SAGAN: Maximizing Fairness using Semantic Attention Based Generative Adversarial Network

In this work, we present an end-to-end framework with a predictor model that provides classifier outputs from given input features and an adversary that tries to predict protected or sensitive features in order to mitigate intrinsic biases with respect to sensitive features (e.g., race, sex). Our proposed model increases the predictor’s capacity to produce correct predictions, while decreasing the adversary’s ability to anticipate sensitive features. We include a novel Semantic Attention (SA) module to the framework and demonstrate that our SA based Generative Adversarial Network (SAGAN) is able to significantly minimize the adversary’s capability to predict sensitive features, while retaining the predictor’s predictive accuracy. UCI Adult (Census) dataset was used a benchmark dataset for the testing. Our results demonstrate that the predictive model does not lose much accuracy, while achieving a Disparate Impact (DI) score very close to 1. The flexibility of the method makes it fitting to be applicable to a broad spectrum of gradient-based learning models, including both regression and classification tasks as well as different definitions of fairness. The source code for the implementation is available on github [1].