A multi-dimensional industrial network behavior anomaly detection method
Through the multi-dimensional industrial network behavior abnormality detection method, multi-view correlation analysis and graph neural network are used to solve the problem of insufficient multi-dimensional recognition capabilities in high-dimensional data processing, and comprehensive detection and abnormal detection performance of industrial control network behavior is achieved. .
Patent Information
- Application Number
- CN202410163731.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-05
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-02-05
AI Technical Summary
The prior art lacks multi-dimensional recognition and understanding capabilities when processing high-dimensional network behavior data, resulting in poor efficiency and being unable to fully capture behavior patterns in complex network environments.
Using the multi-dimensional industrial network behavior anomaly detection method, by constructing a multi-view correlation analysis model and a multi-dimensional cross-domain shared learning model, combining graph neural networks, cross-domain feature representation and single-domain specific learning, strengthening the embedded information of each dimension, and based on predictions The deviation between behavior and actual behavior is calculated and the abnormal score is detected.
It realizes multi-dimensional anomaly detection of industrial control network behavior, improves detection performance and generalization capabilities, is significantly better than other methods, and provides more comprehensive information and more efficient anomaly detection results.
Smart Images

Figure CN117879968B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a network behavior anomaly detection method, and in particular to a multi-dimensional industrial network behavior anomaly detection method, belonging to the technical field of network behavior detection. Background Art
[0002] Existing technologies usually use anomaly detection methods based on deep learning or network behavior auditing methods based on statistical analysis to detect network behavior.
[0003] Deep learning-based anomaly detection methods take advantage of the advanced capabilities of deep learning technology in data analysis and pattern recognition. The advantages of deep learning-based anomaly detection are: (1) Multi-level feature learning: Deep learning networks gradually extract high-level features of data through multiple levels and are able to capture complex data patterns. This hierarchical feature learning is particularly effective in understanding the deep structure of data, especially when dealing with large-scale data sets. (2) Automatic feature extraction: Unlike traditional machine learning methods, deep learning methods can automatically learn and extract features from data without the need for manual design or selection of features. This automated feature extraction reduces the reliance on expertise and enhances the generalization ability of the model. (3) Nonlinear modeling capabilities: Deep learning models can capture complex nonlinear relationships in data through nonlinear activation functions. This is crucial for simulating complex patterns commonly seen in the real world. (4) Big data driven: Deep learning methods perform well when processing large amounts of data, which makes them particularly suitable for today's data-rich environment. As the amount of data increases, the performance of deep learning models generally improves. (5) End-to-end learning: Deep learning can achieve end-to-end learning, directly from raw data to final results, reducing the need for preprocessing and feature engineering. This simplifies the modeling process and may improve overall performance. (6) Multiple network architectures: Deep learning provides a variety of network architectures, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), autoencoders, etc., each of which is suitable for different types of data and tasks. This flexibility allows customized solutions to meet specific anomaly detection needs. (7) Time series analysis capabilities: For time series data, such as financial market data or network traffic, deep learning models (especially RNNs and their variants) can effectively capture temporal dependencies. This is crucial for predicting and detecting anomalies in time series. (8) Imbalanced data processing: Anomaly detection often faces the challenge of unbalanced data, that is, normal data far outnumbers abnormal data. Deep learning methods can effectively handle data imbalance problems through various techniques (such as resampling, cost-sensitive learning, etc.). These technical features of deep learning-based anomaly detection methods enable them to show strong performance in a variety of application scenarios, especially when processing high-dimensional and complex data sets. However, deep learning-based anomaly detection methods may face the curse of dimensionality when processing high-dimensional data, resulting in high computational complexity and inefficient model training;
[0004] The network behavior audit method based on statistical analysis is a traditional and effective technology for monitoring and analyzing network traffic and user behavior. The advantages of the network behavior audit method based on statistical analysis are: (1) Data aggregation and preprocessing: Before statistical analysis, network data usually needs to be aggregated and preprocessed to reduce noise and irrelevant information. Data preprocessing includes steps such as data cleaning, normalization, and variable transformation to ensure data quality and consistency. (2) Descriptive statistical analysis: Use basic statistical measures (such as mean, variance, and standard deviation) to describe the overall characteristics of network behavior. Descriptive statistics help understand the distribution and main trends of data. (3) Anomaly detection: By establishing normal patterns (benchmarks) of network behavior and then comparing real-time data with these benchmarks to identify abnormal behavior. Anomalies are usually defined as significant differences from normal behavior patterns. (4) Time series analysis: Analyze the time series characteristics of network behavior data to identify periodicity, trends, or sudden events. Time series analysis can be used to predict future network behavior or detect sudden behavioral changes. (5) Statistical hypothesis testing: Use statistical hypothesis testing to determine whether the observed data patterns are statistically significant. These tests help distinguish between random changes and true changes in behavior patterns. (6) Clustering and classification: Use cluster analysis to identify natural groups or patterns in network behavior. Classification methods can be used to distinguish normal behavior from abnormal behavior. (7) Association rule mining: Discover association rules in network behavior data and identify associations and sequence patterns between events. This can be used to understand the cause and effect of network behavior. (8) Visualization: Visualize network behavior data intuitively through visualization tools such as charts and heat maps. However, network behavior audit methods based on statistical analysis rely on a single or limited data dimension, such as only considering time or traffic. This limits the comprehensive understanding of behavior patterns in complex network environments. Summary of the invention
[0005] A brief overview of the present invention is provided below in order to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to identify key or important parts of the present invention, nor is it intended to limit the scope of the present invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to a more detailed description discussed later.
