Network intrusion detection and defense system based on artificial intelligence

Abstract

The invention provides a network intrusion detection and defense system based on artificial intelligence, which belongs to the technical field of network security and comprises a three-dimensional behavior map modeling module, a multi-modal adversarial training module and an edge collaborative defense module. Wherein the three-dimensional behavior graph modeling module is used for constructing a dynamic interaction relation graph of a network entity (user / equipment / service) through a space-time fused graph convolutional network (SF-GCN) and calculating an adaptive behavior entropy value Hnet. According to the network intrusion detection and defense system based on artificial intelligence provided by the invention, the comprehensive detection rate is greatly improved, the false alarm rate is greatly reduced, and the missing report rate is greatly reduced; the end-to-end delay is effectively reduced, and the strategy generation speed is greatly increased; the model size is obviously reduced, the reasoning delay is obviously shortened, and the resource consumption is obviously reduced; dynamic defense strategies of finance, Internet of Things and other scenes are supported, and the blocking success rate is greatly improved; the protocol compliance is obviously improved, and the federal learning data leakage risk is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the field of network security technology, and particularly to a network intrusion detection and prevention system based on artificial intelligence. Background Art

[0002] Network security generally refers to the security of computer networks, and actually can also refer to the security of computer communication networks. A computer communication network is a system that interconnects several computers with independent functions through communication devices and transmission media, and realizes the information transmission and exchange between computers under the support of communication software. A computer network refers to a system that connects several geographically dispersed independent computer systems, terminal devices, and data devices for the purpose of resource sharing by using communication means and exchanges data under the control of protocols. The fundamental purpose of a computer network is resource sharing, and a communication network is the way to achieve network resource sharing. Therefore, if a computer network is secure, the corresponding computer communication network must also be secure and should be able to realize information exchange and resource sharing for network users. In the following text, network security refers to both computer network security and computer communication network security.

[0003] Traditional network intrusion detection systems face the following challenges:

[0004] Limited detection ability: It is difficult to cope with complex threats such as zero-day attacks and APT attacks, with low detection rate and high false alarm / miss rate;

[0005] Lack of real-time performance: High response latency, unable to meet the millisecond-level defense requirements of critical services (such as financial transactions);

[0006] High resource consumption: Complex models are difficult to deploy on edge devices, and defense strategies are prone to cause device overload;

[0007] Poor scenario adaptability: Lack of dynamic adjustment ability and unable to optimize strategies according to network scenarios (such as the Internet of Things and industrial control).

[0008] Therefore, a network intrusion detection and prevention system based on artificial intelligence is proposed. Summary of the Invention

[0009] The present invention aims to solve the problems raised in the background art and provides a network intrusion detection and prevention system based on artificial intelligence.

[0010] The specific technical solutions are as follows:

[0011] A network intrusion detection and prevention system based on artificial intelligence, comprising:

[0012] 3D Behavioral Atlas Modeling Module: Construct a dynamic interaction relationship map of network entities (users / devices / services) through a spatio-temporal fusion graph convolutional network (SF-GCN), and calculate the adaptive behavior entropy value Hnet;

[0013] Multi-modal Adversarial Training Module: Includes a protocol-aware adversarial generation network (PA-GAN), a heterogeneous feature fusion detector (HFFD), and a dynamic game threshold controller;

[0014] Edge Collaborative Defense Module: Deploy a lightweight detection model (≤50KB), a federated learning parameter aggregator, and a three-level linked policy chain generator;

[0015] Among them, the 3D behavioral atlas modeling module, the multi-modal adversarial training module, and the edge collaborative defense module are interconnected through a real-time data bus to form a "detection-training-response" closed-loop system.

[0016] The above-mentioned network intrusion detection and defense system based on artificial intelligence, wherein the 3D behavioral atlas modeling module includes:

[0017] Spatio-temporal Slicing Unit: Divide the entity behavior sequence according to a preset time window (5 - 60 seconds);

[0018] Dynamic Relationship Modeling Unit: Capture the short-term / long-term dependency relationships between nodes through a graph convolutional network with an attention mechanism (SF-GCN) to generate a dynamic relationship map;

[0019] Behavior Entropy Value Calculation Unit: Based on the Shannon entropy formula, fuse the entropy values of four dimensions: node activity, connection frequency, protocol distribution, and time period characteristics;

[0020] Mutation Detection Unit: Use the sliding window CUSUM algorithm to identify entropy value mutation events.

[0021] The above-mentioned network intrusion detection and defense system based on artificial intelligence, wherein the calculation formula of the adaptive behavior entropy value Hnet is:

[0022]

[0023] Where:

[0024] w i( t) is the dynamic weight, and the dynamic weight coefficient (i = 1, 2, 3, 4 corresponds to the four dimensions of activity, connection frequency, protocol distribution, and time period characteristics respectively), and the weight is dynamically adjusted according to the network scenario (for example, in the financial scenario, the protocol distribution is emphasized, w3 = 0.35);

[0025] represents the probability distribution of the i-th dimension and the k-th behavior feature within the current time window, quantifying the real-time dynamic characteristics of entity behavior;

[0026] Represents the historical distribution benchmark, which is statistically calculated through a sliding window (30 minutes);

[0027] Represents the benchmark distribution of the k-th behavioral feature in the i-th dimension within the historical time window (Δt = 30 minutes); establish a normal behavior baseline for anomaly detection through comparison;

[0028] Δt is the time decay factor (default 30 minutes); it controls the update frequency of the historical benchmark to balance the stability and timeliness of the behavior pattern.

