Measuring the gender and ethnicity bias in deep models for face recognition

Abstract

We explore the importance of gender and ethnic attributes in the decision-making of face recognition technologies. Our work is in part motivated by the new European regulation for personal data protection, which forces data controllers to avoid discriminative hazards while managing sensitive data like biometric data. The experiments in this paper are aimed to study what extent sensitive data like gender or ethnic origin attributes are present in the most common face recognition networks. For this, our experiments include two popular pre-trained networks: VGGFace and Resnet50. Both pre-trained models are able to classify gender and ethnicity easily (over 95% of performance) even suppressing 80% of the neurons in their embedding layers. The experimentation is conducted on a publicly available database known as Labeled Faces in the Wild with more than 13000 images of faces with a huge range of poses, ages, races and nationalities.