[0006] In view of this, in order to solve the technical problems existing in the prior art of lack of multi-dimensional identification and understanding of network behavior and poor efficiency in processing high-dimensional data, the present invention provides a multi-dimensional industrial network behavior anomaly detection method.
[0007] Solution 1: A multi-dimensional industrial network behavior anomaly detection method, comprising the following steps:
[0008] S1. Construct a behavioral analysis model for multi-view association analysis;
[0009] S2. Construct a multi-dimensional cross-domain shared learning model, taking the multi-dimensional view of the industrial control network and the shared node embedding as input and the new shared node embedding as output;
[0010] S3. Build a single domain specific learning model to learn node features within a dimension on a single dimension view of each industrial control system, so that the model can evaluate and detect anomalies in multiple dimensions;
[0011] S4. Perform single-view specific learning on multiple dimensions, strengthen the learning of embedded information on each dimension, reconstruct the predicted behavior value based on the embedded features of each dimension, and calculate the anomaly score on each dimension based on the deviation between the predicted behavior and the actual behavior;
[0012] S5. Set an abnormal score threshold. When the abnormal score is within the abnormal score threshold, an abnormal alarm is issued.
[0013] Preferably, S1 specifically comprises the following steps:
[0014] S11. Obtain initial data of the industrial control network, preprocess the data, and extract multi-dimensional features;
[0015] The initial data of the industrial control network consists of the sensor characteristic values of N sensors at the Ttrain time and network data; the sensor characteristic value is expressed as:
[0016]
[0017] Among them, s train represents the initial training data vector; represents the characteristic values of N sensors in the industrial control system at time scale t;
[0018] At each time tick t, the sensor value Form an N-dimensional vector, representing the characteristic values of N sensors in the industrial control system;
[0019] For each domain d, d = {1, ...., D}, construct an undirected weighted graph Since the information of each domain is interrelated and shares the same set of nodes, the cross-domain graph is constructed as an undirected weighted multigraph. It contains a node set V with N nodes and an edge set ε with D different types, that is, ε = {ε 1 ,...,ε D};
[0020] The multi-dimensional features include: control behavior sequence, control behavior invariant, data behavior sequence, data behavior invariant, and multi-view data change ratio;
[0021] S12. Constructing a single-dimensional view of the industrial control network and node embedding based on multi-dimensional features;
[0022] A weighted directed connectivity graph is constructed by mutual transmission of data between sensors in the physical dimension to represent the initial association between nodes in the physical dimension. A weighted directed connectivity graph is constructed in the network dimension by mutual transmission of network data packets between nodes, and the initial node embeddings of each dimension are generated through the processed data in the two dimensions.
[0023] S13. Combining multiple single-dimensional views of the industrial control network into multiple views to obtain a multi-dimensional view of the industrial control network and embedding shared nodes;
[0024] After obtaining the single graph representation and node embedding in two dimensions, a shared graph representation that simultaneously represents the node associations in two dimensions is constructed through the common node set characteristics.
[0025] Preferably, S2 specifically comprises the following steps:
[0026] S21. Obtain the weight matrix through multiple graphs, perform convolution with the weight matrix and shared node embedding as input, and iterate multi-layer graph convolution to obtain the learned new node shared embedding;
[0027] S22. embedding the new shared node through the output layer to obtain the reconstructed multi-dimensional node feature prediction value;
[0028] S23. Split the composite multi-dimensional node embedding vector into two vectors according to the ratio, corresponding to the node embedding in the physical dimension and the network dimension respectively, and input them into two specific domain convolution layers respectively with the specific graph representation in the two dimensions; use the residual network idea to feed the input back to the model output, wherein the cosine value is calculated to obtain the graph representation of each dimension which will be used as the input of the next round of specific domain convolution layer respectively, and the node embedding of each dimension in its output value will be reconstructed into the predicted value and used as the final output result of the training model;
[0029] S24. In each round, a new graph representation structure is solved based on the new shared node embedding to perform a new round of embedding learning.
