Uncovering demographic information on deep-face features
Rafael O. Ribeiro · Journal of Engineering Research · 2023
Facial recognition is one of the most successful applications of Deep Learning, with the advent of Convolution Neural Networks (CNN) being associated with break-through results in the last few years.The novel aspect of the use of CNNs has not changed the basic facial recognition pipeline, though: facial detection, pre-processing, feature extraction, and comparison/recognition.This paper investigates the possibility of inferring demographic information from facial features generated by CNNs trained for facial recognition.Pre-trained models on three different architectures (ArcFace, DeepFace and FaceNet) are used to extract features from faces of five distinct datasets: Fairface, UTKFaces, Labeled Faces in the Wild, Racial Faces in the Wild, and CelebA.Features and labels from the Fairface dataset are used to train neural networks classifiers for gender (female and male) and ethnicity (african, asian, caucasian and indian).Differences in the performance of the classifiers were observed, depending on which facial recognition model/ architecture is used as feature extractor.Some of the trained classifiers showed an improvement in performance compared to results in the literature and very low variance among demographic groups. PREVIOUS WORKSWe briefly review some works related to attribute analysis based on deep neural networks.In 2014, (ZHANG et al., 2014) hair style, clothes style, expression, action proposed a method that combines partbased models and deep learning by training pose-normalized CNNs to estimate various attributes, such as gender, hair style, clothes style, expression, and action, using the whole image as input.(LIU et al., 2015) introduced LNet+ANet in 2015, with a focus on facial attributes -it takes only facial images as input.This work also introduced two new datasets,