Network node importance quantitative evaluation method based on deep learning

The network node graph structure model is constructed through deep learning methods, and the importance of network nodes is evaluated using graph convolution neural networks, which solves the problems of sensitivity to network structure changes and high computational complexity in the existing technology, and realizes quantitative evaluation of network node importance and accuracy of system security assessment.

CN119052108APending Publication Date: 2024-11-29GUIZHOU AEROSPACE INST OF MEASURING & TESTING TECH
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411170663.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The prior art is sensitive to network structure changes in the evaluation of network node importance, has high computational complexity, cannot fully reflect the status of nodes, and requires a large amount of computing resources.

Method used

Using deep learning methods, a network node graph structure model is constructed, a network node importance evaluation model is formed through graph convolution neural network, node importance evaluation model is formed, nodes are deleted to form reconstruction diagrams, calculation of the difference between the original diagram and reconstruction diagram, and establishing a network node importance sequence , quantitatively evaluate the importance of nodes.

Benefits of technology

It effectively reduces the computational complexity, realizes quantitative evaluation of network node importance, provides a basis for system network security evaluation, and ensures the effectiveness, authenticity and accuracy of the evaluation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119052108A_ABST
    Figure CN119052108A_ABST
Patent Text Reader

Abstract

The invention discloses a network node importance quantitative evaluation method based on deep learning, and the method comprises the steps: constructing a network node graph structure model of a tested system, and constructing a network node importance comprehensive evaluation index; forming a network node importance evaluation model by using a graph convolutional neural network framework, forming a network node reconstruction graph by deleting network nodes, calculating the difference between the network node reconstruction graph and an original graph, and establishing a network node importance degree sequence according to the difference change condition; the problems that network node importance evaluation in the prior art is sensitive to changes of a network structure, the calculation complexity is high, the status of nodes in the whole network cannot be comprehensively reflected and cannot be quantitatively expressed, and a network node importance evaluation technical scheme method is easily influenced by the complexity of a tested system and cannot be used for evaluating the network node importance are solved. And a large number of computing resources are needed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and in particular relates to a method for quantitatively evaluating the importance of network nodes based on deep learning. Background Art

[0002] In a complex system, the importance of a network node determines the security protection strength of the network node, especially in a protection environment with limited resources or low-latency communication requirements. Therefore, accurate and reasonable quantitative assessment of the importance of network nodes is crucial.

[0003] Traditional methods for quantitatively evaluating the importance of network nodes include degree centrality, betweenness centrality, and eigenvector centrality, which are mainly based on network topology and mathematical models. As the scale and complexity of networks increase, these traditional methods face many challenges in processing high-dimensional, heterogeneous, and dynamic network data. The rapid development of deep learning technology has brought new opportunities for evaluating the importance of network nodes. Deep learning can automatically extract high-level features from data through multiple layers of nonlinear transformations, especially when processing graph-structured data. Graph neural networks are deep learning models used to process graph-structured data. They update the embedding vector of each node in an iterative manner to capture the complex relationship between the node and its neighboring nodes. Graph convolutional networks can effectively extract local structural features of nodes by extending traditional convolution operations to graph data. Graph attention networks introduce an attention mechanism that can dynamically adjust the weights of neighboring nodes based on the similarity between nodes.

[0004] Through the development of the above background technologies, the quantitative evaluation methods of network node importance based on deep learning have made significant progress and applications in many fields. These technologies not only improve the accuracy and efficiency of evaluation, but also provide powerful tools for in-depth research and application of networks.

