An evaluation study of non-contact fingerprint presentation attack detection

Non-contact based fingerprint recognition systems are emerging at rapid pace due the fast growing need in the era of contagious disease transmission i.e. COVID-19. This growth also necessitates an effective security system against possible Presentation Attacks (PAs). In recent years, widely popular Convolution Neural Networks (CNNs) have been deployed vastly for the contact based fingerprint Presentation Attack Detection (PAD) systems. In this work, we have evaluated and compared three widely popular deep CNN models namely, DenseNet121, ResNet50, and EfficientNet for the PAD in non-contact fingerprint recognition systems. The study was conducted on a new publicly available Contactless Fingerprint Spoof (COLFISPOOF) and Contact to Contactless (C2CL) dataset. The experiments were conducted using the standard baseline and four Leave-One-Out (LOO) protocols, to observe the generalization capabilities of the considered deep CNN models for unseen data. Overall results indicates that DenseNet121 model is able to generalize well for different evaluation protocols and provide average Attack Presentation Classification Error Rate (APCER) of $\mathbf{2. 8 1 \%}$ and Bona fide Presentation Classification Error Rate (BPCER) of 0.25%.