Demo: CARLA-based Adversarial Attack Assessment on Autonomous Vehicles

Autonomous vehicles rely on a combination of sensors and sophisticated artificial intelligence (AI) systems to perceive their surroundings. The increasing use of AI in autonomous driving technology has brought to our attention the concerns of the implications of AI failure. In this work, we chose an object detector (OD) as an entry point to study the robustness against adversarial attacks like malicious traffic signs. We design and implement CARLA-A3 (CARLA-based Adversarial Attack Assessment), which is a toolkit aimed to streamline the simulation of adversarial conditions and evaluation of OD with several robustness metrics. The toolkit can serve to rapidly and quantitatively evaluate the effects of a malicious sign presented to the OD.