value of the strategy is positive and the first effect component is higher than a preset component threshold, the weight of the device health score in the overall trust score is increased. The component threshold is preferably set to 85%. A positive net effect value indicates that the current strategy has produced a positive security effect, and the first effect component being higher than the component threshold indicates that behavioral factors are the dominant factor in the strategy's effect. Increasing the weight of the device health score in this case can enhance the sensitivity to changes in the target entity's behavior. For example, if the original weight of the device health score is 25%, it can be increased to 30% when the above condition is met.
[0092] When the net effect value of the strategy is negative and the first effect component is lower than the preset component threshold, the weight of the network environment risk index in the overall trust score is reduced. A negative net effect value indicates that the current strategy has produced a negative security effect (such as an excessively high false positive rate), and the first effect component being lower than the component threshold indicates that behavioral factors are not the dominant factor in the strategy's effect (i.e., environmental factors play a dominant role). In this case, reducing the weight of the network environment risk index can reduce the interference of environmental fluctuations on the trust score. For example, if the original weight of the network environment risk index is 20%, it can be reduced to 15% when the above conditions are met.
[0093] The weights of each dimension are adjusted using normalization to ensure that the sum of the weights of all dimensions remains 100% after adjustment. The preferred step size for weight adjustment is 1% to 5%, and the magnitude of a single adjustment should not exceed 20% of the original weight.
[0094] Example 9 As an optional implementation, the monitoring in step S4 adopts an event-driven asynchronous monitoring method. Event-driven asynchronous monitoring means that the monitoring system does not actively collect data using a polling method, but instead, the policy execution node actively pushes event messages to the monitoring system when a security event occurs (such as a blocking decision, an allowing decision, or an anomaly alarm). The push uses a message queue (such as Apache Kafka or RabbitMQ) as a buffer channel to ensure that event data is not lost in high-concurrency scenarios.
[0095] The actual effect data includes the number of successful blocking attempts, the number of false blocking attempts, access response latency, and business continuity metrics after the policy execution. Successful blocking attempts refer to the number of malicious connection requests or malicious operations successfully blocked by the policy; false blocking attempts refer to the number of normal business connection requests or normal operations incorrectly blocked by the policy; access response latency refers to the time interval from when the target entity initiates an access request to when the policy decision result is obtained; business continuity metrics are a binary indicator of whether core business functions are affected by policy execution (100% indicates no impact, 0% indicates complete interruption).
[0096] The external environmental state parameters are collected in parallel using independent sensor probes at a fixed sampling period. The sensor probes are lightweight acquisition programs deployed on the execution node, decoupled from the policy execution logic, and run independently to ensure that environmental state acquisition is not interrupted due to policy execution failures. The sampling period is shorter than the preset time window to ensure that multiple environmental state samples can be collected within a single time window. Preferably, the sampling period is set to 5 to 15 seconds, and the preset time window is set to 180 to 600 seconds, meaning that 12 to 120 environmental state samples can be collected within each time window.
[0097] For example, in one specific embodiment, the preset time window is 300 seconds, the sampling period of the sensor probe is 10 seconds, and 30 environmental state samples are collected within each time window. The monitoring system uses Apache Kafka as the event message queue. Every time the policy execution node makes a blocking or allowing decision, it pushes an event message to Kafka. The message content includes a timestamp, target entity identifier, decision type, decision basis, and execution result.
[0098] Example 10 This invention also provides a closed-loop optimization system for a zero-trust micro-segmentation strategy based on dynamic evolution of behavioral fingerprints with environmental decoupling, comprising the following modules: The data acquisition module is used to collect multi-dimensional behavioral data of the target entity within a preset time window. Deployed at key nodes in the network (such as zero-trust gateways, terminal agents, and network traffic probes), the data acquisition module obtains the target entity's network access logs, system operation logs, and application access logs through bypass mirroring or embedded acquisition methods. The collected raw data is then formatted and output to the behavior fingerprint generation module.
