Method for discovering abnormal nodes and truth values of Internet of Vehicles based on graph convolution under zero trust
By using the graph convolution method based on zero trust in the Internet of Vehicles, the trust support of vehicle nodes and the generation of new node characteristics are solved, and the problems of untimely transmission of trust relationships and difficulty in topology updates in the Internet of Vehicles are achieved, and more accurate abnormal node identification and data quality improvement.
Patent Information
- Application Number
- CN202411031655.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-15
AI Technical Summary
The existing truth value discovery method cannot fully reflect the trust relationship between vehicles in the Internet of Vehicles, resulting in the untimely transmission of perceived data and trust relationships between vehicle nodes, which limits the performance of the Internet of Vehicles, and the vehicle topology needs to be dynamically updated to adapt to dynamic data. and the change in the relationship between trust.
Using the graph convolution method based on zero trust, the dynamic relationship topology graph of vehicle nodes is constructed, node feature similarity is evaluated to calculate trust support, abnormal nodes are identified, and new node features are generated through graph convolution and attention mechanisms. Implement dynamic updates of vehicle topology.
This method can more precisely identify abnormal nodes, reduce false positives and missed reports, improve the accuracy and timeliness of trust relationship evaluation, ensure data quality and credibility, dynamically update the topology, and improve the performance and communication efficiency of Internet of Vehicles.
Smart Images

Figure CN118972257A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of Internet big data and Internet of Vehicles, and specifically to a method for discovering abnormal nodes and true values of Internet of Vehicles based on graph convolution under zero trust. Background Art
[0002] As the functions of vehicle-mounted sensor equipment continue to become more powerful, a large amount of data is continuously generated through the perception of the road environment, the identification of surrounding vehicles, and the recording of the vehicle's own driving status. These data are widely used in the Internet of Vehicles, such as traffic condition assessment, traffic situation prediction, and traffic flow control. Compared with relying on roadside facilities to perceive the road environment, using vehicles to collect and share data can better expand the perception range and improve the sensitivity of the surrounding environment.
[0003] However, even data describing the same environmental road conditions or object vehicles may have multiple sources, and because data from different sources are collected and processed by a single vehicle equipped with different sensors and different data processing algorithms, there may be data conflicts, unreality, or deviations from the ground truth due to hardware device errors, lack of records, environmental influences, or even malicious tampering. Therefore, it is necessary to correctly identify and clean multiple different or even conflicting data information collected from different on-board sensor devices, and aggregate data information about the same object event from different vehicle sources to obtain the true road condition facts.
[0004] Truth discovery technology aims to extract real and reliable information from potentially inaccurate data to ensure the normal operation of the Internet of Vehicles system and the accuracy of decision-making. Most of the existing truth discovery methods are based on iteration, optimization and probability graph methods. However, existing methods often convert the truth discovery problem into the relationship between source reliability and perception data credibility through simple functions. When mining the relationship between road vehicles, they cannot fully reflect the trust relationship between vehicles, resulting in the inability to transmit perception data and trust relationships between vehicle nodes in a timely manner, limiting the performance of the Internet of Vehicles. In addition, in the field of transportation, data such as vehicle traffic conditions are mostly dynamic data, which causes the relationship between vehicles to change dynamically, so the topology of vehicles also needs to be dynamically updated. Therefore, how to effectively evaluate the trust relationship between vehicles and dynamically update the vehicle topology is a technical problem that needs to be solved urgently. Summary of the invention
[0005] In view of the deficiencies of the above-mentioned prior art, the technical problem to be solved by the present invention is: how to provide a method for discovering abnormal nodes and true values in the Internet of Vehicles based on graph convolution under zero trust, evaluate the trust support of each vehicle node based on the similarity of node features between vehicle nodes, and comprehensively reflect the trust relationship between vehicles; calculate the anomaly score of vehicle nodes with abnormal node features, and accurately identify abnormal nodes; generate new node features, i.e., true values, for abnormal nodes through graph convolution and attention mechanism, and realize dynamic update of vehicle topology.
