Large Crowdcollected Facial Anti-Spoofing Dataset

Abstract

The study about the vulnerabilities of biometric systems against spoofing has been a very active field of research in recent years. In this particular research we are focusing on one of the most difficult types of attack — video replay. We have noticed that currently most of face replay anti-spoofing databases focus on data with little variations of the devices used for replay and record. This fact may limit the generalization performance of trained models since potential attacks in the real world are probably more complex. In this review we present a face anti-spoofing database, which covers a huge range of different devices used for recording and for the video playback. The database contains 1942 genuine images, and 16885 fake faces are made from high quality records of the genuine faces. The database was collected using Amazon Mechanical Turk and Yandex Toloka services. The database was manually checked and the test protocol was provided. Some methods are also provided to be used as a baseline for future research. We hope that database as such can serve as an evaluation platform for the future studies in the literature.