DiffFinger: Advancing Synthetic Fingerprint Generation through Denoising Diffusion Probabilistic Models

TL;DR

This work tackles privacy and dataset diversity challenges in fingerprint biometrics by introducing DiffFinger, a DDPM-based framework for synthetic fingerprint generation. It models fingerprint synthesis with a forward diffusion $x_t = \sqrt{\alpha_t}\, x_0 + \sqrt{1-\alpha_t}\, \epsilon$ and a backward denoiser $\epsilon_{\theta}(x_t,t)$ trained to minimize $\mathcal{L} = \mathbb{E}_{t,\epsilon}[|| \epsilon - \epsilon_{\theta}(x_t,t) ||^2]$, enabling iterative refinement from noise to realistic prints. DiffFinger demonstrates the ability to produce infinite varied fingerprints, multiple impressions per identity, and a synthetic dataset that rival real data in quality as measured by $NFIQ^2$, while achieving improved diversity per identity as reflected in Bozorth3 and FID analyses. The findings suggest DDPMs can meaningfully advance biometric synthesis, enabling safer, more diverse datasets for training and evaluation of fingerprint identification systems, with potential implications for privacy-preserving biometrics and system robustness.

Abstract

This study explores the generation of synthesized fingerprint images using Denoising Diffusion Probabilistic Models (DDPMs). The significant obstacles in collecting real biometric data, such as privacy concerns and the demand for diverse datasets, underscore the imperative for synthetic biometric alternatives that are both realistic and varied. Despite the strides made with Generative Adversarial Networks (GANs) in producing realistic fingerprint images, their limitations prompt us to propose DDPMs as a promising alternative. DDPMs are capable of generating images with increasing clarity and realism while maintaining diversity. Our results reveal that DiffFinger not only competes with authentic training set data in quality but also provides a richer set of biometric data, reflecting true-to-life variability. These findings mark a promising stride in biometric synthesis, showcasing the potential of DDPMs to advance the landscape of fingerprint identification and authentication systems.

Figures (4)

• Figure 1: Example of the fingerprint generation process using the backward process of the DDPM
• Figure 2: Impressions of the same identity from our training dataset
• Figure 3: A comparison of the Bozorth3 scores distribution for each training set to it's synthetic counterpart
• Figure 4: The cumulative density function values of Bozorth3 scores for original vs synthetic impressions, scores over 40 indicate the same identity