[Paper Review] PrintsGAN: Synthetic Fingerprint Generator

Joshua J. Engelsma, Steven A. Grosz+1 more|arXiv (Cornell University)|Jan 10, 2022

Biometric Identification and Security4 citations

TL;DR

PrintsGAN is a GAN-based synthetic fingerprint generator that produces high-quality, realistic 512×512 rolled fingerprints with multiple impressions per finger, enabling the creation of a large-scale synthetic dataset of 525K fingerprints (35K unique fingers, 15 impressions each). When used to pre-train a deep fingerprint embedding network (DeepPrint), it boosts performance to 87.03% TAR at FAR=0.01% on NIST SD4, significantly outperforming training on real data alone.

ABSTRACT

A major impediment to researchers working in the area of fingerprint recognition is the lack of publicly available, large-scale, fingerprint datasets. The publicly available datasets that do exist contain very few identities and impressions per finger. This limits research on a number of topics, including e.g., using deep networks to learn fixed length fingerprint embeddings. Therefore, we propose PrintsGAN, a synthetic fingerprint generator capable of generating unique fingerprints along with multiple impressions for a given fingerprint. Using PrintsGAN, we synthesize a database of 525k fingerprints (35K distinct fingers, each with 15 impressions). Next, we show the utility of the PrintsGAN generated dataset by training a deep network to extract a fixed-length embedding from a fingerprint. In particular, an embedding model trained on our synthetic fingerprints and fine-tuned on a small number of publicly available real fingerprints (25K prints from NIST SD302) obtains a TAR of 87.03% @ FAR=0.01% on the NIST SD4 database (a boost from TAR=73.37% when only trained on NIST SD302). Prevailing synthetic fingerprint generation methods do not enable such performance gains due to i) lack of realism or ii) inability to generate multiple impressions per finger. We plan to release our database of synthetic fingerprints to the public.

Motivation & Objective

• To address the critical lack of publicly available, large-scale fingerprint datasets with sufficient identities and multiple impressions per finger.
• To develop a synthetic fingerprint generation method that produces realistic, diverse, and unique fingerprints with controlled intra-class variation.
• To demonstrate the utility of synthetic fingerprints in improving deep learning models for fixed-length fingerprint embedding and large-scale identification.
• To enable public release of a large synthetic fingerprint database without privacy risks by ensuring no information leakage from the real training data.

Proposed method

• PrintsGAN employs a multi-stage GAN-based pipeline combining domain-knowledge-driven synthesis with deep generative modeling to produce high-fidelity fingerprints.
• The method first generates a 'master print' using a conditional GAN, then applies geometric warping and cropping to simulate different finger positions and orientations.
• Textural rendering is applied in a third stage to add realistic skin texture and ridge details, ensuring fidelity to real fingerprint characteristics.
• The model is trained on a real operational forensic fingerprint database, leveraging both GAN loss and perceptual losses to enhance realism and structural accuracy.
• The system enables control over fingerprint identity and impression variation, allowing generation of multiple distinct impressions per unique finger.
• The entire pipeline is implemented in TensorFlow, enabling on-the-fly generation during model training to reduce storage overhead.

Experimental results

Research questions

• RQ1 Can a GAN-based method generate synthetic fingerprints that are realistic enough to improve deep learning models trained on real data?
• RQ2 Can synthetic fingerprints with multiple impressions per finger enhance the performance of fixed-length fingerprint embedding networks?
• RQ3 Does the synthetic fingerprint generation process preserve sufficient intra-class variation to support robust fingerprint recognition?
• RQ4 Can a large-scale synthetic fingerprint database be publicly released without privacy concerns, given the lack of information leakage from the real training data?

Key findings

• Training a DeepPrint embedding model on PrintsGAN-generated synthetic fingerprints and fine-tuning on 25K real prints from NIST SD302 achieved a 13.66% improvement in TAR (87.03% at FAR=0.01%) on the NIST SD4 benchmark compared to training only on NIST SD302 (73.37%).
• The synthetic fingerprints enabled a rank-1 identification rate of 92.05% on NIST SD4 when used to pre-train the DeepPrint model, outperforming the 85.90% rate achieved with NIST SD302-only training.
• Augmenting the gallery with 100K PrintsGAN-generated fingerprints yielded identification performance comparable to augmenting with 100K real fingerprints (92.05% vs. 86.80%)
• Domain experts rated the realism of PrintsGAN-generated fingerprints as comparable to real fingerprints, confirming high perceptual quality.
• The synthesis process is computationally efficient, generating one 512×512 fingerprint in approximately 99.2ms using two RTX 2080 Ti GPUs.
• The method does not leak significant information from the real training database, enabling safe public release of the 525K synthetic fingerprint dataset.