| With the development of the Internet,people are more and more inclined to use the Internet for daily communication,how to protect the privacy and security of people’s communication in the public network has become a topic of increasing concern.Steganography mainly studies how to embed secret information into ordinary information carriers and hide its existence,thereby protecting its privacy and security.Text has become a common type of information in people’s daily life,and various text data need to be transmitted in social networks every day.In such a context,textbased steganographic methods are facilitated.Steganographic methods based on text generation can not only obtain grammatically correct and semantically coherent texts,but also have good concealment and high embedding rate.Most of the current neural network-based text generative steganography methods use autoregressive iterative generation to generate text sequences from left to right.This property makes it impossible to parallelize predictions at different locations,resulting in relatively high latency in inference。In addition,the current text generative steganography method uses the language model as the underlying generative model.Since the language model only learns the patterns and laws of the language based on the known text,it does not have the knowledge and ability to understand the language,so the generated steganography Text is semantically random and uncontrollable.Based on the research background,this paper conducts related research on the generation-based text steganography method.The main work is as follows:(1)A semantically consistent text steganography method based on generation is proposed.The model is trained with a large amount of abstract data,so that the model can generate steganographic text with sufficient information and consistent context.Through this method,the steganographic text can be controlled.Semantic content,making steganographic text more secure on public channels.In this paper,Bert and CRF are used to build the underlying summary generation model,and Bert learns the semantic features of the input text to calculate in parallel the scores of all possible labels at each moment,that is,the emission matrix and the corresponding emission scores,and obtains the transition matrix through CRF model learning optimization,in the decoding stage,the nonautoregress ive generation method is used,and the encoding path is calculated through the transfer matrix and the emission matrix under the control of secret information,and then the steganographic text with consistent semantics is generated.(2)An embedding algorithm suitable for non-autoregressive generative models is proposed to generate high-quality steganographic text by enlarging the search space as much as possible in the decoding stage.The proposed embedding algorithm is implemented based on the dynamic programming algorithm,and it still has very efficient time performance while expanding the search space.The proposed method embeds secret information in the CRF decoding stage.Based on the dynamic programming algorithm,it is used to find the optimal text encoding path under the current secret control.The embedding algorithm controls the label transfer between adjacent moments according to the binary bit stream,and then starts from the end label Backtrack to reconstruct the final target sequence.The feasibility and effectiveness of the method are proved through experimental verification and theoretical analysis.Experimental results show that the embedding method can improve the fluency of steganographic text,and has better security and concealment while ensuring semantic consistency and time performance. |