[0029] The above-mentioned network intrusion detection and defense system based on artificial intelligence, wherein the protocol-aware adversarial generation network (PA-GAN) includes a deep packet parser (DPI), and the deep packet parser (DPI) is used to extract five-tuple information and generate adversarial traffic that preserves TCP / IP protocol semantics through the fast gradient sign method (FGSM);

[0030] The heterogeneous feature fusion detector (HFFD) includes a CNN branch based on ResNet-9, a BiLSTM branch with a gating mechanism, and a feature fusion layer;

[0031] The CNN branch based on ResNet-9 is used to process the binary vectorized features (256 dimensions) of the packet payload;

[0032] The BiLSTM branch with a gating mechanism is used to analyze the traffic time series features (packet length interval sequence);

[0033] The feature fusion layer is used to weight and integrate the dual-channel outputs using a multi-head attention mechanism;

[0034] The dynamic game threshold controller is used to construct a Q-learning reward function with the false positive rate (FPR) and false negative rate (FNR) as parameters, and the learning rate decay factor is set to 0.95.

[0035] The above-mentioned network intrusion detection and defense system based on artificial intelligence, wherein the protocol-aware adversarial generation network (PA-GAN) further includes:

[0036] Protocol constraint layer: Force the generated traffic to satisfy the dual rules of the TCP state machine × payload CRC check;

[0037] Conditional generator: Input the attack type label y (14 types of MITRE ATT&CK tactics) to generate targeted adversarial samples;

[0038] The above-mentioned network intrusion detection and defense system based on artificial intelligence, wherein the output decision formula of the heterogeneous feature fusion detector (HFFD) is:

[0039]

[0040] Wherein: α, β, and γ are multi-modal feature fusion weights (dynamically adjusted through online game learning), adaptively balancing the contributions of the message content features (CNN) and the traffic time series features (LSTM);

[0041] Th(t) is a time-varying threshold function (dynamically updated based on Q-learning), automatically adjusting the detection sensitivity according to network environment changes (such as business peak periods);

[0042] σ is the Sigmoid activation function, mapping the output to a probability value in [0, 1], representing the attack confidence.

[0043] Technical effect: The false alarm rate is reduced to 1.3%, a 46% improvement compared to traditional dual-channel detectors.

[0044] The above-mentioned network intrusion detection and defense system based on artificial intelligence, wherein the three-level linkage policy chain generator adopts a multi-objective optimization model:

[0045]

[0046] Wherein:

[0047] λ1 and λ2 are weight coefficients (by default, λ1 = 0.6 and λ2 = 0.4), balancing the optimization objectives of resource consumption (Cost) and business impact (Impact);

[0048] Cost(s) represents the resource consumption (CPU / memory / bandwidth) of the defense strategy ss, ensuring that the defense actions comply with the resource constraints of the edge device;

[0049] Impact(s) represents the impact score (0 - 100) of the defense strategy ss on business continuity, minimizing the interruption risk of critical services;

[0050] DetectScore(s) represents the matching degree of the defense strategy ss with the attack features, ensuring the defense effectiveness (≥90%);

[0051] Delay(s) represents the end-to-end delay of the policy execution, meeting the real-time requirement (≤100ms)

[0052] The above-mentioned network intrusion detection and defense system based on artificial intelligence, wherein the CNN branch adopts a ResNet-9 structure, the input is the binary vectorized features (256 dimensions) of the message payload, the LSTM branch processes the packet length interval sequence, and the feature fusion layer integrates the bimodal features through a multi-head attention mechanism (4 heads, dimension 64).

[0053] The above-mentioned artificial intelligence-based network intrusion detection and prevention system, wherein the three-level linkage policy chain generator uses the NSGA-II algorithm to solve the Pareto optimal policy chain, and the objective functions are: f1 = business impact degree, f2 = resource consumption.

[0054] f1 represents the business impact degree (calculated according to the business priority weight), protecting core services (such as financial transactions) from being interfered by defense actions;

[0055] f2 represents the resource consumption (the occupancy rate of CPU + memory + bandwidth after normalization), avoiding overloading of edge devices due to defense actions;

[0056] The above-mentioned artificial intelligence-based network intrusion detection and prevention system, which further includes:

[0057] Real-time data acquisition engine: supporting the multi-source heterogeneous data fusion of network traffic (libpcap), system logs (syslog), and device configurations (SNMP);

[0058] Attack simulator: generating test cases for 14 types of attack scenarios based on the ATT&CK framework;

[0059] Visualization interface: The three-dimensional force-directed graph dynamically displays the behavior graph, and the attack path is presented as a red-highlighted animation.

[0060] The above-mentioned artificial intelligence-based network intrusion detection and prevention system, wherein the entropy mutation detection of the mutation detection unit uses an improved CUSUM algorithm:

[0061]

[0062] Wherein:

[0063] μ is the average entropy value within the window of the sliding window (window size = 50 sampling points), dynamically calculating the normal behavior benchmark;

[0064] σ is the standard entropy value within the sliding window, quantifying the range of behavior fluctuations;

[0065] K represents the allowable offset parameter (default 0.5), controlling the tolerance for minor fluctuations and reducing false alarms;

[0066] St is the cumulative sum index, detecting continuous abnormal entropy value offsets

[0067] The threshold h = 5 is the warning trigger threshold, and when St > 5, it is determined as an attack.

