Supplementing Real Data with Transferable Textures on Face Presentation Attack Detection

Face Presentation Attack Detection (PAD) is the branch of biometrics that detects spoofing in identity verification systems. Therefore, face PAD aims to stop attacks from bad agents trying to impersonate other people, or conceal their identity, for personal gain. One of the problems facing this technology is that current Deep-Learning based methods need thousands of images of real faces and attacks to train properly. However, obtaining that amount of images can be a very difficult and time consuming task. In this work, we aim to find out whether the Transferable-Textures technique can be used in face PAD to artificially increase the number of images in the dataset with similar performance to capturing new images. For this purpose, we trained and tested Siamese Networks on the wildly adopted CASIA-FASD dataset. Our results indicate that networks trained with Transferable Textures produced similar performance to those trained with captured images only. This demonstrates that this technique can be used to supplement manually captured attacks without any performance loss.