LinkedIn's PPRE: Privacy-preserving fairness measurement for race/ethnicity

LinkedIn researchers have introduced a novel approach to fairness measurement that respects strict privacy constraints. Their paper, "Productionized Fairness Measurement Under Privacy Constraints," details the Privacy-Preserving Probabilistic Race/Ethnicity Estimation (PPRE) method, which allows for disaggregated evaluations of AI systems by race and ethnicity without directly accessing or exposing sensitive demographic data. PPRE layers multiple privacy technologies—secure two-party computation, differential privacy, and additive homomorphic encryption—on top of two demographic signal sources: the Bayesian Improved Surname Geocoding (BISG) estimator and a sparse golden survey set of self-reported demographics. This enables LinkedIn to conduct fairne ss audits for its U.S. members while legally and ethically safeguarding personally identifiable information.

The system is designed to power both candidate-side and viewer-side fairness measurements, meaning it can evaluate whether AI models perform equitably for different demographic groups either as subjects of the model or as users interacting with it. The researchers detail the formal privacy guarantees and demonstrate real-world application on LinkedIn's platform. They also provide a transferable framework for other institutions aiming to implement similar privacy-preserving measurement infrastructure. This work addresses the inherent tension between the need for transparency in AI fairness and the legal/cultural constraints around collecting and using race/ethnicity data, offering a practical blueprint for responsible AI governance at scale.