Graduation Semester and Year

Summer 2026

Language

English Language

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Science

Department

Computer Science and Engineering

First Advisor

Habeeb Olufowobi

Second Advisor

Farhad Kamangar

Third Advisor

Manfred Huber

Fourth Advisor

Ming Li

Fifth Advisor

Mohammad Atiqul Islam

Abstract

Graph Neural Networks (GNNs) extend the capabilities of deep learning beyond Euclidean domains by enabling the representation and learning of complex interactions in graph-structured data. These capabilities have led to significant advances in applications such as communication networks, molecular design, recommendation systems, and trajectory forecasting. Despite this success, the potential of GNNs beyond naturally graph-structured domains remains underexplored, particularly in applications where constructing relational representations from conventional data may reveal dependencies that are difficult to capture using traditional learning methods. Moreover, the reliance of GNNs on neighborhood aggregation introduces substantial computational and memory overheads that hinder deployment in resource-constrained and latency-sensitive environments. As GNNs continue to be adopted in safety-critical and privacy-sensitive settings, understanding their vulnerability to adversarial manipulation and information leakage has also become increasingly important.

This dissertation addresses these challenges by demonstrating the applicability of GNNs beyond naturally graph-structured domains through the transformation of conventional vehicular data into relational representations that capture dependencies among driving behaviors, enabling driver identification and maneuver classification for security and safety applications. In addition, this dissertation develops collaborative learning and knowledge distillation techniques that improve the quality of GNN representations and transfer relational knowledge to lightweight graph-less models, enabling low-latency inference without direct access to graph structure at deployment time. Finally, this dissertation systematically investigates the security and privacy landscape of GNNs through a unified characterization of attacks and defenses and develops novel methodologies for auditing privacy and structural information leakage, revealing vulnerabilities including membership inference, link extraction, and information leakage through distilled student models.

These contributions advance the development of secure and efficient graph learning systems and provide practical insights for deploying GNNs in mission-critical applications.

Keywords

Graph Neural Networks, Security, Privacy, Knowledge Distillation, Drivers, Membership Inference Attacks, Link Stealing, Property Inference Attacks, Vehicle, Theft

Disciplines

Artificial Intelligence and Robotics | Cybersecurity | Information Security

Download

Available for download on Wednesday, July 26, 2028

Share

COinS