GAN-TStega: Text Steganography Based on Generative Adversarial Networks

Steganography based on text auto-generation technology is a current topic with great promise and challenges. It has the advantages of large information hiding capacity compared with the modification-based text steganographic methods. The biggest challenge faced by previous methods is that they can hardly generate fluent steganographic texts, and only pay attention to the statistical distribution of individual sentences without considering the overall statistical distribution of all generated texts. This paper proposes a text steganography called GAN-TStega which based on generative adversarial networks (GANs). Firstly, we use strategy update algorithm to solve the problem that traditional GANs are difficult to generate discrete data. Through antagonistic training on different types of text datasets, GAN-TStega can generate high quality texts. Then, by encoding the conditional probability distribution of generator’s output at each iteration, GAN-TStega can achieve secret information hiding. Through this method, we achieve the statistical distribution fitting at the sentence level, thus enhancing the security of steganography system. Experiments show that our method has good performance.