A controllable face forgery framework to enrich face-privacy-protection datasets

Abstract

• Analyze the identity information in the latent features and construct an identity latent space. • Propose an adaptive identity mapping network to edit the latent codes. • Experiments show the effectiveness and the controllability of our work. • Our proposed face forgery framework can enrich the face-privacy-protection dataset. Deep learning not only brings convenience to people, but also promotes the development of facial forgery technology. Considering the current personal portrait security issues, the tampering and forgery of facial data has attracted more and more attention. In order to solve the above issues, we try to implement from another novel angle, that is, enrich the face-privacy-protection dataset to improve the detection ability of forgery faces. Therefore, we propose a controllable face forgery framework. In this work, we firstly analyze the identity information in the latent features and construct an identity latent space based on StyleGAN’s w + latent space. Then, we propose an adaptive identity mapping network to edit the latent codes of the image through encoder and realize the identity transform. Finally, we further enhance the authenticity of the image through post-processing. In order to verify the superiority of our proposed method, we design extensive experiments. Experiments show the effectiveness of identity latent space and the controllability of our model. At the same time, it also shows that our proposed network can generate photo-level results and achieve excellent results in the comparison of other face swapping methods.