[0030] Preferably, S24 specifically comprises calculating the similarity between the embedding vector of node i and its candidate neighbor node j; selecting the first k candidate neighbor nodes with the largest normalized dot product with node i as neighbors of node i, and all nodes and their neighbor nodes together constitute an adjacency matrix A; wherein the value of k is selected based on expectations and prior knowledge;
[0031] Compute the similarity between the embedding vector of node i and its candidate neighbor node j:
[0032] A ji=1{j∈TopK({e ii :k∈C i})} (3)
[0033] Among them, A ji represents the neighbor matrix of node i; C i represents the set of candidate neighbor nodes of node i; e ki represents the normalized dot product between the embedding vector of node i and the embedding vector of node j;
[0034] The working process of the shared convolutional layer is:
[0035]
[0036] in, represents the shared embedding of node i at the l+1th layer on the multigraph G, Represents the input shared feature matrix of node i; Represents the adjacency matrix A learned from the multi-graph G s The neighbor set of node i is obtained; W represents the shared weight matrix for training; α i,j represents the attention coefficient of nodes i and j; Represents the shared node embeddings of the lth layer output of the cross-domain shared graph convolutional layer.
[0037] Preferably, S3 comprises the following steps:
[0038] S31. The single domain specific learning model takes the single graph structure and specific node embedding in each dimension as input. In each dimension, the weight matrix is obtained through the graph structure, and the weight matrix and node embedding are used as input for CGN convolution. The information of the node is fused with its neighbors based on the learned graph structure, and the fused features are used as the final node embedding.
[0039] S32. The model uses the ReLu function as an activation function to increase the nonlinear relationship between each layer of the neural network; and embeds the learned nodes on each specific domain into a new round of specific domain graph representation structures obtained by calculation, and repeatedly iterates multi-layer graph convolution and feature extraction operations to obtain the learned final node embedding until the model converges or reaches a predetermined number of training times;
[0040] Based on the cross-domain shared learning module, the convolutional layer of the single-domain specific learning model is defined as:
[0041]
[0042] in, represents the node embedding of the output of the l+1th domain-specific convolutional layer in a single view; Represents the node embedding matrix of each dimension of the lth layer, which is composed of the shared node embedding matrix of the lth layer Calculate and obtain; From a single view The adjacency matrix A learned above s The neighbor set of node i is obtained; W d represents the shared weight matrix for training; α i,j represents the attention coefficient of nodes i and j; It represents the output of the lth specific graph convolutional layer, that is, the set of node embeddings of each dimension.
[0043] Preferably, S4 specifically converts the new shared node embedding learned in each dimension into a time series prediction value reconstructed according to the embedded features through the output layer linear mapping method, the prediction value is directly cross-compared with the current industrial control network data flow value in multiple dimensions, and anomaly assessment is performed based on the deviation between the predicted value and the actual value in each dimension and the threshold, and the assessment results in each dimension are used as the final anomaly detection result for the industrial control system.
[0044] Preferably, it also includes S6 multi-gradient descent optimization based on multi-task learning, including the following steps: S61. During each training process, the model calculates the relative weight of each task according to the parameter information of each task; the weight is dynamically adjusted during training to balance the learning of the model among multiple tasks; the Pareto optimal solution of the original problem is found by calculating the gradient of each task and continuously updating the parameter vector through the gradient until it meets the KKT condition;
[0045] S62. Introducing relative weights into the loss function so that the loss function of each task is weighted to reflect the importance of each task; performing reverse gradient propagation on the model using the weighted loss function to update the parameters in the model;
[0046] S63.Multiple iterative optimization, multi-gradient descent optimizer adaptively weighs the weights between multiple tasks.
[0047] Solution 2: An electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of a multi-dimensional industrial network behavior anomaly detection method described in Solution 1 when executing the computer program.
[0048] Solution 3: A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the multi-dimensional industrial network behavior anomaly detection method described in Solution 1 is implemented.
[0049] The beneficial effects of the present invention are as follows:
[0050] 1. This paper combines multi-dimensional information for anomaly detection and proposes a cross-domain multi-graph representation method for ICS. By analyzing the behaviors between ICS elements in different dimensions, the cross-domain features between ICS elements in different dimensions are represented on a multi-graph structure, which provides more comprehensive information for subsequent anomaly detection models.
