On the generalisation capabilities of Fisher vector-based face presentation attack detection

Abstract

In past decades, the broad development experienced by biometric systems has unveiled several threats that may decrease their trustworthiness. These are attack presentations that can be easily carried out by a non-authorised subject to gain access to the biometric system. To mitigate those security concerns, most face presentation attack detection techniques have reported good detection performance when they are evaluated on known presentation attack instruments (PAIs) and under acquisition conditions. In contrast, for more realistic scenarios, existing algorithms face difficulties in detecting unknown PAI species which are only included in the test set. A feature space based on Fisher Vectors computed from compact binarised statistical image features histograms, which allows discovering semantic feature subsets from known samples to enhance the detection of unknown attacks is presented. This representation, which is evaluated for challenging unknown attacks taken from freely available facial databases, reports promising results. A Bona Fide Presentation Classification Error Rate 100 under 17% together with an area under the receiving operating characteristic curve of over 98% can be achieved in the presence of unknown attacks. In addition, by training a limited number of parameters, the method is able to achieve state-of-the-art deep learning-based approaches for cross-dataset scenarios.