Network traceability data processing method, system, equipment and medium

By preprocessing and feature extraction of network data, building a dynamic attack map, using quantum-derived genetic algorithms and temporal graph convolution networks, the problems of missing data dimensions and lagging correlation analysis in APT attacks are solved, and the rapid identification and positioning of APT attacks are achieved, and network security protection capabilities are improved.

CN120342751APending Publication Date: 2025-07-18GUANGXI POWER GRID CORP
View PDF 0 Cites 9 Cited by

Patent Information

Application Number
CN202510665069.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-22
Publication Date
2025-07-18

AI Technical Summary

Technical Problem

When responding to advanced persistent threats (APTs), existing network traceability technologies cannot effectively capture hidden payloads in encrypted channels, resulting in missing key attack behaviors and difficult to adapt to the dynamic evolution of APT attack tactics, technologies and programs, resulting in serious delay in attack identification.

Method used

By collecting network data for preprocessing, extracting network behavior characteristics, building dynamic attack maps, and using quantum-derived genetic algorithms and temporal graph convolutional networks to make traceability decisions, achieving rapid identification and positioning of APT attacks.

Benefits of technology

The detection time of APT attacks has been significantly improved from an average of 78 days to 9.2 hours, the false alarm rate has been reduced to 0.13%, and it can accurately identify multiple hidden threats and abnormal traffic, improving network security protection capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120342751A_ABST
    Figure CN120342751A_ABST
Patent Text Reader

Abstract

The invention discloses a network traceability data processing method, system and device and a medium, and the method comprises the steps: collecting network data, carrying out the preprocessing of the network data, and generating the preprocessed network data; extracting behavior characteristics from the preprocessed network data to obtain network behavior characteristics; constructing an attack graph according to the obtained network behavior characteristics, and forming a dynamic graph representing an attack path; and carrying out traceability decision on the dynamic atlas, and positioning an attack source and an attack path through the traceability decision. Through the technical means of six-dimensional feature vector construction, quantum derivation genetic algorithm optimization detection and attack atlas construction through the tense graph convolutional network, the problems of data dimension missing, correlation analysis lagging and insufficient traceability precision are effectively solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline detection, and particularly to a method, system, device and medium for processing network traceability data. Background Art

[0002] An Advanced Persistent Threat (APT) is a complex and covert form of cyber attack, characterized by long-term latency and multi-stage attacks. Such attacks are typically carried out by attackers with abundant resources and sophisticated techniques, targeting specific targets for continuous penetration. As an important means to counter APT attacks, network traceability technology aims to trace and identify the attack source and attack path by analyzing network data, providing decision-making support for network defense. Traditional network traceability technologies mainly rely on methods such as network traffic analysis, log auditing, and signature matching. By collecting and analyzing attack traces, they attempt to reconstruct attack behaviors and determine the attack sources.

[0003] However, existing network traceability technologies face three key deficiencies when dealing with APT threats: Firstly, there is a serious lack of data dimensions. Traditional methods rely on a single data source and cannot effectively capture the hidden payloads in encrypted channels, resulting in the omission of key attack behaviors. Secondly, there is an obvious lag in correlation analysis. Existing technologies mostly adopt static correlation rules and are difficult to adapt to the dynamic evolution of APT attack tactics, techniques, and procedures, leading to a serious delay in attack identification. These deficiencies severely restrict the effective response ability of existing traceability technologies to APT threats and urgently require innovative technical solutions to solve them. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method, system, device and medium for processing network traceability data, which solves the problem that traditional methods cannot effectively capture the hidden payloads in encrypted channels, resulting in the omission of key attack behaviors; secondly, existing technologies are difficult to adapt to the dynamic evolution of APT attack tactics, techniques, and procedures, leading to a serious delay in attack identification.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for processing network traceability data, including:

[0008] Collecting network data and preprocessing the network data to generate preprocessed network data;

[0009] Extracting behavioral features from the preprocessed network data to obtain network behavioral features;

[0010] Construct an attack graph based on the obtained network behavior characteristics to form a dynamic graph representing the attack path;

[0011] Make a traceability decision on the dynamic graph, and locate the attack source and attack path through the traceability decision.

[0012] As a preferred solution of the network traceability data processing method described in the present invention, wherein: the steps of extracting the network behavior characteristics include:

[0013] Use an optimization algorithm to detect covert communication and extract covert communication characteristics;

[0014] Build a network behavior benchmark model based on the network data, and identify abnormal network behaviors by comparing the deviation degree between the real-time traffic characteristics and the network behavior benchmark model;

[0015] Fuse the covert communication characteristics and the abnormal behavior characteristics to generate network behavior characteristics.

[0016] The beneficial effect of this preferred technical solution is that by detecting covert communication through an optimization algorithm and combining it with a network behavior benchmark model for anomaly identification, the accuracy of detecting covert attacks is greatly improved.

[0017] As a preferred solution of the network traceability data processing method described in the present invention, wherein: the steps of constructing the attack graph include:

[0018] Determine entity nodes and behavior nodes according to the network behavior characteristics, and construct a dynamic knowledge graph based on the entity nodes and behavior nodes;

[0019] Calculate the association strength between nodes through the network behavior characteristics and historical attack data to form an initial edge weight;

[0020] Adjust the edge weights in the dynamic knowledge graph according to real-time network behaviors;

[0021] Analyze the obtained dynamic knowledge graph, capture the evolution process of attack behaviors, and form a dynamic graph representing the attack path.

