Robustness of Fairness in Machine Learning

Abstract

As machine learning algorithms become widely used in society, certain subgroups are more at risk of being harmed by unfair treatment. Fairness metrics have been proposed to quantify this harm by measuring certain statistics with respect to an evaluation dataset. In this work, we seek to analyze how robust these metrics are. That is, we are interested in whether these metrics give the same ``fairness score'' when measured on different sets of samples from the same distribution. This is important because it gives us insight into how much we can trust the conclusions given by a fairness metric prior to deployment of a model. We design a framework to conduct experiments to test the robustness of a popular fairness metric. We find that, when compared to more traditional performance metrics, it is more sensitive to fluctuations in the evaluation dataset in a variety of settings. Additionally, our work provides a foundation for studying the robustness of fairness metrics in general.