Progressive Transfer Learning for Face Anti-Spoofing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Face anti-spoofing (FAS) is crucial for secure face recognition systems.
• Current FAS methods require extensive labeled spoofing data, which is difficult to obtain comprehensively.
• The diversity and evolving nature of spoofing attacks necessitate adaptive and efficient FAS solutions.

Purpose of the Study:

• To develop an online learning framework for face anti-spoofing using limited labeled data.
• To enhance the robustness of FAS models against unseen spoofing attack variations.
• To reduce the reliance on large, meticulously annotated spoofing datasets.

Main Methods:

• A semi-supervised learning framework that progressively incorporates unlabeled data with pseudo-labels.
• Exploitation of temporal consistency in video streams to validate pseudo-labels.
• An adaptive transfer mechanism to mitigate the impact of novel spoofing techniques.
• Training on both seen and unseen spoofing data domains to bridge the domain gap.

Main Results:

• The proposed method achieves state-of-the-art performance with significantly less labeled data (less than 0.1%).
• Demonstrated effectiveness in both intra-database and cross-database testing scenarios.
• Outperformed fully-supervised methods in cross-domain testing due to progressive learning.
• Successfully reduced the domain gap between seen and unseen spoofing attack types.

Conclusions:

• The semi-supervised, online learning approach offers a practical and data-efficient solution for face anti-spoofing.
• The method's ability to adapt to unseen attacks makes it highly suitable for real-world applications.
• This framework significantly lowers the barrier to entry for deploying effective FAS systems by minimizing data requirements.