[0022] The beneficial effect of this preferred technical solution is that the construction of the dynamic knowledge graph and the edge weight adaptive adjustment mechanism can capture the evolution process of attack behaviors in real time. This dynamic analysis method shortens the correlation analysis lag time from the hour level to the minute level, effectively coping with the dynamic tactics and techniques of APT attacks and improving the accuracy of attack path recognition.

[0023] As a preferred solution of the network traceability data processing method described in the present invention, wherein: making a traceability decision on the dynamic graph includes the steps of:

[0024] Generate multiple candidate attack paths based on the dynamic knowledge graph;

[0025] For the encrypted traffic included in the candidate attack paths, recover the encrypted content through the quantum tunneling effect to complete the refinement of the candidate attack paths;

[0026] Construct a virtual network environment based on the refined candidate attack paths, train adversarial agents for attackers and defenders in the virtual network environment, and verify and optimize the traceability strategy of the candidate attack paths through adversarial training;

[0027] Share the optimized traceability strategy and attack features.

[0028] As a preferred solution of the network traceability data processing method of the present invention, wherein: the optimization algorithm is a quantum-derived genetic algorithm based on quantum computing theory, which includes:

[0029] Use quantum bit pairs to represent genes in chromosomes and construct an initial population;

[0030] Update the quantum bit state through quantum rotation gate operations and perform population evolution on the initial population;

[0031] Evaluate the quality of individuals in the population through a fitness function and improve the accuracy of covert communication detection through iteration.

[0032] The beneficial effect of this preferred technical solution is that the quantum-derived genetic algorithm expands the search space through quantum bit encoding and rotation gate operations, and optimizing the fitness function design significantly improves the performance of the genetic algorithm.

[0033] As a preferred solution of the network traceability data processing method of the present invention, wherein: when analyzing the obtained dynamic knowledge graph, a temporal graph convolutional network is used, and the processing process of the temporal graph convolutional network includes:

[0034] Construct a node feature matrix to represent the feature information of each node in the graph;

[0035] Generate a normalized adjacency matrix to represent the connection relationship between nodes;

[0036] Create a temporal information matrix to capture the dynamic characteristics of node states changing over time;

[0037] Combine the node feature matrix, the normalized adjacency matrix, and the temporal information matrix, and perform spatio-temporal fusion analysis of graph structure data through convolution operations to identify attack paths and evolution patterns.

[0038] As a preferred solution of the network traceability data processing method described in the present invention, where: the steps of collecting network data and preprocessing the network data include:

[0039] Collect the network data through a lightweight probe architecture;

[0040] Apply an improved wavelet denoising algorithm to the collected network data for noise filtering;

[0041] Construct a multi-dimensional feature vector space;

[0042] Among them, the multi-dimensional feature vector space includes a timestamp synchronization dimension, a protocol stack feature dimension, a communication behavior dimension, a resource consumption dimension, a cross-layer behavior dimension, and a session persistence dimension.

[0043] In a second aspect, the present invention provides a network traceability data processing system, including:

[0044] A data processing module, configured to collect network data and preprocess the collected network data to generate preprocessed network data;

[0045] A feature extraction module, configured to extract behavior features from the network data to obtain network behavior features that can be used for subsequent analysis;

[0046] An attack graph construction module, which constructs an attack graph based on the extracted network behavior features to form a dynamic graph that can represent the attack path;

[0047] A traceability decision module, which makes a traceability decision on the constructed dynamic graph.

[0048] In a third aspect, the present invention provides an electronic device, including:

[0049] A memory and a processor;

[0050] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the network traceability data processing method are implemented.

[0051] In a fourth aspect, the present invention provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by a processor, the steps of the network traceability data processing method are implemented.

[0052] Compared with the prior art, the beneficial effects of the present invention:

[0053] Through technical means such as six-dimensional feature vector construction, quantum-derived genetic algorithm optimization for detection, and temporal graph convolutional network construction of attack graphs, the problems of missing data dimensions, lag in correlation analysis, and insufficient traceability accuracy are effectively solved. In the network of large financial institutions, the APT attack detection time is significantly shortened from an average of 78 days to 9.2 hours, and the false alarm rate is reduced to 0.13%, significantly improving the efficiency and accuracy of traceability, being able to quickly discover and locate the attack source, and providing strong support for timely taking defensive measures.

[0054] Functions such as dynamic time warping covert channel detection, adversarial traffic recognition formula, and multi-protocol support can effectively identify various covert threats and abnormal traffic. For example, in the covert channel detection experiment, the accuracy of the DTW algorithm using the optimized weight matrix is increased from 70% to 90%; the recall rate of adversarial traffic recognition is increased from 60% to 85%, and it can accurately identify encrypted proxy protocols such as Shadowsocks / V2Ray, timely discover illegal network proxy activities, and effectively prevent the leakage of sensitive information and malicious network attacks.

[0055] Through targeted data collection, processing, and analysis methods, effective network security protection is achieved. And through technologies such as dynamic weight transfer learning, it supports cross-organization sharing of threat intelligence, improves the overall network security protection ability, and is applicable to various complex network scenarios in different industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0057] Figure 1 It is a schematic diagram of the overall process of the network traceability data processing method described in an embodiment of the present invention.

[0058] Figure 2 It is a overall system diagram of the network traceability data processing method described in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the specific embodiments of the present invention in detail with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0060] Example 1, refer toFigure 1 , which is an embodiment of the present invention, provides a method for processing network traceability data, including:

[0061] S100. Collect network data and preprocess the network data to generate preprocessed network data;

[0062] S200. Extract behavioral features from the preprocessed network data to obtain network behavioral features;

[0063] S300. Construct an attack graph based on the obtained network behavioral features to form a dynamic graph representing the attack path;

[0064] S400. Make a traceability decision on the dynamic graph, and locate the attack source and attack path through the traceability decision.

