In this paper, a new data-driven information hiding scheme called generative steganography by sampling (GSS) is proposed. The stego is directly sampled by a powerful generator without an explicit cover. Secret key shared by both parties is used for message embedding and extraction, respectively. Jensen-Shannon Divergence is introduced as new criteria for evaluation of the security of the generative steganography. Based on these principles, a simple practical generative steganography method is proposed using semantic image inpainting. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated stego images.

Generative Steganography by Sampling