2026-09-21T08:04:52Z https://www.ideals.illinois.edu/oai-pmh

oai:www.ideals.illinois.edu:2142/129900 2025-10-25 col_2142_5131 col_2142_10761 com_2142_5130 com_2142_10755 com_2142_234

application/pdf en text Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-10-20 without embargo terms The student, Fan Wu, accepted the attached license on 2025-05-29 at 18:03. The student, Fan Wu, submitted this Dissertation for approval on 2025-05-29 at 18:11. This Dissertation was approved for publication on 2025-05-30 at 10:15. DSpace SAF Submission Ingestion Package generated from Vireo submission #22316 on 2025-10-20 at 20:14:44 Differential privacy in the era of generative AI: promises and challenges Wu, Fan 2025-05-30 Forsyth, David A. Chandrasekaran, Varun Forsyth, David A Wang, Gang Peng, Hao Kohno, Tadayoshi Differential Privacy Machine Learning Generative Ai Large Language Models eng Large language models (LLMs) are seeing rapid development and widespread deployment. As these models become increasingly capable and are deployed across diverse domains involving sensitive data, privacy concerns have intensified. Their inadvertently memorizing and leaking private information creates significant privacy risks when they are fine-tuned with user data or deployed as interactive agents. This thesis addresses the critical privacy challenges emerging in the era of generative AI, with a particular focus on protecting training data privacy in LLMs across various learning paradigms and application scenarios, as well as understanding what protection we actually offer. As a central tool, we leverage and scrutinize differential privacy (DP). Concretely, we develop a novel DP framework for language model alignment through preference tuning (RLHF), formalize new privacy definitions for multi-user training data scenarios, and critically examine DP-SGD—the workhorse algorithm for DP LLM training—and reveal an alarming variance in its empirical privacy protection. Together, these contributions advance both the practical applications and fundamental understanding of differential privacy in LLMs, providing researchers and practitioners with new tools and insights to navigate the landscape of privacy in generative AI. 2025-08 Text https://hdl.handle.net/2142/129900 Copyright 2025 Fan Wu Siebel School Comp & Data Sci Computer Science University of Illinois Urbana-Champaign Ph.D. Dissertation