On the duality of privacy and fairness (extended abstract)

When a machine learning model operates over data about individuals, there are two common concerns. On one hand, if the model’s output (i.e., its prediction) allows for information inferences about an individual’s sensitive attributes, we have a privacy issue. On the other hand, if the individual’s sensitive attributes can unduly influence the model’s output, we have a fairness issue. Recently, the interplay between these two concerns has gathered growing attention both in the scientific community and in society as a whole. In this work, we extend the framework of quantitative information flow to formally capture fairness and privacy as duals of each other, and give first steps toward a novel characterization of their relationship.