[0099] The behavior fingerprint generation module is used to extract and encode features from the multidimensional behavior data to generate a short-period behavior fingerprint vector for the target entity. The module integrates an unsupervised feature extraction model based on an autoencoder, which is pre-trained using historical behavior data and can compress the input multidimensional behavior data into a dense feature vector of fixed dimensions. The module iteratively updates using a sliding time window approach, regenerating the target entity's behavior fingerprint vector at the end of each iteration (i.e., at the end of each preset time window) and outputting it to the trust evaluation module.
[0100] The trust assessment module is used to acquire dynamic trust assessment parameters of the target entity and calculate a comprehensive trust score for the target entity based on the short-cycle behavioral fingerprint vector and the dynamic trust assessment parameters. The trust assessment module integrates a multi-source data fusion engine, capable of acquiring identity authentication status from the identity authentication system, device health scores from the terminal management system, and network environment risk indices from the network security situation awareness system. It then performs weighted fusion calculations with the behavioral fingerprint vectors to output a comprehensive trust score to the policy generation module.
[0101] The policy generation module compares the comprehensive trust score with a preset multi-level trust threshold range, determines the current trust level of the target entity based on the comparison result, and adaptively generates a micro-isolation policy for the target entity based on the current trust level and the anomaly degree of the short-cycle behavioral fingerprint vector. The policy generation module has a pre-built multi-level policy template library (including allowance policy templates, demotion policy templates, and isolation policy templates). It matches the corresponding policy template according to the trust level and anomaly degree, fills in specific parameters (isolation object, communication permissions, protocol port, effective duration, priority), generates an executable micro-isolation policy, and outputs it to the policy execution and monitoring module.
[0102] The strategy execution and monitoring module is used to distribute the micro-segmentation strategy to the execution nodes for execution, and continuously monitor the actual effect data of the target entity and the external environment status parameters of the execution nodes after strategy execution. The strategy execution and monitoring module communicates with each execution node through a standardized interface (such as RESTful API or gRPC), receives event messages pushed by the execution nodes using an event-driven asynchronous monitoring method, and collects external environment status parameters at a fixed sampling period using independent sensor probes. The actual effect data and environmental status parameters are then output to the environment decoupling module respectively.
[0103] The environment decoupling module is used to input the external environment state parameters into the environment state encoder to generate the current environment state anchor vector, and to decompose the actual effect data into a first effect component associated with the target entity's behavior and a second effect component associated with the external environment state. The environment decoupling module integrates an environment state encoder based on a multilayer perceptron and an effect decomposition engine based on statistical correlation analysis, enabling the quantitative separation of environmental factors and behavioral factors.
[0104] The strategy effect stripping module is used to calculate the environmental state offset between the current environmental state anchor vector and the historical environmental state anchor vector, determine the environmental attribution weight of the second effect component based on the environmental state offset, and use the environmental attribution weight to strip the environmental contribution value from the second effect component to obtain the net strategy effect value. The strategy effect stripping module internally implements an environmental state offset calculation unit and an environmental attribution weight calculation unit, according to the aforementioned formula. and Perform the calculation.
[0105] The closed-loop correction module is used to generate a feedback correction signal based on the net effect value of the strategy and the first effect component, and to use the feedback correction signal to correct the generation parameters of the short-cycle behavioral fingerprint vector and / or the calculation weights of the comprehensive trust score. Internally, the closed-loop correction module includes a feedback correction signal generation unit and a parameter correction execution unit. It can generate positive or negative correction signals based on the comparison result between the net effect value of the strategy and a preset threshold, and drive the behavioral fingerprint generation module to adjust the time window parameters and drive the trust assessment module to adjust the weight parameters of each dimension, forming a complete closed-loop self-optimization mechanism.
[0106] The data flow between modules is as follows: Data acquisition module → Behavioral fingerprint generation module → Trust assessment module → Policy generation module → Policy execution and monitoring module → Environment decoupling module → Policy effect stripping module → Closed-loop correction module → (feedback to) Behavioral fingerprint generation module and Trust assessment module.
