Attacks and Defenses for Free-Riders in Multi-Discriminator GAN

Zilong Zhao, Jiyue Huang, Stefanie Roos, Lydia Y. ChenarXiv 2022 · arXiv

Introduction: The Free-Rider Problem in Distributed GANs

Generative Adversarial Networks (GANs) have become a cornerstone of synthetic data generation. However, privacy regulations often necessitate distributed training, where data remains siloed. Multi-Discriminator GANs (MD-GANs) address this by employing a central generator that interacts with multiple distributed discriminators, each holding private datasets.

A significant vulnerability in this architecture is the presence of "free-riders." These participants aim to benefit from the final generator model without contributing real data or computational effort. In an MD-GAN, a free-rider typically uses a randomly initialized discriminator. Unlike passive leeches in other systems, these actors provide noisy, random feedback that actively degrades the generator's performance, leading to poor image quality and higher Fréchet Inception Distance (FID) scores.

Architecture of an MD-GAN showing the interaction between a central generator and distributed benign/malicious discriminators.

Proposed Defenses: DFG and DFG+

To mitigate these attacks, the authors propose two defense mechanisms: Discriminator Filtering for Generator (DFG) and its extension, DFG+.

DFG: Discriminator Filtering for Generator

The DFG mechanism relies on periodic probing and clustering:

  1. Probing: The generator periodically sends a set of synthetic images to all participating discriminators.
  2. Local Detector: The generator maintains a local "detector"—a discriminator intentionally initialized with random weights to mimic free-rider behavior.
  3. Clustering: The generator collects classification responses from all participants and applies 2-means clustering. Benign discriminators, having been trained on real data, produce similar responses that differ significantly from the random responses of the detector and free-riders. The generator then identifies and ignores the cluster containing its own detector.

DFG+: Securing Peer Swapping

DFG+ extends this defense to MD-GAN variants where discriminators swap model weights to prevent overfitting. It enables benign discriminators to identify and reject swaps with suspected free-riders, preventing "model leakage" where attackers might steal well-trained models from peers.

Experimental Evaluation

The defenses were evaluated on MNIST, CIFAR-10, and CIFAR-100 datasets with 5 benign clients and varying numbers of free-riders (0 to 5).

Key Results

  • Detection Accuracy: DFG and DFG+ achieved 100% precision and recall across all CIFAR-10 and CIFAR-100 scenarios, perfectly identifying free-riders without penalizing benign participants.
  • Performance Restoration: On CIFAR-10, the defense improved FID scores by 5.22% to 11.53%. On the more complex CIFAR-100, improvements ranged from 5.79% to 13.22%.
  • Computational Efficiency: The 2-means clustering on a small probing set introduces negligible overhead compared to the primary training tasks.

Key Insights

The research highlights that free-riding is not merely a resource issue but a destructive force. In systems utilizing peer-swapping, free-riders can effectively "poison" the training pool, causing the generator to produce images that are heavily degraded by noise.

Example of degraded synthetic images resulting from free-rider interference in an MD-GAN.

The use of a local "detector" as an anchor is a particularly effective strategy. By creating a known "bad" reference point, the system avoids the need to define "good" behavior explicitly, instead identifying outliers relative to a known malicious baseline.

Conclusion and Future Directions

DFG and DFG+ provide a robust, efficient defense against free-riders in distributed GAN training. While highly effective against random-initialization attacks, future work should address adaptive adversaries who might attempt to mimic benign responses, as well as the potential for collusion among multiple free-riders. Testing on higher-resolution datasets like ImageNet will also be crucial for validating scalability.

Summary updated

Abstract

Generative Adversarial Networks (GANs) are increasingly adopted by the industry to synthesize realistic images. Due to data not being centrally available, Multi-Discriminator (MD)-GANs training framework employs multiple discriminators that have direct access to the real data. Distributedly training a joint GAN model entails the risk of free-riders, i.e., participants that aim to benefit from the common model while only pretending to participate in the training process. In this paper, we conduct the first characterization study of the impact of free-riders on MD-GAN. Based on two production prototypes of MD-GAN, we find that free-riders drastically reduce the ability of MD-GANs to produce images that are indistinguishable from real data, i.e., they increase the FID score -- the standard measure to assess the quality of generated images. To mitigate the model degradation, we propose a defense strategy against free-riders in MD-GAN, termed DFG. DFG distinguishes free-riders and benign participants through periodic probing and clustering of discriminators' responses based on a reference response of free-riders, which then allows the generator to exclude the detected free-riders from the training. Furthermore, we extend our defense, termed DFG+, to enable discriminators to filter out free-riders at the variant of MD-GAN that allows peer exchanges of discriminators networks. Extensive evaluation on various scenarios of free-riders, MD-GAN architecture, and three datasets show that our defenses effectively detect free-riders. With 1 to 5 free-riders, DFG and DFG+ averagely decreases FID by 5.22% to 11.53% for CIFAR10 and 5.79% to 13.22% for CIFAR100 in comparison to an attack without defense. In a shell, the proposed DFG(+) can effectively defend against free-riders without affecting benign clients at a negligible computation overhead.
BibTeX
@misc{zhao2022attacks,
  author = {Zilong Zhao and Jiyue Huang and Stefanie Roos and Lydia Y. Chen},
  title = {{Attacks and Defenses for Free-Riders in Multi-Discriminator GAN}},
  year = {2022},
  eprint = {2201.09967},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2201.09967},
}
Figure