Face anti-spoofing is detecting and preventing attacks on facial security system using photos, videos, or masks.

Domain Generalization is required to enhance robustness against variations, such as different devices and environments.

Model Reliability is crucial because unstable predictions can lead to unauthorized access or denial of legitimate users, undermining overall security system trustworthiness.

Domain-invariant : consistent for a task (e.g. class) regardless of domains
Domain-specific  : consistent for a domain (e.g. environment) regardless of the task

Prior work: Aligns only each classifier’s weights for consistent spoof → live direction across domains.

Problem: Output scores remain inconsistent (bias terms misalignment) ⇒ degraded generality

Cause: Classifier bias terms are shaped by the degree of alignment between invariant and specific features, acting as spurious correlations

Challenge: Weights and biases can be aligned across domains, but performance on unknown targets is hard to guarantee

Representation Learning: learns both spaces, aligns bias & weight

Orthogonal Decomposition: invariant & specific features no longer aligned with each other

Image-Text Similarity Loss via Group-wise Scaling Risk Minimization

Bias term alignment in domain-invariant space ⇒ output score consistency regardless of domains.

Feature Orthogonal Decomposition Loss

Weight alignment in domain-specific space ⇒ consistent spoof→live direction across domains.

Benchmark

Reliability with ECE

2D Visualization