End side AI intelligent agent network intrusion behavior identification method and system

Abstract

The invention discloses an end-side AI agent network intrusion behavior identification method and system, and relates to the technical field of network intrusion behavior identification, and the method comprises the steps: obtaining equipment network configuration information according to an end-side AI agent; collecting network behavior monitoring data including network packet length distribution, transmission protocol type and data transmission rate; based on the safety reference knowledge base, generating an initial identification configuration set corresponding to the end-side AI intelligent agent; and continuously monitoring the network connection state of the end-side AI intelligent agent to form a self-adaptive intrusion behavior identification strategy. According to the invention, the technical problems in the prior art that the detection precision and the false alarm rate of a network intrusion detection model are difficult to balance and the configuration cannot dynamically adapt to the environment and threat change under the condition that the resources of the end-side equipment are limited are solved, and self-adaptive identification and dynamic configuration optimization of the network intrusion behavior by the end-side equipment under the resource constraint are realized. The detection precision is improved; and the false alarm rate is reduced.

Description

Technical Field

[0001] This invention relates to the field of network intrusion behavior recognition technology, specifically to a method and system for identifying network intrusion behavior using an edge AI agent. Background Technology

[0002] With the rapid development of the Internet of Things (IoT) and the Industrial Internet, network connections for edge devices, such as smart home terminals and industrial control nodes, are becoming increasingly widespread. However, edge devices generally suffer from hardware constraints such as limited computing power, tight memory resources, and energy sensitivity. Traditional network intrusion detection methods often rely on centralized cloud processing or complex deep learning models, which not only struggle to adapt to edge resource limitations, leading to excessively high inference latency and resource overflow, but also present challenges in balancing detection accuracy and false alarm rate. Furthermore, they cannot dynamically respond to changes in the network environment and new attack patterns, leaving edge devices facing severe network security threats.

[0003] In existing technologies, when resources on the mid-end devices are limited, it is difficult to balance the detection accuracy and false alarm rate of network intrusion detection models, and the configuration cannot be dynamically adapted to changes in the environment and threats. Summary of the Invention

[0004] This application provides a method and system for identifying network intrusion behavior by an edge AI agent, which addresses the technical problem in existing technologies where, in scenarios with limited edge device resources, it is difficult to balance the detection accuracy and false alarm rate of network intrusion detection models, and the configuration cannot be dynamically adapted to changes in the environment and threats.

[0005] In view of the above problems, this application provides a method and system for identifying network intrusion behavior of edge AI intelligent agents.

[0006] A first aspect of this application provides a method for identifying network intrusion behavior of an edge AI agent, the method comprising: Based on the edge AI agent, device network configuration information is obtained. The edge AI agent includes a multi-level lightweight deep learning network driven by the same inference engine, and the pruning rate and quantization bit width of each level of the lightweight deep learning network can be independently adjusted. Network behavior monitoring data, including network packet length distribution, transmission protocol type, and data transmission rate, are collected. Based on a security reference knowledge base and combined with the device network configuration information, an initial identification configuration set corresponding to the edge AI agent is generated. Based on the network behavior monitoring data and the initial identification configuration set, the network connection status of the edge AI agent is continuously monitored to form an adaptive intrusion behavior identification strategy.

[0007] A second aspect of this application provides an edge-side AI agent network intrusion behavior recognition system, the system comprising: The configuration information acquisition module is used to acquire device network configuration information based on the edge AI agent. The edge AI agent includes a multi-level lightweight deep learning network driven by the same inference engine, and the pruning rate and quantization bit width of each level of the lightweight deep learning network can be independently adjusted. The monitoring data acquisition module is used to collect network behavior monitoring data, including network packet length distribution, transmission protocol type, and data transmission rate. The initial identification configuration set generation module is used to generate an initial identification configuration set corresponding to the edge AI agent based on a security reference knowledge base and the device network configuration information. The intrusion behavior identification strategy formation module is used to continuously monitor the network connection status of the edge AI agent based on the network behavior monitoring data and the initial identification configuration set, and form an adaptive intrusion behavior identification strategy.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: Based on the edge AI agent, device network configuration information is acquired; network behavior monitoring data, including network packet length distribution, transmission protocol type, and data transmission rate, is collected; based on a security reference knowledge base and the device network configuration information, an initial identification configuration set corresponding to the edge AI agent is generated; the network connection status of the edge AI agent is continuously monitored to form an adaptive intrusion behavior identification strategy. This achieves the technical effect of enabling edge devices to adaptively identify and dynamically optimize network intrusion behavior under resource constraints, improving detection accuracy and reducing false alarm rate. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A schematic flowchart of a method for identifying network intrusion behavior of an edge AI agent provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an edge-side AI agent network intrusion behavior recognition system provided in an embodiment of this application.

[0011] Figure labeling: Configuration information acquisition module 10, monitoring data acquisition module 20, initial identification configuration set generation module 30, intrusion behavior identification strategy formation module 40. Detailed Implementation

[0012] This application provides a method and system for identifying network intrusion behavior by an edge AI intelligent agent, which addresses the technical problems in existing technologies where, under scenarios with limited edge device resources, it is difficult to balance the detection accuracy and false alarm rate of network intrusion detection models, and the configuration cannot dynamically adapt to changes in the environment and threats.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] Example 1, as Figure 1 As shown, this application provides a method for identifying network intrusion behavior of edge AI agents, the method comprising: Step S100: Obtain device network configuration information based on the edge AI agent. The edge AI agent includes a multi-level lightweight deep learning network driven by the same inference engine, and the pruning rate and quantization bit width of each level of the lightweight deep learning network can be adjusted independently.