[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:
[0007] The method for discovering abnormal nodes and truth values in the Internet of Vehicles based on graph convolution under zero trust includes:
[0008] S1: Construct an adjacency matrix with vehicles in the Internet of Vehicles as nodes; obtain relevant data of each vehicle node;
[0009] S2: Input the adjacency matrix and the relevant data of each vehicle node into the trained graph convolutional network, detect abnormal nodes and update the true value of the abnormal nodes;
[0010] The processing steps for training a graph convolutional network are as follows:
[0011] S201: Construct the network structure of graph convolutional network;
[0012] S202: using the adjacency matrix as training data and the relevant data and true value of each vehicle node as input of the graph convolutional network;
[0013] S203: constructing a dynamic relationship topology graph with vehicles as nodes and relationships between vehicles as edges;
[0014] S204: Evaluate the trust support of each vehicle node based on the similarity of node features between the vehicle nodes, and determine that the vehicle nodes whose trust support is lower than the set trust support threshold have abnormal node features;
[0015] S205: Calculate the abnormality score of the vehicle node with abnormal node characteristics, and when the abnormality score exceeds the set abnormality score threshold, determine that the vehicle node is an abnormal node;
[0016] S206: For abnormal nodes: delete their node features, aggregate the node features of the surrounding vehicle nodes through graph convolution and attention mechanism to generate new node features of the abnormal node, i.e., true value, and then update the dynamic relationship topology graph;
[0017] S207: Calculate the loss function by updating the true value and actual value of all vehicle nodes, and perform back propagation and parameter update on the graph convolution network;
[0018] S208: Repeat steps S202 to S207 until the graph convolutional network converges or reaches a preset number of iterations.
[0019] Preferably, in step S201, the graph convolutional network includes an input layer, a convolutional layer and an output layer; wherein the convolutional layer includes two sub-convolutional layers, and the output of the first sub-convolutional layer serves as the input of the second sub-convolutional layer;
[0020] The formula for the two-layer sub-convolutional layer is expressed as:
[0021]
[0022] Where: H (1) , H (2) They represent the outputs of the first and second sub-convolutional layers respectively; X represents the input of the first sub-convolutional layer, i.e., the input of the graph convolutional network; represents the adjacency matrix with self-connection added; W (0) and W (1) represents the learned weight matrix; σ represents the activation function; represents the degree matrix.
[0023] Preferably, in step S203, the node features of the dynamic relationship topology graph are the perception data of the vehicle nodes; and the edge features of the dynamic relationship topology graph are the physical relationships and trust relationships between vehicles.
[0024] Preferably, in step S204, each vehicle node evaluates whether the node characteristics of other vehicle nodes are credible based on its own node characteristics and the node characteristics of other vehicle nodes;
[0025] The trust support of vehicle node i is the trust evaluation result of all other vehicle nodes on vehicle node i;
[0026] The calculation formula of trust support is expressed as:
[0027]
[0028] Where: represents the trust support of vehicle node i; α v , α d represents the attenuation factor; θ kv ,θ kd represents the difference ratio of the speed and distance between vehicle nodes i and k, where v represents the speed and d represents the distance; R r-1,k represents the trust of vehicle node k at the previous time point; t represents the time point.
[0029] Preferably, in step S205, the abnormality score of the vehicle node is calculated by the following formula:
[0030]
[0031] Where: s i represents the abnormal score of vehicle node i; h i represents the node characteristics of vehicle node i; h j represents the node characteristics of vehicle node j; Ν(i) represents the number of vehicle nodes.
[0032] Preferably, in step S205, the abnormal score threshold corresponding to the vehicle node is dynamically updated;
[0033] The abnormal score threshold of the vehicle node is calculated by the following formula:
[0034]
[0035] Where: represents the abnormal score threshold of vehicle node i; It indicates that the theoretical measurement value of vehicle node i at time t is predicted by its historical node features; It indicates that the theoretical measurement value of vehicle node i at time t is predicted by the node characteristics of the vehicle nodes around it; δ i Represents the error accuracy of vehicle node i.
[0036] Preferably, in step S206, when generating a true value for an abnormal node: the abnormal node and its surrounding vehicle nodes are aggregated through a graph convolution operation; the importance weights between the abnormal node and its surrounding vehicle nodes are calculated through an attention mechanism; the node features of the surrounding vehicle nodes are weighted based on the importance weights to generate updated node features of the abnormal node; the updated node features of the abnormal node are input into a fully connected layer, and the output is a new node feature of the abnormal node, i.e., the true value.
