Unified Generative Adversarial Networks for Multidomain Fingerprint Presentation Attack Detection.

Area of Science:

• Biometrics and Security
• Computer Vision
• Machine Learning

Background:

• Fingerprint biometric systems face security vulnerabilities, necessitating improved presentation attack detection (PAD).
• Existing PAD methods struggle with generalization across different sensors and materials.
• Limited labeled data in target domains hinders model adaptation.

Purpose of the Study:

• To develop a robust fingerprint PAD method that addresses limited sample issues in multiple target domains.
• To enhance the generalization ability of PAD algorithms across diverse sensor and material settings.
• To leverage knowledge transfer from a well-resourced source domain to under-resourced target domains.

Main Methods:

• A unified generative adversarial network (UGAN) was trained for multi-domain image conversion, generating synthetic data to bridge domain gaps.
• A scale-efficient network (EfficientNetV2) with multiple classifiers was trained on source and translated target domain data.
• An additional fusion layer with learnable weights aggregated classifier outputs for improved PAD performance.

Main Results:

• The proposed method demonstrated significant improvements in average classification accuracy on the LivDet2015 dataset.
• Accuracy increased from 67.80% to 80.44% across twelve classification scenarios after adaptation.
• The UGAN effectively reduced the distribution shift between fingerprint representations from different sensors.

Conclusions:

• The proposed UGAN-based approach effectively enhances fingerprint PAD generalization across diverse domains.
• Knowledge transfer from a source domain significantly improves performance in target domains with limited data.
• This methodology offers a promising solution for building more secure and reliable fingerprint biometric systems.