Cybersecurity in Connected and Automated Transportation Systems

Abstract

The last decade has been a witness to critical advancements in the field of intelligent transportation systems (ITS), with a great volume of research focusing on connected and automated vehicles (CAVs). A CAV leverages connected vehicle (CV) and automated vehicle (AV) technologies to enable automation and cooperation between vehicles and/or infrastructures to improve mobility, safety, efficiency, and sustainability in transportation systems. Despite these benefits, both connectivity and automation increase potential attack surfaces on CAVS, thereby introducing unprecedented cybersecurity challenges. This dissertation addresses these challenges by developing a comprehensive cybersecurity framework, which consists of two major components.

The first component of the framework focuses on anomaly detection in CAV sensors. A CAV requires high fidelity data for automated and cooperative driving tasks. However, oftentimes there exist uncertainties in the CAV sensor measurements, including noise and time delay. Therefore, the first part of the dissertation proposes a signal filtering-detection framework to estimate the CAV sensor state in the presence of potential cyberattacks. This dissertation considers sensor anomaly as a result of either malicious attacks or sensor faults, and proposes a series of data-driven detectors in conjunction with extended Kalman filter (EKF)-based algorithms for both signal filtering and anomaly detection. The first part of the dissertation consists of studies under different driving scenarios, including in car following and platooning modes. The study on the platooning mode further analyzes platoon stability under cybersecurity uncertainties, where a new class of string stability measures, namely pseudo string stability, is introduced to analyze the stability of a vehicle string under various types of attacks and imperfect detectors.

Despite the necessity of anomaly detection, a well-developed security monitor should also consider the dynamically changing environment faced by a CAV and the strategic behavior of the adversary to ensure both security and energy efficiency of the CAV. Therefore, the second component of the framework focuses on a macroscopic perspective to address security monitoring challenges faced by a CAV. First, the dissertation addresses the attack profile prediction and sensor selection problem under a security resource constraint. Toward this goal, the dissertation considers a sequential two-player game involving an attacker and a defender. This dissertation thereby develops an online learning algorithm for the defender to solve a variant of the multi-armed bandit (MAB) problem, namely, the multi-armed bandit with variable plays (MAB-VP). The dissertation provides a sublinear regret bound of the proposed algorithm, and derives the conditions under which a Nash equilibrium of the strategic game exists. Based on the analysis of the asymptotic expected reward of the two players, the dissertation assesses the effectiveness of the following two defense strategies from a game theoretical perspective: (1) increasing security resources for monitoring, and (2) improving the performance of the detector.

Considering the fact that continuously monitoring all sensors aboard a CAV can be increasingly energy-intensive and therefore impractical, this dissertation introduces a dynamic security resource allocation problem to selectively monitor a subset of sensors for potential cyberattacks. This is accomplished by providing a mathematical framework based on a partially observable Markov decision process (POMDP), which prescribes a policy to dynamically assign security resource by balancing the trade-off between detection performance and energy-efficiency. This dynamic resource allocation problem serves as a complementary supplement to the sensor selection study, and together they form the last component of the proposed framework.