Privacy-Conscious Synthetic Fingerprint Generation for Generalizable Presentation Attack Detection

Abstract

Fingerprint Presentation Attack Detection (FPAD) is essential for securing biometric systems against spoofing attacks. However, the development of robust models is often hindered by the scarcity and limited material diversity of existing spoof datasets. This paper presents a comprehensive empirical study on the utility of material-specific synthetic fingerprints for enhancing FPAD generalization. Specifically, we employ StyleGAN2-ADA to generate a large-scale dataset of 7,200 spoof images across 72 distinct material categories, modeled after real-world textural and optical characteristics. We evaluate the proposed dataset using a ResNet-50 backbone under an unseen-material protocol, shifting the evaluation focus from standard accuracy to security-critical metrics, including APCER, BPCER, and ACER (ISO/IEC 30107-3). Our findings demonstrate that while unconditional generative models like StyleGAN2-ADA introduce more challenging variations that may lower in-distribution F1-scores compared to paired style-transfer methods, they provide superior structural integrity and closer distributional alignment with real fingerprints relative to the style-transfer baseline. These results indicate that high-fidelity synthetic data serves as a scalable, privacy-conscious alternative to manual collection, enabling the detection of novel presentation attacks in real-world security applications.