[0037] Preferably, in step S206, the importance weight is calculated by the following formula:
[0038] e ij =LeakyReLU(a T [Wh i ||Wh j ]);
[0039]
[0040] Where: α ij represents the importance weight between abnormal node i and vehicle node j; e ij represents the attention score between abnormal node i and vehicle node j; h i represents the node characteristics of abnormal node i; h jrepresents the node features of vehicle node j; W represents the weight matrix; a is the weight vector of the attention mechanism.
[0041] Preferably, in step S206, the updated node feature of the abnormal node is calculated by the following formula:
[0042]
[0043] Where: h i ′ represents the updated node feature of abnormal node i; h j Represents the node characteristics of vehicle node j; α ij represents the importance weight between abnormal node i and vehicle node j; Ν(i) represents the number of vehicle nodes; θ k represents the parameters of the k-th order Chebyshev polynomial; represents the k-th order Chebyshev polynomial; where the Chebyshev polynomial is the convolution kernel in the graph convolutional network.
[0044] The new node features of abnormal nodes are calculated by the following formula:
[0045]
[0046] Where: represents the new node feature of abnormal node i, i.e., the true value; b represents the bias term.
[0047] Preferably, in step S207, the loss function of the graph convolutional network is calculated by the following formula:
[0048]
[0049] Where: Z i represents the updated node feature of vehicle node i, i.e., the true value; Y i represents the true value of vehicle node i; N represents the number of vehicle nodes.
[0050] Compared with the prior art, the method for discovering abnormal nodes and true values of Internet of Vehicles based on graph convolution under zero trust in the present invention has the following beneficial effects:
[0051] After the dynamic relationship topology graph is constructed by the graph convolution network in the present invention, the trust support of each vehicle node is evaluated based on the similarity of the node features between the vehicle nodes, and the vehicle nodes whose trust support is lower than the set trust support threshold are judged to have abnormal node features. First, by evaluating the feature similarity between the vehicle nodes to quantify the trust support, the trust relationship between the vehicles can be fully reflected, and then those abnormal nodes with significantly different features from other vehicle nodes can be effectively identified. This method is more sophisticated than simply judging based on a single feature threshold, and can reduce false positives and false negatives, thereby improving the accuracy of abnormal node detection. Secondly, the characteristics of vehicle nodes will change with time, environment and other factors. The present invention can dynamically adjust the threshold of trust support by continuously evaluating the feature similarity between nodes to adapt to the ever-changing network environment and improve the timeliness and accuracy of detection. Finally, nodes with low trust support often mean that their data are unreliable. Timely identification of these nodes and exclusion of their data can improve the quality and credibility of data in the entire Internet of Vehicles, and provide a more solid foundation for subsequent data analysis and decision support.
[0052] The present invention further calculates the abnormality score of the vehicle node with abnormal node characteristics based on the trust support to determine whether the vehicle node has abnormal node characteristics. When the abnormality score exceeds the set abnormality score threshold, the vehicle node is judged to be an abnormal node. First, by calculating the abnormality score and setting the threshold, those vehicle nodes that really have problems, namely abnormal nodes, can be more accurately identified. This method is more quantitative than simple trust support judgment and can reduce the error caused by subjective judgment. Secondly, in different application scenarios or network conditions, the threshold of the abnormality score can be adjusted according to actual needs. This flexibility enables the system to better adapt to different environments and needs, thereby improving the adaptability and effectiveness of abnormal node detection.
[0053] On the basis of detecting abnormal nodes, the present invention further generates new node features of the abnormal nodes, i.e., true values, by combining the node feature aggregation of the surrounding vehicle nodes through graph convolution and attention mechanism, and then updates the dynamic relationship topology map. First, by aggregating the feature information of the surrounding nodes, new and more accurate node features (i.e., true values) can be generated for the abnormal nodes, thereby restoring their normal functions in the network, helping to maintain the integrity and connectivity of the Internet of Vehicles, and thus improving the overall performance of the vehicle network system. Secondly, the updated node features reflect the common cognition and judgment of the surrounding nodes, which helps to eliminate the impact of abnormal data on the consistency of network data, thereby improving the accuracy and reliability of subsequent data analysis and decision support. Finally, with the update of abnormal node features and the adjustment of the dynamic relationship topology map, the vehicle topology structure of the entire Internet of Vehicles will be optimized accordingly, that is, the dynamic update of the vehicle topology is realized, which helps to reduce redundant connections and abnormal interactions, thereby improving the communication efficiency and stability of the vehicle network. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:
[0055] Figure 1 This is the logical block diagram of the abnormal node and truth value discovery method of the Internet of Vehicles based on graph convolution under zero trust. DETAILED DESCRIPTION
[0056] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.
