<p>With the continuous advancement of deep learning-based generation models, the problem of facial forgery has become increasingly prevalent. Therefore, deep forgery detection has become a major research focus, attracting widespread attention from scholars. However, the complex and diverse technologies employed in facial forgery and generation bring tremendous challenges for the generalization of detection models. For this, we propose a new forgery feature extraction network, FDPNet, with the goal of training a model that can effectively extract forgery traces from various media types and accurately predict the forgery regions to distinguish between genuine and fake content. To improve the model’s accuracy, this work also proposes an adversarial data augmentation technique, AMSM. This method aims to improve the detection model’s generalization ability by diversifying the types of forgery and strengthening self-supervised tasks that are sensitive to specific forgery configurations. Experimental results show that our model’s performance on the dataset was improved by 0.09%. The model’s generalization ability was significantly enhanced in cross-dataset testing, with improvements of 0.66%, 0.43%, and 1.15%. These innovations significantly enhance the accuracy and generalization of forgery detection.</p>