Differentially Private Diffusion Probabilistic Models for Image Synthesis

Watch an 11-minute conference presentation from USENIX Security '24 exploring dp-promise, a novel approach to differentially private image synthesis using diffusion models. Learn how researchers address privacy concerns in deep learning models trained on sensitive images like human faces by developing a method that leverages diffusion model noise during the forward process to guarantee approximate differential privacy. Discover how this innovative approach outperforms existing solutions by avoiding unnecessary differential privacy noise injection while maintaining high image quality across standard metrics and datasets. The presentation features insights from researchers at Nanjing University of Science and Technology, James Cook University, and CSIRO's Data61 who demonstrate how dp-promise effectively balances privacy preservation with model utility in deep generative models.