Network boundary extension sensing method and system based on AI model
By using multimodal data acquisition and dynamic topology mapping of AI models, combined with three-channel threat inference and closed-loop self-optimization, the problem of insufficient network boundary awareness in existing technologies is solved, achieving efficient and accurate threat detection and rapid response, adapting to the dynamic expansion and complex threats of enterprise networks.
Patent Information
- Application Number
- CN202511796682.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-03-10
AI Technical Summary
Existing technologies lack a full-domain dynamic boundary awareness architecture, suffer from fragmented multi-source data, and have insufficient threat detection capabilities. They cannot adapt to the dynamic expansion of enterprise network boundaries and complex threats, resulting in a high rate of missed detection of unknown threats and large protection blind spots.
An AI-based network boundary extension perception method is adopted, which achieves comprehensive perception of network boundaries and efficient threat handling through multimodal data acquisition, feature decoupling and fusion, dynamic topology mapping and three-channel threat inference, combined with a closed-loop self-optimization mechanism.
The accuracy of device authentication has been improved to 99.2%, the threat detection latency is less than 1 second, the false alarm rate has been reduced by 65%, the APT attack detection rate has been increased to 92%, and the threat handling time has been shortened from minutes to seconds, enhancing the system's resilience and adaptability.
Smart Images

Figure CN121644165A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network security technology, and in particular to a network boundary extension perception method and system based on an AI model. Background Technology
[0002] With the widespread adoption of cloud computing, mobile office, and the Internet of Things (IoT), enterprise network boundaries have shifted from fixed physical architectures to dynamically expanding digital spaces. The attack surface has significantly expanded due to factors such as multi-device access and open API interfaces, greatly increasing the difficulty of security protection. Traditional protection solutions have significant shortcomings: traditional firewalls rely on static rules, making it impossible to identify application-layer threats in encrypted traffic, and resulting in a delayed response to emerging threats such as AI-automated attacks and supply chain penetration; single-dimensional detection technologies lack the ability to fuse multi-source data, making it difficult to integrate heterogeneous information such as traffic, behavior, and device fingerprints to build a comprehensive risk profile, leading to a high rate of missed detections for unknown threats.
[0003] Existing systems generally suffer from passive defense limitations, lacking proactive early warning and cross-domain collaboration capabilities, and exhibiting severe data silos, making them unsuitable for the dynamic boundary protection needs of hybrid cloud and remote work scenarios. Meanwhile, threats such as ransomware and zero-day attacks mutate rapidly, making it difficult for traditional rule bases to keep up with their update frequency. Furthermore, network latency and weak infrastructure in special scenarios such as border areas and cross-cloud offices further exacerbate protection blind spots, necessitating breakthroughs in new boundary protection technologies with intelligent sensing and dynamic adaptation capabilities. Summary of the Invention
[0004] The purpose of this invention is to provide a network boundary extension perception method and system based on an AI model, aiming to solve the technical problems of existing technologies such as lack of a full-domain dynamic boundary perception architecture, fragmentation of multi-source data, and insufficient threat detection capabilities.
[0005] To address the aforementioned problems, according to one aspect of this application, an embodiment of the present invention provides a network boundary extension sensing method based on an AI model, comprising the following steps:
[0006] S1. Multimodal data acquisition: The perception agent module deployed at the network edge continuously collects the physical layer fingerprint information, protocol layer characteristics and behavioral layer traffic timing patterns of access devices to form a multi-dimensional raw dataset;
[0007] S2. Feature decoupling and fusion: The original dataset is processed by dividing it into frames according to a preset time window. After standardization and normalization, it is input into the feature decoupling encoder to extract the device identity association feature vector and the network behavior abnormal feature vector. The device-behavior joint representation vector is generated by weighted fusion through an attention mechanism.
[0008] S3. Dynamic Topology Mapping: Based on the spatiotemporal graph neural network, the joint representation vector flow uploaded by the network-wide perception agent is modeled, and a temporal topology graph sequence is constructed with node attributes as joint representation vectors and edge weights calculated by dual factors of communication frequency and protocol consistency, dynamically tracking the network boundary evolution state;
[0009] S4. Three-channel threat reasoning: Through three parallel channels—rule matching detection, unsupervised anomaly detection, and semi-supervised threat hunting—the temporal topology graph sequence and joint representation vector are collaboratively analyzed to generate multi-dimensional threat detection results.
[0010] S5. Decision Fusion and Response: A dynamic weighted voting mechanism is used to fuse the detection results of the three channels to generate a threat score. When the threat score exceeds a preset threshold, the corresponding automated defense response action is triggered.
[0011] S6. Closed-loop self-optimization: Collect feedback data after defense response, update threat inference model parameters and rule base through incremental learning algorithm, and synchronize to all perception agent modules to form a self-optimization closed loop.
[0012] In some implementations, in step S1, the physical layer fingerprint information includes a device hardware identifier, a radio frequency signal intensity distribution curve, power supply ripple spectrum characteristics, and electromagnetic leakage frequency domain characteristics. The electromagnetic leakage frequency domain characteristics are obtained by capturing the electromagnetic leakage signal when the device is working through a wideband antenna array and extracting the frequency domain characteristic peak position and amplitude distribution through Fourier transform.
[0013] In some implementations, in step S2, the feature decoupling encoder adopts a dual-branch convolutional neural network structure. The first branch extracts the device identity association feature vector, and the second branch extracts the network behavior abnormal feature vector. The attention mechanism is a channel attention mechanism, which achieves accurate fusion of the features of the two branches through adaptive weight allocation.
[0014] In some implementations, in step S3, the spatiotemporal graph neural network uses a gated recurrent unit structure to process temporal dependencies, and its adjacency matrix update mechanism introduces a decay factor, so that the influence of historical communication records on the current topology decreases exponentially over time.
