Elamur Research⌕
☰
home / papers / 2026 / Auto-Stega: An Agent-Driven System for Lifelong Strategy Evolution in LLM-Based Text Steganography

Auto-Stega: An Agent-Driven System for Lifelong Strategy Evolution in LLM-Based Text Steganography

Paper · 2026 · Jiuan Zhou, Yu Cheng, Yuan Xie et al.

Summary

Text steganography hides secret bits inside generated text. It must balance capacity, natural-looking output and resistance to automated detectors, and this "trilemma" gets harder as embedding rates rise. Existing LLM-based schemes such as ADG, Discop and LLM-Stega depend on fixed, manually designed strategies. Auto-Stega replaces these with an agent system that works only at inference time, with no training, fine-tuning or parameter access. A web searcher mines recent literature for candidate strategies and turns them into library entries. The system then retrieves and ranks strategies for each request, a steganography LLM writes the stego text, and a scorer LLM checks it against several criteria. A summarizer LLM records successful strategies back into the library, and on the receiving end a decoding LLM extracts the message using the shared strategy. For high embedding rates, the authors add PC-DNTE, an optional module that keeps token sampling close to the base model's distribution. On News, Movie and Tweet corpora, Auto-Stega lowers normalized perplexity by 42.2% and makes steganalysis about 1.2% less accurate than the strongest baseline. This shows that self-improving strategy libraries can ease the trilemma.

Key points

  • The closed loop runs in four stages: generate, evaluate, summarize, update. Library entries carry an "Applicable Scenarios" field used to filter candidates, and new candidates join only if their score exceeds a threshold S_T.
  • PC-DNTE is a plug-and-play mapping that aligns probability mass per bin, so quality holds up at high embedding rates.
  • Against steganalyzers trained on 10,000/1,000/1,000 cover–stego pairs, detection accuracy on Auto-Stega text stays near chance: SANet 51.05%, BiLSTM-Dense 49.65%, Bert-FT 52.35%.
  • In BERT-based sentence-level KLD, Auto-Stega scores lowest on News and second lowest on Movie and Tweet. On those two corpora, ADG and Discop are the respective leaders.
  • Human raters scored Auto-Stega above all three baselines for fluency, coherence and relevance. Planned future work covers multilingual and multimodal covers, adaptive steganalyzers and lower latency.

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

🎓East China Normal UniversityShanghai · China69 works · 2,490 citations🎓Shanghai Innovation InstituteShanghai · China34 works · 325 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksComputer Vision and Pattern Recognition · 3,486 works

Cite

Cite as Jiuan Zhou, Yu Cheng, Yuan Xie and Zhaoxia Yin. 2026. Auto-Stega: An Agent-Driven System for Lifelong Strategy Evolution in LLM-Based Text Steganography. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1612

@inproceedings{zhou2026auto,
  title = {{Auto-Stega: An Agent-Driven System for Lifelong Strategy Evolution in LLM-Based Text Steganography}},
  author = {Jiuan Zhou, Yu Cheng, Yuan Xie and Zhaoxia Yin},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1612},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1612}
}

