Gender shades: intersectional accuracy disparities in commercial gender classi cation

Abstract

Recent studies demonstrate that machine learning algorithms can discriminate based on classes like race and gender. In this work, we present an approach to evaluate bias present in automated facial analysis algorithms and datasets with respect to phenotypic subgroups. Using the dermatologist approved Fitzpatrick Skin Type classi cation system, we characterize the gender and skin type distribution of two facial analysis benchmarks, IJB-A and Adience. We ndthat these datasets are overwhelmingly composed of lighter-skinned subjects (796% for IJB-A and 862% for Adience) and introduce a new facial analysis dataset which is balanced by gender and skin type. We evaluate 3 commercial gender classi cation systems using our dataset and show that darker-skinned females are the most misclassi ed group (with error rates of up to 347%). The maximum error rate for lighter-skinned males is 08%. The substantial disparities in the accuracy of classifying darker females, lighter females, darker males, and lighter males in gender classi cation systems require urgent attention if commercial companies are to build genuinely fair, transparent and accountable facial analysis algorithms.