[0015] In some implementations, in step S4, the rule base of the rule matching detection channel covers OWASP TopTen vulnerability exploitation features, MITRE ATT&CK tactical phase mapping table, and industrial control system protocol anomaly opcode set; the unsupervised anomaly detection channel uses a variational autoencoder to establish a probability distribution model of normal communication patterns and judges anomalies through reconstruction errors; the semi-supervised threat hunting channel uses a graph attention network and introduces three types of meta-path templates: "device-protocol-port", "user-session-resource", and "process-file-network" to mine potential attack chains.
[0016] In some implementations, in step S5, the initial weight of each channel in the dynamic weighted voting mechanism is set to 1 / 3, and the adaptive adjustment of the weight follows the following rules: for each false alarm event, the corresponding channel weight is multiplied by a decay coefficient of 0.95, and for each successful interception of an attack event, the corresponding channel weight is multiplied by a growth coefficient of 1.05. The weight adjustment cycle is 24 hours.
[0017] In some implementations, in step S6, the incremental learning algorithm employs elastic weight solidification technology to apply regularization constraints to important neuron connections when updating model parameters, preventing catastrophic forgetting; the feedback data includes false alarm samples, missed alarm events, and records of timely handling.
[0018] This invention also provides a network boundary extension perception system based on an AI model, used to implement any of the methods described above, including:
[0019] The edge perception layer consists of distributed intelligent perception agents. Each intelligent perception agent includes a data acquisition submodule, a local preprocessing submodule, and an encrypted transmission submodule. The data acquisition submodule integrates a hardware fingerprint sensor, a protocol parser, a traffic probe, and a wideband antenna array.
[0020] The central analysis layer includes a data aggregation server, a feature decoupling encoding cluster, a dynamic topology mapping server, a threat inference engine cluster, and a decision fusion controller.
[0021] The policy enforcement layer consists of the Network Functions Virtualization Manager, the Software-Defined Network Controller, and the Endpoint Security Management Agent.
[0022] The knowledge evolution module is located between the central analysis layer and the strategy execution layer; the model trust assurance unit is embedded in the threat inference engine cluster and includes an adversarial sample detector and a model interpreter.
[0023] In some implementations, the adversarial sample detector of the model trust assurance unit identifies perturbation patterns in the input data through gradient sensitivity analysis, and sets a first threshold to trigger an early warning mode and a second threshold to activate an emergency response; the model interpreter uses a layer correlation propagation algorithm to generate a visual heatmap of the threat determination criteria.
[0024] In some implementations, the central analysis layer also integrates a causal reasoning engine for constructing a dynamic causal graph representing the causal relationships between threat events and performing threat attribution analysis based on counterfactual reasoning. The system uses a cloud-edge-device collaborative reasoning framework, where the cloud server performs complex reasoning of the global model, and the edge nodes handle local reasoning tasks with high real-time requirements.
[0025] Compared with existing technologies, the network boundary extension perception method and system based on AI models of the present invention have at least the following beneficial effects:
[0026] Multimodal data acquisition improves the accuracy of device authentication to 99.2%;
[0027] Spatiotemporal graph neural networks achieve boundary change perception latency of less than 1 second;
[0028] The three-channel detection mechanism reduces the false alarm rate by 65% and increases the APT attack detection rate to 92%.
[0029] Automated response reduces threat handling time from minutes to seconds;
[0030] Closed-loop self-optimization and reliable assurance enhance system resilience.
[0031] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description
[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0033] Figure 1 System overall architecture diagram;
[0034] Figure 2 : The principle of dynamic topology mapping;
[0035] Figure 3 Three-channel threat reasoning flowchart;
[0036] Figure 4 Data acquisition and feature decoupling flowchart;
[0037] Figure 5 : Closed-loop self-optimizing interaction relationship diagram;
[0038] Figure 6 Diagram of the model's credibility assurance mechanism. Detailed Implementation
[0039] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the specific embodiments, structures, features, and effects according to the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "an embodiment" or "an embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0040] In the description of this invention, it should be clearly stated that the terms "first," "second," etc., in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence; the terms "vertical," "lateral," "longitudinal," "front," "rear," "left," "right," "up," "down," "horizontal," etc., indicate orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, and are merely for the convenience of describing this invention, and do not mean that the device or element referred to must have a specific orientation or position, and therefore should not be construed as a limitation of this invention.
[0041] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0042] like Figures 1-6 As shown, this embodiment of the invention provides a network boundary extension perception method based on an AI model, including the following steps:
[0043] S1. Multimodal data acquisition: The perception agent module deployed at the network edge continuously collects the physical layer fingerprint information, protocol layer characteristics and behavioral layer traffic timing patterns of access devices to form a multi-dimensional raw dataset;
[0044] S2. Feature decoupling and fusion: The original dataset is processed by dividing it into frames according to a preset time window. After standardization and normalization, it is input into the feature decoupling encoder to extract the device identity association feature vector and the network behavior abnormal feature vector. The device-behavior joint representation vector is generated by weighted fusion through an attention mechanism.
[0045] S3. Dynamic Topology Mapping: Based on the spatiotemporal graph neural network, the joint representation vector flow uploaded by the network-wide perception agent is modeled, and a temporal topology graph sequence is constructed with node attributes as joint representation vectors and edge weights calculated by dual factors of communication frequency and protocol consistency, dynamically tracking the network boundary evolution state;
[0046] S4. Three-channel threat reasoning: Through three parallel channels—rule matching detection, unsupervised anomaly detection, and semi-supervised threat hunting—the temporal topology graph sequence and joint representation vector are collaboratively analyzed to generate multi-dimensional threat detection results.
