Enhancing Fingerprint Image Synthesis with GANs, Diffusion Models, and Style Transfer Techniques

TL;DR

The paper addresses data-scarce fingerprint synthesis by combining denoising diffusion probabilistic models (DDPMs) and Wasserstein GANs with gradient penalty (WGAN-GP) to generate high-quality live fingerprints and, via CycleWGAN-GP style transfer, plausible spoof fingerprints. DDPM-based live synthesis (notably DDPM-v2) achieves the strongest similarity to real fingerprints with $FID$ as low as $15.78$, while WGAN-GP variants offer robust uniqueness as measured by $FAR$. The authors also demonstrate fingerprint-to-fingerprint transformation under limited data, showing material-dependent spoof realism and highlighting spoofiness as a key factor for transformation success. Metrics such as $FID$, $KID$, $PRDC$, and $FAR$ are used to quantify realism, diversity, and security implications, providing a comprehensive evaluation framework for synthetic fingerprint generation with potential GDPR-compliant data-sharing pathways. Overall, the work advances patch-based fingerprint synthesis by integrating diffusion and style-transfer methods to enable realistic, diverse, and controllable live/spoof generation.

Abstract

We present novel approaches involving generative adversarial networks and diffusion models in order to synthesize high quality, live and spoof fingerprint images while preserving features such as uniqueness and diversity. We generate live fingerprints from noise with a variety of methods, and we use image translation techniques to translate live fingerprint images to spoof. To generate different types of spoof images based on limited training data we incorporate style transfer techniques through a cycle autoencoder equipped with a Wasserstein metric along with Gradient Penalty (CycleWGAN-GP) in order to avoid mode collapse and instability. We find that when the spoof training data includes distinct spoof characteristics, it leads to improved live-to-spoof translation. We assess the diversity and realism of the generated live fingerprint images mainly through the Fréchet Inception Distance (FID) and the False Acceptance Rate (FAR). Our best diffusion model achieved a FID of 15.78. The comparable WGAN-GP model achieved slightly higher FID while performing better in the uniqueness assessment due to a slightly lower FAR when matched against the training data, indicating better creativity. Moreover, we give example images showing that a DDPM model clearly can generate realistic fingerprint images.

Figures (4)

• Figure 1: The forward and reverse processes involved in DDPMs. In the forward process, Gaussian noise is constantly added in each step. This becomes an isotropic Gaussian noise image in the end. In the reverse process, a random noise sample drawn from an isotropic Gaussian distribution is denoised at each level according to a deep learning model and, in theory, reconstituted to a clear image.
• Figure 2: Evaluation based on a matching function applied to pairs of real fingerprint images and/or generated fingerprints from the DDPM-v2 and WGAN-GP-v1 models. Left: Cumulative distribution of the matching scores of real impostors (gray) and assumed impostors (blue and yellow), where the assumed impostors are random pairs of generated images. Right: Synthetic FAR for assumed impostors that are either pairs of generated images or a generated image paired with a real image. The threshold for false accepts is set to achieve a baseline FAR on real impostors (x-axis). The synthetic FARs for matching random pairs of generated images are shown in blue and yellow, and the corresponding FARs for matching random generated images with random real images from the training data are shown in red and green. The dashed grey line represents identical synthetic and baseline FAR. The purple dashed line represents the FAR that is achieved for random image pairs from the training data, i.e. a mix of genuines (images from same finger) and impostors.
• Figure 3: Generated live fingerprints at epochs 1, 50, 100, 150, 200, 400, 800 (last row - left) and 1000 (last row -right), from model structure DDPM-v2 trained on the S3-L dataset.
• Figure 4: The spoof detection scores between Real live (RL) and Cycle live (CL), Real spoof (RS) and Generated spoof (GS). The subfigures (a) and (b) come from material 1, (c) and (d) come from material 2, and so on. All left-side figures illustrate the similarity between RL and CL to verify the consistency of CycleWGAN-GP. All right-side figures illustrate the similarity between RS and GS.