On the Robustness of Intrusion Detection Systems for Vehicles Against Adversarial Attacks

Abstract

Because connected cars typically have several communication capabilities (through 5G, WiFi, and Bluetooth), and third-party applications can be installed on the cars, it would be essential to deploy intrusion detection systems (IDS) to prevent attacks from external attackers or malicious applications. Therefore, many IDS proposals have been presented to protect the controller area network (CAN) in a vehicle. Some studies showed that deep neural network models could be effectively used to detect various attacks on the CAN bus. However, it is still questionable whether such an IDS is sufficiently robust against adversarial attacks that are crafted aiming to target the IDS. In this paper, we present a genetic algorithm to generate adversarial CAN attack messages for Denial-of-Service (DoS), fuzzy, and spoofing attacks to target the state-of-the-art deep learning-based IDS for CAN. The experimental results demonstrate that the state-of-the-art IDS is not effective in detecting the generated adversarial CAN attack messages. The detection rates of the IDS were significantly decreased from 99.27%, 96.40%, and 99.63% to 2.24%, 11.59%, and 0.01% for DoS, fuzzy, and spoofing attacks, respectively.