[0051] 2. This paper proposes ICS anomaly detection based on graph neural network, which combines the graph structure to simultaneously learn the node representations of shared domains and specific domains, so as to perform joint training and anomaly detection in different dimensions; it follows the mechanism of multi-task learning, and therefore, through the multi-gradient descent optimization algorithm, multiple tasks are combined into a complete and efficient model, which makes this model more universal and generalizable;
[0052] 3. The proposed method is evaluated on the SWaT dataset containing multi-dimensional data to demonstrate its effectiveness and advantages compared with other baseline methods. The experimental results show that the proposed model significantly outperforms other state-of-the-art methods in anomaly detection performance, indicating its superior performance, which provides new ideas for research and application in the field of ICS anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0054] Figure 1 It is a flowchart of a multi-dimensional industrial network behavior anomaly detection method;
[0055] Figure 2 A schematic diagram of the behavior analysis model structure for multi-view correlation analysis;
[0056] Figure 3 It is a schematic diagram of the multi-graph representation structure;
[0057] Figure 4 Represents a schematic diagram of the structure for multiple multi-graphs;
[0058] Figure 5 This is a schematic diagram of a multi-dimensional cross-domain shared learning model;
[0059] Figure 6 A flowchart for the interaction between the multi-dimensional cross-domain shared learning model and the single-domain specific learning model;
[0060] Figure 7 Schematic diagram of the multi-gradient descent optimization algorithm flow. DETAILED DESCRIPTION
[0061] In order to make the technical solutions and advantages of the embodiments of the present invention more clearly understood, the exemplary embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.
[0062] Example 1: Reference Figure 1-Figure 7 This embodiment describes a multi-dimensional industrial network behavior anomaly detection method. Industrial control networks are different from traditional networks. They have more complex features, involve a wide variety of equipment, have denser network points, and are relatively fragile protocols, which leads to more security risks. Therefore, in view of the characteristics of industrial control networks, such as high dimensionality, strong correlation, high real-time, limited resources, and difficult updates, the present invention targets attack behaviors in industrial control networks, studies network behavior analysis and audit methods based on industrial characteristics, breaks through key technologies such as protocol parsing, industrial feature extraction, and abnormal behavior detection for industrial control network traffic, and realizes behavioral analysis and auditing of network behaviors in typical process industrial environments.
[0063] The research on behavioral analysis and auditing system for industrial Internet of Things obtains traditional industrial network data through two channels: industrial network data stream files and real-time monitoring, and parses the data packets layer by layer through protocol analysis algorithms. At least 25 industrial control network-specific protocols can be analyzed at the application layer. The preliminary protocol analysis results will be further used for multi-dimensional feature extraction to extract the required feature information from five main dimensions: control behavior sequence, control behavior invariant, data behavior sequence, data behavior invariant and multi-view combination. Further, behavioral prediction modeling is performed on the information learning results. The predicted features on multiple dimensions and the actual feature differences and threshold cross-comparison results will serve as the basis for the final anomaly detection and behavioral auditing. Based on the characteristics of industrial control networks, the present invention comprehensively explores the potential nonlinear associations of nodes in industrial control networks from the two dimensions of physical space and cyberspace, uses the complementary information of related domains to alleviate the sparsity problem, learns cross-domain shared representations and specific representations respectively, and predicts the future behavior of the industrial control system based on the reconstructed node relationship graph. Finally, based on the prediction, the abnormal alarm is responded and the deviation is explained. It can not only extract the intrinsic correlation in complex and high-dimensional industrial control data, but also improve the accuracy of the anomaly detection model without losing too much efficiency, and can provide a more reliable explanation of the abnormal deviation based on the reconstructed association graph, so as to effectively ensure the safety of the industrial control system.
[0064] The specific steps include:
[0065] S1. Construct a behavioral analysis model for multi-view association analysis;
[0066] S11. Obtain initial data of the industrial control network, preprocess the data, and extract multi-dimensional features;
[0067] The initial data of the industrial control network consists of the sensor characteristic values of N sensors at the Ttrain time (i.e. the time series of the read physical values) and network data; the sensor characteristic values are expressed as formula 2-1:
[0068]
[0069] Among them, s train represents the initial training data vector; represents the characteristic values of N sensors in the industrial control system at time scale t;
[0070] At each time tick t, the sensor value Form an N-dimensional vector, representing the characteristic values of N sensors in the industrial control system;
[0071] For each domain d, d = {1, ...., D}, construct an undirected weighted graph Since the information of each domain is interrelated and shares the same set of nodes, the cross-domain graph is constructed as an undirected weighted multigraph. It contains a node set V with N nodes and an edge set ε with D different types, that is, ε = {ε 1 ,...,ε D};
[0072] N sensor values, N-dimensional vectors, N nodes. In actual applications, the sensor values, vectors, and node values are in a corresponding relationship, that is, the number of sensors is consistent with the vector dimension and the number of nodes;
[0073] The multi-dimensional features include: control behavior sequence, control behavior invariant, data behavior sequence, data behavior invariant, and multi-view data change ratio;
[0074] S12. Constructing a single-dimensional view of the industrial control network and node embedding based on multi-dimensional features;
[0075] By transmitting data between sensors in the physical dimension, a weighted directed connectivity graph is constructed to represent the initial association between nodes in the physical dimension. By transmitting network data packets between nodes in the network dimension, a weighted directed connectivity graph is constructed in the network dimension, and the processed data in the two dimensions are used to generate their respective initial node embeddings. Figure 3 ;
[0076] S13. Combining multiple single-dimensional views of the industrial control network into multiple views to obtain a multi-dimensional view of the industrial control network and embedding shared nodes;
[0077] After obtaining the single graph representation and node embedding in two dimensions, a shared graph representation that simultaneously represents the node associations in two dimensions is constructed through the feature of common node sets.
