Mfad: A Multi-Modality Face Anti-Spoofing Dataset

Face anti-spoofing plays an important role in face recognition system to prevent security vulnerability. It acts as an important step to select the face image to the face recognition system. Previous works have provided many databases for face anti-spoofing, but they either contain too few subjects or contain a single modal. Moreover, due to the limited information of RGB images, the effect of face anti-spoofing is difficult to be further improved, High Dynamic Range (HDR) can be a good choice while many devices now support capturing HDR images. To facilitate further studies on face anti-spoofing, we in this work build a dataset with 50 subjects for face anti-spoofing containing 2100 videos, which containing two different modalities, HDR and Near Infrared Ray (NIR). We further study the influences of different modalities on the performance of face anti-spoofing by extensively experimenting with different combinations of modalities with an end-to-end deep learning model, and find that the HDR information contributes most among different modalities. Finally, we verify these findings on the CASIA-FASD dataset which demonstrates better performance of the proposed model.