Continuous Presentation Attack Detection in Face Biometrics Based on Heart Rate

In this paper we study face Presentation Attack Detection (PAD) against realistic 3D mask and high quality photo attacks in dynamic scenarios. We perform a comparison between a new pulse-based PAD approach based on a combination of a skin detector and a chrominance method, and the system used in our previous works (based on Blind Source Separation techniques, BSS). We also propose and study heuristical and statistical approaches for performing continuous PAD with low latency and false non-match rate. Results are reported using the 3D Mask Attack Database (3DMAD), and a self-collected dataset called BiDA Heart Rate Database (BiDA HR) including different video durations, resolutions, frame rates and attack artifacts. Several conclusions can be drawn from this work: (1) chrominance and BSS methods perform similarly under the controlled and favorable conditions found in 3DMAD and BiDA HR, (2) combining pulse information extracted from short-time sequences (e.g. 3 s) can be discriminant enough for performing the PAD task, (3) a high increase in PAD performance can be achieved with simple PAD score combination, and (4) the statistical method for continuous PAD outperforms the simple PAD score combination but it needs more data for building the statistical models.