Learning Smooth and Fair Representations

Xavier Gitiaux, Huzefa Rangwala

Abstract

Organizations that own data face increasing legal liability for its discriminatory use against protected demographic groups, extending to contractual transactions involving third parties access and use of the data. This is problematic, since the original data owner cannot ex-ante anticipate all its future uses by downstream users. This paper explores the upstream ability to preemptively remove the correlations between features and sensitive attributes by mapping features to a fair representation space. Our main result shows that the fairness measured by the demographic parity of the representation distribution can be certified from a finite sample if and only if the chi-squared mutual information between features and representations is finite. Empirically, we find that smoothing the representation distribution provides generalization guarantees of fairness certificates, which improves upon existing fair representation learning approaches. Moreover, we do not observe that smoothing the representation distribution degrades the accuracy of downstream tasks compared to state-of-the-art methods in fair representation learning.

Introduction

In the modern machine learning lifecycle, a critical distinction exists between data controllers (who collect and own data) and data processors (who utilize that data to build models). Under regulations like GDPR, controllers are increasingly held legally responsible for ensuring their data is not used discriminatively, even when handed off to third parties.

The standard mitigation strategy is "fair representation learning": transforming raw data into a latent format that retains utility while remaining statistically independent of sensitive attributes. However, current methods often fail to generalize; they typically guarantee fairness only against a specific "adversary" or auditor seen during training. If a downstream processor employs a more complex model than the one used during the audit, they can often "leak" sensitive information. This paper addresses how to create representations that are "universally fair" for any future user.

The Smooth and Fair Framework

The authors introduce a "Smooth and Fair" framework that ensures fairness guarantees hold up under real-world conditions. The core innovation is the addition of an Additive White Gaussian Noise (AWGN) channel immediately following the encoder.

The proposed framework adding an AWGN channel after the encoder to ensure certifiable fairness for downstream processors.

Instead of releasing a raw representation $Z = t(X)$, the system releases $Z_\sigma = t(X) + \epsilon$, where $\epsilon$ is Gaussian noise. This smoothing step is theoretically grounded: the authors prove that the ability to "certify" fairness (specifically Demographic Parity) from a finite sample depends on the $\chi^2$ mutual information between the features and the representation. If this information is finite, the fairness certificate converges reliably at a rate of $O(n^{-1/2})$.

Unlike standard adversarial approaches that train a second neural network to "guess" sensitive attributes, this framework uses a plug-in auditor. Because the noise makes the distribution a mixture of Gaussians, the authors can use Monte Carlo integration to estimate density directly, making the training process more stable and theoretically sound.

Experimental Evaluation and Key Results

The researchers evaluated their approach on synthetic benchmarks (3D "Swiss Roll") and real-world datasets (UCI Adult, Health Heritage).

Comparison of representation spaces for the Swiss Roll dataset across different fairness constraints, demonstrating the smoothing effect of the AWGN method.

Key Insights

  • Complexity vs. Fairness: A major insight is that fairness isn't just about removing correlations; it's about the complexity of the mapping. If an encoder creates a highly complex, "sharp" mapping, a small sample might appear fair, but a powerful downstream model can still reconstruct sensitive data.
  • Noise as a Generalization Tool: By smoothing the distribution, the authors ensure that samples from different demographic groups are not "too far away" in the representation space. This prevents downstream models from easily distinguishing between them, as seen in the visual comparison of the AWGN method against standard adversarial techniques.
  • Honest Certificates: On the Swiss Roll dataset, the gap between the "certified" fairness and the actual unfairness achieved by a downstream processor was only 0.05 for the AWGN method, compared to 0.74 for standard adversarial cross-entropy (AdvCE). This means the AWGN certificate is a much more reliable indicator of true protection.
  • No "Fairness Tax": Analysis shows that the AWGN method achieves an accuracy-fairness trade-off comparable to state-of-the-art methods, proving that smoothing does not necessarily destroy the utility of the data for downstream tasks.

Conclusion

This work demonstrates that smoothing the representation distribution provides essential generalization guarantees for fairness. By controlling the $\chi^2$ mutual information, data controllers can provide "honest" certificates that protect against unknown future adversaries. Future research may extend these proofs to other metrics like Equalized Odds and explore the potential overlap with Differential Privacy benefits.

Summary updated

Abstract

Organizations that own data face increasing legal liability for its discriminatory use against protected demographic groups, extending to contractual transactions involving third parties access and use of the data. This is problematic, since the original data owner cannot ex-ante anticipate all its future uses by downstream users. This paper explores the upstream ability to preemptively remove the correlations between features and sensitive attributes by mapping features to a fair representation space. Our main result shows that the fairness measured by the demographic parity of the representation distribution can be certified from a finite sample if and only if the chi-squared mutual information between features and representations is finite. Empirically, we find that smoothing the representation distribution provides generalization guarantees of fairness certificates, which improves upon existing fair representation learning approaches. Moreover, we do not observe that smoothing the representation distribution degrades the accuracy of downstream tasks compared to state-of-the-art methods in fair representation learning.
BibTeX
@inproceedings{gitiaux2021learning,
  author = {Xavier Gitiaux and Huzefa Rangwala},
  title = {{Learning Smooth and Fair Representations}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2021},
  url = {https://openreview.net/forum?id=fV726shEdI},
}
Figure