Balancing Biases and Preserving Privacy on Balanced Faces in the Wild

Demographic biases exist in current models used for facial recognition (FR). Our Balanced Faces in the Wild (BFW) dataset is a proxy to measure bias across ethnicity and gender subgroups, allowing one to characterize FR performances per subgroup. We show that results are non-optimal when a single score threshold determines whether sample pairs are genuine or imposters. Furthermore, within subgroups, performance often varies significantly from the global average. Thus, specific error rates only hold for populations matching the validation data. We mitigate the imbalanced performances using a novel domain adaptation learning scheme on the facial features extracted from state-of-the-art neural networks, boosting the average performance. The proposed method also preserves identity information while removing demographic knowledge. The removal of demographic knowledge prevents potential biases from being injected into decision-making and protects privacy since demographic information is no longer available. We explore the proposed method and show that subgroup classifiers can no longer learn from the features projected using our domain adaptation scheme. For source code and data, see https://github.com/visionjo/facerec-bias-bfw.