Linguistic Steganography Using Automatically Generated Paraphrases

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 591–599, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Linguistic Steganography Using Automatically Generated Paraphrases Ching-Yun Chang University of Cambridge Computer Laboratory Ching-Yun.Chang@cl.cam.ac.uk Stephen Clark University of Cambridge Computer Laboratory Stephen.Clark@cl.cam.ac.uk Abstract This paper describes a method for checking the acceptability of paraphrases in context. We use the Google n-gram data and a CCG parser to certify the paraphrasing grammati- cality and fluency. We collect a corpus of hu- man judgements to evaluate our system. The ultimate goal of our work is to integrate text paraphrasing into a Linguistic Steganography system, by using paraphrases to hide informa- tion in a cover text. We propose automati- cally generated paraphrases as a new and use- ful source of transformations for Linguistic Steganography, and show that our method for checking paraphrases is effective at maintain- ing a high level of imperceptibility, which is crucial for effective steganography. 1 Introduction Steganography is concerned with hiding informa- tion in some cover medium, by manipulating prop- erties of the medium in such a way that the hidden information is not easily detectable by an observer (Fridrich, 2009). The covert communication is such that the very act of communication is to be kept se- cret from outside observers. A related area is Wa- termarking, in which modifications are made to a cover medium in order to identify it, for example for the purposes of copyright. Here the changes may be known to an observer, and the task is to make the changes in such a way that the watermark cannot easily be removed. There is a large literature on image steganogra- phy and watermarking, in which images are mod- ified to encode a hidden message or watermark. Image stegosystems exploit the redundancy in an image representation together with limitations of the human visual system. For example, a stan- dard image stegosystem uses the least-significant-bit (LSB) substitution technique. Since the difference between 11111111 and 11111110 in the value for red/green/blue intensity is likely to be undetectable by the human eye, the LSB can be used to hide infor- mation other than colour, without being perceptable by a human observer.1 A key question for any steganography system is the choice of cover medium. Given the ubiqui- tous nature of natural languages and electronic text, text is an obvious medium to consider. However, the literature on Linguistic Steganography, in which linguistic properties of a text are modified to hide information, is small compared with other media (Bergmair, 2007). The likely reason is that it is easier to make changes to images and other non- linguistic media which are undetectable by an ob- server. Language has the property that even small local changes to a text, e.g. replacing a word by a word with similar meaning, may result in text which is anomalous at the document level, or anomalous with respect to the state of the world. Hence find- ing linguistic transformations which …

Content truncated to 500 words