Title: Machine Unlearning in Diffusion Model
Abstract: Diffusion models have achieved impressive generative capabilities, but they may also memorize or reproduce copyrighted, private, or harmful content. Machine unlearning provides an efficient alternative to retraining by removing unwanted concepts while preserving the model’s general generation ability. In this seminar, I will present our recent work on diffusion model unlearning, focusing on multi-concept, training-free, scalable, and verifiable unlearning. We study how concept interference and parameter entanglement make forgetting difficult, and develop methods for selective parameter updates, concept-sensitive neuron pruning, scalable closed-form editing, and auditing-aware erasure. Together, these approaches aim to make diffusion model unlearning more efficient, precise, scalable, and reliable.
Short bio: Xiaolong Ma is an Assistant Professor in the School of Electrical, Computing, and Software Engineering at the University of Arizona. He received his Ph.D. in Computer Engineering from Northeastern University in 2022 and was previously an Assistant Professor at Clemson University. His research focuses on scalable, efficient, and trustworthy AI, with particular interests in computer vision, large language models, efficient learning, and trustworthy generative AI. His work has been published in leading venues including NeurIPS, ICML, CVPR, ICLR, IEEE S&P, ASPLOS, ISCA, MICRO, and IEEE TPAMI. He is a recipient of the NSF CAREER Award, and his research is supported by NSF and NASA.