Machine Unlearning for Generative AI

Abstract: As generative AI advances, the ability to selectively remove information from large language models (LLMs)—known as unlearning—is crucial for regulatory compliance, ethical safeguards, and mitigating harmful content retention. This talk presents recent advances in LLM unlearning from a model-data-optimization perspective, focusing on two key directions. First, from a model perspective, I will present surgical unlearning via weight attribution, a principled approach that strategically identifies and modifies influential model weights to enhance unlearning effectiveness while preserving model utility. Second, from an optimization perspective, I will discuss the critical challenge of robust unlearning through smoothness optimization, specifically tackling relearning attacks, where removed knowledge could be recovered from small amounts of forgotten data. Drawing inspiration from adversarial robustness, I will establish a connection between robust unlearning and sharpness-aware minimization (SAM), demonstrating how smoothness optimization can enhance resistance to relearning attacks. Lastly, this talk will conclude with some open challenges and future directions for integrating unlearning into the AI lifecycle, ensuring long-term trustworthiness and compliance.

Bio: Dr. Sijia Liu is an Assistant Professor in the Department of Computer Science and Engineering at Michigan State University, where he leads the Optimization and Trustworthy ML (OPTML) lab. He also serves as an Affiliated Professor at IBM Research. His research focuses on trustworthy and scalable ML, bridging foundational areas such as optimization and learning theory with applied research in vision and language modeling. Recently, his work has emphasized machine unlearning for generative AI. Dr. Liu has been recognized with numerous prestigious honors, including the NSF CAREER Award in 2024, the Best Paper Runner-Up Award at the Conference on Uncertainty in Artificial Intelligence (UAI) in 2022, and the Best Student Paper Award at the 42nd IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) in 2017. Additionally, he has been the lead organizer of the New Frontiers in Adversarial ML (AdvML-Frontiers) workshop series from 2022 to 2024.

*** This Event will take place in a hybrid format. The location for in-person attendance will be room 1311 EECS. Attendance will also be available via Zoom.

Join Zoom Meeting: https://umich.zoom.us/j/93679028340

Meeting ID: 936 7902 8340

Passcode: XXXXXX (Will be sent via e-mail to attendees)

Zoom Passcode information is also available upon request to Kristi Rieger([email protected])