MA-ViT: Modality-Agnostic Vision Transformers for Face Anti-Spoofing