Authors: Marcel Neunhoeffer, Zhiwei Steven Wu, Cynthia Dwork
Associated Paper: Private Post-GAN Boosting (ICLR 2021)
Overview
This repository contains the replication code for our ICLR 2021 paper “Private Post-GAN Boosting.” The method addresses a fundamental challenge in differentially private synthetic data generation: GAN training often fails to converge due to privacy-protective noise, leading to poor quality synthetic data.
The Method
Private Post-GAN Boosting (Private PGB) combines samples from the full sequence of generators obtained during GAN training—not just the final generator. Using a game-theoretic formulation with multiplicative weights, the method finds an equilibrium between a synthetic data player and a distinguisher.
Key Insight: While individual generators during training may be poor, a weighted mixture from different training epochs can produce high-quality synthetic data with formal differential privacy guarantees.
Experiments Included
- Toy data (Gaussian mixtures)
- MNIST
- US Census data
- Titanic dataset
Core Components
- Private Multiplicative Weights method (Hardt and Rothblum, 2010)
- Discriminator Rejection Sampling (Azadi et al., 2019)
- Game-theoretic reweighting of generated samples