Differentially Private Diffusion Models

Explore the intersection of privacy and generative AI in this 45-minute Google TechTalk presented by Tim Dockhorn from the University of Waterloo. Gain insights into Differentially Private Diffusion Models (DPDMs), a novel approach that combines the power of diffusion models with differential privacy techniques. Learn how DPDMs address the challenge of limited data in privacy-sensitive domains by generating synthetic data while preserving privacy. Discover the optimal design space for diffusion models under differential privacy constraints and understand the concept of noise multiplicity, a tailored modification of differentially private stochastic gradient descent. Delve into recent advancements that leverage public pre-training to enhance DPDM performance, offering a comprehensive overview of this cutting-edge research in machine learning and data privacy.