Spoof Trace Disentanglement for Generic Face Anti-Spoofing

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Face anti-spoofing relies on detecting subtle "spoof traces" like color distortion and Moiré patterns.
• Existing methods struggle with spoof trace diversity and lack of ground truth, limiting generalization and interpretability.

Purpose of the Study:

• To propose a novel adversarial learning framework for explicit spoof trace estimation in face anti-spoofing.
• To enhance model generalization and interpretability by focusing on interpretable spoof traces.

Main Methods:

• A two-step adversarial learning framework (additive and inpainting) to disentangle spoof faces into spoof traces and live counterparts.
• Utilizing disentangled spoof traces for data augmentation to address long-tail spoof types.
• Applying frequency-based image decomposition to input and traces for low-level visual cue analysis.

Main Results:

• Superior spoof detection performance across known, unknown, and open-set attack scenarios.
• Visually convincing estimation and visualization of spoof traces.
• Improved model generalization and interpretability.

Conclusions:

• The proposed adversarial framework effectively estimates spoof traces, leading to robust face anti-spoofing.
• The method enhances performance against diverse and unseen spoofing attacks.
• Explicit spoof trace modeling offers a promising direction for interpretable and generalizable face anti-spoofing solutions.