Conditional Synthetic Live and Spoof Fingerprint Generation

TL;DR

This work addresses privacy and data collection challenges in fingerprint biometrics by introducing a conditional generative framework that creates high-resolution live fingerprints conditioned on finger class (1–10) using StyleGAN2-ADA and StyleGAN3, and eight CycleGANs to translate these into spoof fingerprints across eight materials. Two synthetic datasets, DB2 and DB3, each contain 1,500 live fingerprints and 1,500 spoof variants, enabling robust evaluation of biometric realism, uniqueness, and privacy preservation. The models achieve a Fréchet Inception Distance as low as 5 (StyleGAN3) and high true-acceptance rates at very low false-acceptance rates (e.g., TAR ≈ 99.47% at FAR = 0.01%), while preserving privacy with no significant identity leakage. Furthermore, augmenting spoof-detection training with these synthetic datasets substantially improves performance on unseen data, demonstrating practical utility for PAD research and privacy-preserving biometric data sharing.

Abstract

Large fingerprint datasets, while important for training and evaluation, are time-consuming and expensive to collect and require strict privacy measures. Researchers are exploring the use of synthetic fingerprint data to address these issues. This paper presents a novel approach for generating synthetic fingerprint images (both spoof and live), addressing concerns related to privacy, cost, and accessibility in biometric data collection. Our approach utilizes conditional StyleGAN2-ADA and StyleGAN3 architectures to produce high-resolution synthetic live fingerprints, conditioned on specific finger identities (thumb through little finger). Additionally, we employ CycleGANs to translate these into realistic spoof fingerprints, simulating a variety of presentation attack materials (e.g., EcoFlex, Play-Doh). These synthetic spoof fingerprints are crucial for developing robust spoof detection systems. Through these generative models, we created two synthetic datasets (DB2 and DB3), each containing 1,500 fingerprint images of all ten fingers with multiple impressions per finger, and including corresponding spoofs in eight material types. The results indicate robust performance: our StyleGAN3 model achieves a Fréchet Inception Distance (FID) as low as 5, and the generated fingerprints achieve a True Accept Rate of 99.47% at a 0.01% False Accept Rate. The StyleGAN2-ADA model achieved a TAR of 98.67% at the same 0.01% FAR. We assess fingerprint quality using standard metrics (NFIQ2, MINDTCT), and notably, matching experiments confirm strong privacy preservation, with no significant evidence of identity leakage, confirming the strong privacy-preserving properties of our synthetic datasets.

Figures (9)

• Figure 1: Synthetic fingerprints, each sized $512 \times 512$ pixels, generated using our newly developed model described in Section \ref{['subsubsec:stylegan3_training']}, which leverages the architecture of StyleGAN3. The first row shows fingerprints for the Left-Index, Left-Middle, Left-Ring, Left-Little, and Left-Thumb, while the second row shows fingerprints for the Right-Index, Right-Middle, Right-Ring, Right-Little, and Right-Thumb.
• Figure 2: Synthetic fingerprints, each sized $512 \times 512$ pixels, generated using our newly developed model described in Section \ref{['subsubsec:stylegan2_training']}, which leverages the architecture of StyleGAN2-ADA. The first row shows fingerprints for the Left-Index, Left-Middle, Left-Ring, Left-Little, and Left-Thumb, while the second row shows fingerprints for the Right-Index, Right-Middle, Right-Ring, Right-Little, and Right-Thumb.
• Figure 3: The figure illustrates the overall architecture of our proposed fingerprint generation framework. Users are required to provide the finger/class number, material type (live or one of the eight spoof types), and specify the desired number of fingerprint samples along with the number of multiple impressions per finger.
• Figure 4: Three impressions of the same finger from our DB3 synthetic dataset.
• Figure 5: The VeriFinger matcher analyzed imposter score distributions. Three experiments evaluated uniqueness: DB1 Non-mated scores, DB2 Non-mated scores, and DB3 Non-mated scores. The resulting distributions of match scores (DB2 Non-mated and DB3 Non-mated) closely resemble the distribution of imposter scores for the training images (DB1 Non-mated scores). This indicates that the fingerprints in the generated images exhibit a similar level of uniqueness to those in the training images.