Fingerprint Synthesis Using Deep Generative Models

Abstract: The advancements in biometric technology have amplified the need for more robust fingerprint synthesis techniques. In this thesis, we first explored the application of synthesizing normal fingerprint images in high fidelity using deep generative models (e.g., generative adversarial networks and diffusion models) and created synthetic fingerprints that retain the uniqueness and complexity of the original samples. Thereafter, by employing style transfer techniques (e.g., cycleGAN and cycleWGAN-GP), we effectively blended the global structure of one fingerprint with the local features of another fingerprint to generate a new fingerprint that is both visually realistic and distinctive, such as generating a one-to-one spoof fingerprint that is highly consistent with the corresponding normal fingerprint from the training set. We employed, in the field of image generation, the state-of-the-art metrics assessing the quality of synthetic fingerprints, mainly from the perspectives of statistical analysis and subjective evaluation, and conducted a comprehensive evaluation of synthetic normal and spoof fingerprints. Our best diffusion model achieves a promising Fréchet Inception Distance (FID) score of 15.78 and is capable of generating normal fingerprints that obtain even smaller False Acceptance Rate (FAR) than the real normal fingerprints. Our best style transfer model - cycleWGAN-GP is capable of generating spoof fingerprints in high quality from real normal fingerprints, and the distribution of these synthetic spoof fingerprints closely resembles that of real spoof fingerprints. Our results demonstrate the potential of these methods in generating high-quality fingerprints and can be used for various applications such as security enhancement, template protection, and biometric system evaluation.