Adversarial Attacks in Intrusion Detection Systems: Triggering False Alarms in Connected and Autonomous Vehicles

As connected and autonomous vehicles (CAVs) proliferate, securing their internal vehicle networks (IVNs) against cyber threats is paramount. Current research focuses on developing intrusion detection systems (IDSs) using machine learning (ML) models to handle diverse threats. However, ML-based IDSs introduce significant risks from adversarial attacks. This paper investigates the vulnerability of ML-based IDSs in IVNs to such attacks. It shifts focus from manipulating malicious frames to appear benign to exploring IDS susceptibility to benign frames appearing malicious, potentially triggering false alarms. In critical safety applications like CAVs, these alarms can compromise safety and operational integrity. We studied IVN traffic and designed adversarial samples simulating potential threats. Our experiments, using five ML algorithms and four state-of-the-art adversarial methods, demonstrate an attack success rate of up to 89%. This underscores the urgent necessity to address this vulnerability, as neglecting it renders IDSs ineffective and increases the risk of vehicle manipulation.