[0015] Specifically, a multi-level lightweight deep learning network architecture driven by the same inference engine is first deployed on the edge device. The independent adjustment function of the pruning rate and quantization bit width of each level of the network is enabled through the model configuration interface, which supports subsequent on-demand adaptation to device resources. Then, the network management API and SNMP protocol of the edge device are called to establish a secure communication link and collect core network configuration data of the device in a targeted manner, including IP address range allocation, subnet mask parameters, gateway configuration, port open list, routing forwarding rules, network interface rate limit and access permission policy, etc. At the same time, hardware resource information such as CPU computing power limit and memory capacity are obtained through device system diagnostic tools. All collected data is packaged in a standardized format and transmitted to the configuration storage unit of the edge AI agent, providing basic data support for the generation of the initial identification configuration set and the device network environment and resource carrying capacity.

[0016] Step S200: Collect network behavior monitoring data including network packet length distribution, transmission protocol type, and data transmission rate.

[0017] Specifically, by integrating a real-time data stream processing module and a dedicated monitoring and acquisition unit into the edge device, a comprehensive network behavior data acquisition process is initiated, focusing on capturing three core types of monitoring data: network packet length distribution data, covering the maximum, minimum, and average length of individual network packets, as well as the distribution frequency of different length intervals; transmission protocol type data, accurately identifying the usage type, communication frequency, and protocol field integrity of various interaction protocols such as TCP, UDP, HTTP, and FTP; and data transmission rate data, providing real-time statistics on uplink / downlink data transmission volume, rate fluctuation amplitude, peak rate, and duration percentage per unit time. During the acquisition process, the raw data is simultaneously deduplicated, format-standardized, and outlier-filtered to ensure that the data meets the analysis requirements for subsequent intrusion behavior identification, providing high-quality and timely network behavior data support for the formation of adaptive identification strategies.

[0018] Step S300: Based on the security reference knowledge base and combined with the device network configuration information, generate an initial identification configuration set corresponding to the edge AI agent.

[0019] Specifically, the system first invokes a pre-defined security reference knowledge base to extract reference computing power allocation standards required to achieve target detection accuracy under different network scenarios. This includes the computing power ratio of each network level, inference resource quotas, and reference false alarm evolution trajectory data, including historical false alarm types, false alarm rate change patterns, and optimization directions. Then, combined with the acquired device network configuration information, such as IP allocation, port open status, routing rules, and hardware computing power limits, the system analyzes the security protection requirements and actual resource carrying capacity of the device network architecture. By dynamically adjusting the pruning rate of each level of the lightweight deep learning network in the edge AI agent, such as adjusting the convolutional layer channel retention ratio and quantization bit width (e.g., switching between 8-bit and 16-bit precision), and using a security-aware inference performance evaluation model, the system comprehensively considers detection accuracy, inference latency, and resource utilization to quantitatively evaluate different configuration combinations. The optimal pruning rate and quantization bit width combination that balances detection performance and resource adaptability is selected. Finally, based on this optimal combination, the system generates the initial recognition configuration set corresponding to the edge AI agent.

[0020] Step S400: Based on the network behavior monitoring data and the initial identification configuration set, continuously monitor the network connection status of the edge AI agent to form an adaptive intrusion behavior identification strategy.

[0021] Specifically, the edge AI agent first initiates real-time continuous monitoring of network connection status based on collected network behavior monitoring data, including network packet length distribution, transmission protocol type, data transmission rate, etc., and an initial identification configuration set. It then performs inference analysis on the monitoring data using a multi-level lightweight deep learning network. Simultaneously, it optimizes the initial identification configuration set using a Pareto balance between abnormal traffic detection accuracy and model inference latency. Furthermore, it dynamically updates the identification configuration by considering dynamic changes in security threat levels and referencing system resource usage information such as current edge device inference power consumption and memory usage. During this process, it receives edge scheduling command signals containing the network layer sequence to be adjusted and the upper limit of the model update rate, setting the model update... The new concurrency window, based on system resource data such as CPU utilization, memory usage, and inference queue latency, assesses the inference latency and energy consumption impact index in conjunction with the computational and storage requirements of model updates, and dynamically adjusts the maximum number of concurrent update channels. For the N-level lightweight deep learning network to be optimized, if the number of network layers in the current update state is in a low-load concurrency range, the pruning rate and quantization bit width parameters of any level of the network to be optimized are directly updated. If it is in a high-load concurrency range, the parameter update of the first level is started according to the priority of the network layer sequence to be adjusted, after the number of existing update tasks is reduced by 1. Through the above dynamic adjustment mechanism, an adaptive intrusion behavior identification strategy is formed that adapts to the real-time network state and device resource constraints.

[0022] In one possible implementation, step S400 further includes: Step S410: Optimize the initial identification configuration set using the Pareto balance metric of abnormal traffic detection accuracy and model inference latency.

[0023] Step S420: Simultaneously, based on the dynamic changes in the security threat level, determine the optimized identification configuration set based on the current inference energy consumption and memory usage rate, which are dynamically updated.