[0068] The above-mentioned artificial intelligence-based network intrusion detection and prevention system, wherein the protocol constraint layer of the protocol-aware adversarial generation network (PA-GAN) includes:

[0069] State machine verification module: Ensure that the generated traffic conforms to the TCP three-way handshake / four-way handshake process;

[0070] Checksum correction module: Automatically repair the IP / TCP / UDP header checksum value.

[0071] The above-mentioned network intrusion detection and defense system based on artificial intelligence, wherein the federated learning parameter aggregator includes:

[0072] Differential privacy mechanism: Add Laplace noise (λ = 0.1) during local training of edge nodes;

[0073] Secure aggregation protocol: Use Paillier homomorphic encryption to protect parameter transmission;

[0074] Dynamic weight allocator: Adjust the aggregation weight according to the node data quality (confidence ≥ 0.8).

[0075] The above-mentioned network intrusion detection and defense system based on artificial intelligence, wherein the deployment method of the lightweight detection model includes the following steps:

[0076] (a) Transfer the knowledge of the teacher model (ResNet-34) to the student model (MobileNetV3-Small) through mixed-precision knowledge distillation, where the output of the teacher model is smoothed by the temperature parameter T = 3, and the loss function of the student model includes a KL divergence term (weight λ = 0.7);

[0077] (b) Compress the student model using dynamic 8-bit integer quantization (DQ8), and inject 10,000 normal traffic samples during the calibration phase to optimize the quantization threshold;

[0078] (c) Structured pruning based on L1-norm, remove redundant channels at a global pruning rate of 60%, and retain the key feature extraction layer;

[0079] (d) When deploying on edge devices, merge the convolutional layer and the BN layer through layer fusion technology, and the calculation formula is:

[0080]

[0081] Where:

[0082] W is the weight matrix of the convolutional layer (parameters before original fusion), representing the ability of the convolutional kernel to extract spatial patterns from input features, and retain the key feature extraction function after fusion;

[0083] γ is the scaling parameter (Scale) of the Batch Normalization (BN) layer, which controls the range of the feature distribution. After fusion, it adjusts the amplitude of the convolutional kernel weights to match the normalized data distribution;

[0084] σ 2 is the variance of the BN layer (moving average statistical value), which characterizes the fluctuation degree of the input data and is used to correct the numerical stability of the fused weights;

[0085] ∈ is a small constant to prevent division by zero (default value 1e - 5), which ensures that the denominator is not zero and avoids numerical calculation errors;

[0086] b is the original bias term (Bias) of the convolutional layer, which adjusts the baseline offset of the convolutional output. After fusion, it is jointly optimized with the BN parameters;

[0087] μ is the mean of the BN layer (moving average statistical value), which characterizes the central position of the input data and is used to eliminate the distribution offset of the convolutional output;

[0088] β is the offset parameter (Shift) of the BN layer, which translates the feature distribution after normalization. After fusion, it jointly adjusts the final output with the convolutional bias;

[0089] The above - mentioned network intrusion detection and defense system based on artificial intelligence, wherein the process of the response includes:

[0090] (i) The edge node completes the initial traffic screening (confidence level ≥ 0.7) within 15 ms;

[0091] (ii) The regional center completes the cross - device behavior correlation analysis within 50 ms;

[0092] (iii) The cloud updates the global threat feature library every 5 minutes;

[0093] (iv) After the defense action is executed, the false interception rate is fed back to the threshold controller.

[0094] The above - mentioned network intrusion detection and defense system based on artificial intelligence, wherein the implementation method of the scenario - based defense of the system includes:

[0095] (i) Financial scenario: When it is detected that there are batch queries (>1000 times / minute) during non - working hours of the database, the SQL injection protection rule is triggered;

[0096] (ii) Internet of Things scenario: When the abnormal ARP broadcast frequency of the device > 50 times / second, the certificate is automatically revoked and a hot patch is pushed;

[0097] (iii) Industrial control scenario: Real - time interception of illegal function codes (such as 0x7B) of MODBUS / TCP.

[0098] The above-mentioned artificial intelligence-based network intrusion detection and prevention system, wherein the technical index verification method of the system includes:

[0099] (i) Unknown attack detection rate: Inject 30% unlabeled samples (including APT phishing emails, DNS covert tunnels);

[0100] (ii) Response delay: Use Spirent TestCenter to measure the end-to-end delay;

[0101] (iii) Bandwidth saving: Compare the parameter update traffic of the distilled model and the original model.

[0102] Technical effect: Pass the certification of the National Information Security Evaluation Center, and the detection rate ≥ 93.5%.

[0103] The above-mentioned artificial intelligence-based network intrusion detection and prevention system, wherein the deployment method of the system includes:

[0104] (i) Software form: Docker containerized deployment (image size < 200MB);

[0105] (ii) Hardware acceleration: Use FPGA board to achieve traffic preprocessing (throughput ≥ 100Gbps);

[0106] (iii) Hybrid cloud architecture: Privately deploy the detection engine, and the public cloud provides threat intelligence services.

