Privacy Issues in Diffusion Models: Unveiling Memorization through the Sharpness of Probability Landscapes
13 Aug 2025Dongjae Jeon - AI-ISL, CS, Yonsei University
In this talk, I will begin with a brief overview of my recent research topics, followed by an intuitive introduction to diffusion models to provide the necessary background. I will then present my recent work, featured as a spotlight poster at ICML 2025, which examines privacy risks in diffusion models by analyzing memorization through the sharpness of probability landscapes. The presentation will cover the theoretical foundations, key empirical findings, and a practical inference-time mitigation strategy, demonstrating how our approach enables early detection and effective reduction of memorization without compromising generation quality.
Catering Courtesy of CIP Lab