[0024] Specifically, the non-dominated sorting genetic algorithm NSGA-II is used as the core optimization algorithm to construct a dual-objective Pareto optimization model for abnormal traffic detection accuracy and model inference latency. The first objective function is to maximize the abnormal traffic detection accuracy by comparing network behavior monitoring data with the intrusion feature database in the security reference knowledge base, counting the number of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN), and quantizing the accuracy using the formula (accuracy = (TP + TN) / (TP + TN + FP + FN)). The second objective function is to minimize the model inference latency by using a high-precision timing module on the edge to collect the entire process time from input data to output recognition results of the multi-level lightweight deep learning network, accurate to the millisecond level. During the algorithm initialization phase, the pruning rate of each level of the lightweight deep learning network in the initial recognition configuration set, ranging from 10% to 80%, is used as the decision variable. Discretization and quantization bit width are performed with a 5% gradient, with 8-bit, 16-bit, and 32-bit options available. This generates an initial population of 100-200 individuals, with each individual corresponding to a complete set of data. The network configuration parameters are then iteratively optimized through selection, crossover, and mutation genetic operations. The selection operation uses a roulette wheel algorithm combined with crowding ranking to prioritize individuals near the Pareto front and with a uniform distribution. The crossover operation performs segmented crossover and recombination on the pruning rate and quantization bit width parameters of different individuals. The mutation operation randomly adjusts some configuration parameters of individuals to ensure population diversity. During each iteration, the network configuration corresponding to each individual in the population is inferred and verified by the edge simulation inference unit to calculate its corresponding accuracy and latency indicators. Then, based on the non-dominated ranking criterion, individuals are divided into different levels to eliminate dominated and inefficient configurations. At the same time, the uniform distribution of the Pareto front is maintained by crowding calculation. After 30 to 50 iterations, the Pareto optimal solution set is converged. From this set, the configuration scheme that satisfies the following conditions is selected: accuracy ≥ preset target threshold, such as 95%; latency ≤ edge device tolerance limit, such as 50ms; and optimal crowding. This is used to update and optimize the initial recognition configuration set, ultimately forming an intermediate configuration set that achieves a dynamic balance between detection performance and real-time performance.

[0025] By connecting to a network security threat awareness platform, the system obtains the real-time security threat level of the current network environment, classifying it into four levels: low, medium, high, and urgent, with weighting coefficients of 0.2, 0.4, 0.7, and 1.0 respectively. Simultaneously, utilizing the performance monitoring API and energy consumption acquisition module of the edge device, it accurately collects inference energy consumption data, calculated based on the voltage-current-time integral formula: energy consumption = ∫U(t)I(t)dt, with units of mAh. It also collects memory occupancy data, quantified as used memory / total memory × 100%. Subsequently, a three-input, single-output fuzzy logic inference model is constructed, using security threat level weights, inference energy consumption, and memory occupancy as input variables, and the adaptation degree of optimized identification configuration as the output variable. The input variables are fuzzified using a triangular membership function; for example, inference energy consumption is divided into three fuzzy subsets: low, medium, and high energy consumption, and memory occupancy is divided into three fuzzy subsets: low, medium, and high occupancy. This is based on security... Based on historical configuration optimization experience in the knowledge base, a fuzzy inference rule base is formulated. For example, if the threat level is urgent and the energy consumption is low and the memory usage is medium, the fit is high, so the quantization bit width is increased and the pruning rate is reduced. If the threat level is low and the energy consumption is high and the memory usage is high, the fit is low, so the pruning rate is increased and 8-bit quantization is used. The model performs fuzzy inference using the Mamdani inference method, and then defuzzifies using the centroid method to obtain the configuration fit quantization value. Combined with the obtained Pareto optimal solution set, the configuration scheme with the highest fit is selected. When the threat level is high, the focus is on ensuring detection accuracy, so the quantization bit width is increased to 16-bit / 32-bit and the pruning rate is reduced to 10%~30%. When the energy consumption / memory usage is high, the focus is on reducing resource consumption, so the quantization bit width is switched to 8-bit and the pruning rate is increased to 50%~80%. Finally, an optimized identification configuration set is determined based on the dynamic matching of the current inference energy consumption and memory usage, so as to achieve a dynamic balance between security protection requirements and edge resource constraints.

[0026] In one possible implementation, step S300 further includes: Step S310: Based on the security reference knowledge base, determine the reference computing power allocation and reference false alarm evolution trajectory required to achieve the target detection accuracy.

[0027] Step S320: Using the device network configuration information, combined with the reference computing power allocation and the reference false alarm evolution trajectory, obtain the initial identification configuration set corresponding to the edge AI agent.

[0028] Specifically, the system first invokes a pre-defined security reference knowledge base, which integrates different network scenarios, such as industrial control networks, smart home networks, intrusion behavior feature libraries corresponding to device types, historical detection datasets, target detection accuracy and computing power matching models, and false alarm optimization case libraries. After clarifying the current target detection accuracy threshold for the edge device, such as 95% accuracy and 90% recall, the system retrieves historical configuration data matching this accuracy target from the knowledge base. It then extracts the reference computing power allocation information required to achieve this accuracy, specifically including the computing power proportions of the feature extraction layer, feature fusion layer, and classification decision layer in a multi-level lightweight deep learning network, such as 40%, 30%, and 30% respectively, as well as inference resource quotas and constraints on computing power allocation at each level. Simultaneously, the system retrieves reference false alarm evolution trajectory data corresponding to this accuracy target from the knowledge base, covering the distribution of false alarm types, such as protocol identification false alarms, traffic threshold false alarms, trend curves of false alarm rate changes with pruning rate / quantization bit width, optimization space for false alarm rate under different configuration parameters, and key adjustment nodes. This provides a quantitative reference for the subsequent adaptation and optimization of the initial identification configuration.

