FR-GAN: Fair and Robust Training

We consider the problem of fair and robust model training in the presence of data poisoning. Ensuring fairness usually involves a tradeoff against accuracy, so if the data poisoning is mistakenly viewed as additional bias to be fixed, the accuracy will be sacrificed even more. We demonstrate that this phenomenon indeed holds for state-of-the-art model fairness techniques. We then propose FR-GAN, which holistically performs fair and robust model training using generative adversarial networks (GANs). We first use a generator that attempts to classify examples as accurately as possible. In addition, we deploy two discriminators: (1) a fairness discriminator that predicts the sensitive attribute from classification results and (2) a robustness discriminator that distinguishes examples and predictions from a clean validation set. Our framework respects all the prominent fairness measures: disparate impact, equalized odds, and equal opportunity. Also, FR-GAN optimizes fairness without requiring the knowledge of prior statistics of the sensitive attributes. In our experiments, FR-GAN shows almost no decrease in fairness and accuracy in the presence of data poisoning unlike other state-of-the-art fairness methods, which are vulnerable. In addition, FR-GAN can be adjusted using parameters to maintain reasonable accuracy and fairness even if the validation set is too small or unavailable.