Author ORCID Identifier
https://orcid.org/my-orcid?orcid=0000-0002-5913-6573
Defense Date
2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Computer Science
First Advisor
Tamer Nadeem
Second Advisor
Eyuphan Bulut
Third Advisor
Irfan Ahmed
Fourth Advisor
Sherif Abdelwahed
Fifth Advisor
Yanxiao Zhao
Abstract
Edge devices increasingly run high-stakes machine learning workloads under strict latency, power, memory, and communication constraints. Deep neural networks provide the accuracy needed for tasks such as network intrusion detection, health monitoring, autonomous sensing, and human activity recognition, but conventional inference executes the same computational path for every input. Early-exit neural networks reduce average inference cost by allowing confident samples to terminate before the final layer. However, existing early-exit systems are largely class-agnostic: they allocate computation according to confidence or uncertainty, without considering that different output classes may carry different operational urgency. This dissertation develops class-aware early-exit inference, a framework in which computational depth is shaped by both model confidence and class-level priority. The central thesis is that edge inference should optimize priority-weighted computational depth rather than average computation alone. The work begins with ClassyNet, which encodes class priority during training through Bag-of-Classes assignments and cost-sensitive loss matrices. It then validates class-aware early exits in network intrusion detection, extends fixed priority assignments to runtime-reconfigurable control through RACENet, adapts class-aware training to federated and heterogeneous clients through CAFED, studies the security implications of dynamic exit behavior through DDAS, and extends that security analysis to self-speculative decoder inference under prompt-level adversarial pressure. Across these contributions, the dissertation shows that class-aware early-exit inference can align resource use with application urgency while preserving practical accuracy. ClassyNet achieves up to 9.41$\times$ speedup and a 4$\times$ reduction in average inference latency for high-priority classes with less than 1\% accuracy penalty. Its network intrusion detection application achieves up to 3.57$\times$ faster detection for priority threat classes while keeping overall accuracy within 1\% of the baseline. RACENet reduces high-priority class latency by up to roughly 28\% relative to a class-agnostic early-exit baseline in the tested configurations while supporting priority changes without retraining. CAFED extends the framework to federated learning, reaching average accuracies of 84.3\% on CIFAR-10 and 89.2\% on WISDM across heterogeneous clients while improving substantially over a FedAvg+ClassyNet baseline. DDAS demonstrates that the exit mechanism itself is a security surface: adversarial inputs can increase power consumption by 128\% and inference time by 55\% while preserving enough accuracy to evade accuracy-only monitoring. The final study shows that LayerSkip-style self-speculative decoding can lose most of its clean acceleration under task-preserving adversarial suffixes, and that runtime guarding contains the attack primarily by falling back toward full-depth execution. Together, these results establish class-aware early-exit inference as a practical approach for edge intelligence systems whose outputs have unequal consequences. The dissertation shows that efficient edge inference is not only a matter of reducing computation, but also of deciding where computation should be spent, how that decision can adapt after deployment, how it can be trained under distributed data ownership, and how it can remain robust when dynamic computation becomes an adversarial target in both classifiers and generative models.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
7-13-2026