Can we Generalize and Distribute Private Representation Learning?

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

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.

Introduction

In the era of ubiquitous data collection, the tension between data utility and individual privacy has reached a critical point. Organizations often need to share data for beneficial "ally" tasks (e.g., medical research or service personalization) while ensuring that sensitive "adversary" attributes (e.g., race, gender, or health status) remain protected. Private Representation Learning (PRL) aims to solve this by transforming raw data into a representation that preserves useful information for allies while obfuscating it for adversaries.

However, existing PRL solutions face two significant hurdles: a lack of generality in handling multiple attributes and a reliance on centralized data pooling. Most current models are designed for a single target and a single sensitive attribute, failing to capture the complexity of real-world datasets. Furthermore, legal frameworks like GDPR and HIPAA often prevent the centralization of raw data, creating "data silos."

Proposed Methods: EIGAN and D-EIGAN

To address these limitations, we introduce two novel architectures designed for generalized and distributed privacy preservation.

EIGAN: Centralized Multi-Attribute PRL

The Exclusion-Inclusion Generative Adversarial Network (EIGAN) utilizes a minimax game to optimize the data representation. Unlike previous models, EIGAN's encoder is trained to simultaneously satisfy multiple ally networks while fooling multiple adversary networks. A key technical innovation is the use of Cross-Entropy (CE) loss rather than KL divergence. CE loss provides greater training stability and effectively pushes adversary predictions toward a uniform distribution, essentially forcing the adversary into pure guesswork.

D-EIGAN: Distributed Privacy Preservation

D-EIGAN extends this framework to a federated learning environment. In this setup, multiple nodes train local encoders on their respective siloed datasets. A central parameter server aggregates these weights to form a global privacy-preserving transformation. To optimize communication efficiency, we employ sparsification, where nodes only transmit a random fraction of their model parameters.

The EIGAN (a) and D-EIGAN (b) architectures. EIGAN handles multiple ally and adversary networks, while D-EIGAN enables distributed training via a parameter server.

Experimental Evaluation

We evaluated EIGAN and D-EIGAN across diverse benchmarks, including MNIST, the MIMIC-III clinical dataset, the Adult census dataset, and FaceScrub celebrity images.

Key Insights

  • The Uniformity Goal: Theoretically, the optimal encoder forces the adversary to output a uniform distribution. For a binary sensitive attribute like gender, EIGAN aims for a 50/50 prediction regardless of the input.
  • Handling Task Overlap: When ally and adversary tasks are correlated (e.g., predicting height while hiding gender), the model balances these competing interests using importance weights ($\alpha$).
  • The Sparsification Paradox: In distributed settings with heterogeneous (non-i.i.d.) data, we observed that sending fewer parameters can actually improve performance by acting as a regularizer, preventing the global model from over-fitting to local biases.

Results and Discussion

EIGAN demonstrates significant improvements over existing state-of-the-art methods. On the FaceScrub dataset, EIGAN achieved a 47% improvement in ally accuracy while maintaining robust privacy.

Comparison of EIGAN against baselines (Autoencoder, PCA, DP). (a) Adversary log loss (higher is better for privacy) and (b) Ally log loss (lower is better for utility).

As shown in the performance plots, EIGAN consistently maintains high adversary loss (indicating strong privacy) while achieving the lowest ally loss (indicating high utility) compared to standard techniques like Differential Privacy (DP) or PCA. Furthermore, D-EIGAN's performance remains on par with the centralized EIGAN, proving that high-quality private representations can be learned without pooling raw data.

Conclusion

EIGAN and D-EIGAN provide a robust, scalable framework for private representation learning. By supporting multiple attributes and distributed data sources, these models bridge the gap between theoretical privacy research and practical real-world constraints. Future work will investigate more complex adversary scenarios, such as collusion between multiple attackers attempting to reconstruct sensitive information from the learned representations.

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{azam2022can,
  author = {Sheikh Shams Azam and Taejin Kim and Seyyedali Hosseinalipour and Carlee Joe‐Wong and Saurabh Bagchi and Christopher G. Brinton},
  title = {{Can we Generalize and Distribute Private Representation Learning?}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2022},
  url = {https://openreview.net/forum?id=mMeWqc1DyX},
}
Figure