Date of Award

12-24-2025

Date Published

January 2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering and Computer Science

Advisor(s)

Asif Salekin

Keywords

Healthcare intervention;Model Compression;Personalization;Robustness;Verified Robustness

Subject Categories

Computer Sciences | Physical Sciences and Mathematics

Abstract

Advances in machine learning have enabled significant progress in generating robust solutions. However, designing solutions that balance robustness while meeting the practical constraints of diverse applications, such as limited resources and data, remains a challenge. We explored three key aspects of this interplay: verified robust compressed neural networks, context-wise robust personalization, and reliable counterfactual interventions for healthcare. First, we introduce VeriCompress, a novel framework to streamline the synthesis of compressed neural networks with formal guarantees of adversarial robustness. This enables the deployment of reliable and efficient models in resource-constrained environments, such as smartphones. Second, we developed CRoP (Context-wise Robust Static Personalization), addresses the challenge of context-wise robustness of personalized models. By employing a static personalization approach that incorporates robustness towards external changes, CRoP enhances model robustness without requiring continuous updates or data collection. This method optimizes performance in privacy-sensitive and resource-constrained scenarios. Another variation of this problem was explored in a few-shot learning scenario in collaboration with a fellow PhD student. Finally, we propose a novel mechanism for generating reliable counterfactual examples for, but not restricted to, healthcare interventions. Counterfactual examples suggest minor changes in the input that can result in a different model inference. Our approach ensures counterfactuals have a minimum confidence margin, enhancing their reliability. To demonstrate robustness, we evaluate the transferability of these counterfactuals across different models. Together, these contributions aim to generate robust AI solutions, paving the way for their broader adoption in critical real-world applications.

Access

Open Access

Share

COinS