Avoiding detection on twitter: embedding strategies for linguistic steganography

Abstract

Any serious steganography system should make use of coding. Here, we investigate the performance of our prior linguistic steganographic method for tweets, combined with perfect coding. We propose distortion measures for linguistic steganography, the first of their kind, and investigate the best embedding strategy for the steganographer. These distortion measures are tested with fully automatically generated stego objects, as well as stego tweets filtered by a human operator. We also observed a square root law of capacity in this linguistic stegosystem.

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

“…Edit-based approaches used to dominate the research on linguistic steganography. Arguably, the most effective approach was synonym substitution (Chapman et al, 2001;Bolshakov, 2005;Taskiran 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'.…”

Generating Steganographic Text with LSTMs

“…The state-of-the-art Twitter stegosystem hides 2.8 bits of per tweet (Wil-son and Ker, 2016). Assuming 16.04 words per tweet 5 , our 4-bin system hides 32 bits per tweet, over 11 times higher than (Wilson and Ker, 2016).…”

Near-imperceptible Neural Linguistic Steganography via Self-Adjusting Arithmetic Coding

“…One major limitation of edit-based methods is that they cannot encode information efficiently. (Wilson and Ker, 2016) show the popular Cover-Tweet system (Wilson et al, 2014) can encode only two bits information in each transformed tweet on average. To address this limitation, generationbased methods try directly output the cover text based on the secret message.…”

“…Although being able to maintain the grammatical correctness of output text, those edit-based methods cannot encode information efficiently. For example, the popular CoverTweet system (Wilson and Ker, 2016) can only encode two bits of information in each tweet on average. Recent advances in neural language models (LMs) (Józefowicz et al, 2016;Radford et al, 2019;Yang et al, 2019a) have enabled a diagram shift from edit-based methods to generation-based methods which directly output a cover text by encoding the message reversibly in the choices of tokens.…”