Date of Award
Doctor of Philosophy (PhD)
The transportation system in the US is transforming into an intelligent cyber-physical-systems (CPS) through the inclusion of technology and digital infrastructure. One of the core components of the transportation CPS (TCPS) is connected and automated vehicles (CAV). Advancement in vehicular network technologies and vehicle automation is an important factor in the evolution of CAVs. Improving the security of in-vehicle networks and automated vehicle applications, and reliability of vehicle-to-everything (V2X) communication are critical areas that need further attention from the research community. In this dissertation, the author demonstrates methods and models to improve the security and reliability of CAV communication and applications under different scenarios using state-of-the-art technologies, such as software-defined networking, artificial intelligence, and edge computing. This dissertation is composed of three interrelated articles. The first article demonstrates the development of an anomaly detection model for the in-vehicle controller area network (CAN) of a vehicle. The model is tested for two real-vehicle CAN datasets for two different types of anomalies. The results show improvement in detection accuracy over baseline models. In the second article, an edge-centric handover management system is developed for the internet of vehicles. The system architecture is based on multiple layers of edge devices and distributed computation for managing the handover of CAVs. The system uses a deep learning model for predicting the future movement of vehicles and software-defined networking for implementing the handover of connected vehicles. Analysis shows that an edge-centric handover system is superior to a decentralized individual handover system for handover latency, communication delay, packet loss, and throughput. The third article focuses on the security of automated vehicle applications. This article presents a hybrid defense method that protects deep learning models for traffic sign classification against adversarial attacks. This method uses random filtering, ensembling, and local feature mapping to improve the resilience of the classifier. Analysis shows that this defense method improves upon baseline defense strategies in making the model resilient against different types of adversarial attacks and demonstrates its general applicability for any future adversarial attacks against traffic sign classifiers.
Khan, MD Zadid, "Security and Reliability of Vehicular Networks and Autonomous Vehicle Applications Using Artificial Intelligence and Edge Computing in a Cyber-Physical Systems Environment" (2021). All Dissertations. 2886.