Race Bias Analysis of Bona Fide Errors in Face Anti-spoofing

The study of bias in Machine Learning is receiving a lot of attention in recent years, however, few only papers deal explicitly with the problem of race bias in face anti-spoofing. In this paper, we present a systematic study of race bias in face anti-spoofing with three key features: we focus on the classifier's bona fide errors, where the most significant ethical and legal issues lie; we analyse both the scalar responses of the classifier and its final binary outcomes; the threshold determining the operating point of the classifier is treated as a variable. We apply the proposed bias analysis framework on a VQ-VAE-based face anti-spoofing algorithm. Our main conclusion is that race bias should not necessarily be attributed to different mean values of the response distributions over the various demographics. Instead, it can be better understood as the combined effect of several possible characteristics of these distributions: different means; different variances; bimodal behaviour; the existence of outliers.