Verifying Robustness of Neural Networks in Vision-Based End-to-End Autonomous Driving

This paper addresses the verification of neural network robustness against perturbations of input data in vision-based end-to-end autonomous driving systems. The main contributions of this work are: i) We provide a comprehensive analysis of current neural-network-based perception and decisionmaking components in autonomous vehicles, highlighting their susceptibility to attacks or perturbation. ii) We develop and implement a novel framework for systematically evaluating and verifying the robustness of neural networks against such perturbations, focusing on imperceptible image modifications. iii) We present empirical results demonstrating our verification framework’s effectiveness in identifying weaknesses and providing probabilistic guarantees on network behaviour. Our work underscores the critical need for robust verification methods to ensure the reliability and safety of autonomous driving systems, paving the way for safer integration of neural networks in autonomous vehicles.