Exploring Latent Fingerprint Synthesis with Diffusion Probabilistic Models

Latent fingerprints provide reliable scientific evidence in forensic investigations. However, latent fingerprint data exhibits more scarcity and diversity than regular fingerprints due to variations in crime environments and scenarios. Consequently, leveraging generative techniques to supplement the insufficient real data becomes crucial for downstream tasks while protecting privacy. Previous studies primarily employed Generative Adversarial Networks (GAN) for synthesizing latent fingerprints, leading to issues like mode collapse and unstable training. Recently, Diffusion Probabilistic Models like DDPM have garnered widespread attention, showcasing superior distribution learning capabilities. This study is the first to comprehensively compare GANs and DDPM based on 10,000 high-resolution latent fingerprint images. Frechet Inception Distance (FID) scores and user visual assessment showcase that the samples from DDPM have a smaller distribution discrepancy from the given dataset, along with superior visual diversity and authenticity. These findings establish the status of Diffusion Probabilistic Models across various generative models for real-world fingerprint synthesis, showcasing advantages in stability, distribution learning, and visual quality.