Examining the Use of Neural Networks for Intrusion Detection in Controller Area Networks

In the light of the recently reported attacks, in-vehicle security has become a major concern. Intrusion detection systems, common in computer networks, have been recently proposed for the in-vehicle buses as well. In this work we examine the performance of neural networks in detecting intrusions on the CAN bus. For the experiments we use a CAN trace that is extracted from a CANoe simulation for the commercial vehicle bus J1939 as well as a publicly available CAN dataset. Our results show good performance in detecting both replay and injection attacks, the former being harder to detect to their obvious similarity with the regular CAN frames. Nonetheless we discuss possibilities for integrating such detection mechanisms on automotive-grade embedded devices. The experimental results show that embedding the neural-network based intrusion detection mechanism on automotive-grade controllers is quite challenging due to large memory requirements and computational time. This suggests that dedicated hardware may be required for deploying such solutions in real-world vehicles.