[0029] First, the acquired device network configuration information is analyzed to extract core resource and environmental parameters such as the device's CPU computing power limit, memory capacity, network interface speed, port open type, and routing forwarding capability, clarifying the computing power capacity threshold and network security protection requirements of the edge devices. Then, combined with the determined reference computing power allocation, including the computing power proportion, inference resource quota, and reference false alarm evolution trajectory of each level of lightweight deep learning network, including false alarm type distribution and parameter adjustment optimization rules, to meet the core objectives of target detection accuracy, adapting to device resource constraints, and reducing the false alarm rate, a dynamic adjustment process for the configuration parameters of each level of lightweight deep learning network is initiated. The first-level network, within an adjustable range of pruning rate (10%~0%) and quantization bit width (8-bit / 16-bit / 32-bit), sequentially matches the resource requirements corresponding to the reference computing power allocation, while avoiding high-false-report risk parameter combinations clearly defined in the reference false-report evolution trajectory. During this process, a comprehensive quantitative evaluation is performed using a security-aware inference performance evaluation model. This model is a three-input, single-output, multi-dimensional weighted evaluation architecture, with the core construction logic as follows: The input layer clearly defines three core evaluation indicators: detection accuracy, false-report rate, and resource utilization. Detection accuracy is calculated by comparing with the intrusion feature database in the security reference knowledge base, focusing on the various... The system assesses the coverage of network intrusion behavior identification, quantifies the detection effectiveness of network attacks, and calculates the false positive rate based on the proportion of false positives in non-intrusion traffic. The formula is: False positive rate = False positive / (False positive + True negative). Resource utilization is quantified by combining CPU utilization (used computing power / total computing power × 100%) and memory utilization (used memory / total memory × 100%). The middle layer uses the Analytic Hierarchy Process (AHP) to determine the weights of the indicators. Combining the core requirements of accurate identification, low false positives, and lightweight operation in edge scenarios, the system sets the weights for detection accuracy (0.4), false positive rate (0.3), and resource utilization (0.3). Simultaneously, linear normalization is used to reflect the weights of each indicator. The signal is directed to the [0, 1] interval to eliminate dimensional differences. The output layer calculates the comprehensive performance score using a weighted summation algorithm. The formula is: Comprehensive score = Detection accuracy × 0.4 + (1 - False alarm rate) × 0.3 + (1 - Resource utilization rate) × 0.3. A higher score indicates better overall configuration performance. Based on this model, each set of adjusted configurations is comprehensively evaluated to select the optimal combination of pruning rate and quantization bit width that balances computing power adaptability, low false alarm characteristics, and target detection accuracy. Finally, based on this optimal combination, the initial recognition configuration set corresponding to the edge AI agent is encapsulated to provide accurate and adaptable basic configuration support for subsequent network intrusion behavior recognition.

[0030] In one possible implementation, step S320 further includes: Step S321: Using the device network configuration information, combined with the reference computing power allocation and reference false alarm evolution trajectory, dynamically adjust the pruning rate and quantization bit width of each level of lightweight deep learning network, evaluate the security awareness inference performance under the current configuration, and determine the optimal pruning rate and quantization bit width combination corresponding to the multi-level lightweight deep learning network.

[0031] Step S322: Based on the optimal pruning rate and quantization bit width combination corresponding to the multi-level lightweight deep learning network, obtain the initial recognition configuration set corresponding to the edge AI agent.

[0032] Specifically, the core resource parameters such as CPU computing power limit, memory capacity, and network interface processing speed are first extracted from the device network configuration information to clarify the resource allocation constraints of each level of lightweight deep learning network. Then, based on the reference computing power allocation, including the computing power ratio of each level of network and the inference resource quota, and referencing the false alarm evolution trajectory, including the high false alarm risk parameter range and the false alarm rate optimization law, dynamic adjustment of pruning rate and quantization bit width is initiated. The pruning rate is discretized in the range of 10% to 80% with a 5% gradient, and redundant convolution kernels of each level of network are pruned through the channel pruning algorithm to adapt to the computing power allocation requirements and avoid high false alarm pruning ratios. The quantization bit width is switched between 8 bits, 16 bits, and 32 bits. A linear quantization algorithm is used to quantize the network weights and activation values. Low computing power devices are preferentially matched with low quantization bit width to reduce resource consumption, while medium and high computing power devices select high quantization bit width in combination with false alarm control requirements to ensure feature expression accuracy. For each adjusted configuration, a multi-dimensional quantitative evaluation is conducted using a security-aware inference performance evaluation model. This model uses the abnormal traffic detection accuracy (calculated by comparing with intrusion features in the security reference knowledge base, using the formula Accuracy = (True Positives + True Negatives) / Total Number of Samples), False Alarm Rate (calculated by False Alarm Rate = False Positives / (False Positives + True Negatives), and Resource Utilization (including CPU utilization and memory usage ratio) as core evaluation indicators. The Analytic Hierarchy Process (AHP) is used to determine the weights of each indicator, with accuracy of 0.4, false alarm rate of 0.3, and resource utilization of 0.3, to calculate the overall performance score. Finally, by traversing all configuration combinations, the configuration scheme with the highest overall performance score is selected to determine the optimal pruning rate and quantization bit width combination for each of the multi-level lightweight deep learning networks. This ensures that the combination meets the resource constraints of the reference computing power allocation, conforms to the low false alarm requirement of the reference false alarm evolution trajectory, and adapts to the actual resource carrying capacity of the edge devices.

[0033] First, the optimal configuration combination of the determined multi-level lightweight deep learning network is standardized. Following the network hierarchy, such as feature extraction layer, feature fusion layer, and classification decision layer, the optimal pruning rate for each level is determined, clarifying specific gradient ratios (e.g., 30% for feature extraction layer, 20% for classification decision layer) and quantization bit width, specifying 8-bit / 16-bit / 32-bit selections. Simultaneously, the computational power adaptation parameters, false alarm control thresholds, and inference engine adaptation rules corresponding to each configuration are recorded. Then, combining the architecture specifications of the edge AI agent, the above optimal configuration parameters are integrated with the basic recognition rules in the security reference knowledge base. The resource constraints in the device network configuration information are integrated and encapsulated into structured configuration data containing core fields such as network level identifier, pruning rate parameter, quantization bit width parameter, computing power allocation ratio, false alarm optimization threshold, and inference execution priority. Finally, the structured configuration data is format-validated and integrity-verified to ensure that the network configuration parameters at all levels are conflict-free and complete, and fully adapted to the resource carrying capacity and target detection accuracy requirements of the end-side device. This results in the generation of the initial identification configuration set corresponding to the end-side AI agent, providing standardized and directly callable basic configuration support for the real-time identification of subsequent network intrusion behaviors.

