Time Analysis of Pulse-Based Face Anti-Spoofing in Visible and NIR

Abstract

In this paper we study Presentation Attack Detection (PAD) in face recognition systems against realistic artifacts such as 3D masks or good quality of photo attacks. In recent works, pulse detection based on remote photoplethysmography (rPPG) has shown to be a effective countermeasure in concrete setups, but still there is a need for a deeper understanding of when and how this kind of PAD works in various practical conditions. Related works analyze full video sequences (usually over 60 seconds) to distinguish between attacks and legitimate accesses. However, existing approaches may not be as effective as it has been claimed in the literature in time variable scenarios. In this paper we evaluate the performance of an existent state-of-the-art PAD scheme based on rPPG when analyzing short-time video sequences extracted from a longer video. Results are reported using the 3D Mask Attack Database (3DMAD), and a self-collected dataset called Heart Rate Database (HR), including different video durations, spectrum bands, resolutions and frame rates. Several conclusions can be drawn from this work: a) PAD performance based on rPPG varies significantly with the length of the analyzed video, b) rPPG information extracted from short-time sequences (over 5 seconds) can be discriminant enough for performing the PAD task, c) in general, videos using the NIR band perform better than those using the RGB band, and d) the temporal resolution is more valuable for rPPG signal extraction than the spatial resolution.