Visual Content Generation

Researchers generated synthetic fingerprint images to test weaknesses in fingerprint matching systems.

Business impact

• subjects spoofed at 0.1% false match rate — The generated DeepMasterPrints spoofed 23% of subjects at a 0.1% false match rate.
• subjects spoofed at 1% false match rate — The generated DeepMasterPrints spoofed 77% of subjects at a 1% false match rate.

Data requirements

• fingerprint images from NIST Special Database 9 (image) — Used as rolled fingerprint data for training and evaluation.
• FingerPass DB7 fingerprint dataset (image) — Used as partial capacitive fingerprint samples for training and evaluation.

AI methods and techniques

• generative neural networks — Used to learn how to produce synthetic fingerprint images.
• evolutionary optimization — Used to search the latent space for images that maximize match success.
• latent variable search — Used to explore the generator input space for a DeepMasterPrint.

AI models and model families

WGAN, CMA-ES

Industries

Real-world evidence

1 documented case study on record.

Companies using this: New York University.