Towards Enabling Network Secure Sensor Communication and Perception-Based Lane Detection in Autonomous Vehicles.

Autonomous vehicles (AVs) critically depend on the integrity of their sensor networks and the robustness of their software algorithms to ensure safe and effective navigation. Sensors such as LiDAR, RADAR, GPS, and cameras serve as the primary means for environmental perception, making the reliability and security of their data essential. This creates a pressing need for Intrusion Detection Systems (IDSs) capable of defending against a wide range of cyberattacks. Simultaneously, AVs rely on advanced software for tasks like lane detection, navigation, path planning, object classification, and collision avoidance. This thesis addresses two core challenges in AV development: designing machine learning (ML)-based IDSs that are resilient to both known and zero-day attacks, and developing a robust lane detection system. On the network security front, it explores deep neural networks and various network traffic datasets, proposing a novel IDS framework that demonstrates strong transferability, improved rare attack detection, and broad applicability across backbones and datasets. It also introduces a privacy-aware design using federated learning to secure IoT devices in realistic scenarios. In parallel, the computer vision component focuses on enabling Purdue University’s AV to achieve accurate lane detection in racing environments. It presents a fully annotated racing dataset and a baseline model that outperforms existing methods, offering a valuable resource for advancing research in autonomous racing.