Enhanced verification of safety and security for advanced driver assistance systems

Abstract

The safe operation of advanced driver assistance systems (ADAS) plays a critical role in autonomous vehicles. Rigorous methods such as formal verification are typically used to provide safety guarantees for ADAS. However, they can become overly conservative in the presence of cyberattacks, which introduce additional uncertainties and system vulnerabilities. To address this challenge, this paper enhances formal verification by incorporating verification and falsification into each other for improved safety and security. The verification process of our method describes ADAS-equipped vehicles using hybrid automata, while attacks are over-approximated as bounded inputs. When verification is inconclusive due to over-approximations, a falsification process leverages deep reinforcement learning (DRL) to explore potential attack paths, with rewards shaped by the verification results to uncover vulnerabilities. Finally, comprehensive high-fidelity simulations are conducted to demonstrate the proposed method through Flow* and CARLA/Scenic platforms.