[0107] Example 11 To verify the technical effectiveness of the method proposed in this invention, a comparative experiment was conducted in a simulated enterprise-level hybrid cloud environment.
[0108] 1. Experimental conditions Hardware environment: Control node: 1 server, configured with an Intel Xeon Gold 6248 processor (2.5GHz, 20 cores), 128GB DDR4 memory, 512GB NVMe SSD, and the operating system is Ubuntu 22.04 LTS; Execution nodes: 10 servers, each configured with an Intel Xeon Silver 4214 processor (2.2GHz, 12 cores), 64GB DDR4 memory, 256GB SSD, and CentOS 8 operating system; Network equipment: 2 10 Gigabit switches, supporting OpenFlow protocol; Software environment: Zero-trust policy engine: developed based on Python 3.10, integrating TensorFlow 2.15 for behavioral fingerprint encoding models; Policy execution agent: Implemented based on eBPF technology and deployed on each execution node; Environmental sensor probe: Custom-developed based on Prometheus node_exporter; Comparison scheme: A comparative system based on the existing document (CN122247725A - An AI Zero Trust Adaptive Access Control System for Sensitive Data Access); Parameter settings: Preset time window: 300 seconds; Behavioral fingerprint vector dimension: 128-dimensional; Environmental state anchor vector dimension: 8-dimensional; Environmental state encoder: Multilayer perceptron, structure 4-32-16-8; First trust threshold: 80 points; Second trust threshold: 60 points; Maximum allowable environmental offset. : 0.8; Attenuation coefficient 0.85; Net effect upper limit threshold: 92%; Net effect lower limit threshold: 80%; Component threshold: 85%; Sensor sampling period: 10 seconds; 2. Experimental subjects The experiment uses a simulated enterprise intranet traffic dataset, containing the following three types of target entities: Table 1 Experimental subjects
[0109] The experimental dataset contains 7 days of continuous behavioral data, with the first 5 days used to establish a behavioral baseline and train the model, and the last 2 days used for testing. During testing, the following three types of abnormal scenarios were simulated using a traffic injection tool: Scenario A (Transient Anomaly): Data packet loss and retransmission caused by network congestion (lasting 5-10 minutes); Scenario B (Gradual Anomaly): An application server gradually increases the frequency of database queries (from 1 times the normal value to 3 times, lasting for 2 hours). Scenario C (Persistent Anomaly): A terminal device continuously attempts to access the core database during non-working hours (2:00 AM - 4:00 AM); 3. Experimental Procedure (1) Before the test begins, use the historical data from the previous 5 days to train the behavioral fingerprint encoding model and the environmental state encoder, and establish the behavioral baseline and historical environmental state anchor vector of each target entity.
[0110] (2) Starting at 0:00 on day 6, the method of the present invention (experimental group) and the method described in the existing documents (control group) were run, with both groups using the same initial strategy configuration and the same test traffic.
[0111] (3) At 9:30 on day 6, scenario A (transient anomaly) was injected for 8 minutes; at 14:00 on day 6, scenario B (gradual anomaly) was injected for 2 hours; at 2:00 on day 7, scenario C (persistent anomaly) was injected for 2 hours.
[0112] (4) Continuously monitor and record the following evaluation indicators: Net effect value of strategy (a unique indicator of this invention); False blocking rate (number of legitimate requests blocked incorrectly / total number of legitimate requests); Missed blocking rate (number of unblocked malicious requests / total number of malicious requests); Policy response latency (the time interval from when an anomaly occurs to when the policy takes effect); Model stability (standard deviation of the overall trust score during environmental fluctuations); 4. Experimental Results Data Table 2 Experimental Results
[0113] 5. Experimental Data Analysis (1) False blocking rate analysis: The overall false blocking rate of the method of the present invention is 1.8%, which is significantly lower than the 4.7% of the existing documents, a reduction of 61.7%. During scenario A (transient anomaly), network congestion caused data packet loss. The existing document method misjudged a large number of normal requests that were retransmitted due to network timeout as abnormal behavior and blocked them, causing the false blocking rate to soar to 8.2%. However, the method of the present invention, through the environment decoupling mechanism, attributes the effect of network congestion to the second effect component (environmental correlation). After removing the environmental contribution value, the net effect value of the strategy still remains above 92%, effectively avoiding false blocking caused by environmental fluctuations.
