End-to-End Learning for Autonomous Driving in Secured Smart Cities

Autonomous driving is an indispensable component in the future secured smart cities. The benefits of autonomous driving are numerous, including improving road traffic safety, reducing traffic-related economic loss, reducing traffic congestion, and enabling new vehicle applications. With the recent development of deep learning and sensor technologies, the autonomous vehicle becomes a highly complex networked system that heavily relies on sensor data to perceive the surrounding environment and make the correct decision. Such a system inevitably exposes a large attack surface and multiple attacks have been developed. It is thus crucial to protect the data and use secured machine learning algorithms to prevent, detect, and mitigate these attacks while keeping the autonomous driving system low-cost, low-latency, high-accuracy, and high-reliability. The proposed chapter presents an overview of research in autonomous driving, focusing on using end-to-end deep learning technologies for enhancing performance and security in autonomous vehicles in dynamic, adversarial environments. The chapter introduces autonomous driving paradigms, associated deep-learning methods for end-to-end learning, and the defenses against adversarial attacks. A new method utilizing temporal information for secured autonomous driving is presented; its design and implementation using CNN-LSTM include defenses against adversarial attacks. Experiments and performance demonstrating its success prediction rates are illustrated. Future research directions are described, which include both improving the autonomous driving system and enhancing its security defenses.