Improving Safety Mobility and Cybersecurity with Infrastructure-Assisted Cooperative Driving

Traffic safety and mobility are two primary concerns in urban transportation systems. Connected and autonomous vehicles (CAV) are considered to have great potential to improve urban traffic operations. However, single CAV-based decision-making has two major limitations: 1) inaccurate and restricted perception capabilities. 2) decisions are made to optimize individual objectives rather than achieve system optimal. To address these limitations, CAVs can exchange real-time information collected from onboard sensors with transportation infrastructure through V2X communication and can be controlled collectively under the guidance of infrastructure. The cooperation between infrastructure and CAVs can reduce potential crashes and improve mobility. On the other hand, onboard sensors and V2X communication also bring risks in cyberspace. As the future transportation system will transition into a cyber-physical system, the impact of cyber-attacks is not only limited to vehicle operation but also influences the transportation system as a whole. Modeling and evaluating the cyber-attack impact and reducing cyber risks in transportation system is an urgent research need.

To address the above-mentioned challenges, this dissertation proposes an infrastructure-assisted cooperative driving framework to improve transportation system safety, mobility and reduce cyber risks. The proposed framework starts with a full CAV trajectory planning model (Chapter 2) based on imitation learning that integrates both longitudinal and lateral driving behaviors. Unlike most of the existing microscopic traffic flow models, the new model mimics realistic human driving behaviors and integrates car-following and continuous lane-changing maneuvers. In Chapter 3, this model is extended to an infrastructure-assisted cooperative driving framework that optimizes CAV trajectories and signal control parameters simultaneously. While lane changing process is ignored or simplified in many existing studies, this new framework integrates complete lane changing behaviors in both signal timing optimization and trajectory planning process. The second part of this dissertation focuses on modeling, evaluating, and mitigating data spoofing attacks towards the transportation system. Chapter 4 aims to design a generic anomaly detection model that can be used to identify abnormal trajectories from both known and unknown attacks. First, the attack behaviors of two representative known data spoofing attacks (i.e., estimated time of arrival (ETA) attack, and multi-sensor fusion (MSF) attack) are modeled. Then a machine learning-based anomaly detection framework is proposed to identify falsified trajectories generated by various unknown attacks. The anomaly detection model combines a domain knowledge-based feature extractor, and an anomaly classifier trained with trajectories from the two known attack models. Finally, a novel traffic-informed symbolic regression (TI-SR) approach is developed in Chapter 5 that resembles the output of cyber-attacks in explicit mathematical expressions for transportation system level performance evaluation. In contrast to many existing machine learning-based approaches, the proposed TI-SR model is able to accurately capture both trajectory level dynamics and traffic level safety impacts from cyber-attacks with strong interpretability.