[0047] S5. Decision Fusion and Response: A dynamic weighted voting mechanism is used to fuse the detection results of the three channels to generate a threat score. When the threat score exceeds a preset threshold, the corresponding automated defense response action is triggered.
[0048] S6. Closed-loop self-optimization: Collect feedback data after defense response, update threat inference model parameters and rule base through incremental learning algorithm, and synchronize to all perception agent modules to form a self-optimization closed loop.
[0049] In this embodiment, the perception agent module deployed at the network edge continuously captures the physical layer fingerprint information, protocol layer features, and behavioral layer traffic time-series patterns of access devices, integrating them to form a multi-dimensional raw dataset. This raw data is processed in frames according to a preset time window, standardized, and normalized before being input into a feature decoupling encoder. Device identity association feature vectors and network behavior anomaly feature vectors are extracted, and weighted fusion using an attention mechanism yields a device-behavior joint representation vector. A spatiotemporal graph neural network is used to model the joint representation vector stream uploaded by the perception agents across the entire network, constructing a temporal topology sequence where node attributes are joint representation vectors and edge weights are calculated from both communication frequency and protocol consistency factors, dynamically tracking the network boundary evolution. The temporal topology sequence and joint representation vectors are collaboratively analyzed through three parallel channels: rule matching detection, unsupervised anomaly detection, and semi-supervised threat hunting, generating multi-dimensional threat detection results. A dynamic weighted voting mechanism is used to fuse the three-channel detection results and generate a threat score. When the score exceeds a preset threshold, an automated defense response is triggered. Feedback data after the defense response is collected, and the threat inference model parameters and rule base are updated using an incremental learning algorithm and synchronized to all perception agent modules, forming a self-optimizing closed loop. This process enables comprehensive perception and efficient threat handling of dynamic network boundaries, significantly improving the comprehensiveness and timeliness of threat detection, and adapting to the new protection needs of dynamic and fragmented network boundaries.
[0050] In some implementations, in step S1, the physical layer fingerprint information includes a device hardware identifier, a radio frequency signal intensity distribution curve, power supply ripple spectrum characteristics, and electromagnetic leakage frequency domain characteristics. The electromagnetic leakage frequency domain characteristics are obtained by capturing the electromagnetic leakage signal when the device is working through a wideband antenna array and extracting the frequency domain characteristic peak position and amplitude distribution through Fourier transform.
[0051] In this embodiment, when collecting physical layer fingerprint information of access devices, the perception agent module, in addition to acquiring device hardware identifiers, radio frequency signal strength distribution curves, and power supply ripple spectrum characteristics, also captures electromagnetic leakage signals during device operation through a wideband antenna array. It then extracts the frequency domain feature peak positions and amplitude distributions via Fourier transform, thus enriching the dimensions of the physical layer fingerprint information. This rich physical layer fingerprint information provides a more comprehensive and unique basis for device authentication, effectively avoiding the problem of single fingerprint information being easily forged or tampered with. This makes device identification more accurate and reliable, laying a solid foundation for subsequent threat detection and enhancing the security and anti-attack capabilities of the entire network boundary perception system.
[0052] In some implementations, in step S2, the feature decoupling encoder adopts a dual-branch convolutional neural network structure. The first branch extracts the device identity association feature vector, and the second branch extracts the network behavior abnormal feature vector. The attention mechanism is a channel attention mechanism, which achieves accurate fusion of the features of the two branches through adaptive weight allocation.
[0053] In this embodiment, the feature decoupling encoder employs a dual-branch convolutional neural network structure. The first branch focuses on extracting device identity-related feature vectors, while the second branch specifically extracts network behavior anomaly feature vectors. A channel attention mechanism is introduced, using adaptive weight allocation to precisely fuse the feature vectors output from the two branches, generating a device-behavior joint representation vector. This combination of a dual-branch structure and attention mechanism achieves effective separation and efficient fusion of device identity and behavioral features. This ensures accurate device identity recognition while precisely capturing abnormal signs in network behavior, allowing the joint representation vector to more comprehensively reflect device status and network behavior. This provides high-quality feature input for subsequent threat inference and improves the accuracy of threat detection.
[0054] In some implementations, in step S3, the spatiotemporal graph neural network uses a gated recurrent unit structure to process temporal dependencies, and its adjacency matrix update mechanism introduces a decay factor, so that the influence of historical communication records on the current topology decreases exponentially over time.
[0055] In this embodiment, the spatiotemporal graph neural network employs a gated recurrent unit structure to handle temporal dependencies. A decay factor is introduced into the adjacency matrix update mechanism, causing the influence of historical communication records on the current topology to decrease exponentially over time. Based on this neural network, a temporal topology graph sequence is constructed by modeling the joint representation vector flow uploaded by the network-wide perceptual agents, dynamically tracking the evolution of the network boundary. The gated recurrent unit structure ensures effective capture of temporal data dependencies, and the introduction of the decay factor ensures that the topology graph sequence reflects the latest state of the network boundary in real time, avoiding interference from historical data in current boundary judgments. This allows the system to accurately adapt to the dynamic changes of the network boundary and promptly detect potential threats during boundary expansion or contraction.
[0056] In some implementations, in step S4, the rule base of the rule matching detection channel covers OWASP TopTen vulnerability exploitation features, MITRE ATT&CK tactical phase mapping table, and industrial control system protocol anomaly opcode set; the unsupervised anomaly detection channel uses a variational autoencoder to establish a probability distribution model of normal communication patterns and judges anomalies through reconstruction errors; the semi-supervised threat hunting channel uses a graph attention network and introduces three types of meta-path templates: "device-protocol-port", "user-session-resource", and "process-file-network" to mine potential attack chains.
