Does Capture Background Influence the Accuracy of the Deep Learning Based Fingerphoto Presentation Attack Detection Techniques?

The rapid evolution of modern smartphone techniques has made biometric authentication applications feasible using smartphone cameras. FingerPhoto verification offers the benefits of scalability, reliability, and user convenience. Similar to traditional contact-based fingerprint verification methods, the widespread deployment offingerphoto authen-tication applications has raised concerns regarding the sys-tem being attacked (or spoofed). In this work, we not only study and discuss the generalizability of eight different pre-trained deep learning models against unseen attacks but also present an analysis of how the background of the cap-tured fingerphoto and attack samples will affect the Pre-sentation Attack Detection (PAD) performance. To exper-imentally benchmark the PAD performance with different types of background extractors, we present three different studies: full background, segmenting only the background, and extracting the Region Of Interest (ROI) that pertains to the fingerphoto region. We present an extensive evaluation of three different types of background extraction methods using eight different pre-trained deep learning techniques. The obtained results on the publicly available fingerphoto datasets indicate that by removing the background noise or extracting the ROI regions, the deep learning models will become more reliable for fingerphoto presentation attack detection.