Fourier-Based Frequency Space Disentanglement and Augmentation for Generalizable Face Anti-Spoofing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Generalizing face anti-spoofing (FAS) models to new data distributions is difficult due to domain shifts.
• Existing domain generalization (DG) methods for FAS often focus on spatial features, potentially missing subtle spoofing patterns.
• There is a need for FAS methods that are robust to variations in both data domains and spoofing techniques.

Purpose of the Study:

• To propose a novel approach, Frequency Space Disentanglement and Augmentation (FSDA), for improving the generalization of face anti-spoofing models.
• To leverage frequency domain analysis to better capture spoofing traces.
• To enhance the robustness of FAS models against unseen domain and spoof type variations.

Main Methods:

• Utilized Fourier transformation to analyze face images in the frequency space, separating amplitude (texture) and phase (content) spectrums.
• Disentangled the amplitude spectrum into domain-related and spoof-related components.
• Developed a frequency space augmentation technique by mixing disentangled components and applied distillation and consistency losses.

Main Results:

• The proposed FSDA method demonstrated superior performance in improving the generalization ability of FAS models.
• Experiments on four FAS datasets confirmed the effectiveness of the frequency space approach in capturing robust spoof patterns.
• The method showed improved robustness against various unseen scenarios and spoofing types.

Conclusions:

• Analyzing face images in the frequency space, particularly low-level texture information in the amplitude spectrum, is crucial for effective face anti-spoofing.
• FSDA offers a promising direction for developing highly generalizable face anti-spoofing systems.
• The proposed method effectively addresses the challenge of domain generalization in face anti-spoofing.