2026-05-23T09:22:53Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-2024162025-08-18T22:22:09Zoai_pmh:alloai_pmh:repo_items

202416 https://hdl.handle.net/2286/R.2.N.202416 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2025 199 pages Doctoral Dissertation Academic theses en Welfert, Monica Sankar, Lalitha Kosut, Oliver Dasarathy, Gautam Zou, Shaofeng Arizona State University Partial requirement for: Ph.D., Arizona State University, 2025 Field of study: Computer Engineering Advances in machine learning have driven significant progress in domains ranging from high-fidelity image synthesis to real-world decision-making systems. Despite these successes, modern models face persistent challenges in three critical areas: generating realistic synthetic data, ensuring fairness for underrepresented subpopulations, and safeguarding against the inference of sensitive attributes. This dissertation addresses these challenges through the development of adversarial frameworks for data generation and evaluation, with a focus on stability, fairness, and privacy. The first part investigates training instabilities in generative adversarial networks (GANs), a widely used class of models for synthetic data generation that operates as a two-player minimax game between a generator and a discriminator. While powerful in theory, GANs are often hindered by unstable training dynamics. To address this, a unified loss-function perspective is developed and a novel dual-objective formulation is introduced, employing tunable loss functions for both players. This approach improves convergence guarantees and enhances training stability. The second part turns to fairness in classification tasks with imbalanced subpopulations. It provides a theoretical analysis of two simple yet effective data augmentation strategies, downsampling and upweighting, within the setting of linear last-layer retraining. This setup allows for tractable insights into their impact on worst-group accuracy, a fairness metric that captures performance on the most disadvantaged subpopulation. The analysis is further extended to account for noisy labels, a common challenge in real-world datasets. The final part develops a framework for evaluating the risk of sensitive attribute inference via adversarial estimation. Leveraging tools from statistical estimation theory, the framework establishes lower bounds on the minimum mean-squared error of inference attacks, based on finite-sample training and validation errors. These bounds enable quantifiable assessments of privacy risk under capacity constraints and limited data. Together, these contributions provide principled advances in adversarial methods for synthetic data generation, fairness-aware learning, and privacy risk evaluation, offering both theoretical foundations and practical tools to improve reliability and accountability in machine learning systems. Computer Engineering Adversarial methods Algorithmic fairness data augmentation Generative Adversarial Networks Machine learning Privacy risk evaluation Adversarial Approaches for Generating Synthetic Data and Evaluating Fairness and Inference Risks