Domain Generalization for Face Anti-Spoofing with Joint Frequency-Domain Phase and Spatial-Domain Modulation
Most existing domain generalization methods forface anti-spoofing rely only on spatial-domain features, makingthem vulnerable to performance drops when facing differentenvironments, attack materials, or camera devices. To solve thisproblem, we propose a new Spatial and Spectrum Dual-DomainGeneralization (SSDDG) approach that effectively combines bothspatial and frequency information. Our key insight reveals thatphase components in the frequency domain inherently providestable, domain-invariant features. Building upon this discovery,we develop the SSDDG framework comprising three core compo-nents: (1) A frequency branch with an adaptive filter to extractand adjust important phase features, (2) A spatial branch usingmulti-scale transformers to identify live face characteristics, and(3) A multi-scale interpolation alignment module operates on bothspatial and spectral domains to achieve precise feature alignmentacross multiple scales. We also design two auxiliary supervisions- an interaction loss for cross-domain feature fusion and anadversarial loss for compact live feature distribution learning.Extensive experiments show that our SSDDG approach is effec-tive, competitive with recent domain generalization methods onfour public datasets, and particularly effective in limited-sourcetransfer without target-domain access. The code will be made publicly available upon acceptance of the paper.
Authors
- Tongming Wan (ORCID: https://orcid.org/0009-0009-6489-2048)
- weihua liu
- shu liu
Institutions
- Central South University (CN)
- State Grid Corporation of China (China) (CN)
- Quanzhou University of Information Engineering (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11777889/v1
- Primary Topic
- Biometric Identification and Security
- Type
- preprint