[0065] In the field of network security, the core technical problem faced by APT attack traceability lies in the scattered attack traces, weak correlation, and high concealment. Traditional traceability methods often work independently and lack a systematic data correlation analysis framework. For example, a financial institution once suffered an APT attack. Although the security system captured a large amount of network data, due to the lack of data dimensions, it was unable to detect the hidden command and control traffic in TLS encrypted communication. At the same time, due to the lack of dynamic correlation analysis ability, the evolution path of attack behavior could not be effectively identified, resulting in a delay in attack detection. In addition, traditional traceability methods based on IP addresses have a high false positive rate in a hybrid cloud environment and cannot accurately locate the attack source. The method of the present invention forms a complete traceability process through four key steps: first, collect multi-source heterogeneous network data and perform noise reduction and feature extraction preprocessing; then extract hidden communication and abnormal behavior features based on an optimization algorithm; next, construct a dynamic knowledge graph and apply a temporal graph convolutional network to analyze attack evolution; finally, use quantum computing and adversarial training to achieve accurate traceability decisions, thereby systematically solving the above technical problems.

[0066] Through four consecutive and progressive processing steps, the core technical problems of APT threat traceability can be solved individually. First, the data collection and preprocessing link realizes multi-dimensional and low-interference data acquisition, solving the problem of missing data dimensions in traditional methods; second, the behavioral feature extraction link effectively identifies hidden communication and abnormal behavior through an optimization algorithm, making up for the deficiency of traditional detection methods in encrypted traffic analysis; third, the attack graph construction link effectively solves the problem of lagging correlation analysis through a dynamic knowledge graph and a real-time weight adjustment mechanism, and can accurately capture the evolution process of attack behavior; finally, the traceability decision link comprehensively uses a variety of advanced technologies to generate and verify candidate attack paths, significantly improving the traceability accuracy. This processing flow not only realizes the rapid detection and accurate traceability of APT attacks, but also improves the overall network security protection ability through a threat intelligence sharing mechanism.

[0067] Example 2, referring to Figures 1 to 2 , which is an embodiment of the present invention. Based on the above embodiments, a method for processing network traceability data is provided.

[0068] In the implementation manner of the present application, in step S100, the steps of collecting network data and preprocessing the network data include A1 to A3:

[0069] A1: Collect network data through a lightweight probe architecture;

[0070] A2: Apply an improved wavelet denoising algorithm to the collected network data for noise filtering;

[0071] A3: Construct a multi-dimensional feature vector space;

[0072] Among them, the multi-dimensional feature vector space includes a timestamp synchronization dimension, a protocol stack feature dimension, a communication behavior dimension, a resource consumption dimension, a cross-layer behavior dimension, and a session persistence dimension.

[0073] Specifically, in A1, the lightweight probe architecture is based on the cross-platform Berkeley Packet Filter technology to achieve real-time data collection with low resource occupancy. This architecture supports x86 and ARM architectures, realizes efficient collection with a CPU occupancy rate lower than 3%, and has a dynamic load balancing mechanism to automatically adjust the sampling rate according to the link bandwidth, with a range from 1:1000 to 1:100000.

[0074] In an optional implementation manner, the lightweight probe architecture in step A1 can also achieve data collection through a distributed deployment method. Deploy multiple probes working collaboratively at key nodes of the network, adopt a master-slave structure for probe management, ensure the consistency of cross-probe data through time synchronization, and achieve comprehensive monitoring of a large and complex network environment.

[0075] In another optional implementation manner, the lightweight probe architecture in step A1 can also achieve data collection through kernel-level hook technology, directly capture network packets at the operating system kernel level, bypass the application layer interface, improve the efficiency and integrity of data collection, and is particularly suitable for high-security environments that require in-depth analysis.

[0076] In step A2, the specific implementation of the improved wavelet denoising algorithm includes:

[0077]

[0078] In the formula, λ is the adaptive threshold, σ is the noise level estimated based on the generalized Gaussian distribution, and n is the number of data sampling points; the improved wavelet denoising algorithm uses the Symlets-8 wavelet basis for 5-layer decomposition and optimizes the signal-to-noise ratio through a high-frequency coefficient retention strategy.

[0079] In step A3, the construction of the multi-dimensional feature vector space is to achieve high-precision time synchronization based on the PTPv2 (IEEE 1588) protocol, with an accuracy of ±100 nanoseconds. It is calculated through the cross-device timestamp alignment algorithm:

[0080] t global = t local + δ(t) + ∈(t);

[0081] Among them, t global is the globally unified timestamp, t local is the timestamp recorded by the local device, δ(t) is the crystal oscillator drift compensation, and ∈(t) is the network delay correction.

[0082] Extract protocol deep features such as the TLS fingerprint feature set (including but not limited to the combination of client Hello extension fields, certificate serial number entropy value), the QUIC fingerprint feature set (including but not limited to initial packet version identification, connection ID entropy value), etc. Calculate the packet direction entropy value based on the direction distribution of information theory:

[0083]

[0084] Among them, ω is the packet direction entropy value, quantifying the uncertainty of the communication direction, N is the number of time units within the time window, log2 p i is the logarithmic function with base 2, and p i is the forward / backward traffic ratio within the i-th second, and the sliding window size is set to 500 ms;

[0085] Calculate the differential change rate:

[0086]

[0087] Among them, γ is the differential change rate of resource consumption, is the derivative operation with respect to time; CPU(t) is the CPU usage at time t; CPU max is the maximum CPU capacity; Memory(t) is the memory usage at time t; Memor ytotal is the total memory capacity;

[0088] And perform multi-granularity monitoring of the kernel-level interrupt context switching frequency, the change in the process-level file descriptor usage, and the thread-level mutex contention rate.

