Towards Generalized and Distributed Privacy-Preserving Representation Learning

Sheikh Shams Azam, Taejin Kim, Seyyedali Hosseinalipour, Christopher G. Brinton, Carlee Joe-Wong, Saurabh Bagchi

Introduction

Privacy-Preserving Representation Learning (PRL) has emerged as a critical field for balancing the utility of data sharing with the necessity of protecting sensitive information. This paper introduces EIGAN (Exclusion-Inclusion Generative Adversarial Network) and its distributed counterpart, D-EIGAN, representing the first frameworks capable of handling multiple sensitive attributes across siloed, distributed datasets. By reformulating PRL as a minimax game optimized via cross-entropy loss, the authors demonstrate significant improvements in both utility and privacy robustness.

The Challenge of Real-World Privacy

Modern data ecosystems face two primary hurdles in privacy preservation. First is the Complexity Gap: while most existing PRL methods focus on a single ally task and one sensitive attribute, real-world scenarios like healthcare often involve multiple concurrent objectives and various sensitive features (e.g., race, gender, age). Second is the Distribution Gap: data is frequently siloed across institutions that cannot share raw information due to regulatory constraints. Existing centralized PRL techniques are often inapplicable in these federated environments.

Methodology: EIGAN and D-EIGAN

The proposed EIGAN architecture utilizes a minimax optimization framework. An Encoder generates a data representation that a set of Ally networks uses to predict intended targets, while a set of Adversary networks attempts to recover sensitive attributes from the same representation.

The EIGAN architecture (a) and its distributed extension D-EIGAN (b).

Unlike previous approaches relying on KL-divergence, EIGAN employs Cross-Entropy (CE) loss. The authors argue that CE provides superior stability and more effectively pushes adversaries toward a uniform "random guess" distribution, particularly when data distributions are disjoint.

To address siloed data, D-EIGAN extends this framework to a federated learning setting. Local nodes train individual encoders and periodically synchronize parameters with a central server. To optimize communication, the framework employs sparsification—sharing only a fraction $\phi$ of parameters—and adjustable synchronization frequencies ($\delta$).

Key Results and Insights

The framework was rigorously evaluated across diverse datasets, including MNIST, MIMIC-III, Adult, and FaceScrub.

Utility and Privacy Performance

EIGAN achieves substantial performance gains. On the FaceScrub dataset, it improved ally accuracy by 47% compared to the previous state-of-the-art (Bertran-PRL) while maintaining equivalent privacy levels. On the Adult census dataset, EIGAN matched the performance of optimal closed-form linear solutions while offering the flexibility of non-linear neural architectures.

Ally log loss over training epochs, showing EIGAN's superior utility performance (lower is better).

The Sparsification Paradox

A notable finding in the distributed setting is the "Sparsification Paradox." In non-i.i.d. scenarios where data distributions vary significantly between nodes, sharing fewer parameters (high sparsification) can actually improve global model performance by reducing the propagation of local data biases.

Architecture Agnosticism

The study also confirms that representations trained against one type of adversary (e.g., ResNet) remain private even when attacked by different architectures (e.g., Vision Transformers). This suggests that the privacy guarantees are fundamentally embedded within the learned representation itself.

Conclusion

EIGAN and D-EIGAN provide a robust, scalable solution for generalized privacy-preserving representation learning. By successfully navigating the trade-offs between utility and privacy in both centralized and distributed environments, this work sets a new benchmark for practical PRL applications. Future research will explore more complex non-i.i.d. distributions and the behavior of these representations in continual learning settings.

Summary updated

Abstract

We study the problem of learning representations that are private yet informative, i.e., provide information about intended "ally" targets while hiding sensitive "adversary" attributes. We propose Exclusion-Inclusion Generative Adversarial Network (EIGAN), a generalized private representation learning (PRL) architecture that accounts for multiple ally and adversary attributes unlike existing PRL solutions. While centrally-aggregated dataset is a prerequisite for most PRL techniques, data in real-world is often siloed across multiple distributed nodes unwilling to share the raw data because of privacy concerns. We address this practical constraint by developing D-EIGAN, the first distributed PRL method that learns representations at each node without transmitting the source data. We theoretically analyze the behavior of adversaries under the optimal EIGAN and D-EIGAN encoders and the impact of dependencies among ally and adversary tasks on the optimization objective. Our experiments on various datasets demonstrate the advantages of EIGAN in terms of performance, robustness, and scalability. In particular, EIGAN outperforms the previous state-of-the-art by a significant accuracy margin (47% improvement), and D-EIGAN's performance is consistently on par with EIGAN under different network settings.
BibTeX
@inproceedings{azam2022towards,
  author = {Sheikh Shams Azam and Taejin Kim and Seyyedali Hosseinalipour and Christopher G. Brinton and Carlee Joe-Wong and Saurabh Bagchi},
  title = {{Towards Generalized and Distributed Privacy-Preserving Representation Learning}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2022},
  url = {https://openreview.net/forum?id=mMeWqc1DyX},
}
Figure