Research On Intrusion Detection Technology In Internet Of Vehicles Based On Graph Neural Network

The Internet of Vehicles(Io V)is the key component of intelligent transportation systems.Its open wireless communication methods and dynamic communication network topology make Io Vs more vulnerable to malicious attacks such as disinformation,creating potential risks for cyber security.Intrusion Detection Systems(IDS)based on deep learning has received a lot of attention from scholars in recent years as an important cyber security technology to provide protection for Io V.However,existing deep learning IDSs are limited to Euclidean space and fail to fully learn the information in the traffic data resulting in poor detection performance.We introduce techniques such as graph theory,graph neural networks and contrastive self-supervised learning to investigate Io V intrusion detection systems.To address the problem that most existing IDSs for Io V are based on deep learning techniques such as convolutional neural networks or recurrent neural networks,which model Io V network traffic as European-style data structures such as images or sequences,ignoring the fact that they are essentially graph data structures leading to low detection rates and high false alarm rates,we propose Graph neural network Deep Learning-based Intrusion Detection Methods(GDL-IDM).Io V network traffic is modelled as the graph data structure by considering the correlation between traffic flows.Firstly,the Io V communication topology is generated into a traffic correlation graph,and then the graph neural networks are used to fully learn the traffic structure features and traffic attribute features to transform the Io V intrusion detection task into the node classification problem on the attribute graph.The GDL-IDM is deployed on mobile edge computing servers in order to quickly obtain detection results in response to highly dynamic network.extensive experiments on the UNSW-NB15 datasets and Ve Re Mi extended datasets were used to validate the effectiveness of GDL-IDM on intrusion detection tasks,with overall performance outperforming traditional deep learning intrusion detection models.To address the problem that it is difficult to obtain labeled Io V network traffic data in practical application scenarios,and existing unsupervised learning methods,mainly based on reconstructed error-based autoencoders,aim to reconstruct original data to learn potential representations without directly targeting intrusion detection tasks leading to unsatisfactory detection performance.we propose Contrastive Self-supervised Learning-based Intrusion Detection Methods(CSL-IDM),which transform the Io V intrusion detection task into the anomaly detection problem on an attribute graph.Considering that anomalies are reflected in the consistency relationship between node and its neighbouring substructures,designs contrastive instance pairs of target nodes and local subgraphs,uses graph neural networks as encoders to compress highdimensional features of local subgraphs to extract information,and discriminators to calculate the matching score of each instance pair to assess the anomalies of the nodes.Ablation experiments and parameter sensitivity experiments are designed on the UNSW-NB15 datasets and Ve Re Mi extended datasets to find the optimal parameters and verify the effectiveness of CSL-IDM on intrusion detection tasks through comparative experiments.