A Deep Learning Approach For Intelligent And Connected Vehicle Network Intrusion Detection

The development of a new generation of intelligent and connected vehicle has brought a more convenient and comfortable driving experience,at the same time the vehicle has evolved from a closed and independent system into an important node in the Internet of Vehicles.With the significant increase in external communication interfaces and the high complexity of invehicle network systems,the risk of cyber attacks on vehicles is at a very high level.The controller area network(CAN)bus,one of the most widely used in-vehicle network buses,has become a key target for hackers due to its own security weaknesses.The frequent occurrence of cyber attacks on vehicles in recent years has greatly threatened the lives and property of drivers and passengers.Therefore,how to secure the network of intelligent and connected vehicle has become one of the main research issues for information security researchers.This paper takes securing the vehicle network as the starting point,analyzes in detail the structure of Internet of Vehicles system,the structure of the in-vehicle network system of the intelligent and connected vehicle and the security vulnerabilities of the CAN bus,and proposes a deep learning method for the intelligent and connected vehicle network intrusion detection,the main of our work are as follows:(1)An in-vehicle network intrusion detection method combining convolutional neural networks and model fusion is proposed.Based on the network structures of the representative deep learning models Alex Net and VGG16,two convolutional neural network models My＿Alex Net and My＿VGG were designed and constructed,and the weighted averaging method was used to fuse the two models.Compared to other in-vehicle network intrusion detection methods using deep learning models,the combined convolutional neural network and model fusion approach proposed in this paper has better detection performance.The above method detects 64 consecutive CAN message frames in 0.25 ms and achieves a test accuracy of up to 99.91%.A global average pooling layer is used instead of the fully connected layers.The model improvement resulted in a detection time of 0.21 ms and a test accuracy of up to 99.93%.(2)To address the situation that real vehicle attack datasets are not easily collected in large quantities and deep learning models are highly dependent on the amount of training data,this paper proposes an in-vehicle network intrusion detection method that combines deep transfer learning with support vector machines.In this method,a deep transfer learning model extracts CAN message frame features and a support vector machine implements classification detection.For comparison experiments,different network layers of the models My＿Alex Net and My＿VGG were selected as feature extraction layers.With limited data sets,the combined deep migration learning and support vector machine model approach proposed in this paper still has a good detection performance with a tested accuracy of up to 99.1%.The intrusion detection method combining convolutional neural networks and model fusion and the intrusion detection method combining deep transfer learning and support vector machines proposed in this paper provide new research ideas on how to improve the detection performance of deep learning models for intrusion detection and how to improve the impact of insufficient data.