T-GRU: conTextual Gated Recurrent Unit model for high quality Linguistic Steganography

Linguistic steganography based on text auto-generation technology has been developed greatly in recent years. Most of current researches utilized neural networks to achieve high quality text generation. However, generated text of current researches only depends on the initial word but considers no context imperceptibility, which weaken the steganography imperceptibility severely. In this paper, we propose a conTextual Gated Recurrent Unit (T-GRU) model for text steganography. T-GRU makes use of the good sequence modeling characteristics of Seq2Seq model, inserts GRU into it, and adds an attention mechanism at Decoder to generate steganographic text that conforms to the context. We trained our model with a Chinese lyric dataset which reflects different context information with different singers. We designed several experiments to prove the correctness of this method from the perspective of information hiding efficiency, information imperceptibility and information hiding ability. The experimental results show that the proposed method can generate steganographic texts with more contextual connections and achieve higher imperceptibility.