PrivFair

PrivFair is a library for auditing the fairness of machine learning (ML) models while protecting the privacy of both the model and the data used in the audit. ML models used in healthcare, justice and finance have been found to discriminate based on sensitive attributes such as gender, race or disability. Checking a model for bias requires data about these sensitive characteristics, which is protected by anti-discrimination and data protection law. Existing fairness auditing libraries offer no way to protect the privacy of this audit data. PrivFair addresses this gap. It was published in February 2022.

PrivFair uses Secure Multiparty Computation (MPC), a cryptographic technique that lets several parties jointly compute a result without revealing their own inputs to each other. This supports a common real-world scenario: a company owns a proprietary AI model, and an external investigator, such as a regulator or researcher, holds sensitive data needed to audit it.

With PrivFair, the investigator does not have to disclose their data to anyone in unencrypted form. The company also does not have to reveal its model parameters to anyone in plain text. Both parties keep their information confidential, while the audit still produces a reliable assessment of the model's fairness. The authors demonstrate PrivFair for group fairness audits, which compare model outcomes across demographic groups, using both tabular data and image data.

Auto-discovered on 2026-02-11 by OECD Catalogue Automation