[0034] In one possible implementation, step S400 further includes: Step S430: The edge AI agent receives the edge scheduling instruction signal and dynamically adjusts the pruning rate and quantization bit width of each level of the lightweight deep learning network using the adaptive intrusion behavior recognition strategy.

[0035] Specifically, the edge AI agent establishes a real-time communication link with the scheduling and management module of the edge device, receiving edge scheduling command signals containing the sequence of network levels to be adjusted, clarifying the priority ranking and upper limit of the model update rate of each lightweight deep learning network, and limiting the frequency threshold of parameter adjustment. Combined with the established adaptive intrusion behavior recognition strategy, it initiates a dynamic adjustment process for the configuration of each network level. First, based on the sequence of network levels to be adjusted, the adjustment priority of each network level is clarified. Then, based on the optimization objectives in the strategy that adapt to the current network state and resource constraints, the pruning rate of each lightweight deep learning network level is adjusted as needed within the gradient range of 10% to 80%, with the redundant channel retention ratio and quantization bit width switching between 8-bit, 16-bit, and 32-bit precision for targeted adjustment. During the adjustment process, the upper limit of the model update rate is strictly followed to avoid inference fluctuations caused by frequent adjustments. At the same time, the system resource monitoring module is linked to provide real-time feedback on changes in CPU utilization, memory usage, and inference latency after adjustment, ensuring that the dynamic adjustment of the pruning rate and quantization bit width adapts to the current network intrusion detection requirements without exceeding the resource carrying capacity of the edge device, achieving a dynamic balance between recognition performance and resource consumption.

[0036] In one possible implementation, step S400 further includes: Step S440: The end-side scheduling instruction signal includes the sequence of network layers to be adjusted and the upper limit of the model update rate.

[0037] Step S450: Set a model update concurrency window. Within the model update concurrency window, only M-level lightweight deep learning networks are allowed to update the parameters of pruning rate and quantization bit width simultaneously, and M does not exceed the maximum number of concurrent update channels.

[0038] Specifically, the edge scheduling command signal is generated by the system scheduling module of the edge device based on the network security status and device resource load, and then sent to the edge AI agent. This signal contains two core pieces of information: First, the sequence of network layers to be adjusted, which clearly lists the priority order of the multi-level lightweight deep learning networks that need to have their pruning rate and quantization bit width parameters adjusted. The ranking rules are based on the current key requirements of network intrusion detection, the functional importance of each layer of the network, such as the priority adaptation of the feature extraction layer and the classification decision layer, and the determination of historical adjustment effects, providing a basis for the sequential execution of subsequent parameter updates; Second, the upper limit of the model update rate, which sets a specific parameter update frequency threshold, such as allowing a maximum of 3 parameter adjustments per second, strictly limiting the adjustment frequency of the pruning rate and quantization bit width of each level of lightweight deep learning network, avoiding interruption of the edge device's inference process, surge in resource consumption, or fluctuation in detection accuracy due to frequent parameter updates, and ensuring the stable operation of the network intrusion detection task.

[0039] To address the limited resource capacity of edge devices and the need for orderly parameter updates, a concurrency control mechanism is adopted to set a fixed-duration model update concurrency window, such as a configurable period of 50~200ms. The window duration is preset based on the computing power of the edge device and the computational complexity of the model update, and supports dynamic fine-tuning. After each concurrency window starts, the number of lightweight deep learning network layers that simultaneously perform pruning rate and quantization bit width parameter updates is strictly limited to M levels. The value of M needs to match the maximum number of concurrent update channels in real time. This number of channels is dynamically determined based on the resource capacity of the target operating environment of the edge AI agent, and always satisfies the constraint that M ≤ the maximum number of concurrent update channels. Through this design, CPU computing power conflicts, memory consumption surges, and inference queue blockages caused by simultaneous parameter updates of multiple network layers are avoided. At the same time, it ensures the efficient advancement of network configuration optimization at each level, ensuring that the edge AI agent can still stably maintain the real-time detection capability of network intrusion behavior during the dynamic adjustment of parameters, and realize the coordinated operation of configuration update and detection tasks.

[0040] In one possible implementation, step S450 further includes: Step S451: Based on the target operating environment corresponding to the edge AI agent, obtain system resource usage information, including CPU utilization, memory usage and inference queue latency.

[0041] Step S452: Using the system resource usage information and the computation and storage requirements of the current model update action, evaluate the inference latency and energy consumption impact index.

[0042] Step S453: Dynamically adjust the maximum number of concurrent update channels based on the inference latency and energy consumption impact index.

[0043] Specifically, the target operating environment of the edge AI agent is first defined, such as specific hardware deployment scenarios like smart home terminals and industrial control edge devices. By calling the system monitoring APIs and performance acquisition tools of the edge device, such as the proc file system, the device's built-in resource monitoring components, and inference task scheduling logs, real-time information on core system resource usage is collected: CPU utilization is calculated by statistically analyzing the proportion of CPU power used for model inference, parameter calculation, and system background tasks per unit time, quantified as actual computing power / total CPU computing power × 100%; memory usage is calculated by reading data from the system memory management module, determining the ratio of physical memory used by the model during runtime to the device's total physical memory, i.e., "used memory capacity / total memory capacity × 100%"; inference queue latency is recorded by measuring the waiting time from entering the edge AI agent's inference queue to being started and processed by each level of lightweight deep learning network, accurate to the millisecond level, to comprehensively obtain the current resource load status of the edge device. This provides real-time and accurate basic data support for subsequent evaluation of inference latency and energy consumption impact index, and dynamic adjustment of the maximum number of concurrent update channels.

