Authors: Margarita N. Favorskaya; Andrey I. Pakhirka
Addresses: Reshetnev Siberian State University of Science and Technology, 31 Krasnoyarsky Rabochy Ave., Krasnoyarsk, 660037, Russian Federation ' Reshetnev Siberian State University of Science and Technology, 31 Krasnoyarsky Rabochy Ave., Krasnoyarsk, 660037, Russian Federation
Abstract: Anomaly detection caused by face presentation attacks fundamentally reduces the vulnerability of the face recognition system. Face presentation attacks can be regarded as presentations obtained from secondary sources that are different in nature. This fact means that there is a wide range of presentation attacks, both now and in the future, as well as different approaches to preventing them. We adhere to a one-class classifier approach, but present it via an ensemble classifier based on the analysis of visual maps. During training, we first create a set of maps looking for chromatic colours, specular reflection, blur, and colour diversity features. Second, each of the maps is generalised using CNN stream with a simple architecture followed by an ensemble classifier. Such hybrid detection of face presentation attacks provides fast implementation and protects against unknown attacks in the future. Experiments with public face datasets and own face dataset confirmed our approach.
Keywords: anomaly detection; face recognition; presentation attacks; cue generalisation; ensemble classifier.
International Journal of Reasoning-based Intelligent Systems, 2022 Vol.14 No.1, pp.19 - 26
Received: 19 Sep 2021
Accepted: 20 Nov 2021
Published online: 13 Jun 2022 *