Intrusion Detection for in-vehicle CAN Networks Based on Auxiliary Classifier GANs

With the rapid evolution of artificial intelligence, in-vehicle network and automotive electronic systems are shifting toward intelligence. The existing in-vehicle network protocol CAN could not provide message encryption and authentication mechanism at the beginning of its design so that it is difficult to face increasingly frequent attacks. Existing research efforts on intrusion detection work mostly do not have the ability of novelty detection, and can only detect the known attack. Therefore, this paper proposes an end-to-end intrusion detection method ACGIDS based on auxiliary classifier GANs. It uses the generator to generate negative samples far away from the positive sample distribution in the training process, covering some unknown attack patterns. At the same time, the adversarial training enables the discriminator to learn a clearer decision boundary and avoids the degradation of the discriminator’s discrimination ability through additional double classifiers. We also design a reliable detection process and compare it with other previous methods from the three dimensions, including detection accuracy, storage overhead, and detection speed. The experimental results show that the proposed method is suitable for real-time detection scenarios, maintains competitive comprehensive performance, and has the ability of novelty detection.