Unraveling robustness of deep face anti-spoofing models against pixel attacks

Abstract

In the last few decades, deep-learning-based face verification and recognition systems have had enormous success in solving complex security problems. However, it has been recently shown that such efficient frameworks are vulnerable to face-spoofing attacks, which has led researchers to build proficient anti-facial-spoofing (or liveness detection) models as an additional security layer. In response, increasingly challenging and tricky attacks have been launched to fool these anti-spoofing mechanisms. In this context, this paper presents the results of an analytical study on transfer-learning-based convolutional neural networks (CNNs) for face liveness detection and differential evolution-based adversarial attacks to evaluate the efficiency of face anti-spoofing classifiers against adversarial attacks. Specifically, experiments were conducted under different use-case scenarios on four face anti-spoofing databases to highlight practical criteria that can be used in the development of countermeasures to address face-spoofing issues.