[0057] In this embodiment, the rule-matching detection channel relies on a rule base covering OWASP Top Ten vulnerability exploitation features, the MITREATT&CK tactical phase mapping table, and the set of abnormal opcodes for industrial control system protocols to quickly match known attack patterns. The unsupervised anomaly detection channel uses a variational autoencoder to establish a probability distribution model of normal communication patterns and judges abnormal behavior through reconstruction errors. The semi-supervised threat hunting channel uses a graph attention network and introduces three types of meta-path templates: "device-protocol-port," "user-session-resource," and "process-file-network," to mine potential attack chains. The three channels work in parallel, collaboratively analyzing the temporal topology graph sequence and joint representation vector to generate multi-dimensional threat detection results. This mechanism takes into account the rapid identification of known attacks, the effective discovery of unknown anomalies, and the in-depth mining of potential attack chains, significantly improving the coverage and accuracy of threat detection and effectively addressing various complex network threats.
[0058] In some implementations, in step S5, the initial weight of each channel in the dynamic weighted voting mechanism is set to 1 / 3, and the adaptive adjustment of the weight follows the following rules: for each false alarm event, the corresponding channel weight is multiplied by a decay coefficient of 0.95, and for each successful interception of an attack event, the corresponding channel weight is multiplied by a growth coefficient of 1.05. The weight adjustment cycle is 24 hours.
[0059] In this embodiment, a dynamic weighted voting mechanism assigns an initial weight of 1 / 3 to each of the three channels: rule-matching detection, unsupervised anomaly detection, and semi-supervised threat hunting. The weights are adaptively adjusted according to specific rules: for each false positive, the corresponding channel weight is multiplied by a decay coefficient of 0.95; for each successful interception of an attack, the corresponding channel weight is multiplied by a growth coefficient of 1.05. The weight adjustment cycle is 24 hours. Based on the adjusted weights, the detection results from the three channels are fused to generate a threat score. When the score exceeds a preset threshold, a corresponding automated defense response is triggered. This dynamic weight adjustment mechanism optimizes the decision weights based on the actual detection performance of each channel, improving the rationality and accuracy of decision fusion. The automated response significantly shortens threat handling time and reduces the damage caused by threats to network security.
[0060] In some implementations, in step S6, the incremental learning algorithm employs elastic weight solidification technology to apply regularization constraints to important neuron connections when updating model parameters, preventing catastrophic forgetting; the feedback data includes false alarm samples, missed alarm events, and records of timely handling.
[0061] In this embodiment, feedback data such as false positives, missed detections, and timely handling records are collected after the defense response. The incremental learning algorithm employs elastic weight solidification technology, applying regularization constraints to important neuron connections when updating threat inference model parameters to prevent catastrophic forgetting. The updated model parameters and rule base are synchronized to all perception agent modules, forming a self-optimizing closed loop. This process allows the system to continuously absorb lessons learned from defense practices, constantly optimize model performance and rule base integrity, improve the detection and handling capabilities of new threats, and gradually enhance the system's defense level over time, thereby strengthening the system's long-term adaptability and security.
[0062] This invention also provides a network boundary extension perception system based on an AI model, used to implement any of the methods described above, including:
[0063] The edge perception layer consists of distributed intelligent perception agents. Each intelligent perception agent includes a data acquisition submodule, a local preprocessing submodule, and an encrypted transmission submodule. The data acquisition submodule integrates a hardware fingerprint sensor, a protocol parser, a traffic probe, and a wideband antenna array.
[0064] The central analysis layer includes a data aggregation server, a feature decoupling encoding cluster, a dynamic topology mapping server, a threat inference engine cluster, and a decision fusion controller.
[0065] The policy enforcement layer consists of the Network Functions Virtualization Manager, the Software-Defined Network Controller, and the Endpoint Security Management Agent.
[0066] The knowledge evolution module is located between the central analysis layer and the strategy execution layer; the model trust assurance unit is embedded in the threat inference engine cluster and includes an adversarial sample detector and a model interpreter.
[0067] In this embodiment, the distributed intelligent sensing agent of the edge perception layer collects multi-dimensional data through a hardware fingerprint sensor, protocol parser, traffic probe, and wideband antenna array integrated in the data acquisition submodule. After processing by the local preprocessing submodule, the data is uploaded by the encrypted transmission submodule. The servers and clusters in the central analysis layer aggregate the data, decouple and encode features, perform dynamic topology mapping, threat inference, and decision fusion. The relevant managers and agents in the policy execution layer execute defensive actions according to decision instructions. The knowledge evolution module updates and optimizes the model and rule base between the central analysis layer and the policy execution layer. The model trust assurance unit is embedded in the threat inference engine cluster, providing adversarial sample detection and model interpretation functions. The collaborative work of each layer and module constructs a comprehensive and intelligent network boundary extension perception system, realizing a closed-loop process from data acquisition and analysis inference to defense response and self-optimization, significantly improving the security protection capabilities of the network boundary.
[0068] The encrypted transmission submodule adds timestamp binding and sequence number verification mechanisms to the national cryptographic SM4 algorithm. The receiving end verifies the arrival order and timeliness of data packets and discards data packets with a delay of more than 5 seconds or with discontinuous sequence numbers, effectively preventing replay attacks and man-in-the-middle tampering.
[0069] In some implementations, the adversarial sample detector of the model trust assurance unit identifies perturbation patterns in the input data through gradient sensitivity analysis, and sets a first threshold to trigger an early warning mode and a second threshold to activate an emergency response; the model interpreter uses a layer correlation propagation algorithm to generate a visual heatmap of the threat determination criteria.
