Multi-Domain Feature Alignment for Face Anti-Spoofing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Face anti-spoofing is crucial for secure face recognition systems.
• Current methods often struggle with generalization due to domain discrepancies.
• Domain generalization shows promise but faces challenges with scattered feature spaces.

Purpose of the Study:

• To propose a multi-domain feature alignment framework (MADG) to address poor generalization in face anti-spoofing.
• To improve the robustness of face recognition systems against presentation attacks.
• To enhance the separation between real and fake facial features in the feature space.

Main Methods:

• Developed a multi-domain feature alignment framework (MADG).
• Employed adversarial learning to align features across multiple source domains.
• Incorporated multi-directional triplet loss for enhanced feature separation.

Main Results:

• The MADG framework effectively aligns features from scattered domains.
• Multi-directional triplet loss significantly improves feature space separation.
• Extensive experiments on public datasets demonstrate superior performance.

Conclusions:

• The proposed MADG framework significantly enhances face anti-spoofing generalization.
• The approach effectively mitigates distribution discrepancies across domains.
• Results validate the framework's effectiveness against state-of-the-art methods.