[0005] After searching and analyzing, most of the patents that have been published focus on the network nodes themselves, are sensitive to changes in the network structure, have high computational complexity, cannot fully reflect the status of the nodes in the entire network, and cannot be quantitatively expressed. Especially for expert systems that rely on the construction and maintenance of expert knowledge bases, the applicability is relatively narrow. The original technical solution for evaluating the importance of network nodes is easily affected by the complexity of the system being tested and requires a lot of computing resources; Summary of the invention

[0006] The technical problem to be solved by the present invention is: to provide a quantitative evaluation method for the importance of network nodes based on deep learning, so as to solve the technical problems existing in the prior art of network node importance evaluation, such as being sensitive to changes in network structure, having high computational complexity, being unable to fully reflect the status of the node in the entire network, being unable to express quantitatively, being easily affected by the complexity of the system being tested, and requiring a large amount of computing resources.

[0007] The technical solution of the present invention is:

[0008] A method for quantitatively evaluating the importance of network nodes based on deep learning, the method comprising: constructing a network node graph structure model of a system under test, and constructing a comprehensive evaluation index of the importance of network nodes; using a graph convolutional neural network framework to form a network node importance evaluation model, forming a network node reconstruction graph by deleting network nodes, calculating the difference between the network node reconstruction graph and the original graph, and establishing a network node importance sequence according to the change in the difference.

[0009] The method of constructing the network node graph structure model of the system under test and constructing the comprehensive evaluation index of the importance of the network nodes includes:

[0010] Step 1: Represent the network graph in the form of G(V,E), where G(·) is an undirected and unweighted graph, V = {v1,v2,v3,…,v n} is the set of network nodes, v i (i=1,2,…,n) is a single node in the network, E={e1,e2,e3,…,e m} is the edge between nodes, e i (i=1,2,…,m) is the wireless link for communication between nodes, n and m are the number of nodes and the number of links respectively;

[0011] Step 2: Calculate node degree centrality: If there are n nodes in the network, then node v i The degree centrality of is: Among them, k i For node v i The node degree, C d (v i ) is the node v i The degree centrality of

[0012] Step 3: Calculate node v i Betweenness centrality: non-adjacent nodes v in the network i and v j The shortest path between them will pass through some nodes, a node v i If a node is passed by other shortest paths, it means that the node is very important in the network, and its importance or influence is characterized by the node's betweenness; B i is the node betweenness.

[0013] Step 4. Calculate relative importance: v i To node v j Distance:d ij =|v i v j|. Node v i The relative importance of is, μ m (v i ) is the node starting from the node and returning to the node v through m edges. i The number of circuits;

[0014] Step 5: Node attribute matrix construction: Assume that there are n network nodes in the network, and the corresponding important node solution set is {A1, A2, …, A n}, there are k node importance evaluation indicators, that is, there are k scheme attributes, where k = 4, and the initial decision matrix P is constructed, that is:

[0015]

[0016] Among them, A i (f r )(i=1,2,…,n;r=1,2,3,4) is the node solution; f r (r=1,2,3,4) is the scheme attribute;

[0017] Step 6: Use the coefficient of variation method to determine the normalized weight vector A of each attribute index i w T =[w1,w2,w3,w4], weight of each indicator w i (i=1,2,3,4) is where v i is the coefficient of variation of the i-th indicator, and σ i is the standard deviation coefficient of the i-th indicator; is the average of the i-th indicator; k is the number of indicators;

[0018] Step 7: Calculate the comprehensive evaluation index based on the initialization decision matrix and the weights of the selected indicators such as degree centrality, betweenness centrality, and relative importance of nodes, namely:

[0019]

[0020] Among them, M is the decision coefficient matrix, p ij (i=1,2,…,n;j=1,2,3,4) are the initial decision matrix elements.