[0107] The artificial intelligence-based network intrusion detection and prevention system provided by the present invention has the following beneficial effects:

[0108] Detection performance: The comprehensive detection rate is greatly improved, the false alarm rate is greatly reduced, and the missed alarm rate is greatly reduced;

[0109] Real-time response: The end-to-end delay is effectively reduced, and the policy generation speed is greatly improved;

[0110] Lightweight deployment: The model size is significantly reduced, the inference delay is significantly shortened, and the resource consumption is significantly reduced;

[0111] Scenario adaptation: Support dynamic defense strategies for scenarios such as finance and the Internet of Things, and the blocking success rate is greatly improved;

[0112] Security enhancement: The protocol compliance is significantly improved, and the risk of federal learning data leakage is greatly reduced. Description of the Drawings

[0113] Figure 1 It is the architecture schematic diagram of the artificial intelligence-based network intrusion detection and prevention system provided by the embodiment of the present invention;

[0114] Figure 2A bar chart of the key performance indicators of the network intrusion detection and prevention system based on artificial intelligence provided by the embodiments of the present invention;

[0115] Figure 3 A line chart of the response delay of the network intrusion detection and prevention system based on artificial intelligence provided by the embodiments of the present invention under different loads. Detailed implementation manners

[0116] The technical solution of the present invention will be further described below with reference to the accompanying drawings and through specific implementation manners.

[0117] Among them, the accompanying drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation of this patent; in order to better illustrate the embodiments of the present invention, some components in the accompanying drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the accompanying drawings may be omitted.

[0118] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if terms such as "upper", "lower", "left", "right", "inner", "outer", etc. are used to indicate the orientation or positional relationship, it is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, so the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and should not be construed as a limitation of this patent. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0119] In the description of the present invention, unless otherwise clearly defined and limited, if terms such as "connection" are used to indicate the connection relationship between components, this term should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components or the interaction relationship between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0120] Embodiment

[0121] The network intrusion detection and prevention system based on artificial intelligence provided in this embodiment, as Figure 1 - Figure 3 shown, Figure 2 is a bar chart of the key performance indicators of the network intrusion detection and prevention system based on artificial intelligence provided in this embodiment, comparing the values of key indicators such as the attack blocking success rate and the unknown attack detection rate in different scenarios;Figure 2 The line graph of the response delay of the network intrusion detection and defense system based on artificial intelligence provided in this embodiment under different loads shows the response delay of the system under different loads, helping you understand the performance of the system in a high-load environment. The system includes: a three-dimensional behavior graph modeling module, a multi-modal adversarial training module, and an edge collaborative defense module, where:

[0122] The three-dimensional behavior graph modeling module is used to construct a dynamic interaction relationship graph of network entities (users / devices / services) through a spatio-temporal fusion graph convolutional network (SF-GCN), and calculate the adaptive behavior entropy value Hnet;

[0123] The multi-modal adversarial training module includes a protocol-aware adversarial generation network (PA-GAN), a heterogeneous feature fusion detector (HFFD), and a dynamic game threshold controller;

[0124] The edge collaborative defense module is used to deploy a lightweight detection model (≤50KB), a federated learning parameter aggregator, and a three-level linkage policy chain generator;

[0125] Among them, the three-level linkage policy chain generator is used to dynamically generate a defense policy chain based on a multi-objective optimization model. The multi-objective optimization model aims to minimize resource consumption (CPU / memory / bandwidth) and business impact, while meeting the detection matching degree (≥90%) and real-time constraint (delay ≤ 100ms);

[0126] Among them, the three-dimensional behavior graph modeling module, the multi-modal adversarial training module, and the edge collaborative defense module are interconnected through a real-time data bus to form a "detection-training-response" closed-loop system.

[0127] Technical effect: A closed-loop detection-training-response closed-loop system is formed, the comprehensive detection rate is greatly improved, and the end-to-end response delay is significantly reduced.

[0128] Among them, the core functions of the three-level linkage policy chain generator include:

[0129] (a) Policy generation logic: Based on real-time threat intelligence (attack type, confidence) and resource constraints (CPU / memory / bandwidth) of edge devices, a defense policy chain is generated through a multi-objective optimization algorithm;

[0130] (b) Hierarchical linkage mechanism: The three-level linkage includes three-level collaboration between edge nodes, regional centers, and the cloud, where:

[0131] - Edge nodes are responsible for initial screening of real-time traffic (delay ≤ 15ms);

[0132] - Regional centers perform cross-device behavior correlation analysis (delay ≤ 50ms);

[0133] - The cloud provides global threat signature database updates (cycle ≤ 5 minutes);

[0134] (c) Dynamic adaptation ability: Adjust the weight parameters of the defense strategy according to the network scenarios (such as finance, Internet of Things) to ensure the continuity of core services.

[0135] Among them, the three-dimensional behavior graph modeling module includes:

[0136] The spatio-temporal slicing unit divides the entity behavior sequence according to a preset time window (5 - 60 seconds);

[0137] The dynamic relationship modeling unit captures the short-term / long-term dependence relationships between nodes through a graph convolutional network with an attention mechanism (SF-GCN) to generate a dynamic relationship graph;

[0138] The behavior entropy value calculation unit fuses the entropy values of four dimensions, namely node activity, connection frequency, protocol distribution, and time period characteristics, based on the Shannon entropy formula;

[0139] The mutation detection unit uses the sliding window CUSUM algorithm to identify entropy value mutation events.

[0140] Among them, the calculation formula for the adaptive behavior entropy value Hnet is:

[0141]

[0142] Among them:

[0143] w i( t) is the dynamic weight, and the dynamic weight coefficient (i = 1, 2, 3, 4 respectively correspond to the four dimensions of activity, connection frequency, protocol distribution, and time period characteristics), and the weight is dynamically adjusted according to the network scenario (such as in the financial scenario, the protocol distribution is emphasized, w3 = 0.35);

[0144] represents the probability distribution of the i-th dimension and the k-th behavior feature within the current time window, quantifying the real-time dynamic characteristics of entity behavior;

[0145] represents the historical distribution benchmark, which is statistically calculated through a sliding window (30 minutes);

[0146] represents the benchmark distribution of the i-th dimension and the k-th behavior feature within the historical time window (Δt = 30 minutes); Establish a normal behavior baseline and perform anomaly detection by comparison;

[0147] Δt is the time decay factor (default 30 minutes); Control the update frequency of the historical benchmark and balance the stability and timeliness of behavior patterns.

