CANDIDS: CAN/CAN-FD deep learning-based intrusion detection systems

The automotive industry has experienced a surge in technological advancements, resulting in a significant increase in the connectivity and functionality of modern vehicles. In-vehicle networks, which serve as the backbone of communication between various Electronic Control Units (ECUs), rely on protocols such as the Controller Area Network (CAN) and its successor, CAN with Flexible Data Rate (CAN-FD). However, the lack of robust security measures in these industry-standard protocols has left systems increasingly vulnerable to cyber threats. By implementing Intrusion Detection Systems (IDS), automotive security could be significantly enhanced with minimal disruption to the vehicle's infrastructure. The integration of advanced machine learning algorithms offers significant promise in enhancing the effectiveness of IDS. In this paper, we propose a deep learning-based IDS, named CANDIDS, which is capable of operating effectively not only for the conventional CAN protocol, but also for the most recent CAN-FD one commonly found in modern vehicles. The proposed approach enables the system to identify abnormal network traffic and categorize various types of attack. The experimental results show that CANDIDS achieves 99.47% and 99.87% detection and multiclass classification accuracy for CAN and CAN-FD respectively, illustrating its effectiveness in anomaly detection. Additionally, CANDIDS achieves low latency intrusion detection, completing the process in 2 and 3 milliseconds for CAN and CAN-FD meeting the strict timing demands crucial for the safe and efficient operation of networked automotive and self-driving vehicle applications.