Unknown Presentation Attack Detection with Face RGB Images

Abstract

The vulnerability of biometric identification systems to presentation attacks, also a.k.a. spoof attacks, has received great attention from the biometrics community. Face presentation attacks are particularly easy to fabricate, because face image and videos are easy to obtain from social media, and most face recognition systems rely on 2D RGB image sensors only. To our best knowledge, all the existing face Presentation Attack Detection (PAD) research works focus on solving the problem in a close-set setup, even though it is an open-set problem in real life where Unknown Presentation Attack Detection (UPAD) algorithms are demanded. In this paper, One-Class SVM and AutoEncoder based outliers detection algorithms, are proposed to solve the UPAD problem. It aims to detect the attack presentations from different presentation attack instrument species that are not included in the training and validation set. Both algorithms, together with the state of the art Local Binary Pattern feature are evaluated and compared on four largest face PAD datasets, including the latest Oulu dataset - the largest face PAD dataset so far. Moreover, four UPAD protocols are proposed to evaluate the robustness of the algorithms against environment and device variations. These comprehensive evaluations show that the proposed methods are more accurate on detecting unknown presentation attacks than the state of the art PAD algorithms under complex environment and capturing device variations.