Exploring Generalization Capability for Video Forgery and Detection based on Generative Adversarial Network
Ying Lin, Yanzhen Qu, Yuanpei Li, Zhishen Nie · 2020
With the development of digital image processing technology based on deep learning, the potential risk of using related technologies to threaten the security of multimedia information is increasing. Because the generated human face effect largely depends on the completeness of the input sample set, most of the current deep forgery models have the problem of human side-face collapse. This paper has studied the deep forgery technology of Deepfacelab and Faceswap, and adjusts the original auto-encoder-based model architecture to a generative adversarial network. By using the harmonic mean of cross entropy and mean square error as the loss function, the improved model can reduce the probability of some frames being discarded during training. Meanwhile, by adjusting key characteristics and the weights of features in different frames, it further optimizes the cross-dataset detection performance. Experimental results have shown that the improved model can keep more facial details while still maintain high human face clarity. The detection performance is improved and the cross-dataset average error rate of the deep detection model is about 35%.