Federated Measurement of Demographic Disparities from Quantile Sketches

A team of researchers has developed a breakthrough method for auditing AI fairness across organizations without compromising data privacy. The paper 'Federated Measurement of Demographic Disparities from Quantile Sketches' introduces a communication-efficient protocol that enables multiple organizations to collaboratively measure demographic disparities in their AI systems while keeping sensitive data siloed.

The technical approach centers on measuring disparity as a Wasserstein–Fréchet variance between sensitive-group score distributions. The researchers proved an ANOVA-style decomposition that separates selection-induced mixture effects from cross-silo heterogeneity, providing tight bounds linking local and global fairness metrics. Their one-shot protocol requires each organization to share only group counts and a quantile summary of local score distributions—reducing communication overhead while maintaining accuracy with O(1/k) discretization bias where k represents the number of quantiles.

In practical testing, the method demonstrated remarkable efficiency: experiments on synthetic data and the COMPAS recidivism dataset showed that just a few dozen quantiles suffice to recover global disparity metrics with high accuracy. This addresses a critical challenge in AI governance, where privacy regulations like GDPR often prevent organizations from sharing sensitive demographic data needed for comprehensive fairness audits. The protocol enables what was previously impossible—measuring population-level fairness across multiple data silos while respecting privacy constraints and regulatory requirements.