MobileViT-FocR: MobileViT with Fixed-One-Centre Loss and Gradient Reversal for Generalised Fake Face Detection

Fake face detection is one of the most important face detection technologies, which plays an important role in preventing malicious actors from using generated fake faces for malicious purposes. However, current fake face detection techniques have poor generalisation detection ability to recognise different types of fake face images, which makes it difficult to apply this technology to real-life scenarios. Therefore, it is important to construct a fake face detection model with stronger cross-domain generalisation capabilities. In order to enhance the generalisation detection capability of the model on different face datasets, we propose a MobileViT-FocR model, which uses MobileViT to extract local and global features, and proposes the Fixed-One-Centre(FOC) loss, that is, to select a fixed centre point and focus only on similar features of real face images. The model is tuned with some of the parameters of focal loss to enhance its ability to detect more difficult fake face images. The GRL(Gradient Reversal Layer) is added based on the network to make the model better focus on the generic category differences between fake and real faces rather than the domain differences. Through experimental verification, our model has good detection capability for fake face images of different styles generated by various algorithms. Compared to the original MobileViT model, our model improved by 9.79 \(\%\) , 8.58 \(\%\) , 7.70 \(\%\) , and 8.23 \(\%\) on Internet Celebrity, Celeb-DF, DFDC, and ForgeryNet datasets respectively.