[0021] The method of using the graph convolutional neural network framework to form a network node importance evaluation model includes:

[0022] Step 8: Construct a symmetric normalized Laplace matrix

[0023] Step 9: Let the node feature be X = {x1, x2, ..., x n},σ is a nonlinear activation function, w is a weight parameter matrix,

[0024] Step 10: Introduce attention into the graph structure data to form a graph attention network to learn the edge weights between two adjacent nodes and implement convolution operations between nodes; In the graph structure data, a graph attention network is formed to learn the edge weights between two adjacent nodes and implement convolution operations between nodes. The attention weight is:

[0025]

[0026] Step 11: Update the expression at the graph level: Specific to the expression update on the node:

[0027] N i Represents node v i The neighbor node set, the weight parameter matrix W is used to complete the feature dimension transformation of each node, the parameter α is used to calculate the attention weight between nodes, || represents vector concatenation, α ij represents the node v calculated under α i ,v j , and α(·) represents the nonlinear activation function.

[0028] The method of forming a network node reconstruction graph includes:

[0029] Step 12: In an undirected and unweighted network, the failure of a node means that all edges of the node are invalid. When a node v is deleted from the entire network, i That is, all the edges e related to this node ij Delete, select and delete a node to get the reconstructed graph

[0030] Methods for network node importance sequence include:

[0031] Step 13. Self-supervised learning training: encoder f θ The goal is to learn each node v from the pre-training dataset graph G i The low-dimensional representation (i.e. node embedding) h i ∈H, decoder p θ The learned graph embedding H is used as the input of the auxiliary task for pre-training. θ It can be any neural network for graph learning. The encoder f θ It can be regarded as the encoder f in the fine-tuning stage θinitInitialization of parameters. Afterwards, under the supervision of a specific downstream task, the fine-tuning dataset G t Fine-tune the downstream decoder and the pre-trained encoder f θinit Make fine adjustments.

[0032] Step 14: Calculate the loss: Calculate the loss value of each layer node;

[0033] Step 15, traverse all nodes v: calculate the loss values ​​of all nodes and form an importance sequence;

[0034] Step 16. Obtain importance sequence: The difference between the reconstructed graph and the original graph is used as the importance index of the node, L = (l1, l2, l3, ..., l N ). The greater the difference between the reconstructed graph and the original graph, the greater the impact of the node on the structure and function of the entire network, and the higher the importance of the node. After traversing all the nodes in the network, the importance sequence of the nodes can be obtained.

[0035] Beneficial effects of the present invention:

[0036] The present invention uses a deep learning method to traverse all network nodes to carry out quantitative evaluation of the importance of network nodes, effectively solving the problems of dependency, subjectivity and inapplicability of expert system evaluation schemes. At the same time, the system under test is abstracted as a network node to obtain the adjacency matrix and node feature matrix of the network, simplifying the complexity of the system and retaining the network structure characteristics. The complex network structure is expressed as a network graph, reducing the complexity of the computational algorithm during the deep learning training process.

[0037] The present invention realizes the importance evaluation of network nodes in a quantitative evaluation system. The system under test is abstracted as a network node, and the weighted sum between each node's own information and adjacent information is used as the latest feature of the current node to characterize the flow and propagation of features and messages on the graph network. A graph convolutional neural network is used to iteratively aggregate the features of neighboring nodes, update the features of the current node, and simultaneously learn the structural information of the network and the attribute information of the node to form an original graph to represent the network node. The original graph is reconstructed, and the difference between the original graph and the reconstructed graph is used as the importance index of the node. The greater the difference between the original graph and the reconstructed graph, the greater the impact of the node on the structure and function of the entire network, and the higher the importance of the node. After traversing all nodes in the network, the importance sequence of the nodes can be obtained. The method ensures the characteristics and structure of the network nodes of the system under test, realizes the importance evaluation of network nodes, provides a basis for the weight allocation of system network security evaluation nodes, and ensures that the system network security evaluation is effective, true and accurate.

