MAREPVGG: Multi-attention RepPVGG to Facefake Detection.

The threat posed by the increasing means of face forgery and the lowering of the threshold of use is increasing. Although the detection capability of current detection models is improving, most of them need to consume large computational resources and have complex model architectures. Therefore, in this paper, we propose a new deep learning detection framework MARepVGG, which uses RepVGG as the backbone, combines texture enhancement module and multi-attention module to strengthen the network to learn face forgery features through the idea of heavy parameterization to balance training performance and inference speed. We evaluate our method on the kaggle real and fake face detection dataset, which differs from the computer automatically generated images, where the fake faces are high quality images produced by Photoshop experts. Our method improves the accuracy by 14% on this dataset compared to a baseline of forgery detection by repvgg alone, while the number of parameters is only 8.75 M.