Automated face recognition algorithms generate encodings of face images that are compared to other encodings to compute a similarity score between the two originating face images. These face encodings, also known as feature vectors, contain representations of various facial
features. Some of these facial features, but not all, have been shown to resemble each other across different subjects that happen to share a demographic group assignment, such as having the same race or gender. Recent work has shown that these demographically dependent features can increase similarity scores between different individuals who belong to the same demographic group compared to similarity scores for different individuals in different groups. When one feature vector is compared to many other feature vectors, as in identifications, this effect, referred to as “demographic clustering”, can lead to un-equal false positive identification error rates for different demographic groups. In this study, we propose a method of mitigating this clustering effect from face recognition algorithms to reduce these un-equal error outcomes. Our method presumes that feature space patterns shared within demographic groups can be removed while preserving other distinct features of individuals. In this paper, we prove that this is possible, in principle, by applying linear dimensionality techniques to the feature space of two ArcFace face recognition algorithms. We show this method increases four distinct “fairness” measures while preserving useful true match rates.