AdG-MAE: Adaptive Gated Masked Autoencoder for Domain-Generalised Face Anti-Spoofing

Domain generalisation in Face Anti-Spoofing (FAS) remains a fundamental challenge due to the entangled nature of domain-specific attributes (e.g., lighting, background) and intrinsic spoofing patterns. While existing methods attempt to learn generalised features through complex multi-stream architectures or adversarial training, they often overlook the optimal stage at which frequency information should be integrated. In this paper, we propose the Adaptive Gated Masked Autoencoder (AdG-MAE), a principled framework that decouples structural learning from texture refinement to achieve high generalisation with minimal inference-stage complexity. We introduce a two-stage inductive bias design: First, we demonstrate that a standard spatial reconstruction objective during pre-training establishes a robust structural foundation, superior to frequency-constrained pre-training methods. To enable resource-efficient Vision Transformers (TinyViT) to learn such representations from limited data, we propose a novel Adaptive Gated Self-Attention (AGS) mechanism. This dynamic gate functions as a structural curriculum, transitioning the model from local-dominant processing to global signal amplification. Second, to explicitly bridge the domain gap during fine-tuning, we introduce a Spectral Reweighting Augmentation (SRA) mechanism. By suppressing low-frequency domain identifiers and amplifying high-frequency spoofing patterns, SRA directs the pre-trained encoder toward domain-invariant texture patterns. Extensive experiments across four cross-domain benchmarks demonstrate that AdG-MAE achieves state-of-the-art performance (Average HTER: 5.24%) with significantly lower inference overhead (∼0.65 GFLOPs), validating the effectiveness of our decoupled learning paradigm for resource-constrained deployment.