[0078] Specifically, based on the above work, the probability of node relevance is calculated based on the node embedding matrix of each dimension:
[0079]
[0080] Among them, x i represents the embedding matrix of node i; x j represents the embedding matrix of node i; C i Represents the set of neighbor nodes of node i.
[0081] Apply the Topk threshold to identify the highest probability neighbors of each node and generate a graph structure for each dimension; merge edges of different dimensions into one graph, allowing multiple edges between any two nodes; generate a multi-dimensional graph that can effectively capture complex relationships in the data; in order to form node shared embeddings in the multi-graph structure, append the network dimension embedding vector to the physical dimension embedding vector to obtain a composite multi-dimensional node embedding with more features as the input of the multi-graph learning module; refer to Figure 4 ;
[0082] The ICS element information from different dimensions is integrated into a unified representation. By integrating the information representing multiple behavior domains, it can be used to analyze the correlation and characteristics between different dimensions of ICS behavior, thereby providing a data basis for detecting abnormal behavior in the network.
[0083] S2. Construct a multi-dimensional cross-domain shared learning model, taking the multi-dimensional view of the industrial control network and the shared node embedding as input and the new shared node embedding as output;
[0084] S21. Obtain weight matrices through multiple graphs, perform CGN (Graph Convolutional Networks) convolution with weight matrices and shared node embeddings as input, use Relu function as activation function to increase the nonlinear relationship between neural network layers, map the learned abstract embeddings to node feature matrix representations through a linear output layer, obtain learned node embeddings and graph representation structures obtained by calculating the cosine values of node embeddings, and repeatedly iterate multi-layer graph convolutions to obtain learned new node shared embeddings;
[0085] S22. embedding the new shared node through the output layer to obtain the reconstructed multi-dimensional node feature prediction value;
[0086] S23. The composite multi-dimensional node embedding vector is split into two vectors according to the ratio, corresponding to the node embeddings in the physical dimension and the network dimension respectively, and respectively input into two specific domain convolutional layers with the specific graph representations in the two dimensions; similar to the shared domain convolutional layer, the specific domain convolutional layer also uses the GCN algorithm for convolutional learning, but the residual network idea is used for the specific domain convolutional layer to feed the input back to the model output, wherein the cosine value is calculated to obtain the graph representation of each dimension, which will be used as the input of the next round of specific domain convolutional layer, and the node embeddings of each dimension in its output value will be reconstructed into predicted values and used as the final output result of the training model.
[0087] S24. In order to make the graph structure involved in the convolution better represent the node association, the cosine value calculation method is used to solve the new graph representation structure according to the new shared node embedding in each round;
[0088] Calculate the similarity between the embedding vector of node i and its candidate neighbor node j; select the first k candidate neighbor nodes with the largest normalized dot product with node i as neighbors of node i, and all nodes and their neighbors together form the adjacency matrix A; where the value of k is selected based on expectations and previous knowledge;
[0089] Compute the similarity between the embedding vector of node i and its candidate neighbor node j:
[0090] A ji =1{j∈TopK({e ii :k∈C i})} (3)
[0091] Among them, A ji represents the neighbor matrix of node i; C i represents the set of candidate neighbor nodes of node i; e ki represents the normalized dot product between the embedding vector of node i and the embedding vector of node j;
[0092] The working process of the shared convolutional layer is:
[0093]
[0094] in, represents the shared embedding of node i at the l+1th layer on the multigraph G, Represents the input shared feature matrix of node i; Represents the adjacency matrix A learned from the multi-graph G s The neighbor set of node i is obtained; W represents the shared weight matrix for training; α i,j represents the attention coefficient of nodes i and j; Represents the shared node embedding of the lth layer output of the cross-domain shared graph convolution layer; the structure of the multi-dimensional cross-domain shared learning model refers to Figure 5 ;
[0095] S3. The learning task of the single-domain specific learning module is more specific and fine-grained. It is responsible for learning the node features within the dimension on the single-dimensional view of each industrial control system, so that the model can evaluate and detect anomalies in multiple dimensions; Therefore, a single-domain specific learning model is constructed to learn the node features within the dimension on the single-dimensional view of each industrial control system, so that the model can evaluate and detect anomalies in multiple dimensions;
[0096] S31. The single domain specific learning model takes the single graph structure and specific node embedding in each dimension as input. In each dimension, the weight matrix is obtained through the graph structure, and the weight matrix and node embedding are used as input for CGN convolution. The information of the node is fused with its neighbors based on the learned graph structure, and the fused features are used as the final node embedding.