[0114] (2) Missed Linkage Rate Analysis: The overall missed linkage rate of the method of the present invention is 2.1%, which is lower than the 3.9% of the existing methods, representing a reduction of 46.2%. During Scenario B (gradual anomaly), an application server slowly increases the frequency of database queries. The existing methods failed to identify this gradual behavioral deviation in time, resulting in a missed linkage rate of 5.1%. However, the method of the present invention, through high-frequency updates of short-cycle behavioral fingerprint vectors (regenerated every 300 seconds), can capture subtle changes in behavioral patterns in a timely manner. Combined with the dynamic optimization of behavioral fingerprint generation parameters by feedback correction signals, the missed linkage rate is controlled at 1.9%.
[0115] (3) Response delay analysis: In scenario C (persistent anomaly), the policy response delay of the method of the present invention is 2.8 seconds, which is 55.6% lower than the 6.3 seconds of the existing document. The method of the present invention, through event-driven asynchronous monitoring and real-time environmental decoupling calculation, can quickly complete the entire process of trust score recalculation, policy net effect evaluation and policy correction after the occurrence of abnormal behavior, and realize near real-time policy adaptive adjustment.
[0116] (4) Model stability analysis: During the period of environmental fluctuation (the time period when scenario A and scenario B overlap), the standard deviation of the comprehensive trust score of the method of the present invention was 3.7 points, which was much lower than the 11.2 points of the existing documents, a reduction of 67.0%. This shows that the method of the present invention effectively filters out the interference of environmental factors on the trust score through the environmental decoupling mechanism, so that the trust score more accurately reflects the credibility of the target entity's real behavior, and avoids the vicious cycle of "the more you adjust, the more chaotic it becomes".
[0117] (5) Closed-loop correction effect analysis: During the 24-hour test, the method of this invention triggered a total of 47 closed-loop corrections, of which 23 were time window parameter adjustments (13 shortened and 10 extended), and 24 were weight parameter adjustments (16 increased the weight of device health and 8 decreased the weight of network environment risk). These corrections were all based on the real-time evaluation results of the net effect value of the strategy, realizing the adaptive optimization of the strategy parameters. In contrast, existing document methods, due to the lack of environmental decoupling and differentiated feedback mechanisms, cannot perform similar precise corrections.
[0118] In summary, the method of this invention achieves quantitative separation of "strategy effect" and "environment effect" in strategy effect data through an environment decoupling mechanism, enabling the feedback correction signal to accurately reflect the true effect of the strategy itself, effectively avoiding false feedback and model degradation caused by environmental fluctuations, and significantly outperforming existing technical solutions in key indicators such as false blocking rate, missed blocking rate, response latency, and model stability.