References

41 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jinheon Baek, Sujay Kumar Jauhar, Silviu Cucerzan, and Sung Ju Hwang. 2025. ResearchAgent: Iterative research idea generation over scientific literature with large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Lingu…
  2. Mathias Benedek, Tanja Könen, and Aljoscha C Neubauer. 2012. Associative abilities underlying creativity. Psychology of Aesthetics, Creativity, and the Arts, 6(3):273. find
  3. Yew Ken Chia, Lidong Bing, Soujanya Poria, and Luo Si. 2022. RelationPrompt: Leveraging prompts to generate synthetic data for zero-shot relation triplet extraction. In Findings of the Association for Computational Linguistics: ACL 2022, pages 45–57.
  4. Jinyang Ding, Kejiang Chen, Yaofei Wang, Na Zhao, Weiming Zhang, and Nenghai Yu. 2023. Discop: Provably secure steganography in practice based on" distribution copies". In 2023 IEEE Symposium on Security and Privacy, pages 2238–2255. find
  5. Alec Go, Richa Bhayani, and Lei Huang. 2009. Twitter sentiment classification using distant supervision. CS224N project report, Stanford, 1(12):2009. find
  6. Mingzhi Hu and Hongxia Wang. 2025. Mutual information-optimized steganalysis for generative steganography. IEEE Transactions on Information Forensics and Security, 20:1852–1865. find
  7. Anuj Karpatne, Aryan Deshwal, Xiaowei Jia, Wei Ding, Michael Steinbach, Aidong Zhang, and Vipin Kumar. 2025. Ai-enabled scientific revolution in the age of generative ai: second nsf workshop report. npj Artificial Intelligence, 1(1):18. find
  8. Arthur Koestler. 1964. The act of creation. find
  9. Byung Cheol Lee and Jaeyeon Chung. 2024. An empirical investigation of the impact of chatgpt on creativity. Nature Human Behaviour, 8(10):1906–1914. find
  10. Mengdi Li, Kai Mu, Ping Zhong, Juan Wen, and Yiming Xue. 2019. Generating steganographic image description by dynamic synonym substitution. Signal Processing, 164:193–201. find
  11. Yiming Li, Kejiang Chen, Yaofei Wang, Xin Zhang, Guanjie Wang, Weiming Zhang, and Nenghai Yu. 2025. Coas: Composite audio steganography based on text and speech synthesis. IEEE Transactions on Information Forensics and Security, 20:5978–5991. find
  12. Guorui Liao, Jinshuai Yang, Weizhi Shao, and Yongfeng Huang. 2025. A framework for designing provably secure steganography. In 34th USENIX Security Symposium, pages 6837–6856. find
  13. Xiaogeng Liu, Peiran Li, G. Edward Suh, Yevgeniy Vorobeychik, Zhuoqing Mao, Somesh Jha, Patrick McDaniel, Huan Sun, Bo Li, and Chaowei Xiao. 2025. Autodan-turbo: A lifelong agent for strategy self-exploration to jailbreak llms. In International Conference on Learning Representations. find
  14. Daniil A Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. 2023. Autonomous chemical research with large language models. Nature, 624(7992):570– 578.
  15. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
  16. Yu Cheng, Jiuan Zhou, Jiawei Chen, Zhaoxia Yin, and Xinpeng Zhang. 2025. Rfnns: Robust fixed neural network steganography with popular deep generative models. arXiv preprint arXiv:2505.04116. — [arXiv:2505.04116](https://arxiv.org/abs/2505.04116) find
  17. Ziming Luo, Zonglin Yang, Zexin Xu, Wei Yang, and Xinya Du. 2025. Llm4sr: A survey on large language models for scientific research. arXiv preprint arXiv:2501.04306. 32211 9 find
  18. Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller. 2024. Augmenting large language models with chemistry tools. Nature Machine Intelligence, 6(5):525– 535. find
  19. Andrew Maas, 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.
  20. Laijin Meng, Xinghao Jiang, Qiang Xu, and Tanfeng Sun. 2025. A robust coverless video steganography based on two-level dct features against video attacks. IEEE Transactions on Multimedia, 27:6434–6448. find
  21. Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernockỳ, and Sanjeev Khudanpur. 2010. Recurrent neural network based language model. In Interspeech, volume 2, pages 1045–1048. find
  22. Rishabh Misra. 2022. News category dataset. arXiv preprint arXiv:2209.11429. — [arXiv:2209.11429](https://arxiv.org/abs/2209.11429) find