[0089] In the embodiment of the present application, the steps for extracting network behavior characteristics in step S200 include B1 to B3:

[0090] B1: Use an optimization algorithm to detect covert communication and extract covert communication characteristics;

[0091] B2: Construct a network behavior benchmark model based on network data, and identify abnormal network behaviors by comparing the deviation degree between real-time traffic characteristics and the network behavior benchmark model;

[0092] B3: Integrate the covert communication characteristics and abnormal behavior characteristics to generate network behavior characteristics.

[0093] Specifically, in step B1, the optimization algorithm is the dynamic time warping algorithm optimized by the quantum-derived genetic algorithm. First, the core implementation of the quantum-derived genetic algorithm (QDGA) includes:

[0094] Use quantum bit pairs (α, β) to represent the genes in the chromosome, where α and β satisfy |α| 2 +|β| 2 = 1. Each gene is represented by a quantum bit pair, enabling a single chromosome to simultaneously represent multiple possible states and significantly increasing the coverage of the search space.

[0095] Update the quantum bit state through the rotation matrix U(θ) to achieve population evolution:

[0096]

[0097] Among them, θ is the rotation angle, which is dynamically adjusted by the fitness function to control the evolution direction and rate of the chromosome state.

[0098] Design a fitness function to evaluate the quality of each chromosome. For the detection task of this innovative module, the fitness function considers multiple factors such as detection accuracy, recall rate, and computational complexity. For example:

[0099] Fitness(chromosome) = w1·Accuracy + w2·Recall - w3·Complexity;

[0100] Among them, w1, w2, and w3 are weight coefficients, which are adjusted according to the specific application scenario. The determination basis of the weight coefficients w1, w2, and w3: Parameter tuning is performed through a large amount of experimental data. In different network attack scenario simulation experiments, multiple groups of different weight coefficient combinations are set, and the accuracy rate, recall rate, and computational complexity of the detection task are calculated respectively. According to the experimental results, with the goal of maximizing the fitness function value, optimization algorithms such as grid search or random search are used to finally determine the weight coefficient combination that can make the fitness function perform better in multiple attack scenarios.

[0101] Apply QDGA to the dynamic time warping (DTW) algorithm for covert communication detection:

[0102]

[0103] In the formula, Q = {q1, q2, …, q n} is the query sequence, representing the network traffic feature sequence to be detected; C = {c1, c2, …, c m} is the reference sequence, representing the known covert communication feature template; w i,j is the weight matrix, generated by QDGA optimization, and is used to adjust the importance of different time points and feature elements.

[0104] The covert communication detection process includes: extracting the feature sequence from the preprocessed network data as the query sequence Q; constructing the feature template of the known covert communication as the reference sequence C; using QDGA to optimize the weight matrix; calculating the optimized DTW distance; setting a threshold, and when the DTW distance is less than this threshold, it is determined that there is a covert channel. The basis for determining the DTW distance threshold: the best balance point is determined by drawing the Receiver Operating Characteristic curve (ROC curve). Collect a large number of data samples containing covert channel traffic and normal traffic, test different DTW distance thresholds, and calculate the True Positive Rate (TPR) and False Positive Rate (FPR) under each threshold. Draw the ROC curve with FPR as the abscissa and TPR as the ordinate, and select the threshold corresponding to the point on the curve closest to the point (0, 1) as the best DTW distance threshold. At this threshold, while ensuring a high detection accuracy, the false alarm rate can be effectively controlled. When the DTW distance is less than this threshold, it is determined that there is a covert channel.

[0105] In an alternative embodiment, the algorithm for detecting covert communication in step B1 can also adopt an end-to-end time series anomaly detection method based on deep learning, and directly learn the feature representation of covert communication from the original network traffic through a Long Short-Term Memory network or a Convolutional Neural Network, without manually designing feature engineering.

[0106] In another alternative embodiment, the identification of abnormal network behavior in step B2 can also adopt an unsupervised clustering method, such as DBSCAN or Gaussian mixture model, to automatically discover abnormal clusters in network traffic, which is applicable to the detection scenario of unknown attack types.

[0107] In yet another alternative embodiment, the feature fusion in step B3 can also adopt an ensemble learning method, such as Random Forest or AdaBoost, taking the covert communication detector and the abnormal behavior detector as base learners, and fusing the decision results by voting or weighting to improve the robustness of the detection.

[0108] It should be noted that by combining the quantum-derived genetic algorithm with dynamic time warping, the present invention achieves high-precision detection of covert communication and overcomes the deficiencies of traditional methods in encrypted traffic analysis. At the same time, the multi-level network behavior benchmark model can effectively identify various abnormal behaviors, including zero-day attacks and advanced persistent threats. The fusion of these two types of features uses an attention mechanism, which not only improves the detection accuracy but also enhances the interpretability of traceability by capturing the correlations between features, providing comprehensive and reliable behavior features for the subsequent construction of the attack graph.