[0070] In this embodiment, the adversarial example detector of the model credibility assurance unit identifies perturbation patterns in the input data through gradient sensitivity analysis, and sets a first threshold and a second threshold to trigger the early warning mode and activate the emergency response, respectively. The model interpreter uses a layer correlation propagation algorithm to generate a visual heatmap of the threat determination criteria. The adversarial example detector can effectively identify and respond to adversarial attacks, ensuring the security and stability of model operation; the visual heatmap allows maintenance personnel to clearly understand the AI's decision-making logic, enhancing the credibility and supervisory nature of AI decisions, and facilitating the timely detection and correction of model decision-making biases.
[0071] In some implementations, the central analysis layer also integrates a causal reasoning engine for constructing a dynamic causal graph representing the causal relationships between threat events and performing threat attribution analysis based on counterfactual reasoning. The system uses a cloud-edge-device collaborative reasoning framework, where the cloud server performs complex reasoning of the global model, and the edge nodes handle local reasoning tasks with high real-time requirements.
[0072] In this embodiment, the causal inference engine of the central analysis layer constructs a dynamic causal graph representing the causal relationships between threat events, and performs threat attribution analysis based on counterfactual reasoning. The system adopts a cloud-edge-device collaborative inference framework, where the cloud server executes complex inference tasks for the global model, and edge nodes focus on handling local inference tasks with high real-time requirements. The causal inference engine improves the accuracy of threat attribution, helping to trace the source of threats and take targeted defensive measures; the cloud-edge-device collaborative architecture balances the complexity and real-time performance of inference, ensuring both the depth of global threat analysis and meeting the needs of local real-time threat handling, thus optimizing the overall system operating efficiency and response performance.
[0073] The data aggregation server uses a ring buffer structure to store the raw data stream. The buffer capacity is calculated by multiplying the maximum number of concurrent sensing agents by the peak data rate of a single agent and then by a 10-minute duration. When the buffer occupancy rate reaches 80%, the data compression module is automatically activated. An improved LZ77 algorithm is used to dictionary-encode repeating feature vectors to ensure the real-time performance and integrity of data transmission.
[0074] The physical layer fingerprint information includes the electromagnetic leakage signal captured by the wideband antenna array during device operation, the frequency domain feature peak position and amplitude distribution extracted by Fourier transform; and the addition of thermal fingerprint acquisition based on the device's operating temperature fluctuation curve, recording temperature time series data through an infrared sensing module, combining hardware fingerprint and radio frequency features for triple cross-verification, using Kalman filtering algorithm to remove environmental interference data, and introducing a fingerprint feature timeliness marking mechanism.
[0075] The feature decoupling encoder adopts a dual-branch convolutional neural network structure. The first branch extracts device identity-related feature vectors, and the second branch extracts network behavior abnormal feature vectors. The attention mechanism is a channel attention mechanism. Furthermore, an adversarial training mechanism and Bayesian inference are introduced. By generating adversarial networks, effective features are distinguished from noise. The feature selection threshold is dynamically optimized by combining Shannon entropy values, and a feature stability verification step is added.
[0076] The spatiotemporal graph neural network uses a gated recurrent unit structure to handle temporal dependencies. Its adjacency matrix update mechanism introduces a decay factor, which makes the influence of historical communication records on the current topology decrease exponentially over time. It also adds a dynamic expansion / shrinkage mechanism for topology nodes based on network load thresholds, optimizes resource usage through non-intrusive pruning technology, and combines dynamic relationship graph modeling to model potential associations between nodes, and adds a new topology structure anomaly early warning threshold.
[0077] The rule base of the rule matching detection channel covers OWASP Top Ten vulnerability exploitation features, MITRE ATT&CK tactical phase mapping table, and industrial control system protocol anomaly opcode set; the unsupervised anomaly detection channel uses variational autoencoders to establish a probability distribution model of normal communication patterns and judges anomalies through reconstruction errors; the semi-supervised threat hunting channel uses graph attention networks and introduces three types of meta-path templates: "device-protocol-port", "user-session-resource", and "process-file-network" to mine potential attack chains; and adds a triple weight bias strategy based on threat type, business scenario sensitivity, and attack chain stage to increase the unsupervised detection weight for zero-day threats and add risk correction coefficients for core business scenarios.
[0078] The initial weight of each channel in the dynamic weighted voting mechanism is set to 1 / 3. The adaptive adjustment of the weight follows the following rules: for each false alarm event, the corresponding channel weight is multiplied by a decay coefficient of 0.95; for each successful interception of an attack event, the corresponding channel weight is multiplied by a growth coefficient of 1.05. The weight adjustment cycle is 24 hours. Multi-agent reinforcement learning is introduced to optimize the weight adjustment logic, and a new defense strategy sequence generation module is added to automatically output hierarchical response actions.
[0079] The following are some specific embodiments of the present invention:
[0080] Example 1: Security Protection of Power System Distribution Network
[0081] One hundred intelligent sensing agents are deployed at the edge of the power distribution network to collect physical fingerprints such as hardware identifiers, radio frequency signal strength, and power ripple spectrum of RTU devices, and to capture electromagnetic leakage signals through a wideband antenna array. A protocol parser processes port activity sequences according to the IEC 61850 protocol, and traffic probes collect data packet distribution characteristics.
[0082] A 5-second time window is set for data frame segmentation. After zero-mean unit variance standardization, the data is input into a dual-branch CNN feature decoupling encoder. The first branch outputs 256-dimensional device identity features, and the second branch outputs 128-dimensional behavioral anomaly features. These features are then fused into a 384-dimensional joint representation vector through a channel attention mechanism.
[0083] The dynamic topology mapping server adopts a spatiotemporal graph neural network with a gated cyclic unit structure. The edge weights are calculated based on the communication frequency (60%) and protocol consistency (40%), with a decay factor set to 0.9. The topology graph sequence is updated every 1 second.