[0097] S32. The model uses the ReLu function as an activation function to increase the nonlinear relationship between each layer of the neural network; and embeds the learned nodes on each specific domain into a new round of specific domain graph representation structures obtained by calculation, and repeatedly iterates multi-layer graph convolution and feature extraction operations to obtain the learned final node embedding until the model converges or reaches a predetermined number of training times;
[0098] Based on the cross-domain shared learning module, the convolutional layer of the single-domain specific learning model is defined as:
[0099]
[0100] in, represents the node embedding of the output of the l+1th domain-specific convolutional layer in a single view; Represents the node embedding matrix of each dimension of the lth layer, which is composed of the shared node embedding matrix of the lth layer Calculate and obtain; From a single view The adjacency matrix A learned above s The neighbor set of node i is obtained; W d represents the shared weight matrix for training; α i,j represents the attention coefficient of nodes i and j; Represents the output of the lth specific graph convolutional layer, that is, the set of node embeddings in each dimension; the interaction process between the multi-dimensional cross-domain shared learning model and the single-domain specific learning model refers to Figure 6 ;
[0101] S4. Perform single-view specific learning on multiple dimensions, strengthen the learning of embedded information on each dimension, reconstruct the predicted behavior value based on the embedded features of each dimension, and calculate the anomaly score on each dimension based on the deviation between the predicted behavior and the actual behavior;
[0102] The new shared node embedding learned in each dimension is converted into a time series prediction value reconstructed according to the embedding feature through the linear mapping method of the output layer. The prediction value is directly cross-compared with the current industrial control network data flow value in multiple dimensions. The anomaly assessment is performed based on the deviation between the predicted value and the actual value in each dimension and the threshold, and the assessment results in each dimension are used as the final anomaly detection result for the industrial control system.
[0103] S5. Set an abnormal score threshold. When the abnormal score is within the abnormal score threshold, an abnormal alarm is issued.
[0104] S6. Multiple gradient descent optimization based on multi-task learning; including,
[0105] S61. During each training process, the model calculates the relative weight of each task based on the parameter information of each task. These weights are dynamically adjusted during training to balance the model across multiple tasks.
[0106] By properly planning this ratio, domain-sharing convolution can achieve better learning optimization, which in turn can affect the learning and training effect of the entire model. The process of dynamically adjusting weights is equivalent to solving the Pareto optimal solution of the multi-objective optimization problem, and its necessary condition is the KKT (Karush-Kuhn-Tucker) condition
[37] . Therefore, the parameter vector can be continuously updated by calculating the gradient of each task until it meets the KKT condition, that is, the Pareto optimal solution of the original problem is found.
[0107] S62. These relative weights are introduced into the loss function so that the loss function of each task is weighted to reflect the importance of each task. This weighted loss function is then used to perform back gradient propagation on the model to update the parameters in the model. In this way, the model can optimize multiple tasks at the same time, rather than optimizing each task individually.
[0108] S63. Through multiple iterations of optimization, the multi-gradient descent optimizer can adaptively weigh the weights between multiple tasks, so that the optimization effect of each task can be fully utilized, thereby improving the training efficiency and accuracy of the model. At the same time, since the multi-gradient descent optimizer can calculate the gradients of all tasks in a single backward pass, it can reduce the computational overhead and improve the convergence speed and training speed of the model.
[0109] Since the present invention is mainly composed of a multi-layer neural network and each layer is interrelated, the weights involved that need to be updated are relatively complex, so it is particularly critical to deal with the weight distribution problem among multiple tasks. For this purpose, the method of minimizing the weighted linear combination of multiple tasks is generally adopted. This method is only effective when the tasks are not competitive, but this situation is rare. In particular, in the model of the present invention, there is an obvious competitive relationship between the learning tasks in each dimension, and the distribution of such competitive weights will affect the correctness of the prediction. Therefore, these methods are difficult to achieve the best optimization effect in the learning model of the present invention.
[0110] An optimization algorithm based on the Multiple Gradient Descent Algorithm (MGDA) uses a gradient-based optimization method and provably converges to a point on the Pareto set. It is calculated through a single backward pass without explicit task-specific gradients, making the computational overhead of this method negligible. The multiple gradient descent optimizer will adaptively weigh the weight distribution among multiple tasks, and reasonably allocate the weight of the next round of learning to each task according to the task objectives and training gradients, so that the model can achieve better training results.
[0111] The multi-gradient descent optimizer can reasonably allocate weights for multiple tasks of the multi-view association learning model when performing feed-forward optimization of their respective losses. The multi-gradient descent optimizer takes the parameters of each task after each round of learning as input, and calculates the Pareto optimal solution of each task parameter sequence as the loss value of multiple tasks to calculate their respective weights. This weight is used as a relative proportion and multiplied by the respective loss, and the sum is taken as the loss value of the parent task, that is, the multi-dimensional cross-domain shared learning layer, to perform feed-forward optimization on the entire network.
