3D mask presentation attack detection via high resolution face parts
Oleg Grinchuk, Aleksandr Parkin, Evgenija Glazistova · 2021
3D mask presentation attack detection (PAD) is a long standing challenge in face anti-spoofing due to the high fidelity of attack artifacts and a limited number of samples available for training and evaluation. With the recent release of the large-scale and diverse CASIA-SURF HiFiMask dataset [19], it now becomes possible to address 3D mask PAD with deep neural networks. This paper introduces a new one-shot method for 3D mask PAD that extracts fine-grained information from appropriate parts of the human face and uses it to identify subtle differences between real and fake samples. The proposed method achieves state-of-the-art results of 3% ACER on the CASIA-SURF HiFiMask test set.