Face forgery technology poses significant risks, highlighting the need for effective detection methods. However, existing approaches often focus on specific forgery types and lack generalization. We treat face forgery detection as a zero-shot learning problem, aiming to detect unseen forgery methods. We propose a meta-learning-based model that combines EfficientNet and the Vision Transformer (ViT). The model processes images through two branches: ViT detects global discrepancies, while EfficientNet captures local features. Cross-attention integrates these features, enhancing detection capabilities. And we employ quadruplet loss to enhance the compactness within classes and the separation between different classes. Our model, evaluated on FaceForensics++ and Celeb-DF datasets, shows improved generalization and performance.