[0109] In the embodiment of the present application, the steps for constructing the attack graph in step S300 include C1 to C4:

[0110] C1: Determine entity nodes and behavior nodes based on network behavior features, and construct a dynamic knowledge graph based on the entity nodes and behavior nodes;

[0111] C2: Calculate the association strength between nodes through network behavior features and historical attack data to form an initial edge weight;

[0112] C3: Adjust the edge weights in the dynamic knowledge graph according to real-time network behavior;

[0113] C4: Analyze the obtained dynamic knowledge graph to capture the evolution process of attack behaviors and form a dynamic graph representing the attack path.

[0114] Specifically, in C1, the construction method of the dynamic knowledge graph G=(V, E, W t ) first determines entity nodes and behavior nodes according to network behavior features; the entity node set includes IP nodes, representing the source IP and target IP addresses in network communication; the certificate fingerprint node represents the unique identifier of the digital certificate used in TLS / SSL communication. The behavior node set includes command and control (C2) heartbeat nodes, representing the periodic communication behavior between the infected host and the control server.

[0115] In step C2, the calculation method of the initial edge weight is to calculate the association strength between nodes through network behavior features and historical attack data to form the initial edge weight W0(e ij ); at each time step t, calculate the partial derivative of the loss function L with respect to the edge e ij ; calculate the new edge weight W according to the edge weight update formula t (e ij ); update the edge weights of the dynamic knowledge graph G.

[0116] Use a temporal graph convolutional network (T-GCN) to analyze the dynamic knowledge graph and capture the evolution process of attack behaviors:

[0117]

[0118] Among them, H (l) represents the node feature matrix of the l-th layer, is the normalized adjacency matrix, is 's degree matrix, W (l) is the learnable weight matrix of the l-th layer, σ is the activation function (such as the ReLU function), T (l) is the temporal information matrix, and ⊙ represents element-wise multiplication.

[0119] Normalized adjacency matrix: is calculated as where A is the original adjacency matrix and I is the identity matrix. By using the normalized adjacency matrix, the over-smoothing problem in the graph convolutional network can be solved, and the performance of the model can be improved.

[0120] Temporal information matrix: T (l) is used to fuse information in the time dimension. It can be obtained by analyzing historical node features. For example, a recurrent neural network (RNN) or a long short-term memory network (LSTM) can be used to process time series data to obtain the temporal features at each time step.

[0121] Model training: The training process of the improved TGCN model usually uses the backpropagation algorithm to update the learnable weight matrix W (l) . The training data can be historical network attack data. By annotating different attack stages and attack types, the model can learn the characteristics and patterns of attack behaviors.

[0122] In an alternative implementation, the dynamic knowledge graph construction in step C1 can also introduce more abundant node types, such as domain name nodes, file hash nodes, and user account nodes, to establish a more comprehensive representation of the attack scenario through various entity relationships and enhance the expressive power of the graph.

[0123] In another alternative implementation, the edge weight update mechanism in step C3 can also introduce a reinforcement learning method, regarding the edge weight adjustment as an action and the network security state as the environment, and optimizing the weight update strategy through a reward and punishment mechanism to improve the accuracy of attack prediction.

[0124] In yet another alternative implementation, the dynamic graph analysis in step C4 can also combine a graph neural network and a variational autoencoder to achieve generative modeling of the attack path, which can not only identify the existing attack paths but also predict the potential attack evolution directions.

[0125] It should be noted that by constructing a dynamic knowledge graph and applying a temporal graph convolutional network for analysis, the present invention realizes the accurate capture of the dynamic evolution process of attack behaviors, effectively solving the problem of lag existing in traditional correlation analysis methods. By integrating static network behavior characteristics with time dimension information, a temporal correlation between attack events is established, enabling the identification of potential attack paths at the initial stage of an attack, and significantly shortening the attack detection time.

[0126] In the embodiment of the present application, the traceability decision on the dynamic graph in step S400 includes steps D1 to D4:

[0127] D1: Generate multiple candidate attack paths based on the dynamic graph;

[0128] D2: For the encrypted traffic included in the candidate attack paths, restore the encrypted content through the quantum tunneling effect to complete the refinement of the candidate attack paths;

[0129] D3: Construct a virtual network environment based on the refined candidate attack paths, train adversarial agents for attackers and defenders in the virtual network environment, and verify and optimize the traceability strategy of the candidate attack paths through adversarial training;

[0130] D4: Share the optimized traceability strategy and attack features.

[0131] Specifically, in step D1, the implementation method of generating multiple candidate attack paths based on the dynamic graph is: Use the Monte Carlo tree search algorithm to generate candidate attack paths. This algorithm is based on the node and edge information of the dynamic graph, starting from the target node, and generating possible attack source paths through backward tracing. The search process is divided into four stages: selection, expansion, simulation, and backpropagation.

[0132] Selection stage: Start from the root node (target node), and select the most valuable path for exploration based on the upper confidence bound algorithm. The upper confidence bound value calculation will consider both the known value and exploration potential of the path, ensuring that the algorithm can balance between known high-value paths and under-explored paths.

[0133] Add a new node at the end of the selected path to expand the search tree. The selection of the new node is based on the connection relationship and edge weights in the dynamic graph, and nodes with high connection strength and temporal relationships conforming to the attack chain are preferentially selected. Starting from the newly expanded node, randomly simulate to the end of the path (possible attack source) according to the heuristic strategy. The simulation strategy will consider the security attributes and behavior characteristics of the nodes.

[0134] Update the value estimates of each node on the path according to the simulation results. The value calculation comprehensively considers the possibility, integrity, and degree of compliance with typical attack patterns of the path. By iterating the above process multiple times, finally select the top K paths with the highest value as candidate attack paths. To improve the diversity of the generated paths, the algorithm also introduces a path diversity reward mechanism to ensure that the generated candidate paths can cover different attack possibilities.

