Provably Secure Generative Steganography Based on Autoregressive Model

Synthetic data and generative models have been more and more popular with the rapid development of machine learning and artificial intelligence (AI). Consequently, generative steganography, a novel steganographic method finishing the operation of steganography directly in the process of image generation, tends to get more attention. However, most of the existing generative steganographic methods have more or less shortcomings, such as low security, small capacity or limited to certain images. In this paper, we propose a novel framework for generative steganography based on autoregressive model, or rather, PixelCNN. Theoretical derivation has been taken to prove the security of the framework. A simplified version is also proposed for binary embedding with lower complexity, for which the experiments show that the proposed method can resist the existing steganalytic methods.