With the development of AI and deep learning technology, image and video generation has made significant progress in recent years. However, these technologies have also created serious social and security issues. This study aims to improve the current deepfake detection model’s generalization ability. In the multi-scale training experiment, adding 60% scale images significantly improved the model performance on different resolution images. Adding noise augmentation methods improved the model’s detection ability on different augmented images in the data augmentation experiment. The model combining the best scale and best augmentation methods had a 9% performance improvement in the test of multiple scales and enhanced images. The study results show that this proposed method can effectively improve the performance of deepfake detection technology.