Learning From Easy to Hard: Fingerprint Spoof Detection With Hard Sample Mining

To counter unauthorized access posed by presentation attacks in fingerprint authentication systems, deep learning-based fingerprint spoofing detection methods have been introduced. However, different artificial materials produce non-uniform spoof image representations, leading to a bias toward easy-to-recognize samples while ignoring hard-to-recognize ones. Hard sample mining is beneficial for accurately modeling a discriminative function. In this paper, we propose a novel fingerprint spoof detection method with hard sample mining called EasyHard-FSD. Specifically, EasyHard-FSD introduces a self-training (teacher/student model), where a teacher model is learned using all original samples, and it is used to identify possible hard samples. EasyHard-FSD employs a Hard Mining Module (HMM) to mask out hard samples for the student model. This strategy allows the student model to learn a better discriminative function for recognizing live and spoof fingerprint samples accurately. Experimental results on LiveDet2013 and LiveDet2015 demonstrate that EasyHard-FSD outperforms state-of-the-art methods in fingerprint spoof detection. Our code is publicly available athttps://github.com/lalakitchen/EasyHard-FSD.