A novel approach for bias mitigation of gender classification algorithms using consistency regularization
Abstract
Published research has confirmed the bias of automated face-based gender classification algorithms across gender-racial groups. Specifically, unequal accuracy rates were obtained for women and dark-skinned people for face-based automated gender classification algorithms. To mitigate the bias of gender classification and other facial-analysis-based algorithms in general, the vision community has proposed several techniques. However, most of the existing bias mitigation techniques suffer from a lack of generalizability, need a demographically-annotated training set, are application-specific, and often offer a trade-off between fairness and classification accuracy. This means that fairness is often obtained at the cost of a reduction in the classification accuracy of the best-performing demographic sub-group. In this paper, we propose a novel bias mitigation technique that leverages the power of semantic preserving augmentations at the image- and feature-level in a self-consistency setting for the downstream gender classification task. Thorough experimental validation on gender-annotated facial image datasets confirms the efficacy of our bias mitigation technique in improving overall gender classification accuracy as well as reducing bias across all gender-racial groups over state-of-the-art bias mitigation techniques. Specifically, our proposed technique obtained a reduction in the bias by an average of 30% over existing bias mitigation techniques as well as an improvement in the overall classification accuracy of about 5% over the baseline gender classifier. Therefore, resulting in state-of-the-art generalization performance in the intra- and cross-dataset evaluations. Additionally, our proposed technique operates in the absence of demographic labels and is application agnostic, compared to most of the existing bias mitigation techniques.