Integrating Artificial Intelligence for Transportation Cyber-Physical Systems: From Detection to Cyber Resilience

This dissertation investigates different methods for methane detection using unmanned aerial vehicles (UAVs). The author addresses the fundamental problem of methane emissions, a potent greenhouse gas responsible for global warming. The first part of the dissertation aims to create a cost-effective system for real-time methane detection and intensity prediction using ensemble learning models trained on meteorological parameters. The ensemble learning-based classification models show a high accuracy for methane detection. The second part of the dissertation investigates cyberattack detection for UAVs using a graph-based intrusion detection system (IDS) for controller area networks (CAN). This lightweight, protocol-independent IDS uses graph-based machine learning (GB-ML) algorithms such as graphSAGE and graph-based transformers to detect cyberattacks. By leveraging the capacity of these tools to model complex communication patterns of the CAN network, the IDS provides up to a 29% increase in detection accuracy over traditional long short-term memory (LSTM) models. This strategy can effectively recognize the combination of flooding, fuzzy, and replay attacks and provide a lightweight, scalable, real-time solution that allows UAVs to operate safely in complex environments. The IDS shows a potential for reliable applications in consumer, commercial, and industrial UAVs to support different security and computational requirements. The third part of the dissertation concludes by exploring the application of quantum AI to enhance perception systems for autonomous vehicles (AVs). Hybrid classical-quantum deep learning (HCQ-DL) models are more robust against adversarial attacks than classical deep learning (C-DL) models. By utilizing transfer learning algorithms and pre-trained models such as alexnet and vgg-16 as feature extractors, quantum models with more than 1000 quantum circuit variations were subjected to three well-known untargeted adversarial methods: projected gradient descent (PGD), fast gradient sign attack (FGSA), and gradient attack (GA). HCQ-DL models were accurate at over 95% accuracy in no-attack scenarios and over 91% accuracy during FGSA and GA attacks compared to C-DL models. Notably, during the PGD attack, one of the most effective adversarial attacks used in the study, the alexnet-based HCQ-DL model achieved an 85% accuracy compared to C-DL models, where accuracy went below 21%. These findings suggest using HCQ-DL models to protect AV perception modules and open the door to future improvements in robust AI-assisted driving systems. The dissertation brings together innovative AI technologies that UAVs can use effectively to detect methane while utilizing AI to detect cyberattacks on UAVs. The author also investigates strategies to secure AI in an autonomous vehicle’s perception module against adversarial attacks. This work opens the possibilities for robust AI systems that serve various applications, from climate monitoring to protecting transportation cyber-physical systems (CPS).