LZW-CIE: a high-capacity linguistic steganography based on LZW char index encoding

With the effect of digitalization, the transfer of all text documents over the Internet rather than human transmission has increased, and this situation has revealed the idea that text documents can be used as a carrier that can safely store information. Realizing that methods such as word-line shifting, usage of spaces, replacement of the word with its synonym are fragile against steganalysis, led to new searches and it was determined that deep learning models were more resistant to detecting the presence of hidden words. In this study, the text generation based on the information that is wanted to be hidden without a carrier text, both at word and character level, was performed. Arithmetic coding, perfect tree and Huffman coding methods were used as secret information embedding methods in text generation based on word level. In this part of the study, bidirectional LSTM architecture with attention mechanism was created as language model. In text generation based on character level, a new secret information embedding algorithm is created by combining the LZW compression algorithm with the Char Index (LZW-Char Index Encoding) method. The character-level model is created as a result of using the encoder-decoder architecture together with bidirectional LSTM and Bahdanau attention. The proposed method was evaluated from the perspectives of information embedding efficiency, information imperceptibility and hidden information capacity. As a result of the experiments, it was determined that the method exceeded the state-of-the-art performance and was more resistant to steganalysis.

Automated summary

With the advancement of digitalization, the use of Internet for transmitting text documents instead of human transmission has grown. This has led to the idea that text documents can serve as a secure means of storing information. Researchers have found that deep learning models are more resistant to steganalysis compared to traditional methods such as word-line shifting and synonym replacement. In this study, text generation techniques were employed to hide information both at the word and character level, using methods such as arithmetic coding and Huffman coding. The proposed method outperformed existing techniques in terms of information embedding efficiency and resistance to steganalysis.