关键词: 伪造算法识别, 人脸伪造, 深度学习

Abstract: With the development of deep learning, face forgery methods are gradually increasing. The existing deepfake face detection models are mostly concentrated on the true or false binary classification, with poor generalization, which is difficult to adapt to the current iterative forgery algorithms. What’s more, the researches on the face images with different sharpness and noise interference are not deep enough. Therefore, in order to solve the above problems, this paper proposes a low quality face image forging algorithm recognition method based on pre-training model. Specifically, the entire model is divided into two major modules: pre-training and main training.In the pre-training stage, Perception and Reconstruction-MAE (PR-MAE) based on mask strategy is designed to further excavate mask features to construct perceptual loss and enhance the global perceptual learning ability of the model. And the UDenseNet training network is constructed to improve the ability to capture details by U-shaped dense blocks, in order to adapt to more complex tasks. Meanwhile, the Multi-Noise Fusion (MNF) module is added to improve the model's resistance to multiple types of noise by dynamic injection strategy and enhance the robustness. In the main training process, CLIP supplementary learning module is introduced to perform the comparative learning with the pre-trained UDenseNet to improve the accuracy of the model. The experimental results show that the ACC index of this method can reach 89.12% and 90.25% on the latest MCFF and DF40 classification datasets, the average ACC index can reach 87.22% in the tests of the low definition images, 83.36% in the tests of the various kinds of noise and the index can reach more than 90% at the highest in the generalization test.

Key words: Forgery algorithm recognition, Face forgery, Deep learning