[0057] The following is a further detailed description through specific implementation methods:
[0058] Example:
[0059] This embodiment discloses a method for discovering abnormal nodes and true values in an Internet of Vehicles based on graph convolution under zero trust.
[0060] like Figure 1 As shown in the figure, the method for discovering abnormal nodes and true values of Internet of Vehicles based on graph convolution under zero trust includes:
[0061] S1: Construct an adjacency matrix with vehicles in the Internet of Vehicles as nodes; obtain relevant data of each vehicle node;
[0062] In this embodiment, the relevant data of the vehicle node includes perception data, IoT relationships and trust relationships between vehicles.
[0063] S2: Input the adjacency matrix and the relevant data of each vehicle node into the trained graph convolutional network, detect abnormal nodes and update the true value of the abnormal nodes;
[0064] The processing steps for training a graph convolutional network are as follows:
[0065] S201: Construct the network structure of the graph convolutional network; randomly initialize the weights and bias items of the graph neural network;
[0066] S202: using the adjacency matrix as training data and the relevant data and true value of each vehicle node as input of the graph convolutional network;
[0067] S203: constructing a dynamic relationship topology graph with vehicles as nodes and relationships between vehicles as edges;
[0068] S204: Evaluate the trust support of each vehicle node based on the similarity of node features between the vehicle nodes, and determine that the vehicle nodes whose trust support is lower than the set trust support threshold have abnormal node features;
[0069] S205: Calculate the abnormality score of the vehicle node with abnormal node characteristics. When the abnormality score exceeds the set abnormality score threshold, the vehicle node is judged to be an abnormal node (excessive deviation or malicious injection);
[0070] S206: For the abnormal node: delete its node features, aggregate the node features of its surrounding vehicle nodes through graph convolution and attention mechanism to generate a new node feature of the abnormal node, i.e., the true value, and then update the dynamic relationship topology graph (i.e., replace the original node features of the abnormal node in the dynamic relationship topology graph with the new node features, i.e., the true value);
[0071] S207: Calculate the loss function by updating the true value and actual value of all vehicle nodes, and perform back propagation and parameter update on the graph convolution network;
[0072] S208: Repeat steps S202 to S207 until the graph convolutional network converges or reaches a preset number of iterations.
[0073] After the dynamic relationship topology graph is constructed by the graph convolution network in the present invention, the trust support of each vehicle node is evaluated based on the similarity of the node features between the vehicle nodes, and the vehicle nodes whose trust support is lower than the set trust support threshold are judged to have abnormal node features. First, by evaluating the feature similarity between the vehicle nodes to quantify the trust support, the trust relationship between the vehicles can be fully reflected, and then those abnormal nodes with significantly different features from other vehicle nodes can be effectively identified. This method is more sophisticated than simply judging based on a single feature threshold, and can reduce false positives and false negatives, thereby improving the accuracy of abnormal node detection. Secondly, the characteristics of vehicle nodes will change with time, environment and other factors. The present invention can dynamically adjust the threshold of trust support by continuously evaluating the feature similarity between nodes to adapt to the ever-changing network environment and improve the timeliness and accuracy of detection. Finally, nodes with low trust support often mean that their data are unreliable. Timely identification of these nodes and exclusion of their data can improve the quality and credibility of data in the entire Internet of Vehicles, and provide a more solid foundation for subsequent data analysis and decision support.
[0074] The present invention further calculates the abnormality score of the vehicle node with abnormal node characteristics based on the trust support to determine whether the vehicle node has abnormal node characteristics. When the abnormality score exceeds the set abnormality score threshold, the vehicle node is judged to be an abnormal node. First, by calculating the abnormality score and setting the threshold, those vehicle nodes that really have problems, namely abnormal nodes, can be more accurately identified. This method is more quantitative than simple trust support judgment and can reduce the error caused by subjective judgment. Secondly, in different application scenarios or network conditions, the threshold of the abnormality score can be adjusted according to actual needs. This flexibility enables the system to better adapt to different environments and needs, thereby improving the adaptability and effectiveness of abnormal node detection.
