Elamur Research⌕
☰
home / papers / 2026 / Linguistic Steganography via Self-Adjusting Asymmetric Number System

Linguistic Steganography via Self-Adjusting Asymmetric Number System

Paper · 2026 · Yiting Liu, Chungen Xu, Fei Yang et al.

Summary

Generative linguistic steganography hides a secret bitstream inside text produced by a language model. Existing coding schemes (fixed-length, Huffman, arithmetic coding, ADG) give little control over payload size and tend to sacrifice naturalness for capacity. The authors introduce SA-ANS, which they describe as the first generative stego method built on Asymmetric Numeral Systems (ANS).

At every generation step, the candidate token pool is resized to fit the model's output distribution. A truncated, nucleus-style selection keeps the smallest high-probability set that satisfies an imperceptibility threshold ν. The user sets an embedding precision L, which determines how many bits each token carries. Sender and receiver first share a short context, and the receiver then reverses the ANS state updates to recover the message.

A theoretical analysis argues that the scheme decodes more stably than arithmetic coding and costs less to compute than Huffman coding or ADG. On four benchmark datasets, SA-ANS beats the baselines in perplexity, KL divergence, steganalysis resistance and human ratings. This makes it a practical, model-agnostic option for covert communication.

Key points

  • Core idea: SA-ANS combines ANS-based probabilistic coding with adaptive candidate selection.
    • Two main settings control the trade-off between capacity and naturalness: embedding precision L and threshold ν.
    • The candidate set grows only when needed, so rare, low-ranked tokens are mostly avoided.
  • Three-stage protocol:
    • Preparatory stage: sender and receiver share a context.
    • Embedding stage: a pretrained LM and the self-adjusting ANS encoder produce the stego text.
    • Extraction stage: the receiver decodes in sync with the sender.
    • Pseudocode for embedding and extraction is given in the appendices.
  • Evaluation:
    • Baselines: FLC, HC, AC and ADG.
    • Criteria: embedding efficiency, linguistic quality (PPL/SPPL), statistical similarity (D_KL), robustness to steganalysis, and human judgment.
  • Architecture test: SA-ANS was run with GPT-2, LLaMA-3.1-8B and Qwen-3-8B.
    • The two 8B models gave clearly better SPPL than GPT-2, especially at higher ν.
    • D_KL differences across models were small and inconsistent.
  • Resources: Code is available at github.com/liuyiting900/SAANS. The paper also covers dual-use ethics and recommends watermarking or traceable decoding and access control.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Nanjing University of Science and TechnologyNanjing · China14 works · 1,590 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksComputer Vision and Pattern Recognition · 2,869 works

Cite

Cite as Yiting Liu, Chungen Xu, Fei Yang, Pan Zhang and Linlong Wang. 2026. Linguistic Steganography via Self-Adjusting Asymmetric Number System. Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.1162/coli.a.22

@article{liu2026linguistic,
  title = {{Linguistic Steganography via Self-Adjusting Asymmetric Number System}},
  author = {Yiting Liu, Chungen Xu, Fei Yang, Pan Zhang and Linlong Wang},
  journal = {Computational Linguistics},
  year = {2026},
  doi = {10.1162/coli.a.22},
  url = {https://doi.org/10.1162/coli.a.22}
}