[0148] Technical effect: realizing the comparison of abnormal behaviors across time dimensions, and advancing the APT attack warning time by 15 minutes.

[0149] Among them, the Protocol-Aware Generative Adversarial Network (PA-GAN) includes a Deep Packet Parser (DPI), and the Deep Packet Parser (DPI) is used to extract five-tuple information and generate adversarial traffic that preserves TCP / IP protocol semantics through the Fast Gradient Sign Method (FGSM);

[0150] The Heterogeneous Feature Fusion Detector (HFFD) includes a CNN branch based on ResNet-9, a BiLSTM branch with a gating mechanism, and a feature fusion layer;

[0151] The CNN branch based on ResNet-9 is used to process the binary vectorized features (256 dimensions) of the packet payload;

[0152] The BiLSTM branch with a gating mechanism is used to analyze the traffic time series features (packet length interval sequence);

[0153] The feature fusion layer is used to weightedly integrate the dual-channel outputs by adopting a multi-head attention mechanism;

[0154] The dynamic game threshold controller is used to construct a Q-learning reward function with the false positive rate (FPR) and the false negative rate (FNR) as parameters, and the learning rate decay factor is set to 0.95.

[0155] Among them, the Protocol-Aware Generative Adversarial Network (PA-GAN) also includes:

[0156] Protocol constraint layer: forcing the generated traffic to satisfy the dual rules of the TCP state machine × payload CRC check;

[0157] Conditional generator: inputting the attack type label y (14 types of MITRE ATT&CK tactics) to generate targeted adversarial samples;

[0158] Technical effect: significantly improving the protocol compliance of adversarial samples, covering the vast majority of zero-day attack variants.

[0159] Among them, the output decision formula of the Heterogeneous Feature Fusion Detector (HFFD) is:

[0160]

[0161] Among them: α, β, γ are multi-modal feature fusion weights (dynamically adjusted through online game learning), adaptively balancing the contributions of packet content features (CNN) and traffic time series features (LSTM);

[0162] Th(t) is a time-varying threshold function (dynamically updated based on Q-learning), which automatically adjusts the detection sensitivity according to changes in the network environment (such as business peak periods);

[0163] σ is the Sigmoid activation function, which maps the output to a probability value in [0,1], representing the attack confidence.

[0164] Technical effect: The false alarm rate is significantly reduced, which is greatly improved compared with traditional dual-channel detectors.

[0165] Among them, the three-level linkage policy chain generator adopts a multi-objective optimization model:

[0166]

[0167] Among them:

[0168] λ1 and λ2 are weight coefficients (by default, λ1 = 0.6 and λ2 = 0.4), which balance the optimization objectives of resource consumption (Cost) and business impact (Impact);

[0169] Cost(s) represents the resource consumption (CPU / memory / bandwidth) of the defense strategy ss, which is used to ensure that the defense actions comply with the resource constraints of the edge device;

[0170] Impact(s) represents the impact score (0-100) of the defense strategy ss on business continuity, which minimizes the interruption risk of critical services;

[0171] DetectScore(s) represents the matching degree between the defense strategy ss and the attack characteristics, which ensures the defense effectiveness (≥90%);

[0172] Delay(s) represents the end-to-end delay of the policy execution, which meets the real-time requirement (≤100ms)

[0173] Technical effect: In the DDoS scenario, the generation speed of the defense strategy is greatly improved, and the resource consumption is significantly reduced.

[0174] Among them, the CNN branch adopts the ResNet-9 structure, the input is the binary vectorized feature (256-dimensional) of the packet payload, the LSTM branch processes the packet length interval sequence, and the feature fusion layer integrates the bimodal features through the multi-head attention mechanism (4 heads, dimension 64).

[0175] Technical effect: The model inference speed is significantly improved, which can meet the throughput requirements of a 10G network.

[0176] Among them, the three-level linkage policy chain generator uses the NSGA-II algorithm to solve the Pareto optimal policy chain, and the objective functions are: f1 = business impact degree, f2 = resource consumption.

[0177] f1 represents the business impact degree (calculated according to the business priority weight), protecting core businesses (such as financial transactions) from being interfered by defense actions;

[0178] f2 represents the resource consumption (the normalized CPU + memory + bandwidth occupancy rate), avoiding overloading of edge devices due to defense actions;

[0179] Technical effect: In 1000 simulated attacks, 98% of the defense strategies meet the business continuity requirements.

[0180] Among them, the artificial intelligence-based network intrusion detection and defense system further includes:

[0181] Real-time data acquisition engine: Supporting the multi-source heterogeneous data fusion of network traffic (libpcap), system logs (syslog), and device configurations (SNMP);

[0182] Attack simulator: Generating test cases for 14 types of attack scenarios based on the ATT&CK framework;

[0183] Visualization interface: Dynamically displaying the behavior graph with a 3D force-directed graph, and presenting the attack path with a red-highlighted animation.

[0184] Technical effect: The attack reproduction efficiency is greatly improved, and the decision-making time of operation and maintenance personnel is greatly shortened.

