PROJECT 4

Fair Federated Learning With Private Access to Sensitive Features

Project Leader: Meisam Razaviyayn

Website: https://sites.usc.edu/razaviyayn/

Abstract: As we rely more heavily on artificial intelligence to support human decisions, we are prone to serious societal risks including unfair/discriminatory outcomes and violating individuals’ privacy. To limit such adverse impacts, as a community, we should ensure that machine learning systems comply with specific privacy and fairness standards. To this end, this project studies the fundamental computational and statistical limits and develops algorithms for two questions related to responsible machine learning: 1) Fair machine learning without direct access to sensitive (protected) features: This thrusts studies fundamental limits of learning and develops algorithms for the scenario that sensitive attributes (such as race or gender) are not directly available to the learner; 2) Thrust 2: Fair federated learning with private access to sensitive features. This thrust develops fair federated learning algorithms among multiple parties when only a few of them have access to (private) sensitive attributes.