[0044] Based on the acquired system resource usage information, including CPU utilization, memory usage, and inference queue latency, the core requirements of the current model update action are first quantitatively analyzed: computational requirements are assessed by adjusting the convolutional channel pruning operation corresponding to the pruning rate of each level of lightweight deep learning network and the computational consumption of weight / activation value conversion corresponding to quantization bit width switching; storage requirements are accurately calculated by the storage usage of temporary configuration data and intermediate calculation results during parameter update. Subsequently, a dual-index evaluation model is constructed, in which inference latency is calculated by superimposing the current inference queue latency and the estimated additional time consumption of model update. The additional time consumption is dynamically extrapolated based on the matching degree between computational requirements and current CPU utilization, accurate to the millisecond level; the energy consumption impact index is constructed by constructing an evaluation system using a weighted summation algorithm, with weight coefficients set at 0.4 for CPU utilization, 0.3 for memory usage, and 0.3 for model update computational requirements. After normalizing each index, a weighted calculation is performed to form a quantitative index in the range of 0 to 1. The higher the value, the greater the energy consumption impact, comprehensively evaluating the impact of model update actions on the operating status of edge devices.

[0045] First, a dual-indicator linkage adjustment model is constructed by pre-setting inference latency threshold ranges, such as low latency ≤ 50ms, medium latency 50~100ms, and high latency > 100ms, and energy consumption impact index ranges, such as low impact ≤ 0.3, medium impact 0.3~0.7, and high impact > 0.7. Then, the evaluated inference latency and energy consumption impact index are mapped to the corresponding ranges, and the maximum number of concurrent update channels is dynamically adjusted according to preset rules: if the inference latency is in the low range and the energy consumption impact index is low impact, it indicates that the edge device resources are sufficiently redundant, and the maximum number of concurrent update channels can be adjusted upward by 1~2 levels, such as from level 2 to level 4, to accelerate the model configuration update efficiency. If the inference latency is in the middle range or the energy consumption impact index is in the middle range, maintain the current maximum concurrent update channel count to balance update efficiency and resource consumption. If the inference latency is in the high range or the energy consumption impact index is in the high range, it indicates that the model update has put significant pressure on the operation of the end-side device. The maximum concurrent update channel count needs to be adjusted down by 1 to 2 levels, such as from level 3 to level 1, to avoid inference queue blockage or device energy overload. During the adjustment process, the update results are synchronized to the scheduling module of the end-side AI agent in real time to ensure that the maximum concurrent update channel count always adapts to the resource load status of the end-side device and the model update requirements, and ensures the stable operation of the network intrusion detection task.

[0046] In one possible implementation, step S400 further includes: Based on the initial identification configuration set, an N-level lightweight deep learning network to be optimized is determined.

[0047] Compare the number of network layers currently in the update state. If the number of layers in the update state is within the low-load concurrency range of the maximum concurrent update channel number, then directly start the parameter update of the U-th level lightweight deep learning network. The U-th level lightweight deep learning network is any level in the N-th level lightweight deep learning network.

[0048] Specifically, based on the initial identification configuration set, the core configuration parameters of each level of lightweight deep learning network are extracted, including pruning rate, quantization bit width, computing power allocation ratio, and false alarm control threshold. Combined with the target detection accuracy standard, edge device resource carrying capacity threshold, and historical optimization cases in the security reference knowledge base, a multi-dimensional screening model is constructed. By comparing the actual operating parameters of each level of network in the initial identification configuration set with the preset standard, lightweight deep learning networks that fail to reach the target threshold in detection accuracy, exceed the edge tolerance limit in inference latency, have a false alarm rate higher than the allowable range, or have abnormal resource utilization, such as CPU utilization or memory utilization exceeding the adaptation range, are screened out. The functional importance of the screened network levels is verified and the necessity of optimization is evaluated. Non-substantial non-compliance caused by temporary resource fluctuations is eliminated. Finally, the N-level lightweight deep learning networks that need pruning rate and quantization bit width adjustment are determined, and these N-level networks are sorted according to optimization priority to clarify the core optimization direction of each level of network, providing clear target objects for subsequent parameter updates.

[0049] First, the scheduling and monitoring module of the edge AI agent is used to count in real time the number of lightweight deep learning network layers currently updating pruning rate and quantization bit width parameters. At the same time, based on the preset maximum concurrent update channel number, a low-load concurrency interval is defined, that is, the number of network layers in the current update state is ≤ 50% of the maximum concurrent update channel number. Then, the counted number of layers in the current update state is compared with this low-load concurrency interval. If it is determined to be within this interval, it means that the computing power, memory and other resources of the edge device still have sufficient redundancy to support the new parameter update task without affecting the real-time performance and stability of network intrusion detection. At this time, without waiting for the existing update task to complete, any level is randomly selected from the N-level lightweight deep learning network to be optimized as the U-level lightweight deep learning network. Its pruning rate is started, and the quantization bit width is adjusted according to the adaptation requirements within the gradient range of 10%~80%. The parameter update process switches between 8-bit, 16-bit and 32-bit to ensure the efficient advancement of the initial recognition configuration set optimization, while ensuring the stable operation of the core detection function of the edge AI agent.