[0038] The present invention solves the technical problems that the existing technology for evaluating the importance of network nodes is sensitive to changes in the network structure, has high computational complexity, cannot fully reflect the status of the node in the entire network, cannot express quantitatively, is easily affected by the complexity of the system under test, and requires a large amount of computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0040] A quantitative evaluation method for the importance of network nodes based on deep learning includes constructing a network node graph structure model of the system under test, calculating the network node betweenness centrality, node relative importance, etc., constructing a comprehensive evaluation index of the importance of network nodes, using a graph convolutional neural network framework to learn based on the comprehensive evaluation index of the importance of network nodes, forming a network node importance evaluation model, deleting network nodes to form a network node reconstruction graph, calculating the difference between the network node reconstruction graph and the original graph, establishing a network node importance sequence based on the difference change, quantitatively evaluating the importance of network nodes, providing a basis for the node weight allocation of the system network security evaluation, and ensuring that the system network security evaluation is effective, true and accurate; it specifically includes:

[0041] 1. Node correlation importance calculation: Calculating node correlation is used to construct a comprehensive evaluation index of network node importance.

[0042] Step 1: Represent the network graph in the form of G(V,E), where G(·) is an undirected and unweighted graph, V = {v1,v2,v3,…,v n} is the set of network nodes, v i (i=1,2,…,n) is a single node in the network, E={e1,e2,e3,…,e m} is the edge between nodes, e i (i=1,2,…,m) is the wireless link for communication between nodes, n and m are the number of nodes and the number of links respectively;

[0043] Step 2: Calculate node degree centrality: If there are n nodes in the network, then node v i The degree centrality of is: Among them, k i For node v i The node degree, C d (v i ) is the node v i The degree centrality of

[0044] Step 3: Calculate node v i Betweenness centrality: non-adjacent nodes v in the network i and vj The shortest path between them will pass through some nodes, a node v i If a node is passed by other shortest paths, it means that the node is very important in the network, and its importance or influence is characterized by the node's betweenness; B i is the node betweenness.

[0045] Step 4. Calculate relative importance: v i To node v j Distance:d ij =|v i v j |. Node v i The relative importance of is, μ m (v i ) is the node starting from the node and returning to the node v through m edges. i The number of circuits;

[0046] 2. Comprehensive evaluation of node importance: By constructing comprehensive evaluation indicators, it is used to train the graph convolutional neural network key node detection model.

[0047] Step 5: Node attribute matrix construction: Assume that there are n network nodes in the network, and the corresponding important node solution set is {A1, A2, …, A n}, there are k node importance evaluation indicators, that is, there are k scheme attributes, where k = 4, and the initial decision matrix P is constructed, that is:

[0048]

[0049] Among them, A i (f r )(i=1,2,…,n;r=1,2,3,4) is the node solution; f r (r=1,2,3,4) is the scheme attribute;

[0050] Step 6: Use the coefficient of variation method to determine the normalized weight vector A of each attribute index i w T =[w1,w2,w3,w4], weight of each indicator w i (i=1,2,3,4) is where v i is the coefficient of variation of the i-th indicator, and σ i is the standard deviation coefficient of the i-th indicator; is the average of the i-th indicator; k is the number of indicators;

[0051] Step 7: Calculate the comprehensive evaluation index based on the initialization decision matrix and the weights of the selected indicators such as degree centrality, betweenness centrality, and relative importance of nodes, namely:

[0052]

[0053] Among them, M is the decision coefficient matrix, p ij (i=1,2,…,n;j=1,2,3,4) are the initial decision matrix elements.

[0054] 3. Key node detection model based on graph convolutional neural network: Graph convolutional neural network is based on the perspective of spectrum. Its essence is to perform Laplace smoothing on the network. It takes the weighted sum of each node’s own information and adjacent information as the latest feature of the current node to characterize the flow and propagation of features and messages on the graph network.