[0135] In step D2, the implementation method for recovering the encrypted content of the encrypted traffic included in the candidate attack path through the quantum tunneling effect is as follows:

[0136] Map the encrypted key search problem to the Ising model. The Ising model is a commonly used model in quantum computing and is suitable for solving combinatorial optimization problems. In the encrypted traffic analysis scenario, each possible key bit is represented as a spin variable with values of +1 or -1, corresponding to the bit values 1 or 0.

[0137] Based on the known characteristics of the encryption algorithm, construct the Hamiltonian function so that the energy state corresponding to the correct key is the lowest. The encoding process considers two types of constraints: internal logical constraints (the relationships between key bits) and external verification constraints. Set the interaction energy between spin variables according to the logical relationships between key bits and the characteristics of the encryption algorithm. For example, for key bits with dependencies, assign a higher interaction energy value; for independent bits, assign a lower or zero value. Determine the magnitude and direction of the external magnetic field based on the initial guess value of the key and the constraint conditions of the encryption algorithm. The external magnetic field reflects the tendency of prior knowledge and constraint conditions on the value of a specific key bit.

[0138] Submit the encoded problem to the quantum annealer for processing. The quantum annealer can efficiently explore different states in the energy landscape through the quantum tunneling effect and avoid getting trapped in local optimal solutions. The annealing process starts with high quantum fluctuations, gradually reduces the quantum effects and enhances the classical characteristics, and finally converges to the state with the lowest energy, that is, the possible correct key. Verify the candidate keys returned by the quantum annealer and try to decrypt the traffic samples using conventional computing methods. If the decryption is successful, use this key to recover the encrypted communication content and further refine the attack path information.

[0139] In step D3, the implementation method for training an adversarial agent based on the refined candidate attack paths to construct a virtual network environment is as follows: Based on the refined candidate attack paths, construct a virtual environment that reflects the real network topology and security characteristics. The environment includes various network devices, host systems, and security mechanisms. The environment modeling process also considers the uncertain factors in the real network, such as network latency, data packet loss, and system load fluctuations.

[0140] Design attacker agents and defender agents. The attacker agents simulate real attack behaviors and attempt to implement candidate attack paths; the defender agents attempt to identify and trace back attacks. Both types of agents adopt a deep reinforcement learning architecture, specifically including: the state space contains current network access permissions, acquired information, target distance, etc.; the action space contains various attack techniques; the reward function is designed based on the success degree, concealment, and efficiency of attacks, encouraging the agents to explore more concealed and efficient attack paths. The state space contains observed network behaviors, security alerts, system status, etc.; the action space contains various defense and tracing operations; the reward function is designed based on tracing accuracy, response speed, and resource consumption, encouraging the agents to quickly and accurately identify the attack source and attack path.

[0141] In step D4, the implementation method for sharing the optimized tracing strategy and attack features is as follows: Use blockchain technology and homomorphic encryption technology to build a secure sharing platform, and follow data protection regulations such as GDPR to ensure the security and compliance of threat intelligence during the sharing process.

[0142] In an optional implementation manner, the generation of candidate attack paths in step D1 can also be combined with the analysis from the attacker's perspective. By modeling the knowledge of penetration testing experts or collecting the behavior data of real attackers, the authenticity and integrity of path generation can be enhanced.

[0143] In another optional implementation manner, the recovery of encrypted content in step D2 can also be combined with side-channel analysis techniques. By analyzing side-channel information such as electromagnetic radiation, power consumption fluctuations, or time differences during the encryption process, it can assist in key recovery and is applicable to environments where quantum computing cannot be directly applied.

[0144] In yet another optional implementation manner, the construction of the virtual environment in step D3 can also adopt digital twin technology to create a high-fidelity virtual replica of the physical network, realizing the precise simulation of network behaviors, application systems, and security mechanisms, and improving the authenticity of adversarial training and the applicability of training results.

[0145] In the embodiments of this application, the optimization algorithm is a quantum-derived genetic algorithm based on quantum computing theory, which includes: using quantum bit pairs to represent genes in chromosomes to construct an initial population; updating the quantum bit states through quantum rotation gate operations to perform population evolution on the initial population; evaluating the advantages and disadvantages of individuals in the population through a fitness function, and improving the accuracy of concealed communication detection through iteration.

[0146] In an alternative embodiment, the optimization algorithm is a quantum - derived genetic algorithm based on quantum computing theory, which can also be combined with quantum interference gate operations. By introducing the quantum entanglement effect, the chromosome encoding can represent more complex feature correlations, thus enhancing the algorithm's detection ability for weak patterns in covert communication. Specifically, when implemented, after the standard quantum rotation gate operation, a Hadamard - CNOT sequence is added to create an entangled state, enabling the individual evolution process to consider the interactions between multiple features simultaneously and avoiding local optimal traps.

[0147] In another alternative embodiment, the method of constructing the initial population with quantum bits can also adopt a diversity - enhancement strategy. By performing cluster analysis on the initial population, it is ensured that the initial individuals are distributed in different regions of the solution space, improving the global search ability. At the same time, a topological structure protection mechanism is introduced to maintain the diversity level in the population and prevent premature convergence. This method is particularly suitable for complex and diverse network environments and can explore the potential covert communication feature space more comprehensively.