[0075] On the basis of detecting abnormal nodes, the present invention further generates new node features of the abnormal nodes, i.e., true values, by combining the node feature aggregation of the surrounding vehicle nodes through graph convolution and attention mechanism, and then updates the dynamic relationship topology map. First, by aggregating the feature information of the surrounding nodes, new and more accurate node features (i.e., true values) can be generated for the abnormal nodes, thereby restoring their normal functions in the network, helping to maintain the integrity and connectivity of the Internet of Vehicles, and thus improving the overall performance of the vehicle network system. Secondly, the updated node features reflect the common cognition and judgment of the surrounding nodes, which helps to eliminate the impact of abnormal data on the consistency of network data, thereby improving the accuracy and reliability of subsequent data analysis and decision support. Finally, with the update of abnormal node features and the adjustment of the dynamic relationship topology map, the vehicle topology structure of the entire Internet of Vehicles will be optimized accordingly, that is, the dynamic update of the vehicle topology is realized, which helps to reduce redundant connections and abnormal interactions, thereby improving the communication efficiency and stability of the vehicle network.
[0076] In order to better introduce the technical solution of the present invention, this embodiment is described through the following parts.
[0077] 1. Network structure of graph convolutional network
[0078] In this embodiment, the graph convolutional network includes an input layer, a convolutional layer and an output layer;
[0079] The convolution layer includes two sub-convolution layers, and the output of the first sub-convolution layer serves as the input of the second sub-convolution layer;
[0080] The formula for the two-layer sub-convolutional layer is expressed as:
[0081]
[0082] Where: H (1) , H (2) They represent the outputs of the first and second sub-convolutional layers respectively; X represents the input of the first sub-convolutional layer, i.e., the input of the graph convolutional network; represents the adjacency matrix with self-connection added; W (0) and W (1) represents the learned weight matrix; σ represents the activation function; represents the degree matrix.
[0083] 2. Dynamic relationship topology diagram
[0084] In this embodiment, the node features of the dynamic relationship topology graph are the perception data of the vehicle nodes; the edge features of the dynamic relationship topology graph are the physical relationship and trust relationship between vehicles. The physical relationship refers to the position coordinates and distance relationship between vehicle nodes; the trust relationship refers to the trust of vehicle i in vehicle j, and the trust is calculated based on the node similarity of real-time perception data.
[0085] In this embodiment, let V be the vehicle node set, E be the edge set between vehicles, X be the node feature matrix, A be the adjacency matrix, and F be e is the edge feature matrix. Define the node feature X, each vehicle node v i The characteristics of time series T are the credibility and real-time perception data of time series T-1, such as speed, position, acceleration, etc. Define edge feature F e , where each edge e ij The characteristics of F are the physical relationship and trust relationship between vehicles. eij ={d ij,distance ,v ij,reliability}.
[0086] 3. Trust Support of Vehicle Nodes
[0087] In this embodiment, each vehicle node evaluates whether the node characteristics of other vehicle nodes are credible based on its own node characteristics and the node characteristics of other vehicle nodes;
[0088] The trust support of vehicle node i is the trust evaluation result of all other vehicle nodes on vehicle node i;
[0089] The calculation formula of trust support is expressed as:
[0090]
[0091] Where: represents the trust support of vehicle node i; α v , α d represents the attenuation factor; θ kv ,θ kd represents the difference ratio of the speed and distance between vehicle nodes i and k, where v represents the speed and d represents the distance; R t-1,krepresents the trust of vehicle node k at the previous time point; t represents the time point.
[0092] 4. Anomaly Score Calculation
[0093] In this embodiment, the abnormal score of the vehicle node is calculated by the following formula:
[0094]
[0095] Where: s i represents the abnormal score of vehicle node i; h i represents the node characteristics of vehicle node i; h j represents the node characteristics of vehicle node j; Ν(i) represents the number of vehicle nodes.
[0096] Specifically, considering the default unlabeled background real data in a zero-trust environment, the anomaly score threshold corresponding to the vehicle node is dynamically updated.
[0097] The abnormal score threshold of the vehicle node is calculated by the following formula:
[0098]
[0099] Where: represents the abnormal score threshold of vehicle node i; It indicates that the theoretical measurement value of vehicle node i at time t is predicted through its historical node features (perception data); It indicates that the theoretical measurement value of vehicle node i at time t is predicted through the node characteristics (perception data) of the vehicle nodes around it; δ i Represents the error accuracy of vehicle node i.
