Fingerprint Presentation Attack Detection using Unsupervised Divergence Based Domain Adaptation

Abstract

 The biometric system that uses fingerprints is prone to different types of attacks. The presentation attack is one of the easiest to perform on the fingerprint sensor. In recent years, several Fingerprint Presentation Attack Detection (FPAD) approaches have been proposed. These FPAD approaches have attained fair results on a dataset of different materials (cross-material). However, FPAD method performance degrades up to 30% when training and testing datasets are from different distributions (sensors). So for a robust FPAD method, it must learn domain-independent features to have consistent performance. To mitigate the domain-shift and FPAD, we have proposed unsupervised divergence-based domain adaptation (UDDA) with an Adaptive Loss Function (ALF). The ALF integrates domain divergence loss (DDL) and classification loss. The ALF helps in learning domain-invariant features and accurately classifying live and fake fingerprints in a cross-sensor scenario. The investigational outcomes confirmed that the offered UDDA method reduces the cross-sensor average classification error (ACE) by 19.94% and 19.23% on LivDet 2015 and LivDet 2017, respectively.