DeepFake Face Image Detection based on Improved VGG Convolutional Neural Network
Xu Chang, Jian Wu, Tongfeng Yang, Guorui Feng · 2020
DeepFake can forge high-quality tampered images and videos that are consistent with the distribution of real data. Its rapid development causes people's panic and reflection. In this paper we presents an improved VGG network named NA-VGG to detect DeepFake face image, which was based on image noise and image augmentation. Firstly, In order to learn the tampering artifacts that may not be seen in RGB channels, SRM filter layer is used to highlight the image noise features; Secondly, the image noise map is augmented to weaken the face features. Finally, the augmented noise images are input into the network to train and judge whether the image is forged. The experimental results using the Celeb-DF dataset have shown that NA-VGG made great improvements than other state-of-the-art fake image detectors.