[0100] 5. Truth value update of abnormal nodes
[0101] In this embodiment, when generating a true value for an abnormal node: the abnormal node and its surrounding vehicle nodes are aggregated through a graph convolution operation; the importance weights between the abnormal node and its surrounding vehicle nodes are calculated through an attention mechanism; the node features of the surrounding vehicle nodes are weighted based on the importance weights to generate updated node features of the abnormal node; the updated node features of the abnormal node are input into the fully connected layer, and the output is the new node features of the abnormal node, i.e., the true value.
[0102] Specifically, the importance weight is calculated by the following formula:
[0103] e ij =LeakyReLU(a T [Wh i ||Wh j ]);
[0104]
[0105] Where: α ij represents the importance weight between abnormal node i and vehicle node j; e ij represents the attention score between abnormal node i and vehicle node j; h i represents the node characteristics of abnormal node i; h j represents the node features of vehicle node j; W represents the weight matrix; a is the weight vector of the attention mechanism.
[0106] Specifically, the updated node features of abnormal nodes are calculated using the following formula:
[0107]
[0108] Where: h i ′ represents the updated node feature of abnormal node i; h j Represents the node characteristics of vehicle node j; α ij represents the importance weight between abnormal node i and vehicle node j; Ν(i) represents the number of vehicle nodes; θ k represents the parameters of the k-th order Chebyshev polynomial; represents the k-th order Chebyshev polynomial; where the Chebyshev polynomial is the convolution kernel in the graph convolutional network.
[0109] Specifically, the new node features of the abnormal nodes are calculated using the following formula:
[0110]
[0111] Where: represents the new node feature of abnormal node i, i.e., the true value; b represents the bias term.
[0112] 6. Loss Function
[0113] In this embodiment, the loss function of the graph convolutional network is calculated by the following formula:
[0114]
[0115] Where: Z i represents the updated node feature of vehicle node i, i.e., the true value; Y i represents the true value of vehicle node i; N represents the number of vehicle nodes.
[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.
Claims
1. A method for discovering abnormal nodes and true values in Internet of Vehicles based on graph convolution under zero trust, characterized in that: include: S1: Construct an adjacency matrix with vehicles in the Internet of Vehicles as nodes; Obtain relevant data of each vehicle node; S2: Input the adjacency matrix and the relevant data of each vehicle node into the trained graph convolutional network, detect abnormal nodes and update the true value of the abnormal nodes; The processing steps for training a graph convolutional network are as follows: S201: Construct the network structure of graph convolutional network; S202: using the adjacency matrix as training data and the relevant data and true value of each vehicle node as input of the graph convolutional network; S203: constructing a dynamic relationship topology graph with vehicles as nodes and relationships between vehicles as edges; S204: Evaluate the trust support of each vehicle node based on the similarity of node features between the vehicle nodes, and determine that the vehicle nodes whose trust support is lower than the set trust support threshold have abnormal node features; S205: Calculate the abnormality score of the vehicle node with abnormal node characteristics, and when the abnormality score exceeds the set abnormality score threshold, determine that the vehicle node is an abnormal node; S206: For abnormal nodes: delete their node features, aggregate the node features of the surrounding vehicle nodes through graph convolution and attention mechanism to generate new node features of the abnormal node, i.e., true value, and then update the dynamic relationship topology graph; S207: Calculate the loss function by updating the true value and actual value of all vehicle nodes, and perform back propagation and parameter update on the graph convolution network; S208: Repeat steps S202 to S207 until the graph convolutional network converges or reaches a preset number of iterations.
2. The method for discovering abnormal nodes and truth values of Internet of Vehicles based on graph convolution under zero trust as claimed in claim 1, characterized in that: In step S201, the graph convolutional network includes an input layer, a convolutional layer and an output layer; wherein the convolutional layer includes two sub-convolutional layers, and the output of the first sub-convolutional layer serves as the input of the second sub-convolutional layer; The formula for the two-layer sub-convolutional layer is expressed as: Where: H (1) , H (2) They represent the outputs of the first and second sub-convolutional layers respectively; X represents the input of the first sub-convolutional layer, i.e., the input of the graph convolutional network; represents the adjacency matrix with self-connection added; W (0) and W (1) represents the learned weight matrix; σ represents the activation function; represents the degree matrix.