[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A closed-loop optimization method for a zero-trust micro-segmentation strategy based on dynamic evolution of behavioral fingerprints with environmental decoupling, characterized in that, Includes the following steps: S1. Collect multi-dimensional behavioral data of the target entity within a preset time window, extract and encode features from the multi-dimensional behavioral data, and generate a short-cycle behavioral fingerprint vector of the target entity. S2. Obtain the dynamic trust assessment parameters of the target entity. The dynamic trust assessment parameters include at least the identity authentication status, device health score and network environment risk index. Calculate the comprehensive trust score of the target entity based on the short-cycle behavioral fingerprint vector and the dynamic trust assessment parameters. S3. Compare the comprehensive trust score with the preset multi-level trust threshold range, determine the current trust level of the target entity based on the comparison result, and adaptively generate the micro-isolation strategy of the target entity based on the current trust level and the anomaly degree of the short-cycle behavioral fingerprint vector. S4. The micro-segmentation strategy is sent to the execution node for execution, and the actual effect data of the target entity and the external environment status parameters of the execution node after the strategy is executed are continuously monitored. S5. Input the external environment state parameters into the environment state encoder to generate the current environment state anchor vector; decompose the actual effect data into a first effect component associated with the target entity behavior and a second effect component associated with the external environment state; S6. Calculate the environmental state offset between the current environmental state anchor vector and the historical environmental state anchor vector; determine the environmental attribution weight of the second effect component based on the environmental state offset; use the environmental attribution weight to extract the environmental contribution value from the second effect component to obtain the net effect value of the strategy. S7. Generate a feedback correction signal based on the net effect value of the strategy and the first effect component; use the feedback correction signal to correct the generation parameters of the short-cycle behavioral fingerprint vector and / or the calculation weight of the comprehensive trust score, so as to achieve environmental adaptive closed-loop optimization of the micro-isolation strategy.
2. The closed-loop optimization of the behavior fingerprint dynamic evolution zero-trust micro-segmentation strategy based on environment decoupling according to claim 1, characterized in that, The external environment state parameters mentioned in step S5 include: the real-time bandwidth utilization rate of the network where the execution node is located, the data packet round-trip delay jitter value, the CPU load rate and memory usage rate of the execution node; the environment state encoder is an encoding network based on a multilayer perceptron, which takes the normalized external environment state parameters as input and outputs a floating-point vector as the current environment state anchor vector, and the dimension of the floating-point vector is a preset fixed dimension.
3. The closed-loop optimization of the behavior fingerprint dynamic evolution zero-trust micro-segmentation strategy based on environment decoupling according to claim 1, characterized in that, Step S5 decomposes the actual effect data into a first effect component and a second effect component, specifically including: S51. Obtain the behavioral baseline characteristics of the target entity within a historical reference time period, wherein the behavioral baseline characteristics include the operation frequency baseline, resource access pattern baseline, and communication object baseline. S52. The portion of the actual effect data that has a statistical correlation with the behavioral baseline features is determined as the first effect component, and the portion of the actual effect data that has no statistical correlation with the behavioral baseline features but has a statistical correlation with the external environment state parameters is determined as the second effect component.
4. The closed-loop optimization of the behavior fingerprint dynamic evolution zero-trust micro-segmentation strategy based on environment decoupling according to claim 1, characterized in that, Step S6 calculates the environmental state offset, specifically including: S61. Calculate the difference between the current environmental state anchor vector and the historical environmental state anchor vector in each dimension to obtain the environmental state difference vector. S62. Take the arithmetic square root of the sum of the squares of each dimension of the environmental state difference vector to obtain the environmental state offset. ,in The first anchor point vector of the current environment state dimensional components, The first anchor point vector of historical environment state dimensional components, This represents the total number of dimensions of the anchor vector.
5. The closed-loop optimization of the behavior fingerprint dynamic evolution zero-trust micro-segmentation strategy based on environment decoupling according to claim 4, characterized in that, The environmental attribution weight Calculate using the following formula: ,in This is the preset maximum allowable environmental offset. The preset attenuation coefficient and ;when hour, The net effect value of the strategy ,in The total effect value of the actual effect data. This is the second effect component.
6. The closed-loop optimization of the behavior fingerprint dynamic evolution zero-trust micro-segmentation strategy based on environment decoupling according to claim 1, characterized in that, Step S3, which adaptively generates the micro-segmentation strategy for the target entity, specifically includes: S31. The comprehensive trust score is compared with a first trust threshold and a second trust threshold, wherein the first trust threshold is greater than the second trust threshold; S32. When the comprehensive trust score is greater than or equal to the first trust threshold, a release policy is generated that allows the target entity to access normally according to the current permissions. S33. When the comprehensive trust score is between the second trust threshold and the first trust threshold, a demotion strategy is generated to restrict the target entity's access to sensitive resources. S34. When the comprehensive trust score is less than or equal to the second trust threshold, an isolation strategy is generated to block all unnecessary network connections of the target entity.
