Private Post-GAN Boosting

Marcel Neunhoeffer, Steven Wu, Cynthia DworkICLR 2021 · OpenReview

Generating synthetic data that looks real but protects individual privacy is a major goal for modern data sharing. Differentially Private Generative Adversarial Networks (DP-GANs) are a leading tool for this task, but they have a significant flaw: they are notoriously finicky to train. Because of the noise required to guarantee privacy, DP-GANs often fail to converge, leaving researchers with a final "generator" that produces low-quality, blurry, or biased data.

In their ICLR 2021 paper, "Private Post-GAN Boosting," Marcel Neunhoeffer, Zhiwei Steven Wu, and Cynthia Dwork propose a clever workaround. Instead of trying to fix the training process itself, they suggest we stop throwing away the "history" of the GAN. By using a private boosting scheme, their method selects and combines the best samples from the entire training run to create a high-quality synthetic dataset.

The Challenge of Private GANs

Training a GAN is a high-stakes game of cat-and-mouse between a Generator (trying to create fake data) and a Discriminator (trying to catch the fakes). In a non-private setting, achieving equilibrium is already difficult. When you add Differential Privacy (DP), you must inject noise into every gradient update to ensure no single individual's data can be identified.

This noise acts like a thick fog, making it incredibly difficult for the GAN to find the optimal state. Often, the generator at the very end of training is actually worse than one from a few epochs earlier. It might also suffer from "mode collapse," where it only learns to produce one type of data (e.g., only generating the number "1" in a digit dataset). Currently, researchers spend massive amounts of time and privacy budget fine-tuning architectures to solve this, often with limited success.

Methodology: Private Post-GAN Boosting (PGB)

The authors' core insight is that while the final generator might be mediocre, the collection of generators produced throughout the training history likely contains all the components needed for a great dataset.

  1. The Post-GAN Game: They frame the problem as a two-player zero-sum game. One player (the synthetic data player) wants to pick a distribution over all previously generated samples to "fool" the other player (the distinguisher), who uses a mixture of all previously trained discriminators.
  2. Private Multiplicative Weights (MW): To solve this game privately, they adapt the Multiplicative Weights Exponential Mechanism (MWEM). Usually, MWEM is computationally impossible for high-dimensional data because it maintains a weight for every possible record in the universe. PGB makes this tractable by only maintaining weights over the samples already generated by the GAN.
  3. Privacy Budgeting: They split the total privacy budget ($\epsilon$). Most of it (80-90%) goes to training the GAN, and a small remainder is used for the PGB boosting phase to privately select and reweight the best samples.
  4. Discriminator Rejection Sampling (DRS): As an added bonus, they use the "equilibrium" mixture of discriminators found during the game to perform rejection sampling—essentially a final filter that throws out samples the discriminators are too "sure" are fake.

Experimental Setup

The authors evaluated PGB across a diverse range of data types to prove its versatility:

  • 2D Toy Data: A mixture of 25 Gaussians, which serves as a classic test for mode collapse.
  • MNIST: The standard benchmark of 70,000 handwritten digits.
  • US Census Data: Excerpts from the 1940 and 2010 decennial censuses, focusing on demographic attributes.
  • Titanic Dataset: A small, popular machine learning dataset used to test if synthetic data can train a classifier that works on real-world test data.

Key Results

The improvements provided by PGB were consistent across all evaluated metrics, proving that treating the entire training trajectory as a "committee of experts" is highly effective.

Comparison of generated 2D toy data (25 Gaussians). The bottom row demonstrates how DP PGB and DP PGB+DRS recover the data distribution much better than the DP Last Generator.
  • Visualizing the Improvement: As seen in the 2D toy data experiment above, the standard DP-GAN final generator (bottom row, "DP Last Generator") suffers from severe mode collapse and noise. In contrast, PGB successfully recovers the 25 distinct data modes, closely mirroring the real data distribution.
  • Utility in Downstream Tasks: On the Titanic dataset, a model trained on standard DP-GAN data achieved ~56% accuracy. Using PGB boosted that accuracy to ~64%, much closer to the ~76% achieved with real data.
  • Image Quality: For MNIST, the Inception Score (a measure of image variety and quality) improved from 8.07 (standard DP-GAN) to 8.66 (PGB + DRS).
  • Statistical Accuracy: On Census data, PGB significantly reduced the "Total Variation Distance" for marginal distributions (e.g., race membership), dropping the error from 0.58 to 0.13.

Limitations and Future Work

While highly effective, PGB does have limitations. The mathematical proof for PGB relies on a "support coverage assumption"—meaning the GAN's total output must cover the support of the real data. If the GAN never generates a certain type of record during its entire training run, PGB cannot magically create it; it can only reweight what is already there.

Additionally, deciding exactly how much privacy budget to save for the boosting phase (the 80/20 split) is currently a heuristic, and hyperparameter selection in DP remains a difficult open research area. Future work could explore integrating this boosting during training rather than just as a post-processing step, potentially guiding the GAN to even better convergence.

Summary updated

Abstract

Differentially private GANs have proven to be a promising approach for generating realistic synthetic data without compromising the privacy of individuals. Due to the privacy-protective noise introduced in the training, the convergence of GANs becomes even more elusive, which often leads to poor utility in the output generator at the end of training. We propose Private post-GAN boosting (Private PGB), a differentially private method that combines samples produced by the sequence of generators obtained during GAN training to create a high-quality synthetic dataset. To that end, our method leverages the Private Multiplicative Weights method (Hardt and Rothblum, 2010) to reweight generated samples. We evaluate Private PGB on two dimensional toy data, MNIST images, US Census data and a standard machine learning prediction task. Our experiments show that Private PGB improves upon a standard private GAN approach across a collection of quality measures. We also provide a non-private variant of PGB that improves the data quality of standard GAN training.
BibTeX
@inproceedings{neunhoeffer2021private,
  author = {Marcel Neunhoeffer and Steven Wu and Cynthia Dwork},
  title = {{Private Post-GAN Boosting}},
  booktitle = {International Conference on Learning Representations},
  year = {2021},
  url = {https://openreview.net/forum?id=6isfR3JCbi},
}
Figure