University of Waterloo

Purpose

In the realm of Autonomous Driving (AD), the application of Deep Learning (DL) and Deep Reinforcement Learning (DRL) has been pivotal. However, ensuring the robustness of these technologies against cyber threats remains a critical challenge. This project proposes the development of a high-fidelity virtual platform to assess the certified robustness of DL and DRL algorithms in AD, focusing on cybersecurity vulnerabilities. The proposed framework leverages the capabilities of the Carla open-source simulator to create a realistic virtual environment that mimics real-world cyberattack scenarios on AD systems. This project will have three phases: the first phase will involve identifying and training baseline DL/DRL algorithms for AD and selecting the state-of-the-art methods suitable for various AD scenarios. In the second phase, the focus will be on developing a virtual platform prototype capable of accurately simulating diverse cyberattack scenarios, thereby providing a testbed for evaluating the resilience of DL/DRL-based AVs. The final phase involves training certified defense algorithms designed explicitly for DL/DRLbased AV systems, emphasizing their optimization for AD requirements and testing against a spectrum of cyber threats.