A practical and effective approach to large-scale automated linguistic steganography

Abstract

Several automated techniques exist to transform ciphertext into text that “looks like” natural-language text while retaining the ability to recover the original ciphertext. This transformation changes the ciphertext so that it doesn’t attract undue attention from, for example, attackers or agencies or organizations that might want to detect or censor encrypted communication. Although it is relatively easy to generate a small sample of quality text, it is challenging to be able to generate large texts that are “meaningful” to a human reader and which appear innocuous. This paper expands on a previous approach that used sentence models and large dictionaries of words classified by part-of-speech [7]. By using an “extensible contextual template” approach combined with a synonym-based replacement strategy, much more realistic text is generated than was possible with NICETEXT.