Adversarial Attacks and Mitigations on Scene Segmentation of Autonomous Vehicles

Abstract

In this study, we focus on the effectiveness of adversarial attacks on the scene segmentation function of autonomous driving systems (ADS). We explore both offensive as well as defensive aspects of the attacks in order to gain a comprehensive understanding of the effectiveness of adversarial attacks with respect to semantic segmentation. More specifically, in the offensive aspect, we improved the existing adversarial attack methodology with the idea of momentum. The adversarial examples generated by the improved method show higher transferability in both targeted as well as untargeted attacks. In the defensive aspect, we implemented and analyzed five different mitigation techniques proven to be effective in defending against adversarial attacks in image classification tasks. The image transformation methods such as JPEG compression and low pass filtering showed good performance when used against adversarial attacks in a white box setting.