Transformation Tolerance and Demographic Robustness of Machine-based Face Recognition Systems

Abstract

Face recognition is widely acknowledged to be a very complex visual task for both humans and computers. Previous studies which analyze robustness of facial recognition systems have revealed that the ability to recognize faces becomes worse as the blur levels of face images increases, and that naturalistic color is important for facial recognition at high blur levels. Additionally, previous studies of current state of the art face recognition technologies have found bias in face recognition amongst different races, resulting in a worse recognition performance for people of color. In this study, we evaluate the performance and robustness of a current state-of-the-art facial recognition neural network architecture (ResNet-101) trained on an augmented facial identity dataset (Augmented Casia Webface) and perform a thorough comparison between White, Black and East Asian identities. We created a full-color, a grayscale and many hue-shifted datasets and then Gaussian blurred each dataset at different intensities and compared how AI systems perform relative to humans and amongst the different races.