An Attention-Guided Framework for Explainable Biometric Presentation Attack Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Deep learning excels in facial biometrics but lacks explainability.
• Current methods for facial Presentation Attack Detection (PAD) struggle to justify their decisions.
• Explainability is crucial for trust and security in biometric systems.

Purpose of the Study:

• To develop an explainable deep neural architecture for Facial Biometric Presentation Attack Detection (PAD).
• To generate both visual and verbal explanations for PAD decisions.
• To enhance classification performance by integrating explanation-derived information.

Main Methods:

• Utilized Grad-CAM for visual explanations (saliency maps).
• Employed a Long-Short-Term-Memory (LSTM) network with a modified gate for verbal explanations.
• Integrated spatial and temporal information to detect anomalous visual characteristics.
• Incorporated explanations as additional features to improve classification.

Main Results:

• The proposed framework effectively utilizes spatial and temporal data for spoofing detection.
• Explanations derived from Grad-CAM and LSTM improved overall classification performance.
• The method demonstrated effectiveness across multiple benchmark datasets (CASIA-FA, Replay Attack, MSU-MFSD, HKBU MARs).

Conclusions:

• The novel explainable deep neural architecture enhances Facial Biometric PAD.
• The integration of visual and verbal explanations improves system robustness and trustworthiness.
• This approach addresses the critical need for interpretable AI in biometric security.