Detection of Steganographic Techniques on Twitter

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2564–2569, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Detection of Steganographic Techniques on Twitter Alex Wilson and Phil Blunsom and Andrew D. Ker Department of Computer Science University of Oxford Oxford, OX1 3QD, UK {alex.wilson,phil.blunsom,andrew.ker}@cs.ox.ac.uk Abstract We propose a method to detect hidden data in English text. We target a system pre- viously thought secure, which hides mes- sages in tweets. The method brings ideas from image steganalysis into the linguis- tic domain, including the training of a feature-rich model for detection. To iden- tify Twitter users guilty of steganography, we aggregate evidence; a first, in any do- main. We test our system on a set of 1M steganographic tweets, and show it to be effective. 1 Introduction Consider this: two isolated prisoners, communi- cating by letters scrutinised by the prison warden. They cannot write openly about escape plots, and the warden destroys any written in code. They must hide the true message within the letter, us- ing steganography: the art of hiding information. In cover modification1steganography, a cover object is tweaked so that it carries a hidden mes- sage: this is called embedding, the tweaked ver- sion is the stego object, and the message is the pay- load (Fridrich, 2009). The message should not be detectable to any observer (the standard terminol- ogy for this is the warden, taken from the prisoner metaphor) who knows the system is deployed, but does not know the original cover. We are concerned here with linguistic steganog- raphy, in which the cover is a piece of text, and the message is embedded using textual transfor- mations intended to preserve the meaning of the original: synonym substitutions, syntactical trans- formations, etc. Note that we are not concerned 1There are other steganographic paradigms, not in scope. Translation based methods hide information in the automatic translation of the cover (e.g. Meng et al. (2011)). Cover gen- eration methods automatically produce text containing the payload (e.g. Chapman et al. (2001)). with the subset of linguistic steganography that hides in file formatting (e.g. white space, as in Por et al. (2008)), which has no security against an informed warden (in the case of hiding informa- tion by adding extraneous white space, the warden simply has to look for consistent irregular use of spaces to spot an active steganographer). The field suffers from a number of issues: com- pared to images, text covers have low capac- ity (Chang and Clark, 2010); certain methods are weak against human attackers (Grosvald and Orgun, 2011) (most paraphrase systems cannot guarantee perfectly fluent stego objects); finally, authors are generally concerned with the perfor- mance of the transformation (whether they pro- duce grammatically/semantically correct transfor- mations), rather than whether the generated stego objects are detectable or not (e.g. Chang and Clark (2010)). However, there is a new challenger in the field. We proposed a new linguistic stegosystem (Wilson et al., 2014) and verified its security against hu- man …

Content truncated to 500 words