Homophonic Pun Generation with Lexically Constrained Rewriting

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 303–313, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 303 Near-imperceptible Neural Linguistic Steganography via Self-Adjusting Arithmetic Coding Jiaming Shen, Heng Ji, Jiawei Han Department of Computer Science, University of Illinois Urbana-Champaign, IL, USA {js2, hengji, hanj}@illinois.edu Abstract Linguistic steganography studies how to hide secret messages in natural language cover texts. Traditional methods aim to transform a secret message into an innocent text via lexical substitution or syntactical modification. Re- cently, advances in neural language models (LMs) enable us to directly generate cover text conditioned on the secret message. In this study, we present a new linguistic steganogra- phy method which encodes secret messages us- ing self-adjusting arithmetic coding based on a neural language model. We formally analyze the statistical imperceptibility of this method and empirically show it outperforms the pre- vious state-of-the-art methods on four datasets by 15.3% and 38.9% in terms of bits/word and KL metrics, respectively. Finally, human eval- uations show that 51% of generated cover texts can indeed fool eavesdroppers.1 1 Introduction Privacy is central to modern communication sys- tems such as email services and online social net- works. To protect privacy, two research fields are established: (1) cryptography which encrypts se- cret messages into codes such that an eavesdrop- per is unable to decrypt, and (2) steganography which encodes messages into cover signals such that an eavesdropper is not even aware a secret message exists (Westfeld and Pfitzmann, 1999; bin Mohamed Amin et al., 2003; Chang and Clark, 2014). One useful cover signal for steganography is natural language text because of its prevalence and innocuity in daily life. Traditional linguistic steganography methods are mostly edit-based, i.e., they try to directly edit the secret message and transform it into an innocent text that will not raise the eavesdropper’s suspi- cious eyes. Typical strategies include synonym 1Code and datasets are available at https://github. com/mickeystroller/StegaText. Plaintext Let’s meet in Room 9112A at 2am tonight Cover Text Boris Johnson unveils new COVID-19 lockdown rules A public channel monitored by Eve (eavesdropper) 0 1 0 0 1 0 1 0 1 1 0 1 0 0 1 0 1 1 … Ciphertext encrypt decrypt encode decode Alice (sender) Bob (receiver) Paired Steganography Encoder-Decoder Encoder Decoder Hello world ! 0 1 0 0 1 0 1 0 1 1 0 Hello world ! 0 1 0 0 1 0 1 0 1 1 0 Plaintext Let’s meet in Room 9112A at 2am tonight 0 1 0 0 1 0 1 0 1 1 0 1 0 0 1 0 1 1 … Ciphertext Cover Text Boris Johnson unveils new COVID-19 lockdown rules Figure 1: Linguistic steganography pipeline. substitution (Topkara et al., 2006), paraphrase sub- stitution (Chang and Clark, 2010), and syntactic transformation (Safaka et al., 2016), applied to vari- ous text media such as Email (Tutuncu and Hassan, 2015) and Twitter (Wilson et al., 2014). Although being able to maintain the grammatical correctness of output text, those edit-based …

Content truncated to 500 words