With the rapid iteration of facial forgery techniques， robust detection mechanisms that can handle unseen forgery methods are increasingly crucial. However， current approaches are primarily tailored to specific forgery techniques， posing limitations in addressing this broader detection challenge. To address these issues， this paper proposes a difficulty-aware meta-learning （DAML） method for cross-domain face forgery detection. During the meta-training phase， our method utilizes a model-agnostic meta-learning （MAML）approach， using historical forgery images to optimize the meta-leaner. By leveraging a small amount of data in the target domain， we can adjust the parameters to fit new tasks. To tackle the issue of unstable training in model-agnostic meta-learning methods， we introduce a difficulty-aware mechanism to dynamically adjust the learning weights for different tasks during the training phase. We conduct extensive experiments on several publicly available benchmark datasets. The experimental results demonstrate that our method outperforms several methods， such as： RECCE， Xception， RFM， etc.， achieving better adaptability to unseen target domains.