A DeepFake Face Detection Method Using Vision Transformer with a Convolutional Module

The Artificial Intelligence (AI) industry has developed rapidly in recent years. Plenty of large generative models such as ChatGPT have achieved remarkable accomplishment. AI technology provides much convenience for human beings in various fields, but it also brings much harm. Many lawbreakers utilize this technology to perform DeepFake operations such as replacing one face with another for crucial people, posing a challenge to social security. In this work, we present a novel structure detector for DeepFake examination. The network is a convolutional module of the Convolutional Neural Network embedded in the Vision Transformer, called Convolutional Neural Network Vision Transformer (CNNVT). Where the CNN captures the fine-grained features of the face, ViT implements the context representation and final classification. Subsequently, a series of experiments could prove the validity of our network. We trained our model on the DeepFake dataset and got 98.67% Recall and 98.68% Accuracy. The model has practical significance in preventing the propagation of DeepFake data.