MORE | Spring 2021
Adversarial Attacks on Autonomous Driving with Physically Realizable Patterns
Following the physical adversarial attacks faced by Deep Neural Networks (DNNs), controllers learned using such DNNs with Reinforcement Learning (RL) approach are also vulnerable particularly in risk-sensitive areas such as Autonomous Driving (AD). This research aims to study such attacks on AD mainly in the Spatio-temporal domain in the presence of adversarial physical objects on road. Studying such an under-explored topic can provide impactful outcomes and questions the security of current AD systems. These results could be further utilized to make AD systems even more robust and secure in the future.
Mentor: Yezhou Yang, Yi Ren