3. The method for discovering abnormal nodes and true values of Internet of Vehicles based on graph convolution under zero trust as claimed in claim 1, characterized in that: In step S203, the node features of the dynamic relationship topology graph are the perception data of the vehicle nodes; the edge features of the dynamic relationship topology graph are the physical relationship and trust relationship between vehicles.
4. The method for discovering abnormal nodes and truth values of Internet of Vehicles based on graph convolution under zero trust as claimed in claim 1, characterized in that: In step S204, each vehicle node evaluates whether the node characteristics of other vehicle nodes are credible based on its own node characteristics and the node characteristics of other vehicle nodes; The trust support of vehicle node i is the trust evaluation result of all other vehicle nodes on vehicle node i; The calculation formula of trust support is expressed as: Where: represents the trust support of vehicle node i; α v , α d represents the attenuation factor; θ kv ,θ kd represents the difference ratio of the speed and distance between vehicle nodes i and k, where v represents the speed and d represents the distance; R t-1,k represents the trust of vehicle node k at the previous time point; t represents the time point.
5. The method for discovering abnormal nodes and truth values of Internet of Vehicles based on graph convolution under zero trust as claimed in claim 1, characterized in that: In step S205, the abnormal score of the vehicle node is calculated by the following formula: Where: s i represents the abnormal score of vehicle node i; h i Represents the node characteristics of vehicle node i; h j Represents the node characteristics of vehicle node j; Ν(i) represents the number of vehicle nodes.
6. The method for discovering abnormal nodes and truth values of Internet of Vehicles based on graph convolution under zero trust as claimed in claim 1, characterized in that: In step S205, the abnormal score threshold corresponding to the vehicle node is dynamically updated; The abnormal score threshold of the vehicle node is calculated by the following formula: Where: represents the abnormal score threshold of vehicle node i; It indicates that the theoretical measurement value of vehicle node i at time t is predicted by its historical node features; It indicates that the theoretical measurement value of vehicle node i at time t is predicted by the node characteristics of the vehicle nodes around it; δ i Represents the error accuracy of vehicle node i.
7. The method for discovering abnormal nodes and truth values in Internet of Vehicles based on graph convolution under zero trust as claimed in claim 1, characterized in that: In step S206, when generating a true value for an abnormal node: the abnormal node and its surrounding vehicle nodes are aggregated through a graph convolution operation; The importance weights between abnormal nodes and their surrounding vehicle nodes are calculated through the attention mechanism; Perform weighted calculation on the node features of surrounding vehicle nodes based on importance weights to generate updated node features of abnormal nodes; The updated node features of the abnormal nodes are input into the fully connected layer, and the output is the new node features of the abnormal nodes, i.e., the true value.
8. The method for discovering abnormal nodes and truth values of Internet of Vehicles based on graph convolution under zero trust as claimed in claim 7, characterized in that: In step S206, the importance weight is calculated by the following formula: e ij =LeakyReLU(a T [Wh i ||Wh j ]); Where: α ij represents the importance weight between abnormal node i and vehicle node j; e ij represents the attention score between abnormal node i and vehicle node j; h i Represents the node characteristics of abnormal node i; h j Represents the node characteristics of vehicle node j; W represents the weight matrix; a is the weight vector of the attention mechanism.
9. The method for discovering abnormal nodes and truth values of Internet of Vehicles based on graph convolution under zero trust as claimed in claim 7, characterized in that: In step S206, the updated node feature of the abnormal node is calculated by the following formula: Where: h i ′ represents the updated node features of abnormal node i; h j Represents the node characteristics of vehicle node j; α ij represents the importance weight between abnormal node i and vehicle node j; Ν(i) represents the number of vehicle nodes; θ k represents the parameters of the k-th order Chebyshev polynomial; represents the Chebyshev polynomial of order k; The new node features of abnormal nodes are calculated by the following formula: Where: represents the new node feature of abnormal node i, i.e., the true value; b represents the bias term.
10. The method for discovering abnormal nodes and truth values of Internet of Vehicles based on graph convolution under zero trust as claimed in claim 1, characterized in that: In step S207, the loss function of the graph convolutional network is calculated by the following formula: Where: Z i represents the updated node feature of vehicle node i, i.e., the true value; Y i represents the true value of vehicle node i; N represents the number of vehicle nodes.
