High-Performance Linguistic Steganalysis, Capacity Estimation and Steganographic Positioning

摘要

With the rapid development of natural language processing technology, various linguistic steganographic methods have been proposed increasingly, which may bring great challenges in the governance of cyberspace security. The previous linguistic steganalysis methods based on neural networks with word embedding layer could only extract the context-independent word-level features, which are insufficient for capturing the complex semantic dependencies in sentences, thus may limit the performance of text steganalysis. In this paper, we propose a novel linguistic steganalysis model. We first employ the BERT or Glove component to extract the contextualized association relationships of words in the sentences. Then we put these extracted features into BiLSTM to further get context information. We use the attention mechanism to find out local parts that may be discordant in text. Finally, based on these extracted features, we use the softmax classifier to decide if the input sentence is cover or stego. Experimental results show that the proposed model can achieve currently the best performance of text steganalysis and hidden capacity estimation. Further experiments found that proposed model can even locate where the secret information may be embedded in the text to a certain extent. To the best of our knowledge, we made the first attempt to achieve text steganography positioning in the field of text steganalysis.