DP-SGD vs PATE: Which Has Less Disparate Impact on GANs?

Georgi GanevarXiv 2021 · arXiv

Introduction

Generative Adversarial Networks (GANs) are increasingly utilized to produce synthetic datasets, facilitating data sharing while aiming to preserve individual privacy. To provide formal privacy guarantees, these models are often trained using Differential Privacy (DP) frameworks. However, DP is known to introduce "disparate impact" in discriminative models, where the accuracy of minority subgroups is disproportionately degraded. This study investigates whether similar disparate impacts occur in generative models, specifically comparing the two leading deep learning DP frameworks: DP-SGD and PATE.

Methodology and Experimental Setup

The authors evaluate two privacy-preserving GAN architectures:

  1. DP-WGAN: Employs DP-SGD, which guarantees privacy by clipping and adding noise to individual gradients of the discriminator during training.
  2. PATE-GAN: Utilizes the PATE (Private Aggregation of Teacher Ensembles) framework. Instead of a single discriminator, it uses an ensemble of "teacher" discriminators trained on disjoint partitions of the data; a "student" discriminator is then trained on noisy labels provided by the teachers.

Experiments were conducted on the MNIST handwritten digit dataset, artificially imbalanced to test three scenarios:

  • Minority: Digit "8" is undersampled to 10% or 25% of its original size.
  • Majority: Digit "8" remains balanced while all other classes are undersampled to 10% or 25%.
  • Mixed: A combination where "8" is made the relative minority or majority, and all other classes are uniformly decreased.

The models were tested across various privacy budgets ($\epsilon = 0.5, 5, 15, \infty$). Evaluation metrics included Size RMSE (distribution matching) and Recall Truncated RMSE (the utility of synthetic data for training downstream classifiers).

Key Insights

Privacy as Regularization

A notable finding is that PATE-GAN does not exhibit a monotonic performance decline as privacy constraints tighten. Instead, moderate privacy budgets ($\epsilon = 5$ or $15$) often improve performance over the non-DP baseline. The noise and ensemble structure appear to act as a regularizer, enhancing the model's generalization capabilities.

Divergent Effects on Class Distribution

The two frameworks handle class imbalance in opposing ways. DP-WGAN tends to "homogenize" the dataset by shrinking majority classes and inflating minority ones. In contrast, PATE-GAN tends to exacerbate existing imbalances, further increasing the size of majority classes at the expense of smaller ones.

Results and Disparate Impact

Overall Performance and the Minority Challenge

PATE-GAN generally outperforms DP-WGAN across most settings, maintaining better class counts and higher recall. While DP-WGAN's utility collapses at strict privacy levels ($\epsilon = 0.5$), PATE-GAN remains relatively stable.

Performance comparison in Minority settings (25% and 10% imbalance). Note the failure of PATE-GAN to capture the 10% minority class (digit 8) across all privacy levels.

The Extreme Minority Failure

Despite its overall superiority, PATE-GAN has a significant weakness: it fails to generate any instances of extreme minority classes (e.g., the 10% Minority setting). This failure occurs even without DP, likely because the disjoint data partitioning prevents individual teachers from seeing enough minority samples to provide meaningful feedback. DP-WGAN, while lower in overall utility, successfully captures these extreme minority groups.

Performance comparison in Majority settings (25% and 10% imbalance).

Conclusion and Future Directions

The study highlights a critical trade-off between overall utility and the preservation of extreme minority classes. PATE-GAN offers superior general performance and a favorable privacy-utility trade-off but risks completely erasing small subpopulations. Future research should explore hybrid approaches that leverage PATE's robustness while incorporating DP-SGD's ability to represent rare subgroups, ensuring that synthetic data generation does not inadvertently marginalize underrepresented populations.

Summary updated

Abstract

Generative Adversarial Networks (GANs) are among the most popular approaches to generate synthetic data, especially images, for data sharing purposes. Given the vital importance of preserving the privacy of the individual data points in the original data, GANs are trained utilizing frameworks with robust privacy guarantees such as Differential Privacy (DP). However, these approaches remain widely unstudied beyond single performance metrics when presented with imbalanced datasets. To this end, we systematically compare GANs trained with the two best-known DP frameworks for deep learning, DP-SGD, and PATE, in different data imbalance settings from two perspectives -- the size of the classes in the generated synthetic data and their classification performance. Our analyses show that applying PATE, similarly to DP-SGD, has a disparate effect on the under/over-represented classes but in a much milder magnitude making it more robust. Interestingly, our experiments consistently show that for PATE, unlike DP-SGD, the privacy-utility trade-off is not monotonically decreasing but is much smoother and inverted U-shaped, meaning that adding a small degree of privacy actually helps generalization. However, we have also identified some settings (e.g., large imbalance) where PATE-GAN completely fails to learn some subparts of the training data.
BibTeX
@misc{ganev2021dp,
  author = {Georgi Ganev},
  title = {{DP-SGD vs PATE: Which Has Less Disparate Impact on GANs?}},
  year = {2021},
  eprint = {2111.13617},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2111.13617},
}
Figure