Robustness to Word- and Sentence-Level Rewriting in Text Steganography

Develop robustness mechanisms for current text steganography frameworks that reliably preserve message recovery under practical word- and sentence-level rewriting attacks, including deletion, insertion, and paraphrasing, when adversaries are unaware of the steganographic system’s segmentation rules.

Background

Existing robust linguistic steganography methods primarily defend against token-level perturbations. However, real-world adversaries may modify text at the word or sentence level without knowing the system’s segmentation rules, for example by deleting or inserting words or paraphrasing sentences. Such operations can disrupt token-level defenses and cause decoding failures. The paper presents its sentence-embedding-space framework as an approach to this broader robustness issue, but characterizes the challenge as unresolved for current text steganography frameworks.

This problem concerns establishing reliable message extraction despite semantic-preserving rewriting that may alter word boundaries, sentence structure, or surface realizations. Solving it would improve the practical applicability of linguistic steganography in communication environments where transmitted text can be actively edited or paraphrased.

References

Several limitations remain. The theoretical guarantees are asymptotic and rely on finite-state assumptions, leaving the text lengths needed for reliable detection unresolved. The boundary-achieving test also requires knowledge of the human class and contamination level. Future work could develop finite-sample guarantees and study how these bounds transfer to unseen modification mechanisms.

— Robust Detection of LLM-Generated Text under Contamination  (2609.29935 - Li et al., 24 Sep 2026) in Section 6, Conclusion, page 8

Consequently, while token-level defenses can improve robustness against token-boundary-aligned perturbations, robustness under practical word- and sentence-level rewriting remains a fundamental and unresolved challenge for current text steganography frameworks.

— Robust Coverless Linguistic Steganography via Sentence Embedding Space with Global Resynchronization  (2609.04970 - Xiong et al., 4 Sep 2026) in Section 1, Introduction

These results therefore provide limited evidence of configuration sensitivity: the constitution shifts the trade-off in the intended direction, but reliably separating neighboring priorities and meeting stringent robustness targets remain open challenges.

— Constitution-Guided Watermarking  (2610.09552 - Aremu et al., 7 Oct 2026) in Section 5, subsection “Sensitivity to Constitutional Priorities”