Performance evaluation of face encoding techniques: a case study on the Asian population

Abstract

The purpose of this study was to compare the performance of facial recognition systems in terms of recognition performance, the bias of gender on a face recognition system, and the effectiveness of facial recognition attacks using dictionary attack methods on the Asian population datasets. Typically, the face recognition system consists of three main components: the face detection module, the face embedding module, and the face matching module. The process starts by detecting the face region from the face image. Then, the face region image is converted to a vector (embedding) representation. Lastly, the distance between a test image and a template (an enrolled face image) is computed by calculating the vector similarity and the decision to accept or reject the test image depends on the computed similarity score. The biometrics system performance is then evaluated based on False Acceptance Rate and False Rejection Rate. The models used in this research were 3 pre-trained models: ResNet50, SeNet50, and FaceNet. The systems were evaluated based on an Asian face database comprising 1819 images of 107 individuals. The result indicated no bias in system performance when tested against a facial image with different gender attributes. The model with the best recognition performance was SeNet50. Lastly, using a dictionary attack, researchers examined the attack performance of facial recognition systems, it found that the attack has a high success rate of 22.94%, 21.5%, and 22.72% when 5 photos were used on ResNet50, SeNet50, and FaceNet models, respectively.