Non-reference image quality assessment for fingervein presentation attack detection

Abstract

Non-reference image quality measures are used to distinguish real biometric data from data as used in presentation/sensor spoofing attacks. An experimental study shows that based on a set of 6 such measures, classification of real vs. fake fingervein data is feasible with an accuracy of 99% on one of our datasets. However, we have found that the best quality measure (combination) and classification setting highly depends on the target dataset. Thus, we are unable to provide any other recommendation than to optimise the choice of quality measure and classification setting for each specific application setting. Results also imply, that generalisation to unseen attack types might be difficult due to dataset dependence of the results.