[0050] In one possible implementation, step S400 further includes: If the number of updates is within the high-load concurrency range of the maximum concurrent update channel number, parameter updates of the U-th level lightweight deep learning network are allowed to be initiated after the number of existing update tasks is reduced by 1, wherein the U-th level lightweight deep learning network is at the first position in the sequence of network layers to be adjusted.

[0051] Specifically, the scheduling and monitoring module of the edge AI agent first uses real-time statistics to count the number of network layers currently updating pruning rate and quantization bit width parameters. Simultaneously, based on the dynamically adjusted maximum concurrent update channel count, a high-load concurrency interval is identified: the number of network layers currently updating is greater than 70% of the maximum concurrent update channel count. Then, the current number of updating layers is compared with this high-load concurrency interval. If it is determined to be within this interval, it indicates that the edge device's computing power, memory, and other resources are nearing their capacity limits, and adding new update tasks should be avoided to prevent resource overload or inference queue blockage. At this point, it is necessary to wait for existing update tasks to complete. Once any one step is completed, the number of network layers in the update state is reduced by 1. After ensuring that the current update quantity falls back to a reasonable range that can handle the maximum number of concurrent update channels, the network ranked first in the sequence of network layers to be adjusted is selected as the U-th level lightweight deep learning network. Its pruning rate is then activated, and the bit width is precisely adjusted and quantized within the gradient range of 10% to 80% according to the adaptation requirements. The parameter update process, which switches between 8-bit, 16-bit, and 32-bit as needed, ensures that the network layers to be optimized are configured and updated in an orderly manner according to priority, while strictly controlling the resource consumption of the end devices, thus ensuring low latency and stable operation of the network intrusion detection task.

[0052] Example 2, based on the same inventive concept as the edge AI agent network intrusion behavior recognition method in the foregoing examples, such as... Figure 2 As shown, this application provides an edge-side AI agent network intrusion behavior recognition system. The system and method embodiments in this application are based on the same inventive concept. The system includes: The configuration information acquisition module 10 is used to acquire device network configuration information based on the edge AI agent. The edge AI agent includes a multi-level lightweight deep learning network driven by the same inference engine, and the pruning rate and quantization bit width of each level of the lightweight deep learning network can be adjusted independently.

[0053] The monitoring data acquisition module 20 is used to collect network behavior monitoring data, including network packet length distribution, transmission protocol type, and data transmission rate.

[0054] The initial identification configuration set generation module 30 is used to generate an initial identification configuration set corresponding to the edge AI agent based on the security reference knowledge base and the device network configuration information.

[0055] The intrusion behavior recognition strategy forming module 40 is used to continuously monitor the network connection status of the edge AI agent based on the network behavior monitoring data and the initial recognition configuration set, and form an adaptive intrusion behavior recognition strategy.

[0056] Furthermore, the system also includes: The initial identification configuration set is optimized using the Pareto balance index of abnormal traffic detection accuracy and model inference latency; at the same time, an optimized identification configuration set is determined based on the dynamic changes in security threat level and dynamically updated current inference energy consumption and memory usage.

[0057] Furthermore, the system also includes: Based on the security reference knowledge base, the reference computing power allocation and reference false alarm evolution trajectory required to achieve the target detection accuracy are determined; by combining the device network configuration information with the reference computing power allocation and reference false alarm evolution trajectory, the initial recognition configuration set corresponding to the edge AI agent is obtained.

[0058] Furthermore, the system also includes: By using the device network configuration information, combined with the reference computing power allocation and the reference false alarm evolution trajectory, the pruning rate and quantization bit width of each level of lightweight deep learning network are dynamically adjusted to evaluate the security perception inference performance under the current configuration, and determine the optimal pruning rate and quantization bit width combination corresponding to the multi-level lightweight deep learning network; based on the optimal pruning rate and quantization bit width combination corresponding to the multi-level lightweight deep learning network, the initial recognition configuration set corresponding to the edge AI agent is obtained.

[0059] Furthermore, the system also includes: The edge AI agent receives the edge scheduling command signal and uses the adaptive intrusion behavior recognition strategy to dynamically adjust the pruning rate and quantization bit width of each level of the lightweight deep learning network.

[0060] Furthermore, the system also includes: The edge scheduling command signal includes the sequence of network layers to be adjusted and the upper limit of the model update rate; a model update concurrency window is set, within which only M-level lightweight deep learning networks are allowed to update the parameters of pruning rate and quantization bit width at the same time, and M does not exceed the maximum number of concurrent update channels.

[0061] Furthermore, the system also includes: Based on the target operating environment corresponding to the edge AI agent, system resource usage information is obtained, including CPU utilization, memory usage, and inference queue latency. Using the system resource usage information, combined with the computation and storage requirements of the current model update action, the inference latency and energy consumption impact index are evaluated. Based on the inference latency and energy consumption impact index, the maximum number of concurrent update channels is dynamically adjusted.

[0062] Furthermore, the system also includes: Based on the initial identification configuration set, determine the N-level lightweight deep learning network to be optimized; compare the number of network layers currently in the update state, if the number of update states is in the low-load concurrency range of the maximum concurrent update channel number, then directly start the parameter update of the U-level lightweight deep learning network, where the U-level lightweight deep learning network is any level in the N-level lightweight deep learning network.

[0063] Furthermore, the system also includes: If the number of updates is within the high-load concurrency range of the maximum concurrent update channel number, parameter updates of the U-th level lightweight deep learning network are allowed to be initiated after the number of existing update tasks is reduced by 1, wherein the U-th level lightweight deep learning network is at the first position in the sequence of network layers to be adjusted.

