Machine Learning-Based Intrusion Detection for Automotive CAN Networks on Embedded Platforms

Modern cars are composed of electronic components that communicate using the Controller Area Network (CAN) protocol in most cases. CAN is vulnerable to several attacks, which may be dangerous and cause damage. In this paper, we evaluate five machine learning-based models for intrusion detection for two types of CAN attacks (Denial of Service and impersonation). We train and test the models on a public dataset and on a dataset from a real vehicle, demonstrating the generalization and applicability of the models. The obtained results achieve up to 100% of Fl-Score, better than related works. In addition, we embedded the models in a RISC-V platform and evaluate their execution times, proving the feasibility of applying proposed intrusion detection algorithms in a real-time scenario.