Fair Learning with Private Demographic Data
This addresses the challenge of ensuring fairness in machine learning for real-world applications where demographic data is restricted, offering a practical solution for entities needing to comply with privacy regulations while mitigating bias.
The paper tackles the problem of learning non-discriminatory predictors when sensitive attributes like race are unavailable due to privacy laws, by proposing a scheme that allows private release of such data while maintaining fairness guarantees. It shows how to adapt existing learners to work with privatized attributes and provides theoretical performance assurances, with potential applications to partially available protected data.
Sensitive attributes such as race are rarely available to learners in real world settings as their collection is often restricted by laws and regulations. We give a scheme that allows individuals to release their sensitive information privately while still allowing any downstream entity to learn non-discriminatory predictors. We show how to adapt non-discriminatory learners to work with privatized protected attributes giving theoretical guarantees on performance. Finally, we highlight how the methodology could apply to learning fair predictors in settings where protected attributes are only available for a subset of the data.