[0112] The present invention is compared with GDN, TranAD, DTAAD, USAD, MAD-GAN, LSTM-NDT and other methods, with precision, recall and F1 score as the comparison criteria of the experiment. The results show that the best experimental results of anomaly detection of six-process water treatment industrial control systems on the SWaT dataset are F1 scores up to 83.5%, precision up to 83.4%, and recall up to 84.5%, which have achieved results that are better than other current methods;
[0113] The present invention innovatively extracts features of industrial control networks from multiple dimensions, and uses the idea of cross-fusion to deeply explore the nonlinear correlations inherent in the industrial control network through cross-domain shared learning to more comprehensively construct a graph representation structure and node feature embedding, and further conducts targeted feature learning in each dimension based on the richer and more comprehensive results obtained from cross-domain shared learning, so as to fully learn the potential characteristics of the industrial control network in each dimension, and more accurately predict the behavior of the industrial control network based on this, and judge the current behavior based on the predicted value. Therefore, compared with all other models that only consider industrial control network information in a single dimension, this model has better detection effect.
[0114] Embodiment 2: The computer device of the present invention may be a device including a processor and a memory, such as a single chip microcomputer including a central processing unit, etc. Moreover, the processor is used to implement the steps of the above-mentioned multi-dimensional industrial network behavior anomaly detection method when executing the computer program stored in the memory.
[0115] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.
[0116] The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.
[0117] Embodiment 3: Computer readable storage medium embodiment.
[0118] The computer-readable storage medium of the present invention can be any form of storage medium that can be read by a processor of a computer device, including but not limited to non-volatile memory, volatile memory, ferroelectric memory, etc. A computer program is stored on the computer-readable storage medium. When the processor of the computer device reads and executes the computer program stored in the memory, the steps of the above-mentioned multi-dimensional industrial network behavior anomaly detection method can be implemented.
[0119] The computer program includes computer program code, which may be in source code form, object code form, executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer readable media do not include electric carrier signals and telecommunication signals.
[0120] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments may be envisioned within the scope of the invention thus described. In addition, it should be noted that the language used in this specification is selected primarily for readability and teaching purposes, rather than for explaining or defining the subject matter of the present invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the present invention is illustrative, not restrictive, with respect to the scope of the present invention, which is defined by the appended claims.
Claims
1. A multi-dimensional industrial network behavior anomaly detection method, characterized in that: The following steps are involved: S1. Constructing a behavior analysis model for multi-view association analysis, including the following steps: S11. Obtain initial data of the industrial control network, preprocess the data, and extract multi-dimensional features; The initial data of the industrial control network consists of the sensor characteristic values and network data of N sensors at 𝑇train time; the sensor characteristic value is expressed as: (1) in, represents the initial training data vector; represents the characteristic values of N sensors in the industrial control system at time scale t; At each time scale t, the sensor characteristic value Form an N-dimensional vector, representing the characteristic values of N sensors in the industrial control system; For each domain , construct an undirected weighted graph ; Since the information of each domain is interrelated and shares a completely consistent set of nodes, the cross-domain graph is constructed as an undirected weighted multigraph , which contains a node set V with N nodes and an edge set with D different types ,Right now ; The multi-dimensional features include: control behavior sequence, control behavior invariant, data behavior sequence, data behavior invariant, and multi-view data change ratio; S12. Constructing a single-dimensional view of the industrial control network and node embedding based on multi-dimensional features; A weighted directed connectivity graph is constructed by mutual transmission of data between sensors in the physical dimension to represent the initial association between nodes in the physical dimension. A weighted directed connectivity graph is constructed in the network dimension by mutual transmission of network data packets between nodes, and the initial node embeddings of each dimension are generated through the processed data in the two dimensions. S13. Combine multiple single-dimensional views of the industrial control network into multiple views to obtain a multi-dimensional view of the industrial control network and embed shared nodes; After obtaining the single graph representation and node embedding in two dimensions, a shared graph representation that simultaneously represents the node associations in two dimensions is constructed through the common node set characteristics; S2. Construct a multi-dimensional cross-domain shared learning model, taking the multi-dimensional view of the industrial control network and the shared node embedding as input and the new shared node embedding as output; S3. Build a single domain specific learning model to learn node features within a dimension on a single dimension view of each industrial control system, so that the model can evaluate and detect anomalies in multiple dimensions; S4. Perform single-view specific learning on multiple dimensions, strengthen the learning of embedded information on each dimension, reconstruct the predicted behavior value based on the embedded features of each dimension, and calculate the anomaly score on each dimension based on the deviation between the predicted behavior and the actual behavior; S5. Set an abnormal score threshold. When the abnormal score is within the abnormal score threshold, an abnormal alarm is issued.