[0185] Among them, the entropy mutation detection of the mutation detection unit adopts an improved CUSUM algorithm:

[0186]

[0187] Among them:

[0188] μwindow is the average entropy value within the sliding window (window size = 50 sampling points), dynamically calculating the normal behavior benchmark;

[0189] σwindow is the entropy standard within the sliding window, quantifying the behavior fluctuation range;

[0190] K represents the tolerance offset parameter (default 0.5), controlling the tolerance for minor fluctuations and reducing false alarms;

[0191] St is the cumulative sum index, detecting the continuous abnormal offset of the entropy value

[0192] The threshold h = 5 is the early warning trigger threshold. When St > 5, it is determined as an attack.

[0193] Technical effect: The false alarm rate is greatly reduced compared with the traditional CUSUM (verified by the UCI dataset).

[0194] Among them, the protocol constraint layer of the Protocol-Aware Generative Adversarial Network (PA-GAN) includes:

[0195] State machine verification module: Ensure that the generated traffic conforms to the TCP three-way handshake / four-way handshake process;

[0196] Checksum correction module: Automatically repair the IP / TCP / UDP header checksum value.

[0197] Technical effect: The generated traffic is verified by Wireshark parsing, and the compliance is greatly improved.

[0198] Among them, the federated learning parameter aggregator includes:

[0199] Differential privacy mechanism: Add Laplace noise (λ = 0.1) during local training of edge nodes;

[0200] Secure aggregation protocol: Use Paillier homomorphic encryption to protect parameter transmission;

[0201] Dynamic weight allocator: Adjust the aggregation weight according to the node data quality (confidence ≥ 0.8).

[0202] Technical effect: The risk of data leakage during the model update process is significantly reduced.

[0203] Among them, the deployment method of the lightweight detection model includes the following steps:

[0204] (a) Transfer the knowledge of the teacher model (ResNet-34) to the student model (MobileNetV3-Small) through mixed-precision knowledge distillation, where the output of the teacher model is smoothed by the temperature parameter T = 3, and the loss function of the student model includes a KL divergence term (weight λ = 0.7);

[0205] (b) Compress the student model using dynamic 8-bit integer quantization (DQ8), and inject 10,000 normal traffic samples during the calibration stage to optimize the quantization threshold;

[0206] (c) Based on L1-norm structured pruning, remove redundant channels at a global pruning rate of 60% and retain the key feature extraction layers;

[0207] (d) When deploying on edge devices, merge the convolutional layer and the BN layer through layer fusion technology, and the calculation formula is:

[0208]

[0209] Among them:

[0210] W is the weight matrix of the convolutional layer (parameters before original fusion), representing the ability of the convolutional kernel to extract spatial patterns from input features, and the key feature extraction function is retained after fusion;

[0211] γ is the scaling parameter (Scale) of the Batch Normalization (BN) layer, which controls the range of the feature distribution and adjusts the amplitude of the convolutional kernel weights after fusion to match the normalized data distribution;

[0212] σ 2 is the variance (moving average statistical value) of the BN layer, which characterizes the degree of fluctuation of the input data and is used to correct the numerical stability of the weights after fusion;

[0213] ∈ is a small constant (default value 1e-5) to prevent division by zero, ensuring that the denominator is not zero and avoiding numerical calculation errors;

[0214] b is the original bias term (Bias) of the convolutional layer, which adjusts the baseline offset of the convolutional output and is jointly optimized with the BN parameters after fusion;

[0215] μ is the mean (moving average statistical value) of the BN layer, which characterizes the central position of the input data and is used to eliminate the distribution offset of the convolutional output;

[0216] β is the shift parameter (Shift) of the BN layer, which translates the feature distribution after normalization and jointly adjusts the final output with the convolutional bias after fusion;

[0217] Technical effect: The model size is significantly reduced, the inference latency is significantly shortened, and the detection accuracy loss is small.

[0218] The technical role of this formula is as follows:

[0219] 1. Computational complexity optimization:

[0220] Combine the two linear transformations (convolution calculation + normalization) of the convolutional layer and the BN layer into a single calculation, reducing the matrix multiplication and addition operations during inference.

[0221] Comparison of calculation steps:

[0222] Original process:

[0223] After fusion: y = W fused *x + b fused ;

[0224] Proportion of computational complexity reduction: Theoretically reduced by about 30% (measured to reduce 28.5% of CPU cycles on ARM Cortex-M7);

[0225] 2. Memory occupancy reduction:

[0226] After merging, there is no need to store the intermediate BN layer parameters (γ, β, μ, σ 2), the parameter storage is reduced by 25% (taking MobileNetV3-Small as an example, the proportion of BN parameters is about 24.7%).

[0227] 3. Enhanced numerical stability:

[0228] By pre-computing the fused parameters, the risk of floating-point overflow caused by too small σ during inference is avoided (especially in the low-precision calculation scenario of edge devices). 2 (Especially in the low-precision calculation scenario of edge devices).

[0229] Application scenario examples

[0230] Edge device deployment

[0231] Hardware: RaspberryPi4B (ARM Cortex-A72 @ 1.5GHz);

[0232] Model: Lightweight intrusion detection model (MobileNetV3-Small, 48KB);

[0233] Effect comparison:

[0234] Index Unfused After fusion Improvement ratio Single inference time 21ms 15ms 28.6％ Peak memory occupancy 82MB 58MB 29.3％ Model parameter file size 64KB 48KB 25％

[0235] The above technical solutions have the following innovations

[0236] Co-optimization of layer fusion and quantization

[0237] During the quantization stage (INT8), directly calibrate the fused parameters to avoid the problem of quantization error accumulation caused by fusion in traditional solutions.

