The Influence of Gender and Skin Colour on the Watchlist Imbalance Effect in Facial Identification Scenarios

Abstract

Nowadays it is well known that artificial intelligence can be biased. In biometric recognition, this is a very sensitive topic since biased algorithms often discriminate against specific demographic groups. This can have severe consequences when searching criminal databases or blacklists. In this context, the watchlist imbalance effect might induce additional performance differentials based on the demographic composition of the target database. In this work, we utilise a fairly distributed subset of the FairFace database to evaluate the watchlist imbalance effect when combining the demographic attributes gender and skin colour. The results show that the skin colour has a huge impact on the differential performance to the disadvantage of dark skin tones.