[0148] In yet another alternative embodiment, the design of the fitness function can also adopt a multi - objective optimization method, considering multiple indicators such as detection accuracy, false - alarm rate, and computational resource consumption simultaneously. The non - dominated solution set is selected through the Pareto - optimal principle, and the index weights can be flexibly adjusted in different application scenarios. In addition, the fitness function can introduce a time - varying factor. As the number of algorithm iterations increases, the weights of each index are dynamically adjusted, focusing on global search in the early stage and local optimization in the later stage to accelerate the algorithm convergence.

[0149] In a further alternative embodiment, when the computing environment does not support real quantum computing, the quantum - derived genetic algorithm can also be implemented on a classical computer through a quantum simulator. Although the performance advantages of real quantum computing cannot be obtained, through optimized matrix operations and parallel processing, the basic characteristics of the algorithm can still be maintained, and it is superior to traditional genetic algorithms in detecting complex covert communication. In actual deployment, the most suitable implementation method can be automatically selected according to the available computing resources to ensure applicability in various environments.

[0150] In the embodiments of this application, when analyzing the obtained dynamic knowledge graph, a temporal graph convolutional network is used. The processing process of the temporal graph convolutional network includes E1 - E4:

[0151] E1: Construct a node feature matrix to represent the feature information of each node in the graph;

[0152] E2: Generate a normalized adjacency matrix to represent the connection relationship between nodes;

[0153] E3: Create a temporal information matrix to capture the dynamic characteristics of node states changing over time;

[0154] E4: Combine the node feature matrix, the normalized adjacency matrix, and the temporal information matrix, and perform spatio-temporal fusion analysis on the graph-structured data through convolution operations to identify attack paths and evolution patterns.

[0155] In an alternative embodiment, the construction of the node feature matrix in step E1 can also introduce a graph autoencoder for unsupervised feature learning, automatically extract the hidden features of nodes through the reconstruction task, reduce the dependence on manually designed features, and improve the adaptability to unknown attack features.

[0156] In another alternative embodiment, the creation of the temporal information matrix in step E3 can also use a temporal convolutional network instead of LSTM, utilize causal convolution and dilated convolution techniques to capture temporal dependencies, have better parallel performance and a longer effective receptive field, and are suitable for processing attack behavior sequences with long-term dependencies.

[0157] In yet another alternative embodiment, the spatio-temporal fusion analysis in step E4 can also adopt a graph attention network mechanism, adaptively aggregate neighbor node information by learning the attention weights between nodes, improve the model's attention to key nodes and key relationships, and enhance the ability to identify complex attack chains.

[0158] It should be noted that the temporal graph convolutional network constructs a three-dimensional representation of network attack behaviors by fusing the node connection relationships in the spatial domain and the evolution characteristics in the temporal domain, effectively solving the lag problem existing in traditional correlation analysis methods. At the same time, this method has good interpretability, can clearly present the formation process of attack paths and key nodes, is convenient for security analysts to conduct in-depth investigations and traceability decisions, and provides strong technical support for network security protection.

[0159] The steps of collecting network data and preprocessing the network data include: collecting network data through a lightweight probe architecture; applying an improved wavelet denoising algorithm to the collected network data for noise filtering; constructing a multi-dimensional feature vector space; where the multi-dimensional feature vector space includes a timestamp synchronization dimension, a protocol stack feature dimension, a communication behavior dimension, a resource consumption dimension, a cross-layer behavior dimension, and a session persistence dimension.

[0160] Specifically, the probe deployment adopts a hierarchical distributed architecture, and probes with different functions are deployed on network boundaries, core switching nodes, and key servers to form a comprehensive data collection network. The types of network data collected include IPsec ESP / AH protocol fields at the network layer, TLS1.3 session identification information and QUIC protocol features at the transport layer, and inter-process communication data and memory paging analysis data at the application layer.

[0161] The improved wavelet denoising algorithm uses the Symlets-8 wavelet basis for 5-layer multi-scale decomposition. Through the adaptive threshold adjustment strategy and the high-frequency coefficient retention mechanism, it effectively filters out the random noise in the network data while retaining the characteristics of attack behaviors. Compared with traditional wavelet denoising, while maintaining the integrity of signal characteristics, the PSNR of the denoising effect of this algorithm increases by 2.3 dB, greatly improving the data quality for subsequent analysis.

[0162] The multi-dimensional feature vector space comprehensively characterizes network communication behaviors through six key dimensions: The timestamp synchronization dimension achieves high-precision time synchronization of ±100 nanoseconds based on the PTPv2 protocol; the protocol stack feature dimension extracts TLS fingerprint features and QUIC fingerprint features; the communication behavior dimension quantifies the communication mode through the information entropy theory; the resource consumption dimension monitors the changes of multi-granularity system resources; the cross-layer behavior dimension constructs a benchmark model based on RFC standards and calculates the deviation degree; the session persistence dimension captures the long-term session characteristics through a dynamic decay function.

[0163] It should be noted that the method for constructing this multi-dimensional feature vector space breaks through the limitations of traditional single data sources, realizes the all-round characterization of network behaviors, and provides a high-quality data basis for subsequent behavior feature extraction and attack graph construction. Practical deployment proves that this method can effectively address the problem of missing data dimensions, significantly improve the detection ability for hidden threats in encrypted communications, and lay a solid foundation for accurate network traceability.

[0164] Embodiment 3, the above is a schematic solution of a network traceability data processing method. It should be noted that the technical solution of the network traceability data processing system belongs to the same concept as the technical solution of the above network traceability data processing method. For the details not described in detail in the technical solution of the network traceability data processing system in this embodiment, reference can be made to the description of the technical solution of the above network traceability data processing method.