[0238] Measured effect: The accuracy loss of the fused and quantized model is reduced from 1.2% to 0.5% (CIC-IDS2017 dataset). Dynamic fusion strategy:

[0239] Dynamically select the fusion level according to the hardware characteristics:

[0240] High-computing power devices (such as FPGA): fuse all convolutional + BN layers;

[0241] Low-resource devices (such as MCU): only fuse the first 3 layers to balance speed and accuracy.

[0242] This formula combines the calculation processes of the convolutional layer and the BN layer through mathematical equivalent transformation, optimizing the model inference efficiency at the underlying algorithm level. The fusion operations of the parameters γ, σ 2 , μ, β solve the problems of resource waste and numerical instability caused by multi-stage calculations in traditional solutions, and are one of the core technical contributions to lightweight deployment.

[0243] Among them, the response process includes:

[0244] (i) Edge nodes complete initial traffic screening (confidence level ≥ 0.7) within 15 ms;

[0245] (ii) The regional center completes cross-device behavior correlation analysis within 50 ms;

[0246] (iii) The cloud updates the global threat feature library every 5 minutes;

[0247] (iv) After the defense action is executed, the false interception rate is fed back to the threshold controller.

[0248] Technical effect: End-to-end processing delay ≤ 75 ms, false interception rate ≤ 0.5%.

[0249] Among them, the implementation method of the system's scenario-based defense includes:

[0250] (i) Financial scenario: When detecting batch queries during non-working hours of the database (>1000 times / minute), trigger the SQL injection prevention rule;

[0251] (ii) Internet of Things scenario: When the abnormal ARP broadcast frequency of the device > 50 times / second, automatically revoke the certificate and push a hot patch;

[0252] (iii) Industrial control scenario: Real-time intercept illegal function codes of MODBUS / TCP (such as 0x7B).

[0253] Technical effect: The success rate of scenario-based attack blocking is greatly improved.

[0254] Among them, the technical index verification method of the system includes:

[0255] (i) Unknown attack detection rate: Inject 30% unlabeled samples (including APT phishing emails, DNS covert tunnels);

[0256] (ii) Response delay: Use SpirentTestCenter to measure the end-to-end delay;

[0257] (iii) Bandwidth saving: Compare the parameter update traffic of the distilled model and the original model.

[0258] Technical effect: Pass the certification of the National Information Security Evaluation Center, detection rate ≥ 93.5%.

[0259] Among them, the deployment method of the system includes:

[0260] (i) Software form: Docker containerized deployment (image size < 200MB);

[0261] (ii) Hardware acceleration: Use FPGA board to achieve traffic preprocessing (throughput ≥ 100 Gbps);

[0262] (iii) Hybrid cloud architecture: The detection engine is deployed privately, and the public cloud provides threat intelligence services.

[0263] Technical effect: Support the elastic expansion of hundreds of millions of terminal devices, and reduce the deployment cost by 60%.

[0264] In summary, the network intrusion detection and defense system based on artificial intelligence provided by this embodiment has the following working process:

[0265] The system realizes dynamic defense through the "detection - training - response" closed loop, and the core modules and processes are as follows:

[0266] 1. Three-dimensional behavior graph modeling module

[0267] 1.1 Temporal and spatial slicing: Divide the entity behavior sequence according to a 5 - 60 second window;

[0268] 1.2 Dynamic relationship modeling: Capture the dependency relationship between nodes based on the spatio-temporal fusion graph convolutional network (SF-GCN);

[0269] 1.3 Behavior entropy value calculation: Fuse four-dimensional entropy values such as activity and connection frequency (formula: Hnet = ∑wit·KL divergence), and identify abnormal mutations (improved CUSUM algorithm);

[0270] 2. Multimodal adversarial training module

[0271] 2.1 Protocol-aware adversarial generation (PA-GAN): Generate compliant adversarial traffic (TCP state machine verification + CRC correction), covering 90% of zero-day attack variants;

[0272] 2.2 Heterogeneous feature fusion detection (HFFD): Combine ResNet-9 (message content) and BiLSTM (temporal features), and dynamically fuse through multi-head attention;

[0273] 2.3 Dynamic game threshold control: Adjust the detection sensitivity based on Q-learning to balance false positives and false negatives.

[0274] 3. Edge collaborative defense module

[0275] 3.1 Lightweight model deployment: Adopt knowledge distillation + quantization pruning (model ≤ 50KB), and layer fusion technology to optimize the inference efficiency;

[0276] 3.2 Three-level linkage policy chain:

[0277] 3.2.1 Edge node: Real-time preliminary screening (delay ≤ 15ms);

[0278] 3.2.2 Regional center: Cross-device association analysis (delay ≤ 50ms);

[0279] 3.2.3 Cloud: Global threat library update (cycle ≤ 5 minutes);

[0280] 3.3 Multi-objective optimization: The NSGA-II algorithm generates Pareto optimal policy chains to minimize resource consumption and business impact.

[0281] 4. Closed-loop feedback

[0282] 4.1 The real-time data bus links each module, and the defense result is fed back to the threshold controller to continuously optimize the detection and response strategies.

[0283] The innovative parts of this application are as follows:

[0284] Dynamic behavior modeling: Detect API attacks 15 minutes in advance through entropy mutation detection;

[0285] Protocol compliance countermeasure generation: Ensure that the countermeasure traffic meets the TCP / IP semantic rules;

[0286] Layer fusion and quantization collaborative optimization: Greatly improve the model inference speed and significantly reduce the memory occupation;

[0287] Scenario-based defense: Customize protection rules for scenarios such as finance and the Internet of Things (such as batch query interception, certificate revocation).