[0055] Step 8: Construct a symmetric normalized Laplace matrix

[0056] Step 9: Let the node feature be X = {x1, x2, ..., x n},σ is a nonlinear activation function, w is a weight parameter matrix,

[0057] Step 10: Introduce attention into the graph structure data to form a graph attention network to learn the edge weights between two adjacent nodes and implement convolution operations between nodes; In the graph structure data, a graph attention network is formed to learn the edge weights between two adjacent nodes and implement convolution operations between nodes. The attention weight is:

[0058]

[0059] Step 11: Update the expression at the graph level: Specific to the expression update on the node:

[0060] N i Represents node v i The neighbor node set, the weight parameter matrix W is used to complete the feature dimension transformation of each node, the parameter α is used to calculate the attention weight between nodes, || represents vector concatenation, α ij represents the node v calculated under α i ,v j , and α(·) represents the nonlinear activation function.

[0061] The method of forming a network node reconstruction graph includes:

[0062] Step 12: In an undirected and unweighted network, the failure of a node means that all edges of the node are invalid. When a node v is deleted from the entire network, i That is, all the edges e related to this node ij Delete, select and delete a node to get the reconstructed graph

[0063] Methods for establishing a network node importance sequence include:

[0064] Step 13. Self-supervised learning training: encoder f θ The goal is to learn each node v from the pre-training dataset graph G i The low-dimensional representation (i.e. node embedding) h i ∈H, decoder p θ The learned graph embedding H is used as the input of the auxiliary task for pre-training. θ It can be any neural network for graph learning. The encoder f θ It can be regarded as the encoder f in the fine-tuning stage θinit Initialization of parameters. Afterwards, under the supervision of a specific downstream task, the fine-tuning dataset G t Fine-tune the downstream decoder and the pre-trained encoder f θinit Make fine adjustments.

[0065] Step 14: Calculate the loss: Calculate the loss value of each layer node;

[0066] Step 15, traverse all nodes v: calculate the loss values ​​of all nodes and form an importance sequence;

[0067] Step 16. Obtain importance sequence: The difference between the reconstructed graph and the original graph is used as the importance index of the node, L = (l1, l2, l3, ..., l N ). The greater the difference between the reconstructed graph and the original graph, the greater the impact of the node on the structure and function of the entire network, and the higher the importance of the node. After traversing all the nodes in the network, the importance sequence of the nodes can be obtained.

Claims

1. A quantitative evaluation method of network node importance based on deep learning, Features: The method comprises: constructing a network node graph structure model of the system under test, constructing a comprehensive evaluation index of the importance of network nodes; forming a network node importance evaluation model using a graph convolutional neural network framework, forming a network node reconstruction graph by deleting network nodes, calculating the difference between the network node reconstruction graph and the original graph, and establishing a network node importance sequence according to the change of the difference.

2. According to the method for quantitatively evaluating the importance of network nodes based on deep learning in claim 1, Features: The method of constructing the network node graph structure model of the system under test and constructing the comprehensive evaluation index of the importance of the network nodes includes: Step 1: Represent the network graph in the form of G(V,E), where G(·) is an undirected and unweighted graph, V = {v 1 ,v 2 ,v 3 ,…,v n } is the set of network nodes, v i (i=1,2,…,n) is a single node in the network, E={e 1 ,e 2 ,e 3 ,…,e m } is the edge between nodes, e i (i=1,2,…,m) is the wireless link for communication between nodes, n and m are the number of nodes and the number of links respectively; Step 2: Calculate node degree centrality: If there are n nodes in the network, then node v i The degree centrality of is: Among them, k i For node v i The node degree, C d (v i ) is the node v i The degree centrality of Step 3: Calculate node v i Betweenness centrality: non-adjacent nodes v in the network i and v j The shortest path between them will pass through some nodes, a node v i If a node is passed by other shortest paths, it means that the node is very important in the network, and its importance or influence is characterized by the node's betweenness; B i is the node betweenness. Step 4. Calculate relative importance: v i To node v j Distance:d ij =|v i v j |. Node v i The relative importance of is, μ m (v i ) is the node starting from the node and returning to the node v through m edges. i The number of circuits; Step 5: Node attribute matrix construction: Assume that there are n network nodes in the network, and the corresponding important node solution set is {A 1 ,A 2 ,…,A n }, there are k node importance evaluation indicators, that is, there are k scheme attributes, where k = 4, and the initial decision matrix P is constructed, that is: Among them, A i (f r )(i=1,2,…,n;r=1,2,3,4) is the node solution; f r (r=1,2,3,4) is the scheme attribute; Step 6: Use the coefficient of variation method to determine the normalized weight vector A of each attribute index i w T =[w 1 ,w 2 ,w 3 ,w 4 ], the weight of each indicator w i (i=1,2,3,4) is where v i is the coefficient of variation of the i-th indicator, and σ i is the standard deviation coefficient of the i-th indicator; is the average of the i-th indicator; k is the number of indicators; Step 7: Calculate the comprehensive evaluation index based on the initialization decision matrix and the weights of the selected indicators such as degree centrality, betweenness centrality, and relative importance of nodes, namely: Among them, M is the decision coefficient matrix, p ij (i=1,2,…,n;j=1,2,3,4) are the initial decision matrix elements.

