Explore the cutting-edge research on generating differentially private synthetic data using foundation model APIs in this 56-minute Google TechTalk presented by Sivakanth Gopi. Dive into the challenges and potential of leveraging large foundation models as black boxes, utilizing only their inference APIs to create high-quality synthetic data while maintaining privacy. Learn about the innovative Private Evolution (PE) framework and its promising results, including significant improvements in FID scores for CIFAR10 dataset. Gain insights into the speaker's background, the abstract of the talk, and the comprehensive syllabus covering topics such as generative neural networks, differential privacy, diffusion models, and the intrinsic dimension of data.
Differentially Private Synthetic Data Generation via Foundation Model APIs
Overview
Syllabus
Intro
(Generative) Neural networks can memorize!
Differential Privacy (DP)
Existing Approaches
DP Synthetic Data via APIs
Why can it work?
Diffusion Models
Distance function
Downstream Classification
Private Evolution at work
Stable Diffusion
Non-private convergence of PE
Proof Sketch
Intrinsic dimension
Conclusions
Taught by
Google TechTalks
Related Courses
-
Differentially Private Diffusion Models Generate Useful Synthetic Images
-
Differentially Private Diffusion Models
-
Noise-Aware Statistical Inference with Differentially Private Synthetic Data
-
Marginal-based Methods for Differentially Private Synthetic Data - Differential Privacy for ML Series
-
Differentially Private Sampling from Distributions - Google Algorithms Seminar
