Securing the Lane: Defences Against Patch Attacks on Autonomous Vehicle's Lane Detection

Autonomous vehicles represent a profound leap in transportation, relying on Advanced Driver Assistance Systems (ADAS), which use lane detection to enhance decisionmaking and safety. However, with complexity comes vulnerability. As artificial intelligence continues its infiltration into safety-critical environments, they become vulnerable to patch attacks. The potential disruption of algorithms introduces risks that may not be ignored. In this study, we ask: How resilient are these systems, and can we fortify them? We dive into various defence mechanisms such as Gaussian noise, Gaussian blur and adversarial training using MetaDrive simulator, which offers a realistic simulation environment. Through experiments that use RESA, SCNN, ResNet50 and ERFNet, our findings reveal that Gaussian noise and adversarial training not only enhance robustness but do so without compromising accuracy, significantly reducing attack success rates. On the other hand, Gaussian blur falls short of expectations. These findings imply that improving safety-critical lane detection systems and integrating their defences into real-world systems in a necessity for more complex future scenarios.