  23. Wanli Peng, Jinyu Zhang, Yiming Xue, and Zhenghong Yang. 2021. Real-time text steganalysis based on multi-stage transfer learning. IEEE Signal Processing Letters, 28:1510–1514. find
  24. Yinyin Peng, Yaofei Wang, Donghui Hu, Kejiang Chen, Xianjin Rong, and Weiming Zhang. 2024. Ldstega: Practical and robust generative image steganography based on latent diffusion models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3001–3009. find
  25. Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3982–3992.
  26. Kai Ruan, Xuan Wang, Jixiang Hong, Peng Wang, Yang Liu, and Hao Sun. 2024. Liveideabench: Evaluating llms’ divergent thinking for scientific idea generation with minimal context. arXiv preprint arXiv:2412.17596.
  27. Timo Schick and Hinrich Schütze. 2021. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943– 6951.
  28. Mohammad Shirali-Shahreza. 2008. Text steganography by changing words spelling. In 2008 10th international conference on advanced communication technology, volume 3, pages 1912–1913. find
  29. Jie Wang, Yihao Wang, Yihao Li, Ru Zhang, and Jianyi Liu. 2025a. Multi-classification of linguistic steganography driven by large language models. IEEE Signal Processing Letters, 32:1440–1444. find
  30. Zhuang Wang, Linna Zhou, Xuekai Chen, Zhili Zhou, and Zhongliang Yang. 2025b. Stlc-kg: A social text steganalysis method combining large-scale language models and common-sense knowledge graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 25461–25469. find
  31. Juan Wen, Xuejing Zhou, Ping Zhong, and Yiming Xue. 2019. Convolutional neural network based text steganalysis. IEEE Signal Processing Letters, 26(3):460–464. find
  32. Jiaxuan Wu, Zhengxian Wu, Yiming Xue, Juan Wen, and Wanli Peng. 2024. Generative text steganography with large language model. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10345–10353. find
  33. Yiming Xue, Jiaxuan Wu, Ronghua Ji, Ping Zhong, Juan Wen, and Wanli Peng. 2024. Adaptive domain-invariant feature extraction for cross-domain linguistic steganalysis. IEEE Transactions on Information Forensics and Security, 19:920–933. find
  34. Hao Yang, YongJian Bao, Zhongliang Yang, Sheng Liu, Yongfeng Huang, and Saimei Jiao. 2020. Linguistic steganalysis via densely connected lstm with feature pyramid. In Proceedings of the 2020 ACM Workshop on Information Hiding and Multimedia Security, pages 5–10. find
  35. Zhongliang Yang, Xiaoqing Guo, Ziming Chen, Yongfeng Huang, and Yujin Zhang. 2018. Rnn-stega: Linguistic steganography based on recurrent neural networks. IEEE Transactions on Information Forensics and Security, 14(5):1280–1295. find
  36. Siyu Zhang, Zhongliang Yang, Jinshuai Yang, and Yongfeng Huang. 2021. Provably secure generative linguistic steganography. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3046–3055. find
  37. Xin Zhang, Kejiang Chen, Na Zhao, Weiming Zhang, and Nenghai Yu. 2025a. Provably secure public-key steganography based on admissible encoding. IEEE Transactions on Information Forensics and Security, 20:3161–3175. find
  38. Ziping Zhang, Jiamin Zeng, Yan Xu, Xiaowei Yi, Yun Cao, and Changjun Liu. 2025b. Triple-stage robust audio steganography framework with aac encoding for lossy social media channels. In Proceedings of the ACM Workshop on Information Hiding and Multimedia Security, pages 131–141. find
  39. Andy Zhou, Kevin Wu, Francesco Pinto, Zhaorun Chen, Yi Zeng, Yu Yang, Shuang Yang, Sanmi Koyejo, James Zou, and Bo Li. 2025a. Autoredteamer: Autonomous red teaming with lifelong attack integration. arXiv preprint arXiv:2503.15754. — [arXiv:2503.15754](https://arxiv.org/abs/2503.15754) find
  40. Pengcheng Zhou, Zhengyang Fang, Zhongliang Yang, Zhili Zhou, and Linna Zhou. 2025b. Efficient streaming voice steganalysis in challenging detection scenarios. IEEE Transactions on Information Forensics and Security, 20:5966–5977. 32212 10 find
  41. Qing Zhou, Ping Wei, Zhenxing Qian, Xinpeng Zhang, Sheng Li, and Chuan Qin. 2025c. Conditional flow-based generative steganography. IEEE Transactions on Dependable and Secure Computing, 22(5):5632– 5647. 32213 11 find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhou-2026-auto-stega-agent-driven-system/