AutoFairMask: Enhancing Fairness in Machine Learning with Automated Protected Attributes Detection

Artificial Intelligence (AI) and machine learning (ML) are increasingly integral to decision-making processes in various aspects of our lives. However, their predictions could be biased against protected social groups. Existing bias mitigation techniques often require the manual selection of protected attributes, limiting their generalizability. In this paper, we present AutoFairMask, a model-based extrapolation method for the bias mitigation with the automatic protected attributes detection, comparing three approaches: (1) a base model with no fairness interventions, (2) a base FairMask model with a single manually selected protected attribute, and (3) AutoFairMask, an enhanced FairMask model with the automatic protected attribute detection. Our experimental results show that FairMask with the autodetection performs better at minimizing biases while preserving the predictive performance than both the baseline and manually defined FairMask. Notably, the autodetection approach uncovers additional sources of biases (e.g., police district) that are missed by the manual selection. This emphasizes how crucial adaptive bias reduction strategies are to the development of fairness-aware AI. Our findings suggest that automated identification multi-protected attributes are able to maintain the fairness and accuracy in real-world ML applications.