D2MoN: Detecting and Mitigating Real-Time Safety Violations in Autonomous Driving Systems

Abstract

This paper proposes D2MON, a data-driven real-time safety monitor, to detect and mitigate safety violations of an autonomous vehicle (AV). The key insight is that traffic situations that lead to AV safety violations fall into patterns and can be identified by learning from existing safety violations. Our approach is to use machine learning techniques to model the traffic behaviors that result in safety violations and detect their symptoms in advance before the actual crashes happen. If D2MoN detects surroundings as dangerous, it will take safety actions to mitigate the safety violations so that the AV remains safe in the evolving traffic environment. Our steps are twofold: (1) We use software fuzzing and data augmentation techniques to generate efficient safety violation data for training our ML model. (2) We deploy the model as a plug-and-play module to the AV software, detecting and mitigating safety violations of the AV in runtime. Our evaluation demonstrates our proposed technique is effective in reducing over 99% of safety violations in an industry-level autonomous driving system, Baidu Apollo.