Differentially Private Diffusion Models @GoogleTechTalks
Differentially Private Diffusion Models  @GoogleTechTalks
Uploaded May 2023 | Updated November 2026, 1 day ago
A 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.
Differentially Private Diffusion ModelsNASA CARA - Air Traffic Control in SpaaaaaaaaceWorst-Case Membership Inference of Language ModelsStable Estimators for Fast Private Statistics2022 Blockly Developers Summit: Contributing to BlocklyCharles Hoskinson | CEO of Input Output Global | web3 talks | Apr 4th 2024 | MC: Marlon Ruiz2022 Blockly Developers Summit: Backwards ExecutionPrivacy Amplification for Correlated-Noise Mechanisms via b-Min-Sep SubsamplingCan LLMs Keep a Secret? Testing Privacy Implications of Language ModelsPrivate Adaptations of Large Language ModelsEvan Shapiro | CEO Mina Foundation | web3 talks | Feb 16th 2023 | MC: Marlon RuizA Fast Algorithm for Adaptive Private Mean Estimation
Google TechTalks |

Differentially Private Diffusion Models

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER