Author ORCID Identifier
https://orcid.org/0000-0003-0637-2430
Defense Date
2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Computer Science
First Advisor
Milos Manic
Second Advisor
Preetam Ghosh
Third Advisor
Bridget McInnes
Fourth Advisor
Benny Varghese
Fifth Advisor
Bjorn Vaagensmith
Abstract
Cyber-Physical systems (CPSs) are at the core of modern critical infrastructures. The increasing complexity of CPSs makes them vulnerable to diverse cyber-physical attacks. Due to the safety-critical nature of transportation applications, any failures can result in massive economic losses, disrupt essential services, and even endanger human lives. Consequently, ensuring the security of vehicular transportation CPSs is of utmost importance.
Today’s complex CPSs generate massive amounts of unlabeled data. Labeling these data is an expensive and time-consuming task, and often requires domain expertise. Therefore, the existing supervised algorithms cannot take advantage of the abundance of real-world unlabeled data. Given these concerns, the adoption of unsupervised learning methodologies becomes imperative. Hence, we developed unsupervised deep learning-based anomaly detection systems (ADSs) that can be trained using such unlabeled data.
While traditional neural networks trained on one domain may excel within that specific context, they often fail when applied to data from a similar domain (not domain-adaptable). This is due to data distribution disparities, even when the feature space remains consistent. In such scenarios, learning domain-invariant features is essential for transferring knowledge from one domain to another. Unsupervised transfer learning-based domain-adaptable neural networks provide a great way to achieve this. Therefore, we developed domain adversarial ADSs that can effectively use cross-domain data to develop efficient ADSs.
Time-series unsupervised AD requires learning long-term dependencies from sequential data. Normal behavior data often exhibit these dependencies, while anomalies do not. This distinction aids in effectively differentiating anomalies from normal data. Generative AI (genAI) approaches, such as transformer architectures with attention mechanisms, excel at capturing these long-term interdependencies. This makes genAI approaches ideal for time-series anomaly detection in CPSs. Therefore, we develop novel genAI-based methodologies to improve anomaly detection in CPSs.
Thus, the main objective of this dissertation is to improve the security of CPSs in vehicular transportation applications using unsupervised AD. This main objective is delineated into three sub-objectives: 1) Develop combined cyber and physical anomaly detection frameworks to improve overall system health using unsupervised NNs, 2) Develop unsupervised transfer learning methodologies to improve AD performance in CPSs using cross-domain CPS data, and 3) Develop GenAI-based approaches for effective unsupervised ADSs.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
12-4-2025