References

23 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. AlSabhany, Ahmed A., Ahmed Hussain Ali, Farida Ridzuan, A. H. Azni, and Mohd Rosmadi Mokhtar. 2020. Digital audio steganography: Systematic review, classification, and analysis of the current state of the art. Computer Science Review, 38:100316. https://doi.org/10.1016/j.cosrev.2020.100316 Bolshako… ↗
  2. Duda, Jarek. 2013. Asymmetric numeral systems: Entropy coding combining speed of Huffman coding with compression rate of arithmetic coding. arXiv preprint arXiv:1311.2540. — [arXiv:1311.2540](https://arxiv.org/abs/1311.2540) find
  3. Fang, Tina, Martin Jaggi, and Katerina Argyraki. 2017. Generating steganographic text with LSTMS. arXiv preprint arXiv:1705.10742. https://doi.org/10.18653/v1/P17-3017 ↗
  4. Fridrich, Jessica. 2009. Steganography in Digital Media: Principles, Algorithms, and Applications. Cambridge University Press. https://doi.org/10.1017/CBO9781139192903 Go, Alec, Richa Bhayani, and Lei Huang. 2009. Twitter sentiment classification using distant supervision. CS224N Project Report, St… ↗
  5. Hermann, Karl Moritz, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. Advances in Neural Information Processing Systems, 28.
  6. Huang, Siyuan, Xavier Garcia, Thomas Scialom, et al. 2024. LlaMa 3: Open foundation and instruction-tuned language models. Meta AI. https://llama.meta.com/llama3 ↗
  7. Huang, Yong Feng, Shanyu Tang, and Jian Yuan. 2011. Steganography in inactive frames of VoIP streams encoded by source Codec. IEEE Transactions on Information Forensics and Security, 6(2):296–306. https://doi.org/10.1109/TIFS.2011.2108649 Huffman, David A. 1952. A method for the construction of min… ↗
  8. Kadhim, Inas Jawad, Prashan Premaratne, Peter James Vial, and Brendan Halloran. 2019. Comprehensive survey of image steganography: Techniques, evaluations, and trends in future research. Neurocomputing, 335:299–326. https://doi.org/10.1016/j.neucom.2018 .06.075 Kang, Huixian, Hanzhou Wu, and Xinpen… — [doi:10.1016/j.neucom.2018](https://doi.org/10.1016/j.neucom.2018) ↗
  9. Karpov, Artem, Tinuade Adeleke, Seong Hah Cho, and Natalia Perez-Campanero. 2025. The steganographic potentials of language models. arXiv preprint arXiv:2505.03439. — [arXiv:2505.03439](https://arxiv.org/abs/2505.03439) find
  10. Kuznetsov, Oleksandr, Kyrylo Chernov, Aigul Shaikhanova, Kainizhamal Iklassova, and Dinara Kozhakhmetova. 2025. DeepStego: Privacy-preserving natural language steganography using large language models and advanced neural architectures. Computers, 14(5):165. https://doi.org/10.3390 /computers14050165… find
  11. Liu, Pengfei, Xipeng Qiu, and Xuanjing Huang. 2016. Recurrent neural network for text classification with multi-task learning. arXiv preprint arXiv:1605.05101. find
  12. Low, Steven H., Nicholas F. Maxemchuk, Jack T. Brassil, and Lawrence O’Gorman. 1995. Document marking and identification using both line and word shifting. In Proceedings of INFOCOM’95, volume 2, pages 853–860. https://doi.org/10.1109/INFCOM.1995 .515956 Luo, Yubo, Yongfeng Huang, Fufang Li, and Ch… — [doi:10.1109/infcom.1995](https://doi.org/10.1109/infcom.1995) ↗
  13. Maas, Andrew, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142–150. Mikolov, Tomas, Ma…
  14. Niu, Yan, Juan Wen, Ping Zhong, and Yiming Xue. 2019. A hybrid R-BILSTM-C neural network based text steganalysis. IEEE Signal Processing Letters, 26(12):1907–1911. https://doi.org/10.1109/LSP.2019.2953953 Perry, Neil, Sanket Gupte, Nishant Pitta, and Lior Rotem. 2025. Robust steganography from larg… — [arXiv:2504.08977](https://arxiv.org/abs/2504.08977) ↗
  15. Petitcolas, Fabien A. P. and Stefan Katzenbeisser. 2000. Information Hiding Techniques for Steganography and Digital Watermarking (Artech House Computer Security Series). find
  16. Artech House. Qwen Team. 2024. Qwen3: A new generation of large language models from Alibaba Cloud. https://github.com/QwenLM/Qwen. ↗
  17. Alibaba Cloud. Radford, Alec, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI Blog, 1(8):9. find
  18. Reid, Mark D. and Robert C. Williamson. 2009. Generalised Pinsker inequalities. arXiv preprint arXiv:0906.1244. — [arXiv:0906.1244](https://arxiv.org/abs/0906.1244) find
  19. Rissanen, Jorma and Glen G. Langdon. 1979. Arithmetic coding. IBM Journal of Research and Development, 23(2):149–162. https://doi.org/10.1147/rd .232.0149 Shen, Jiaming, Heng Ji, and Jiawei Han. 2020. Near-imperceptible neural linguistic steganography via self-adjusting arithmetic coding. arXiv prep… find
  20. ArXiv. Wayner, Peter. 1992. Mimic functions. Cryptologia, 16(3):193–214. https://doi.org/10.1080/0161-119291866883 Wen, Juan, Xuejing Zhou, Ping Zhong, and Yiming Xue. 2019. Convolutional neural network based text steganalysis. IEEE Signal Processing Letters, 26(3):460–464. https://doi.org/10.1109/… — [doi:10.1080/0161-119291866883](https://doi.org/10.1080/0161-119291866883) ↗
  21. Yang, Zhongliang, Xueshun Peng, Yongfeng Huang, and Chinchen Chang. 2018c. A novel method of speech information hiding based on 3D-magic matrix. arXiv preprint arXiv:1809.03010. — [arXiv:1809.03010](https://arxiv.org/abs/1809.03010) find
  22. Yang, Zhongliang, Ke Wang, Jian Li, Yongfeng Huang, and Yu-Jin Zhang. 2019. TS-RNN: Text steganalysis based on recurrent neural networks. IEEE Signal Processing Letters, 26(12):1743–1747. https://doi.org/10.18280/ts.360405 Yang, Zhong Liang, Si-Yu Zhang, Yu-Ting Hu, Zhi-Wen Hu, and Yong-Feng Huang.… ↗
  23. Zhou, Xuejing, Wanli Peng, Boya Yang, Juan Wen, Yiming Xue, and Ping Zhong. 2021. Linguistic steganography based on adaptive probability distribution. IEEE Transactions on Dependable and Secure Computing, 19(5):2982–2997. https://doi.org/10.1109/TDSC.2021 .3079957 Ziegler, Zachary M., Yuntian Deng,… ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liu-2026-linguistic-steganography-self-adjusting-asymmetric/