Improving the generalization of face forgery detection via single domain augmentation

In recent years, significant progress has been made in the field of face forgery and face forgery detection. However, the performance of the detection methods in the unknown environment is far beyond satisfactory due to the feature distribution deviation of different fake face generators. In this paper, we adopt the domain generalization theory to improve the generality of fake face detection. The utilized method augments the original image samples by introducing gradient noise yielded during back-propagation, simulating the forgery features in unknown domains. In the construction of the detection network, we propose a multi-scale synthetic artifact trace tracker (MSATT) to enhance the manipulation traces through multi-scale content suppression. Meanwhile, we observed that the synthesized images present a noticeable color abnormality after going through the proposed MSATT module. Therefore, we designed a color difference perception network (CDPNet) to capture this unique feature. Experimental results demonstrate that both the domain augmentation and the proposed CDPNet can effectively improve the performance of the detection network. The proposed method is competitive with the state-of-the-art face forgery detection methods on both intra- and inter-dataset evaluations.