Empirical Evaluation of LBP-Extension Features for Finger Vein Spoofing Detection

Abstract

Biometric systems based upon finger vein images have been shown to be vulnerable to presentation attacks. For this paper, we consider a variety of methods extending local binary patterns (LBP) which can be used to distinguish between fake and real finger vein images. In the experiments, it is not only the accuracy of the respective methods as compared to baseline LBP which is documented, but also the impact of two further criteria: (1) The influence of selecting training & test samples non-randomly, i.e., selecting training samples in person-specific and size-varying manner, and (2) the impact of lowering the feature dimensionality by considering only uniform patterns. Our results show that these two criteria have to be considered if one wants to apply finger vein anti-spoofing mechanisms while the baseline LBP technique turns out to be competitive to almost all of its ``improvements''. As subject specific training data is usually not available, our results underpin the importance of using sufficiently sized training data when aiming for high spoofing detection accuracy.