Security Testing of Visual Perception Module in Autonomous Driving System

In recent years, developments of visual perception techniques based on deep learning have significantly promoted the prosperity of autonomous driving in Internet of vehicles scenarios. However, frequent security issues of autonomous driving systems have raised concerns about the future of autonomous driving. Since the behaviors of deep learning systems lack interpretability, testing the robustness of autonomous driving systems based on deep learning is challenging. The existing efforts on security testing for autonomous driving have limitations in scene description, security defect detection, and defect interpretation. Aiming at testing the security of the visual perception module of autonomous driving, we design and implement a scene-driven security testing system. A flexible scene description method that balances authenticity and richness is proposed. We utilize the real-time rendering engine to generate scenes for autonomous driving security testing. We design an efficient scene search algorithm for nonlinear systems that dynamically schedules search plans based on the testing feedback. We also design a failure analyzer to profile the cause of security issues automatically. We reproduce the latest dynamic automatic driving testing system, which is based on the real-time rendering engine, and test the CILRS system and CIL system with our system and the state-of-the-art system. The experimental results show that our system’s failure discovery rate is 1.4 times that of the state-of-the-art scene-driven dynamic testing system in the same amount of time. Further experiments show that our system can find 1 939 and 1 671 scenes through 3 000 dynamically-searched scenes, respectively, which cause security issues in the CIL and CILRS system’s visual perception module. The searched scenes are in three environments: the fields, the country, and the city, and the average search time for each failure-causing scene is 16.86s. From a statistical perspective, our analyzer finds that objects on both sides of the road, rainy weather and red or yellow objects are more likely to cause the CILRS system to fail.