CANPerFL: Improve In-Vehicle Intrusion Detection Performance by Sharing Knowledge

The Controller Area Network (CAN) is a widely used communication protocol in automobiles, but it is vulnerable to various types of attacks. To address this issue, researchers have been exploring the use of intrusion detection systems (IDS) for the CAN bus. Deep learning and machine learning have been proven to be powerful tools for detecting intrusions accurately and quickly. However, deep learning models require large amounts of data to achieve optimal performance, which can be challenging in the case of a CAN bus IDS. To overcome this challenge, we propose a novel machine learning-based IDS called CANPerFL that uses a personalized federated learning scheme to aggregate datasets from different car models. By building a universal model trained on a small amount of data from each manufacturer, we can provide global knowledge that can be transferred to improve the performance of each participant. To demonstrate the efficiency of the proposed model, we collected a real CAN dataset consisting of three different car models: KIA, BMW, and Tesla. The experimental results show that the proposed model increases F1 scores by 4% overall, compared to baselines. Moreover, the proposed system provides significant advantages when the local dataset of each participant is relatively small. According to our experiments, the proposed models can achieve F1 scores of more than 90% with at least 30k training samples on each client. Finally, we show empirically that each participant takes benefits from joining the CANPerFL system.