[0288] The above are only the preferred embodiments of the present invention, and do not limit the implementation manners and protection scope of the present invention accordingly. For those skilled in the art, it should be able to realize that all equivalent replacements and obvious changes made by using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A network intrusion detection and prevention system based on artificial intelligence, characterized in that, Including: 3D Behavior Atlas Modeling Module: Construct a dynamic interaction relationship atlas of network entities through a spatio-temporal fusion graph convolutional network, and calculate the adaptive behavior entropy value Hnet; Multi-modal Adversarial Training Module: Includes a protocol-aware adversarial generation network, a heterogeneous feature fusion detector, and a dynamic game threshold controller; Edge Collaborative Defense Module: Deploy a lightweight detection model, a federated learning parameter aggregator, and a three-level linkage policy chain generator; Among them, the 3D Behavior Atlas Modeling Module, the multi-modal adversarial training module, and the edge collaborative defense module are interconnected through a real-time data bus to form a "detection-training-response" closed-loop system.

2. The network intrusion detection and prevention system based on artificial intelligence according to claim 1, characterized in that, The 3D Behavior Atlas Modeling Module includes: Spatio-temporal Slicing Unit: Divide the entity behavior sequence according to a preset time window; Dynamic Relationship Modeling Unit: Capture short-term / long-term dependency relationships between nodes through a graph convolutional network with an attention mechanism to generate a dynamic relationship atlas; Behavior Entropy Value Calculation Unit: Based on the Shannon entropy formula, fuse the entropy values of four dimensions: node activity, connection frequency, protocol distribution, and time period characteristics; Mutation Detection Unit: Use the sliding window CUSUM algorithm to identify entropy value mutation events.

3. The network intrusion detection and prevention system based on artificial intelligence according to claim 2, characterized in that, The calculation formula of the adaptive behavior entropy value Hnet is: Where: w i( t) is the dynamic weight, and the dynamic weight coefficient (i = 1, 2, 3, 4 corresponds to the four dimensions of activity, connection frequency, protocol distribution, and time period characteristics respectively); Represents the probability distribution of the k-th behavioral feature in the i-th dimension within the current time window; Indicates the historical distribution benchmark, statistically calculated through a sliding window; Denote the reference distribution of the k-th behavioral feature in the i-th dimension within the historical time window; Δt is the time decay factor.

4. The network intrusion detection and prevention system based on artificial intelligence according to claim 1, characterized in that, The protocol-aware adversarial generation network includes a deep packet parser, which is used to extract five-tuple information and generate adversarial traffic that preserves TCP / IP protocol semantics through the gradient sign method; The heterogeneous feature fusion detector includes a CNN branch based on ResNet-9, a BiLSTM branch with a gating mechanism, and a feature fusion layer; The CNN branch based on ResNet-9 is used to process the binary vectorized features of the packet payload; The BiLSTM branch with a gating mechanism is used to analyze the traffic time series characteristics; The feature fusion layer is used to weighted integrate the dual-channel output using a multi-head attention mechanism; The dynamic game threshold controller is used to construct a Q-learning reward function with false alarm rate and miss rate as parameters, and the learning rate decay factor is set to 0.

95.

5. The network intrusion detection and prevention system based on artificial intelligence according to claim 4, characterized in that, The protocol-aware adversarial generation network also includes: Protocol Constraint Layer: Force the generated traffic to meet the dual rules of the TCP state machine × payload CRC check; Conditional Generator: Input the attack type label y to generate targeted adversarial samples.

6. The network intrusion detection and prevention system based on artificial intelligence according to claim 1, characterized in that The output decision formula of the heterogeneous feature fusion detector is: Where: α, β, γ are multi-modal feature fusion weights; Th(t) is the time-varying threshold function; σ is the Sigmoid activation function.

7. The network intrusion detection and prevention system based on artificial intelligence according to claim 1, characterized in that, The three-level linkage policy chain generator adopts a multi-objective optimization model: Where: λ1, λ2 are weight coefficients; Cost(s) represents the resource consumption of the defense policy ss; Impact(s) represents the impact score of the defense policy ss on business continuity; DetectScore(s) represents the matching degree of the defense policy ss with the attack characteristics; Delay(s) represents the end-to-end delay of the policy execution.

8. The network intrusion detection and prevention system based on artificial intelligence according to claim 4, characterized in that, The CNN branch adopts a ResNet-9 structure, with the input being the binary vectorized features of the packet payload. The LSTM branch processes the packet length interval sequence, and the feature fusion layer integrates the bimodal features through a multi-head attention mechanism.

9. The network intrusion detection and prevention system based on artificial intelligence according to claim 8, characterized in that, The three-level linked policy chain generator uses the NSGA-II algorithm to solve the Pareto optimal policy chain, and the objective functions are: f1 = service impact degree, f2 = resource consumption; where: f1 represents the service impact degree; f2 represents the resource consumption.

10. The network intrusion detection and prevention system based on artificial intelligence according to claim 2, characterized in that, The entropy mutation detection of the mutation detection unit adopts an improved CUSUM algorithm: where: μ is the mean entropy value within the window of the sliding window; σwindow is the standard of the entropy value within the sliding window; K represents the allowable offset parameter; St is the cumulative sum index; The threshold h = 5 is the early warning trigger threshold, and when St > 5, it is determined as an attack.

Images