Productionized Fairness Measurement Under Privacy Constraints

Explainable & Ethical AI

Published: arXiv: 2606.27558v1

Authors

Osonde A. Osoba Yuzi He Saikrishna Badrinarayanan Varun Mithal Sakshi Jain Natesh S. Pillai

Abstract

Fairness measurements in the form of disaggregated evaluations often rely on demographic signals that are legally constrained or culturally sensitive. Race and ethnicity signals are among the more difficult signals to curate and use for this task. This paper presents Privacy-Preserving Probabilistic Race/Ethnicity Estimation (PPRE) as a method for enabling fairness measurements with respect to race/ethnicity for U.S.\ LinkedIn members in a privacy-preserving manner. PPRE applies privacy technologies (specifically: secure two-party computation, differential privacy, and additive homomorphic encryption) on top of two race/ethnicity demographic signal sources (the Bayesian Improved Surname Geocoding estimator and a sparse golden survey set of self-reported demographics) to power a fairness measurement solution with respect to US-based race/ethnicity demographics. We detail its privacy guarantees and demonstrate its application on candidate- and viewer-side fairness measurements. We close with a transferable framework for institutions seeking to implement similar privacy-preserving measurement infrastructure.

Paper Summary

Problem

The main problem this paper addresses is how to measure fairness in AI systems, particularly when it comes to sensitive demographic data such as race and ethnicity. The challenge is that collecting and using this data can be legally sensitive, culturally charged, and often simply unavailable. As a result, institutions committed to monitoring AI systems for fair treatment across protected demographic groups cannot fulfill that commitment without first solving a demographic data problem.

Key Innovation

This paper introduces a new system called PPRE (Privacy-Preserving Race/Ethnicity Fairness Measurement) that allows for fair and accurate measurement of AI system performance across demographic groups without compromising user privacy. PPRE uses a combination of cryptographic primitives and a custom protocol to enable secure and efficient computation of fairness metrics, even when the sensitive demographic data is not directly available.

Practical Impact

The practical impact of this research is significant. By enabling institutions to measure and address fairness in AI systems without compromising user privacy, PPRE can help to promote fairness and equity in decision-making systems. This can have far-reaching implications for various domains, including hiring, lending, and education. Moreover, PPRE can be applied to a wide range of AI systems, from recommendation engines to decision-making algorithms.

Analogy / Intuitive Explanation

Imagine you want to measure the fairness of a restaurant's seating policy. You want to know if the restaurant is seating people of different ethnicities equally or if there are biases in the seating process. However, the restaurant doesn't collect data on ethnicity, and even if they did, it would be sensitive and potentially discriminatory. PPRE is like a magic wand that allows you to measure fairness in this scenario without actually knowing the ethnicity of the customers. It does this by using a combination of cryptographic techniques and a custom protocol to compute fairness metrics in a way that preserves user privacy.

Paper Information

Categories:

cs.LG cs.CR

Published Date:

arXiv ID:

2606.27558v1

Quick Actions

Back to Home