7. The closed-loop optimization of the behavior fingerprint dynamic evolution zero-trust micro-segmentation strategy based on environment decoupling according to claim 1, characterized in that, In step S7, the generation parameters of the short-cycle behavioral fingerprint vector are corrected using the feedback correction signal, including: when the net effect value of the strategy is higher than a preset upper limit threshold of net effect, the length of the preset time window is shortened and the update frequency of the short-cycle behavioral fingerprint vector is increased; when the net effect value of the strategy is lower than a preset lower limit threshold of net effect, the length of the preset time window is extended and the update frequency of the short-cycle behavioral fingerprint vector is reduced.
8. The closed-loop optimization of the behavior fingerprint dynamic evolution zero-trust micro-segmentation strategy based on environment decoupling according to claim 1, characterized in that, Step S7 uses the feedback correction signal to correct the calculation weight of the comprehensive trust score, including: when the net effect value of the strategy is positive and the first effect component is higher than the preset component threshold, increasing the calculation weight of the device health score in the comprehensive trust score; when the net effect value of the strategy is negative and the first effect component is lower than the preset component threshold, decreasing the calculation weight of the network environment risk index in the comprehensive trust score.
9. The closed-loop optimization of the behavior fingerprint dynamic evolution zero-trust micro-segmentation strategy based on environment decoupling according to claim 1, characterized in that, The monitoring in step S4 adopts an event-driven asynchronous monitoring method. The actual effect data includes the number of successful blocking attempts, the number of false blocking attempts, the access response latency, and business continuity indicators after the policy is executed. The external environment status parameters are collected in parallel by independent sensor probes at a fixed sampling period, and the sampling period is shorter than the preset time window.
10. A closed-loop optimization system for behavior fingerprint dynamic evolution based on environment decoupling and zero-trust micro-segmentation strategy, characterized in that, include: The data acquisition module is used to collect multi-dimensional behavioral data of the target entity within a preset time window; The behavior fingerprint generation module is used to extract and encode features from the multidimensional behavior data to generate a short-period behavior fingerprint vector of the target entity. The trust assessment module is used to obtain the dynamic trust assessment parameters of the target entity and calculate the comprehensive trust score of the target entity based on the short-cycle behavioral fingerprint vector and the dynamic trust assessment parameters. The strategy generation module is used to compare the comprehensive trust score with a preset multi-level trust threshold range, determine the current trust level of the target entity based on the comparison result, and adaptively generate a micro-isolation strategy for the target entity based on the current trust level and the anomaly degree of the short-cycle behavioral fingerprint vector. The strategy execution and monitoring module is used to distribute the micro-segmentation strategy to the execution node for execution, and continuously monitor the actual effect data of the target entity and the external environment status parameters of the execution node after the strategy is executed; The environment decoupling module is used to input the external environment state parameters into the environment state encoder to generate the current environment state anchor vector, and decompose the actual effect data into a first effect component associated with the target entity behavior and a second effect component associated with the external environment state; The strategy effect stripping module is used to calculate the environmental state offset between the current environmental state anchor vector and the historical environmental state anchor vector, determine the environmental attribution weight of the second effect component based on the environmental state offset, and use the environmental attribution weight to strip the environmental contribution value from the second effect component to obtain the net strategy effect value. The closed-loop correction module is used to generate a feedback correction signal based on the net effect value of the strategy and the first effect component, and to use the feedback correction signal to correct the generation parameters of the short-cycle behavioral fingerprint vector and / or the calculation weight of the comprehensive trust score.
Citation Information
Patent Citations
Zero-trust network dynamic access control method based on AI behavior portrait
CN120768583A
AI zero-trust adaptive access control system suitable for sensitive data access
CN122247725A
SOAR-driven micro-isolation strategy dynamic arrangement method for zero-trust architecture
CN122496297A
