Differentially Private Diffusion Models
12A Google TechTalk, presented by Tim Dockhorn (University of Waterloo), 2023/04/12
ABSTRACT: While modern machine learning models rely on increasingly large training datasets, data is often limited in privacy-sensitive domains. Generative models trained with differential privacy (DP) on sensitive data can sidestep this challenge, providing access to synthetic data instead. We build on the recent success of diffusion models (DMs) and introduce Differentially Private Diffusion Models (DPDMs), which enforce privacy using differentially private stochastic gradient descent (DP-SGD). In this seminar, I will give an accessible introduction to DMs and explain how their optimal design space changes under the DP paradigm. I will then discuss noise multiplicity: a powerful modification of DP-SGD tailored to the training of DPDMs. Lastly, I will discuss some recent follow-up work that uses public pre-training to improve the performance of DPDMs.