[0165] This embodiment also provides a network traceability data processing system, including:

[0166] A data processing module, configured to collect network data and preprocess the collected network data to generate preprocessed network data;

[0167] A feature extraction module, configured to extract behavior features from the network data to obtain network behavior features that can be used for subsequent analysis;

[0168] An attack graph construction module, configured to construct an attack graph based on the extracted network behavior features to form a dynamic graph that can represent the attack path;

[0169] A traceability decision module, configured to make traceability decisions on the constructed dynamic graph.

[0170] This embodiment also provides an electronic device applicable to the situation of network traceability data processing, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the network traceability data processing method as proposed in the above embodiment.

[0171] This embodiment also provides a storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the network traceability data processing method as proposed in the above embodiment.

[0172] The storage medium proposed in this embodiment and the network traceability data processing method proposed in the above embodiment belong to the same inventive concept. For technical details not described in detail in this embodiment, reference can be made to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0173] From the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software and necessary general-purpose hardware, and of course, it can also be implemented by hardware. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk, or optical disc of a computer, etc., including several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0174] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for processing network traceability data, characterized in that Including: Collect network data and preprocess the network data to generate preprocessed network data; Extract behavioral features from the preprocessed network data to obtain network behavior features; Construct an attack graph based on the obtained network behavior features to form a dynamic graph representing the attack path; Through traceability decision-making on the dynamic graph, locate the attack source and attack path through traceability decision-making.

2. The network traceability data processing method according to claim 1, wherein The extraction steps of the network behavior features include: Use an optimization algorithm to detect covert communication and extract covert communication features; Construct a network behavior benchmark model based on the network data, and identify abnormal network behaviors by comparing the deviation degree between real-time traffic features and the network behavior benchmark model; Fuse the covert communication features and the abnormal behavior features to generate network behavior features.

3. The network traceability data processing method according to claim 2, wherein The construction steps of the attack graph include: Determine entity nodes and behavior nodes according to the network behavior features, and construct a dynamic knowledge graph based on the entity nodes and behavior nodes; Calculate the association strength between nodes through the network behavior features and historical attack data to form an initial edge weight; Adjust the edge weights in the dynamic knowledge graph according to real-time network behaviors; Analyze the obtained dynamic knowledge graph to capture the evolution process of attack behaviors and form a dynamic graph representing the attack path.

4. The network traceability data processing method according to claim 3, characterized in that, Making traceability decisions on the dynamic graph includes the steps of: Generate multiple candidate attack paths based on the dynamic graph; For the encrypted traffic included in the candidate attack paths, recover the encrypted content through the quantum tunneling effect to complete the refinement of the candidate attack paths; Construct a virtual network environment based on the refined candidate attack paths, train attacker and defender adversarial agents in the virtual network environment, and verify and optimize the traceability strategy of the candidate attack paths through adversarial training; Share the optimized traceability strategy and attack features.

5. The network traceability data processing method according to claim 4, characterized in that The optimization algorithm is a quantum-derived genetic algorithm based on quantum computing theory, which includes: Use quantum bit pairs to represent genes in chromosomes and construct an initial population; Update the quantum bit states through quantum rotation gate operations to evolve the initial population; Evaluate the quality of individuals in the population through a fitness function, and improve the accuracy of covert communication detection through iteration.

6. The network traceability data processing method according to claim 5, wherein When analyzing the obtained dynamic knowledge graph, a temporal graph convolutional network is used, and the processing process of the temporal graph convolutional network includes: Construct a node feature matrix to represent the feature information of each node in the graph; Generate a normalized adjacency matrix to represent the connection relationship between nodes; Create a temporal information matrix to capture the dynamic characteristics of node states changing over time; Combine the node feature matrix, the normalized adjacency matrix and the temporal information matrix, and perform spatio-temporal fusion analysis on the graph structure data through convolution operations to identify attack paths and evolution patterns.

7. The network traceability data processing method according to claim 6, wherein The steps of collecting network data and preprocessing the network data include: Collect the network data through a lightweight probe architecture; Apply an improved wavelet denoising algorithm to the collected network data for noise filtering; Construct a multi-dimensional feature vector space; Among them, the multi-dimensional feature vector space includes a timestamp synchronization dimension, a protocol stack feature dimension, a communication behavior dimension, a resource consumption dimension, a cross-layer behavior dimension, and a session persistence dimension.

8. A network traceability data processing system, which applies the method according to any one of claims 1-7, characterized in that Including: A data processing module, configured to collect network data and preprocess the collected network data to generate preprocessed network data; A feature extraction module, configured to extract behavior features from the network data to obtain network behavior features that can be used for subsequent analysis; An attack graph construction module, configured to construct an attack graph based on the extracted network behavior features to form a dynamic graph that can represent the attack path; A traceability decision module, configured to make a traceability decision on the constructed dynamic graph.

9. An electronic device, including: A memory and a processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the network traceability data processing method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by a processor, the steps of the network traceability data processing method according to any one of claims 1 to 7 are implemented.

Citation Information

Cited By

  • Aging network propagation source detection method based on graph neural network

    CN120729628A

  • Safety control method, device and equipment of industrial network and medium

    CN120750650A

  • Heterogeneous atlas-based network attack path prediction method and device, and medium

    CN120785667A

  • Network attack path prediction method and device based on heterogeneous graph, and medium

    CN120785667B

  • Dynamic response method of network attack defense strategy based on multi-head attention fusion

    CN121037108A