[0084] The threat inference engine employs a three-channel parallel processing approach: the rule matching channel loads rule bases such as OWASP Top Ten and MITRE ATT&CK; the unsupervised anomaly detection channel uses a variational autoencoder with a reconstruction error threshold of 0.3; and the semi-supervised threat hunting channel uses a graph attention network to analyze attack chains along three types of meta-path templates.
[0085] The decision fusion controller employs dynamic weighted voting, with initial weights of 1 / 3 for each event. False alarm weights decay by 0.95, while successful interception weights increase by 1.05, with a 24-hour adjustment cycle. When a threat score exceeds 0.7, it triggers responses such as SDN flow table entry isolation and process freezing.
[0086] The knowledge evolution module updates model parameters through elastic weight solidification technology and synchronizes them to all perceptual agents, forming a self-optimizing closed loop.
[0087] Example 2: System Hardware Configuration
[0088] The edge sensing layer uses the FPC1020 hardware fingerprint sensor and STM32H743 microcontroller, supports IEC 61850 and Modbus protocol parsing, and uses the national cryptographic SM4 algorithm for encrypted transmission.
[0089] The central analysis layer is configured with an Intel Xeon Gold 6330 processor, an NVIDIA A100 graphics card cluster, a 60TB circular buffer, and uses a Redis cluster to maintain the topology graph sequence.
[0090] The policy enforcement layer deploys the NFV manager based on the OpenStack cloud platform, and the SDN controller supports the OpenFlow 1.3 protocol.
[0091] The model credibility assurance unit is equipped with a dual-threshold detection mechanism (early warning threshold 0.2, emergency threshold 0.5), and generates a visual heat map through the layer correlation propagation algorithm.
[0092] The software-defined network controller supports the OpenFlow 1.3 protocol. Its flow table update operation includes a matching field that precisely matches the source and destination IP addresses, transport layer protocol type, and destination port number; an action field that includes forwarding to a specified port, modifying the VLAN tag, and setting queue priority; and a timeout field that sets a dual timer with a hard timeout of 30 seconds and an idle timeout of 10 seconds, enabling fine-grained flow control.
[0093] The endpoint security management agent has a built-in lightweight sandbox environment. When it receives a process freeze command, it migrates the target process to the sandbox to continue running and monitors its subsequent behavior. If no further malicious operations are detected within 60 seconds, the freeze is automatically lifted, reducing interference with legitimate business operations and balancing security protection with business continuity requirements.
[0094] The knowledge evolution module verifies and synchronizes new attack signatures to the local rule caches of all edge-aware agents, forming a distributed rule update mechanism. The synchronization process employs differential transmission technology, transmitting only the rule changes to reduce network bandwidth consumption and ensure that edge nodes can promptly acquire the latest threat detection capabilities.
[0095] Example 3: Multimodal fingerprint enhancement acquisition and verification
[0096] Building upon Example 1, the sensing agent module further integrates an infrared sensing module to collect temperature fluctuation curves during device operation, forming a thermal fingerprint. This thermal fingerprint, along with traditional physical fingerprints such as hardware identifiers and radio frequency signals, is input into a triple cross-validation module. This module utilizes a Kalman filter algorithm to smooth the temperature time-series data, effectively eliminating transient environmental interference. Simultaneously, the system adds a timestamp and validity period marker to each fingerprint feature, constituting a fingerprint feature timeliness marking mechanism. When a device fingerprint exceeds its validity period or exhibits abnormal fluctuations, the system will trigger re-authentication. This mechanism significantly enhances the anti-counterfeiting capabilities of device authentication and its robustness in complex industrial environments.
[0097] Example 4: Feature Processing and Adversarial Purification
[0098] In the feature decoupling and fusion stage, the dual-branch CNN encoder introduces an adversarial training mechanism during training. Specifically, a generative adversarial network (GAN) discriminator structure is used to help distinguish effective features from noise in the input data. During training, Bayesian inference is used to model the uncertainty of features and calculate the Shannon entropy of the features. The system dynamically adjusts the feature selection threshold based on the entropy value; features with excessively high entropy values (high feature disorder) are assigned lower weights or eliminated. In addition, a newly added adversarial feature cleansing module cleans feature vectors suspected of being maliciously interfered with before feature fusion. The cleansing threshold of this module is positively correlated with the number of adversarial training iterations, ensuring the stability of the model when facing adversarial examples.
[0099] Example 5: Federated Learning and Privacy Protection Optimization
[0100] In the closed-loop self-optimization phase, in addition to employing Elastic Weight Consolidation (EWC) technology at the center, the system also introduces a federated learning framework to protect edge data privacy. Each edge-aware agent calculates the model gradient locally using its own data, and then splits the gradient into multiple shares using a secret share generation algorithm. Before transmission, the system performs hierarchical noise injection, adding more noise to the gradients corresponding to highly sensitive features. All gradient shares to be uploaded must undergo zero-knowledge auditing to generate evidence, ensuring that the data has not been maliciously tampered with. The central server collects and aggregates the encrypted shares, obtaining only the overall model update, and cannot reverse-engineer the original data. This process uses gradient integrity check codes to prevent parameter tampering during transmission, achieving efficient collaborative learning under privacy protection.
[0101] Example 6: Cloud-Edge-Device Collaboration and Secure Communication
[0102] The system constructs a cloud-edge-device collaborative inference framework. Edge nodes deploy lightweight models optimized through channel pruning, responsible for initial detection and response with high real-time requirements. The cloud deploys a complete, high-precision model incorporating generative AI, responsible for in-depth attribution of complex attacks and global situational analysis. To achieve secure and efficient collaboration, the system deploys encrypted tunnels based on Software Defined Boundaries (SDP) on all communication links. This mechanism employs memory optimization techniques to reduce encryption / decryption overhead and utilizes heterogeneous network adaptation algorithms to dynamically select the optimal transmission path, ensuring low-latency end-to-end encrypted communication across various access methods such as Wi-Fi, 5G, and wired networks. Protocol adaptive switching latency is ≤10ms, guaranteeing the real-time performance and security of command and data transmission.