[0064] It should be noted that the order of the embodiments described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. Specific embodiments of this specification have been described above. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0065] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0066] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for identifying network intrusion behavior of an edge-side AI agent, characterized in that, The method includes: According to the edge AI agent, the device network configuration information is obtained. The edge AI agent includes a multi-level lightweight deep learning network driven by the same inference engine, and the pruning rate and quantization bit width of each level of the lightweight deep learning network can be adjusted independently. Collect network behavior monitoring data, including network packet length distribution, transmission protocol type, and data transmission rate. Based on the security reference knowledge base and combined with the device network configuration information, an initial identification configuration set corresponding to the edge AI agent is generated. Based on the network behavior monitoring data and the initial identification configuration set, the network connection status of the edge AI agent is continuously monitored to form an adaptive intrusion behavior identification strategy.

2. The method for identifying network intrusion behavior of an edge AI agent as described in claim 1, characterized in that, Based on the network behavior monitoring data and the initial identification configuration set, the method further includes continuously monitoring the network connection status of the edge AI agent. The initial identification configuration set is optimized using a Pareto balance metric between abnormal traffic detection accuracy and model inference latency. At the same time, based on the dynamic changes in security threat levels, an optimized identification configuration set is determined based on the current inference energy consumption and memory usage, which are dynamically updated.

3. The method for identifying network intrusion behavior of an edge AI agent as described in claim 1, characterized in that, Based on a security reference knowledge base and combined with the device network configuration information, an initial identification configuration set corresponding to the edge AI agent is generated. The method includes: Based on the aforementioned security reference knowledge base, the reference computing power allocation and reference false alarm evolution trajectory required to achieve the target detection accuracy are determined. By combining the device network configuration information with the reference computing power allocation and the reference false alarm evolution trajectory, the initial identification configuration set corresponding to the edge AI agent is obtained.

4. The method for identifying network intrusion behavior of an edge AI agent as described in claim 3, characterized in that, By combining the reference computing power allocation and the reference false alarm evolution trajectory, an initial recognition configuration set corresponding to the edge AI agent is obtained. The method includes: By using the device network configuration information, combined with the reference computing power allocation and the reference false alarm evolution trajectory, the pruning rate and quantization bit width of each level of lightweight deep learning network are dynamically adjusted to evaluate the security awareness inference performance under the current configuration and determine the optimal pruning rate and quantization bit width combination corresponding to the multi-level lightweight deep learning network. Based on the optimal pruning rate and quantization bit width combination corresponding to the multi-level lightweight deep learning network, the initial recognition configuration set corresponding to the edge AI agent is obtained.

5. The method for identifying network intrusion behavior of an edge AI agent as described in claim 1, characterized in that, The method for forming an adaptive intrusion behavior identification strategy includes: The edge AI agent receives the edge scheduling command signal and uses the adaptive intrusion behavior recognition strategy to dynamically adjust the pruning rate and quantization bit width of each level of the lightweight deep learning network.

6. The method for identifying network intrusion behavior of an edge AI agent as described in claim 5, characterized in that, The end-side scheduling command signal includes the sequence of network layers to be adjusted and the upper limit of the model update rate; Set a model update concurrency window. Within the model update concurrency window, only M-level lightweight deep learning networks are allowed to update parameters such as pruning rate and quantization bit width simultaneously. M does not exceed the maximum number of concurrent update channels.

7. The method for identifying network intrusion behavior of an edge AI agent as described in claim 6, characterized in that, Within the model update concurrency window, only M-level lightweight deep learning networks are allowed to simultaneously update parameters such as pruning rate and quantization bit width. The method includes: Based on the target operating environment corresponding to the edge AI agent, system resource usage information is obtained, including CPU utilization, memory usage and inference queue latency. By using the system resource usage information and combining it with the computation and storage requirements of the current model update action, the inference latency and energy consumption impact index are evaluated. The maximum number of concurrent update channels is dynamically adjusted based on the inference latency and energy consumption impact index.

8. The method for identifying network intrusion behavior of an edge AI agent as described in claim 7, characterized in that, The method includes: Based on the initial identification configuration set, an N-level lightweight deep learning network to be optimized is determined; Compare the number of network layers currently in the update state. If the number of layers in the update state is within the low-load concurrency range of the maximum concurrent update channel number, then directly start the parameter update of the U-th level lightweight deep learning network. The U-th level lightweight deep learning network is any level in the N-th level lightweight deep learning network.

9. The method for identifying network intrusion behavior of an edge AI agent as described in claim 8, characterized in that, The method for comparing the number of network layers currently in an update state includes: If the number of updates is within the high-load concurrency range of the maximum concurrent update channel number, parameter updates of the U-th level lightweight deep learning network are allowed to be initiated after the number of existing update tasks is reduced by 1, wherein the U-th level lightweight deep learning network is at the first position in the sequence of network layers to be adjusted.

10. A device-side AI agent network intrusion behavior recognition system, characterized in that, The system is used to implement the edge AI agent network intrusion behavior identification method according to any one of claims 1-9, the system comprising: The configuration information acquisition module is used to acquire device network configuration information based on the edge AI agent. The edge AI agent includes a multi-level lightweight deep learning network driven by the same inference engine, and the pruning rate and quantization bit width of each level of the lightweight deep learning network can be adjusted independently. The monitoring data acquisition module is used to collect network behavior monitoring data, including network packet length distribution, transmission protocol type, and data transmission rate. The initial identification configuration set generation module is used to generate the initial identification configuration set corresponding to the edge AI agent based on the security reference knowledge base and the device network configuration information. The intrusion behavior recognition strategy formation module is used to continuously monitor the network connection status of the edge AI agent based on the network behavior monitoring data and the initial recognition configuration set, and form an adaptive intrusion behavior recognition strategy.

Images