Deep Learning Model for Face Anti-Spoofing in Overcoming Domain Generalisation with Depth Estimation and Generative Adversarial Network

Abstract

The use of facial biometrics to gain access to a security system is common in communication/computing devices. However, this convenience comes with a vulnerability to security breaches, where facial images can be falsified using photos or videos of someone with access rights. The availability of photos or videos of individuals on social media can exacerbate this. A face anti-spoofing system (FAS) is a crucial component for determining whether an input image is genuine or synthetic in biometric systems that utilize facial image information. Many methods have been used to realise this system, both with a hand-crafted method-based approach and deep learning (DL). However, research on the distribution differences between the test dataset and the training dataset is still rare. This article discusses the use of deep learning (DL)-based models for face anti-spoofing (FAS) applications. This study employs a model that utilizes depth map estimation to identify discriminative features and a generative adversarial network (GAN) to address the issue of distribution differences through a data generation approach. For models implemented with an intra-set simulation scenario, test results for two public datasets, NUAA and CASIA, provided the best results in terms of half total error rate (HTER) metrics, at 2.97% and 2.7% respectively. Meanwhile, simulations comparing the characteristics of the test dataset and the training dataset revealed that applying GAN to enhance the model's generalization ability could reduce the bona fide presentation classification error rate (BPCER) by 9.75%.