[0103] Example 7: Causal Reasoning and Dynamic Attribution
[0104] At the central analysis layer, a causal reasoning engine is integrated. Based on collected threat event data, this engine constructs a dynamic causal graph, where nodes represent security events (such as abnormal logins or suspicious file access), and edges represent the causal probabilities between events. When a security event occurs, the engine performs counterfactual reasoning, such as: "If this device had not been previously implanted with a backdoor, what is the probability of this abnormal external connection?" By comparing the differences between fact and counterfactual, the engine accurately pinpoints the root cause of the threat, enabling threat attribution analysis. This helps security operations personnel not only know "what happened," but also understand "why it happened," thereby developing more targeted defense strategies.
[0105] The algorithm employs a federated learning framework combined with dynamic differential privacy technology and the SM9 key exchange algorithm. It introduces a secret share generation algorithm to split local gradient shares, protects sensitive features through hierarchical noise injection, uploads only desensitized model parameters that have been audited and certified by zero knowledge, adds a gradient integrity check code to prevent parameter tampering, and integrates counterfactual reasoning to verify the rationality of parameters. The verification threshold of counterfactual reasoning is dynamically linked to the differential privacy noise level, with a linkage response delay of ≤20ms.
[0106] The system adds thermal fingerprint acquisition based on the temperature fluctuation curve of the equipment, records temperature time series data through infrared sensing module, performs triple cross-verification by combining hardware fingerprint and radio frequency features, uses Kalman filtering algorithm to remove environmental interference data, introduces fingerprint feature timeliness marking mechanism, and superimposes four-dimensional spatiotemporal feature fusion to improve anti-interference capability. The spatiotemporal fusion weight is adjusted in real time based on the equipment operating conditions, with an adjustment step size of 0.01 to 0.1.
[0107] An adversarial training mechanism and Bayesian inference are introduced. By generating adversarial networks, effective features are distinguished from noise. The feature selection threshold is dynamically optimized by combining Shannon entropy values. A feature stability verification step is added. Computational resources are allocated based on feature importance weights. An adversarial feature purification module is added to remove malicious interference features. The purification threshold is positively correlated with the number of adversarial training iterations. The number of iterations is set to 50-200 rounds.
[0108] A dynamic scaling-up / scaling mechanism for topology nodes based on network load thresholds is added. Resource consumption is optimized through non-intrusive pruning technology. The potential associations between nodes are modeled using dynamic relationship graphs, and the dynamic attributes of node relationships are quantified. An early warning threshold for topology structure anomalies is added. Hypergraph modeling is integrated to characterize multi-dimensional node associations, and the hypergraph edge weights are dynamically assigned based on the attack path probability, with the assignment precision retained to two decimal places.
[0109] It adds least privilege access behavior verification rules based on zero trust architecture and a real-time context analysis engine, embeds a continuous trust assessment mechanism to dynamically update access permissions, optimizes rule matching accuracy by combining risk semantic identifier library, introduces an automatic rule library iteration update module, achieves cross-domain rule adaptation through AI semantic recognition, and the semantic recognition model is fine-tuned and optimized through domain datasets, with fine-tuning dataset coverage ≥85%.
[0110] A triple weighting bias strategy based on threat type, business scenario sensitivity, and attack chain stage is added. The unsupervised detection weight is increased for zero-day threats, and a risk correction coefficient is added for core business scenarios. Multi-agent reinforcement learning is introduced to optimize the weight adjustment logic. The decision priority is optimized by combining the threat pre-allocation model. A new defense strategy sequence generation module is added to automatically output graded response actions, and the response actions are dynamically correlated with business interruption losses, with loss quantification accuracy reaching tens of thousands of yuan.
[0111] We have added robust training for adversarial examples and real-time scanning by a virtual browser, regularly injected simulated attack samples, combined model iteration vulnerability detection mechanism and version difference verification to optimize defense strategies, added quantitative evaluation indicators for optimization effects, introduced human-machine collaborative verification closed loop to correct optimization direction, and prioritized manual verification based on the confidence level of optimization effects, with the confidence level threshold set at 0.75.
[0112] It adds a software-defined boundary (SDP) based encrypted tunnel transmission mechanism and end-to-cloud collaborative communication middleware, adopts memory optimization technology and heterogeneous network adaptation algorithm to improve transmission efficiency, realizes low-latency end-to-end encrypted communication between the perception agent and the central analysis layer, integrates electro-optic phase feedback chaotic encryption to enhance transmission security, and superimposes protocol adaptive technology to adapt to multiple types of network access, with adaptive protocol switching latency ≤10ms and switching success rate ≥99.9%.
[0113] The system adds a compliance verification module and a semantic analysis engine for model decision results, outputs compliance scores against cybersecurity regulations, generates compliance audit reports with logical derivations, verifies the rationality of the model through pre-calculation of fixed values, introduces a pre-trained AI model to build a review logic database, adds AI-generated content trace detection to avoid compliance risks, and links the detection results with the compliance score for correction, with a linkage correction coefficient of 0.8 to 1.2.
[0114] Lightweight models optimized by channel pruning are deployed at edge nodes for real-time preliminary detection, while high-precision models that integrate generative AI are deployed in the cloud for in-depth tracing. By caching hotspot threat features at the edge and collaborating with middleware, gradient data is transmitted in encrypted form using secret shares to improve response speed and security. A dynamic weight allocation mechanism is added to enable effective participation of edge nodes. Cloud-edge-size models are introduced to collaboratively optimize global accuracy, with collaborative inference latency ≤50ms and inference accuracy ≥98%.
