Frustratingly Easy Edit-based Linguistic Steganography with a Masked Language Model

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5486–5492 June 6–11, 2021. ©2021 Association for Computational Linguistics 5486 Frustratingly Easy Edit-based Linguistic Steganography with a Masked Language Model Honai Ueoka Yugo Murawaki Sadao Kurohashi Graduate School of Informatics, Kyoto University hueoka@icn.cce.i.kyoto-u.ac.jp {murawaki, kuro}@i.kyoto-u.ac.jp Abstract With advances in neural language models, the focus of linguistic steganography has shifted from edit-based approaches to generation- based ones. While the latter’s payload capac- ity is impressive, generating genuine-looking texts remains challenging. In this paper, we re- visit edit-based linguistic steganography, with the idea that a masked language model of- fers an off-the-shelf solution. The proposed method eliminates painstaking rule construc- tion and has a high payload capacity for an edit-based model. It is also shown to be more secure against automatic detection than a generation-based method while offering better control of the security/payload capacity trade- off. 1 Introduction Steganography is the practice of concealing a mes- sage in some cover data such that an eavesdropper is not even aware of the existence of the secret message (Simmons, 1984; Anderson and Petitco- las, 1998). While images, videos, and audio have been dominant cover media (Fridrich, 2009), nat- ural language is a promising choice, thanks to the omnipresence of text (Bennett, 2004). Formally, the goal of linguistic steganography is to create a steganographic system (stegosystem) with which the sender Alice encodes a secret mes- sage, usually in the form of a bit sequence, into a text and the receiver Bob decodes the message, with the requirement that the text is so natural that even if transmitted in a public channel, it does not arouse the suspicion of the eavesdropper Eve. For a stegosystem that creates the text through transfor- mation, we refer to the original text as the cover text and the modified text as the stego text. A stegosys- tem has two objectives, security and payload ca- pacity. Security is the degree of how unsuspicious the stego text is while payload capacity is the size of the secret message relative to the size of the stego text. The two objectives generally exhibit a trade-off relationship (Chang and Clark, 2014). Edit-based approaches used to dominate the re- search on linguistic steganography. Arguably, the most effective approach was synonym substitu- tion (Chapman et al., 2001; Bolshakov, 2005; Taski- ran et al., 2006; Chang and Clark, 2014; Wilson and Ker, 2016), where a bit chunk was assigned to each member of a synonym group, for example, ‘0’ to marry and ‘1’ to wed. The cover text She will marry him was then modified to the stego text She will wed him such that the latter carried the secret bit sequence ‘1’. This conceptual simplicity was, however, over- shadowed by the complexity of linguistic phenom- ena such as part-of-speech ambiguity, polysemy, and context sensitivity. For this reason, edit-based approaches were characterized by the painstaking construction of synonym substitution rules, which were tightly coupled with acceptability checking mechanisms …

Content truncated to 500 words