One-Class Fingerprint Presentation Attack Detection Using Auto-Encoder Network.

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Automated Fingerprint Recognition Systems (AFRSs) are vulnerable to Presentation Attacks (PAs).
• Developing effective Presentation Attack Detection (PAD) methods is crucial but challenging due to unpredictable PAs.
• Existing methods often require labeled PA data, limiting their applicability.

Purpose of the Study:

• To propose a novel One-Class PAD (OCPAD) method for fingerprint PA detection using Optical Coherence Technology (OCT) images.
• To develop a PAD model trained exclusively on genuine (bonafide) fingerprint data.
• To enhance reconstruction error accuracy using an activation map-based weighting model.

Main Methods:

• A novel One-Class PAD (OCPAD) model utilizing an auto-encoder network trained on bonafide fingerprints.
• Calculation of spoofness score based on reconstruction error and latent code from the auto-encoder.
• Refinement of reconstruction error accuracy via an activation map-based weighting model.
• Decision-level fusion of statistics and distance measures for final prediction.

Main Results:

• The proposed OCPAD achieved a True Positive Rate (TPR) of 99.43% at a False Positive Rate (FPR) of 10%.
• Achieved a TPR of 96.59% at an FPR of 5%.
• Significantly outperformed feature-based and supervised learning-based PAD models.

Conclusions:

• The OCPAD method offers a robust and effective solution for fingerprint PAD using OCT images.
• The one-class approach eliminates the need for PA samples during training, simplifying deployment.
• The method demonstrates superior performance compared to existing approaches, enhancing AFRS security.