2. A multi-dimensional industrial network behavior anomaly detection method according to claim 1, characterized in that: S2 specifically includes the following steps: S21. Obtain the weight matrix through multiple graphs, perform convolution with the weight matrix and shared node embedding as input, and iterate multi-layer graph convolution to obtain the learned new node shared embedding; S22. embedding the new shared node through the output layer to obtain the reconstructed multi-dimensional node feature prediction value; S23. Split the composite multi-dimensional node embedding vector into two vectors according to the ratio, corresponding to the node embedding in the physical dimension and the network dimension respectively, and input them into two specific domain convolution layers respectively with the specific graph representation in the two dimensions; use the residual network idea to feed the input back to the model output, wherein the cosine value is calculated to obtain the graph representation of each dimension which will be used as the input of the next round of specific domain convolution layer respectively, and the node embedding of each dimension in its output value will be reconstructed into the predicted value and used as the final output result of the training model; S24. In each round, a new graph representation structure is solved based on the new shared node embedding to perform a new round of embedding learning.
3. A multi-dimensional industrial network behavior anomaly detection method according to claim 2, characterized in that: S24 specifically calculates the similarity between the embedding vector of node i and its candidate neighbor node j; selects the first k candidate neighbor nodes with the largest normalized dot product with node i as neighbors of node i, and all nodes and their neighbor nodes together constitute an adjacency matrix A; wherein the value of k is selected based on expectations and prior knowledge; Compute the similarity between the embedding vector of node i and its candidate neighbor node j: (3) in, represents the neighbor matrix of node i; Represents the set of candidate neighbor nodes of node i; represents the normalized dot product between the embedding vector of node i and the embedding vector of node j; The working process of the shared convolutional layer is: (4) in, represents the shared embedding of node i at the l+1th layer on the multigraph G, Represents the input shared feature matrix of node i; Represents the adjacency matrix learned from the multi-graph G Obtain the neighbor set of node i; Represents the shared weight matrix for training; represents the attention coefficient of nodes i and j.
4. A multi-dimensional industrial network behavior anomaly detection method according to claim 1, characterized in that: S3 includes the following steps: S31. The single domain specific learning model takes the single graph structure and specific node embedding in each dimension as input. In each dimension, the weight matrix is obtained through the graph structure, and the weight matrix and node embedding are used as input for CGN convolution. The information of the node is fused with its neighbors based on the learned graph structure, and the fused features are used as the final node embedding. S32. The model uses the ReLu function as the activation function to increase the nonlinear relationship between the layers of the neural network; The learned nodes on each specific domain are embedded into a new round of specific domain graph representation structures obtained through calculation, and the learned final node embedding is obtained by repeated iteration of multi-layer graph convolution and feature extraction operations until the model converges or reaches a predetermined number of training times; Based on the cross-domain shared learning module, the convolutional layer of the single-domain specific learning model is defined as: (5) in, represents the node embedding of the output of the l+1th domain-specific convolutional layer in a single view; Represents the node embedding matrix of each dimension of the lth layer, which is composed of the shared node embedding matrix of the lth layer Calculate and obtain; From a single view The adjacency matrix learned above Obtain the neighbor set of node i; Represents the shared weight matrix for training; represents the attention coefficient of nodes i and j; It represents the output of the lth specific graph convolutional layer, that is, the set of node embeddings of each dimension.
5. A multi-dimensional industrial network behavior anomaly detection method according to claim 1, characterized in that: Specifically, S4 converts the new shared node embedding learned in each dimension into a time series prediction value reconstructed according to the embedded features through the output layer linear mapping method. The prediction value is directly cross-compared with the current industrial control network data flow value in multiple dimensions. Anomaly assessment is performed based on the deviation between the predicted value and the actual value in each dimension and the threshold, and the assessment results in each dimension are used as the final anomaly detection result for the industrial control system.
6. A multi-dimensional industrial network behavior anomaly detection method according to claim 1, characterized in that: It also includes S6's multi-task learning-based multi-gradient descent optimization, including, S61. During each training process, the model calculates the relative weight of each task based on the parameter information of each task; the weight is dynamically adjusted during training to balance the learning of the model among multiple tasks; the Pareto optimal solution of the original problem is found by calculating the gradient of each task and continuously updating the parameter vector through the gradient until it meets the KKT condition; S62. Introducing relative weights into the loss function so that the loss function of each task is weighted to reflect the importance of each task; performing reverse gradient propagation on the model using the weighted loss function to update the parameters in the model; S63.Multiple iterative optimization, multi-gradient descent optimizer adaptively weighs the weights between multiple tasks.
7. An electronic device, characterized in that: It includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of a multi-dimensional industrial network behavior anomaly detection method as described in any one of claims 1 to 6 when executing the computer program.
8. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the multi-dimensional industrial network behavior anomaly detection method described in any one of claims 1-6 is implemented.
Citation Information
Patent Citations
Abnormal behavior detection method based on user behavior analysis
CN107196953A
Behavior anomaly recognition method and device based on mass data full-text retrieval
CN112579728A
Apparatus having engine using artificial intelligence for detecting bot anomalies in a computer network
US11457031B1
