ABSTRACT
Nowadays, reliable and automatic subject authentication has become of the utmost importance in multiple scenarios. Over the last decades, biometric recognition has shown to be a good alternative to password-based systems. In spite of their numerous advantages, biometric systems are vulnerable to presentation attacks (PAs), i.e., attempts to log into the system with a fake biometric characteristic or presentation attack instrument (PAI). These attacks pose a severe threat to the security of the authentication system: any person could eventually fabricate or order a gummy finger to impersonate someone else. Therefore, the development of accurate presentation attack detection (PAD) schemes is key to the wider deployment of secure biometric systems. In this chapter, we present a novel approach for fingerprint PAD based on short-wave infrared (SWIR) images and multi-spectral convolutional neural networks (CNNs). In particular, four samples are acquired at different SWIR wavelengths, which are subsequently fed to five different CNN models: a residual network trained from scratch and four different pre-trained networks (i.e., MobileNet, MobileNetV2, VGG19, and VGGFace). In our proposed approach, we first pre-process the four channels multi-spectral information through a convolutional layer to obtain RGB images (red, green, and blue – three channels) before applying regular CNN models. Over a database comprising 8,214 bona fide samples and 3,310 PA samples stemming from 41 different PAI species, the best score level fusion approach of three CNN models yields an Attack Presentation Classification Error Rate (APCER) of 1.16% for a low Bona fide Presentation Classification Error Rate (BPCER) of 0.2%. On the other hand, a BPCER of 1.10% can be also obtained for an APCER of 0.2% by selecting a different configuration of the fused models.
