TL;DR
Current domain generalizable face anti-spoofing methods fail when encountering unseen environments because they only align the direction (weights) of decision boundaries while ignoring the classification thresholds (biases). This paper introduces GD-FAS, a framework that jointly aligns weights and biases using Feature Orthogonal Decomposition and Group-wise Scaling Risk Minimization. The approach achieves state-of-the-art generalization and significantly improves model calibration on unseen target domains.
What Problem Does This Paper Address?
Face Recognition (FR) systems are everywhere, from airport security to smartphone unlocking. Consequently, Face Anti-Spoofing (FAS)—detecting whether a face presented to a camera is real or a fake photo/video spoof—is a critical defense mechanism.
The real challenge lies in Domain Generalizable FAS (DGFAS): training a model on a few source domains (e.g., specific cameras, lighting conditions) so that it performs flawlessly on entirely unseen target domains.
Existing methods typically try to extract domain-invariant features by aligning the directions (classifier weights) of local decision boundaries. However, the authors point out a glaring omission: they ignore the bias terms. As shown in Figure 1, if the bias terms remain misaligned, the classification thresholds shift across domains, leading to degraded performance and poor reliability when the model faces a new environment.
Methods
To solve this, the authors propose a framework called GD-FAS, built on top of the vision-language model CLIP. It introduces two novel components to jointly align weights and biases:
- Group-wise Scaling Risk Minimization (GS-RM): This module balances learning across multiple domains by adaptively scaling group-wise losses. It penalizes low-loss (easy/highly similar) groups and boosts high-loss (hard/dissimilar) groups during training, forcing the classifier to find a consistent bias threshold that works across all groups.
- Feature Orthogonal Decomposition (FOD): Using the classic Gram-Schmidt orthogonalization process, FOD explicitly splits the image feature space into two distinct, perpendicular subspaces: domain-invariant and domain-specific.
By ensuring complete mathematical orthogonality, the domain-specific features have an inner product of zero with the classifier weights. This prevents domain-specific variations from interfering with the aligned bias terms.
Additionally, the authors introduce Expected Calibration Error (ECE) to the DGFAS field to quantitatively measure how reliable the model's confidence scores are across different domains.
Datasets
The framework was evaluated using a leave-one-out protocol across four widely recognized benchmark FAS datasets:
- CASIA-FA (C)
- OULU-NPU (O)
- Idiap Replay-Attack (I)
- MSU-MFSD (M)
For extended testing, they also utilized CelebA-Spoof as an extra source dataset, and validated generalization on SiW-Mv2, Surf (S), CeFA (C), and WMCA (W).
Key Insights and Ideas
- The Geometry of Bias: If domain-specific features ($f_S$) are perfectly orthogonal to the invariant basis ($bI$), then the term $f_S \cdot w^T$ drops to zero. This mathematically guarantees that the domain bias is consistently aligned as zero across all domains.
- Calibration as a Proxy for Alignment: A well-calibrated model (low ECE) means its predicted probabilities match actual accuracy. The authors show that tracking ECE is an excellent way to evaluate whether bias terms are truly aligned across domain shifts.
- Stability Over Early Stopping: Many recent DGFAS methods (like FLIP) suffer from severe performance degradation if they aren't stopped early during training. GD-FAS achieves exceptional convergence stability, maintaining top-tier performance across extended training epochs.
Key Results
GD-FAS achieves state-of-the-art performance across standard benchmarks, measured by Half Total Error Rate (HTER) and Area Under the Curve (AUC):
- Benchmark Superiority: Without extra data, GD-FAS achieves an average HTER of 1.83%, outperforming recent methods like SAFAS (3.60%) and CFPL (3.60%).
- Perfect Predictions: When utilizing CelebA-Spoof as an extra source dataset, GD-FAS achieves a 0.00% HTER (perfect classification) in the OCI $\rightarrow$ M evaluation scenario.
- Reliability: GD-FAS consistently yields the lowest ECE scores across almost all domain shift scenarios, proving that its threshold alignment strategy works.
Limitations and Future Work
The authors note a slight trade-off between model reliability (ECE) and raw generalization metrics when combining GS-RM and FOD. While using FOD alone reduces ECE to its absolute lowest levels, the full combination of GS-RM and FOD is required to get the best balance of low error rates (HTER) and calibration. Future work could explore refining this trade-off or applying this orthogonal decomposition strategy to other domain generalization tasks outside of face anti-spoofing.
Summary updated