Generating Fair Universal Representations using Adversarial Models

We present a data-driven framework for learning fair universal representations (FUR) that guarantee statistical fairness any task may not be known priori. Our leverages recent advances in adversarial to allow data holder learn which set of sensitive attributes are decoupled from the rest dataset. formulate this as constrained minimax game between an encoder and adversary where constraint ensures measure usefulness (utility) representation. The resulting problem is censoring, i.e., finding representation least informative about given utility constraint. For appropriately chosen loss functions, our censoring precisely clarifies optimal strategy against strong information-theoretic adversaries; it also achieves demographic parity representations. evaluate performance proposed on both synthetic publicly available datasets. these datasets, we use two tradeoff measures: vs. fidelity downstream tasks, amply demonstrate multiple features can effectively censored even ensure accuracy tasks.