Practical IDS on In-vehicle Network Against Diversified Attack Models

A vehicle bus is a specialized internal communication network that interconnects components inside a vehicle. The Controller Area Network (CAN bus), a robust vehicle bus standard, allows micro-controllers and devices to communicate with each other. The community has seen many security breach examples that exploit CAN functionalities and other in-vehicle flaws. Intrusion detection systems (IDSs) on in-vehicle network are advantageous in monitoring CAN traffic and suspicious activities. Whereas, existing IDSs on in-vehicle network only support one or two attack models, and identifying abnormal in-vehicle CAN traffic against diversified attack models with better performance is more expected as can be then implemented practically. In this paper, we propose an intrusion detection system that can detect many different attacks. The method analyzes the CAN traffic generated by the invehicle network in real time and identifies the abnormal state of the vehicle practically. Our proposal fuses the autoencoder trick to the SVM model. More precisely, we introduce to the system an autoencoder that learns to compress CAN traffic data into extracted features (which can be uncompressed to closely match the original data). Then, the support vector machine is trained on the features to detect abnormal traffic. We show detailed model parameter configuration by adopting several concrete attacks. Experimental results demonstrate better detection performance (than existing proposals).