LinkedIn Develops Privacy-Preserving Race/Ethnicity Fairness Measurement

LinkedIn has introduced a new methodology called Privacy-Preserving Probabilistic Race/Ethnicity Estimation (PPRE) to address the challenge of measuring fairness in AI systems while adhering to strict privacy regulations. This method specifically targets race and ethnicity signals for U.S. LinkedIn members, which are often legally constrained or culturally sensitive, making direct collection and use difficult for fairness assessments. PPRE integrates advanced privacy technologies, including secure two-party computation, differential privacy, and additive homomorphic encryption. It combines data from two sources: the Bayesian Improved Surname Geocoding (BISG) estimator and a sparse set of self-reported demographics. By applying these privacy-enhancing techniques, PPRE enables the measurement of fairness with respect to U.S.-based race/ethnicity demographics without compromising individual user privacy. The paper details the privacy guarantees of PPRE and demonstrates its practical application in measuring fairness for both candidate-side and viewer-side scenarios on the LinkedIn platform. It concludes by offering a transferable framework, allowing other institutions to implement similar privacy-preserving measurement infrastructures for their own AI fairness initiatives.