# Rethinking Domain Generalization for Face Anti-spoofing: Separability and Alignment

**Authors:** Yiyou Sun, Yaojie Liu, Xiaoming Liu, Yixuan Li, Wen-Sheng Chu  
**arXiv:** [2303.13662](https://arxiv.org/abs/2303.13662)  
**Date:** 2023-03-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.13662v1/

## Key points

- Generalizing face anti-spoofing (FAS) models across different domains is a challenge
- Previous approaches to remove domain-specific signals from feature representation can backfire when applied to unseen test domains
- A new FAS strategy called Separability and Alignment (SA-FAS) is proposed, which encourages domain separability while aligning the live-to-spoof transition trajectory for all domains
- SA-FAS is formulated as a problem of invariant risk minimization (IRM), enabling the learning of domain-variant feature representation while maintaining a domain-invariant classifier
- The effectiveness of SA-FAS is demonstrated on challenging cross-domain FAS datasets with state-of-the-art performance
- Considering both domain separability and alignment is important in developing robust FAS models that can generalize across diverse domains.

## Summary

This paper explores the challenge of generalizing face anti-spoofing (FAS) models across different domains with variations in image resolution, blurriness, and sensor differences. Previous approaches have attempted to remove domain-specific signals from feature representation through metric learning or adversarial losses; however, this study shows that such efforts can backfire when applied to unseen test domains. Instead of constructing a domain-invariant feature space, the authors propose a new FAS strategy called Separability and Alignment (SA-FAS), which encourages domain separability while aligning the live-to-spoof transition trajectory for all domains. This approach is formulated as a problem of invariant risk minimization (IRM), enabling the learning of domain-variant feature representation while maintaining a domain-invariant classifier. The authors demonstrate the effectiveness of SA-FAS on challenging cross-domain FAS datasets and establish state-of-the-art performance. Overall, this study highlights the importance of considering both domain separability and alignment in developing robust FAS models that can generalize across diverse domains.