Facial Analysis-Driven Regression Model for Age Estimation Using Deep Learning

Abstract

The increasing speed of facial recognition systems, as they transition to artificial intelligence techniques, creates new possibilities and risks related to security in biometric and virtual identity systems. This paper presented a unified multi-task deep learning framework for continuous age estimation, gender classification, and detection of images created by AI using facial images. Traditional methods treat each of these tasks independently; the proposed architecture combines their regression (age) and classification (gender and authenticity) functions in a single model that learns shared representations of data by way of a conjoined set of objectives. A convolutional backbone extracts a sequence of facial features using several layers and optimizes them simultaneously via age prediction (regression) and gender prediction or authenticity classification (binary) across different branches on each layer. A combination of loss functions enables coordinated multi-task training, as well as increased separation between learned features for all three tasks. Experiment results show very little error (mean absolute) in estimating ages, high levels of accuracy in identifying gender, and good performance when identifying faces generated by AI. With an efficient and scalable method to perform authentication, demographic research, and digital forensics in current and future biometric systems, this study proposes an improved framework for conducting biometric system assessments.