[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. An AI model-based network boundary expansion awareness method, characterized by, The method comprises the following steps: S1. Multimodal data acquisition: The perception agent module deployed at the network edge continuously collects physical layer fingerprint information, protocol layer features and behavior layer traffic timing patterns of access devices to form a multidimensional original data set; S2. Feature decoupling and fusion: The original data set is processed by frame according to the preset time window, and after normalization, it is input into the feature decoupling encoder to extract device identity associated feature vectors and network behavior anomaly feature vectors, and a device-behavior joint representation vector is generated by weighted fusion through an attention mechanism; S3. Dynamic topology mapping: Based on a spatio-temporal graph neural network, the joint representation vector stream uploaded by the perception agent of the whole network is modeled, a time sequence topology graph sequence is constructed with the node attribute being the joint representation vector and the edge weight being calculated by the communication frequency and protocol consistency double factors, and the evolution state of the network boundary is dynamically tracked; S4. Three-channel threat reasoning: Through rule matching detection, unsupervised anomaly detection and semi-supervised threat hunting, the time sequence topology graph sequence and the joint representation vector are analyzed cooperatively to generate multidimensional threat detection results; S5. Decision fusion and response: The three-channel detection results are fused by using a dynamic weighted voting mechanism to generate a threat score, and when the threat score exceeds a preset threshold, the corresponding automatic defense response action is triggered; S6. Closed-loop self-optimization: Feedback data after defense response is collected, the threat reasoning model parameters and rule base are updated through an incremental learning algorithm, and are synchronized to all perception agent modules to form a self-optimizing closed loop.
2. The method of claim 1, wherein, In step S1, the physical layer fingerprint information includes device hardware identifier, radio frequency signal strength distribution curve, power supply ripple spectrum feature and electromagnetic leakage frequency domain feature, the electromagnetic leakage frequency domain feature is obtained by capturing the electromagnetic leakage signal of the device during operation through a wideband antenna array, and the frequency domain feature peak position and amplitude distribution are extracted through Fourier transform.
3. The method of claim 1, wherein, In step S2, the feature decoupling encoder adopts a double-branch convolutional neural network structure, the first branch extracts the device identity associated feature vector, the second branch extracts the network behavior anomaly feature vector, and the attention mechanism is a channel attention mechanism, which realizes accurate fusion of the two branch features through adaptive weight distribution.
4. The method of claim 1, wherein, In step S3, the spatio-temporal graph neural network adopts a gated recurrent unit structure to process the time sequence dependence relationship, and the adjacency matrix update mechanism introduces a decay factor, so that the influence of historical communication records on the current topology decreases exponentially over time.
5. The method of claim 1, wherein, In step S4, the rule matching detection channel rule base covers OWASP Top Ten vulnerability feature, MITRE ATT&CK tactic phase mapping table and industrial control system protocol abnormal operation code set; the unsupervised anomaly detection channel adopts a variational autoencoder to establish a probability distribution model of normal communication mode, and judges the anomaly through reconstruction error; the semi-supervised threat hunting channel adopts a graph attention network, introduces three types of meta-path templates of "device-protocol-port", "user-session-resource" and "process-file-network", and mines potential attack chains.
6. The method of claim 1, wherein, In step S5, the initial value of each channel weight of the dynamic weighted voting mechanism is set to 1 / 3, and the adaptive adjustment of the weight follows the following rules: each time a false positive event occurs, the corresponding channel weight is multiplied by a decay coefficient of 0.95, and each time an attack event is successfully intercepted, the corresponding channel weight is multiplied by a growth coefficient of 1.05, and the weight adjustment period is 24 hours.
7. The method of claim 1, wherein, In step S6, the incremental learning algorithm uses the elastic weight solidification technology to apply regularization constraints to important neuron connections when updating model parameters to prevent catastrophic forgetting; the feedback data includes false positive samples, missed events, and handling time limit records.
8. An AI model based network boundary extension awareness system for implementing the method of any one of claims 1-7, characterized by, Comprise: The edge perception layer is composed of distributed intelligent perception agents, each of which includes a data acquisition submodule, a local preprocessing submodule, and an encrypted transmission submodule. The data acquisition submodule integrates hardware fingerprint sensors, protocol analyzers, traffic probes, and wideband antenna arrays. The central analysis layer includes data aggregation servers, feature decoupling coding clusters, dynamic topology mapping servers, threat reasoning engine clusters, and decision fusion controllers. The policy execution layer is composed of network function virtualization managers, software-defined network controllers, and terminal security management agents. The knowledge evolution module is set between the central analysis layer and the policy execution layer. The model trust assurance unit is embedded in the threat reasoning engine cluster and includes an adversarial sample detector and a model interpreter.
9. The system of claim 8, wherein, The adversarial sample detector of the model trust assurance unit identifies disturbance patterns in input data through gradient sensitivity analysis and sets a first threshold to trigger a warning mode and a second threshold to activate an emergency response. The model interpreter uses a layer-dependent propagation algorithm to generate a visual heat map of threat judgment criteria.
10. The system of claim 8, wherein, The central analysis layer also integrates a causal reasoning engine for building a dynamic causal graph representing the causal relationship between threat events and conducting threat attribution analysis based on counterfactual reasoning. The system uses a cloud-edge-end collaborative reasoning framework, in which the cloud server performs complex reasoning of the global model, and the edge node handles local reasoning tasks with high real-time requirements.
Citation Information
Cited By
Laser array adaptive control method and system based on artificial intelligence
CN121863178A