3. According to the method for quantitatively evaluating the importance of network nodes based on deep learning in claim 2, Features: The method of using the graph convolutional neural network framework to form a network node importance evaluation model includes: Step 8: Construct a symmetric normalized Laplace matrix Step 9: Set the node feature to be X = {x 1 ,x 2 ,…,x n },σ is a nonlinear activation function, w is a weight parameter matrix, Step 10: Introduce attention into the graph structure data to form a graph attention network to learn the edge weights between two adjacent nodes and implement convolution operations between nodes; In the graph structure data, a graph attention network is formed to learn the edge weights between two adjacent nodes and implement convolution operations between nodes. The attention weight is: Step 11: Update the expression at the graph level: Specific to the expression update on the node: N i Represents node v i The neighbor node set, the weight parameter matrix W is used to complete the feature dimension transformation of each node, the parameter α is used to calculate the attention weight between nodes, || represents vector concatenation, α ij represents the node v calculated under α i ,v j , and α(·) represents the nonlinear activation function.

4. According to the method for quantitatively evaluating the importance of network nodes based on deep learning in claim 1, Features: The method of forming a network node reconstruction graph includes: Step 12: In an undirected and unweighted network, the failure of a node means that all edges of the node are invalid. When a node v is deleted from the entire network, i That is, all the edges e related to this node ij Delete, select and delete a node to get the reconstructed graph 5. According to the method for quantitatively evaluating the importance of network nodes based on deep learning in claim 1, Features: Methods for network node importance sequence include: Step 13. Self-supervised learning training: encoder f θ The goal is to learn each node v from the pre-training dataset graph G i The low-dimensional representation (i.e. node embedding) h i ∈H, decoder p θ The learned graph embedding H is used as the input of the auxiliary task for pre-training. θ It can be any neural network for graph learning. The encoder f θ It can be regarded as the encoder f in the fine-tuning stage θinit Initialization of parameters. Afterwards, under the supervision of a specific downstream task, the fine-tuning dataset G t Fine-tune the downstream decoder and the pre-trained encoder f θinit Make fine adjustments. Step 14: Calculate the loss: Calculate the loss value of each layer node; Step 15, traverse all nodes v: calculate the loss values ​​of all nodes and form an importance sequence; Step 16. Obtain importance sequence: The difference between the reconstructed graph and the original graph is used as the importance index of the node, L = (l 1 ,l 2 ,l 3 ,......,l N ). The greater the difference between the reconstructed graph and the original graph, the greater the impact of the node on the structure and function of the entire network, and the higher the importance of the node. After traversing all the nodes in the network, the importance sequence of the nodes can be obtained.

Citation Information

Cited By

  • Industrial chain integrity evaluation model, system and method

    CN119692808A

  • Bentweenness centrality estimation method and device based on sampling strategy

    CN120223573A