Source-linked AI summary
Robust Coverless Linguistic Steganography via Sentence Embedding Space with Global Resynchronization
Lizhi Xiong, Yuping Lu, Jun Li, Ziqiang Li, Zhangjie Fu
TL;DR
Existing linguistic steganography methods struggle with word- and sentence-level perturbations and variable-length bit-slippage. The paper proposes sentence-embedding-space coding with hierarchical clustering and GRM, reporting stronger robustness while retaining effective capacity and security. A key trade-off is that greater LT-code redundancy reduces embedded-message capacity.
Problem
Existing methods remain vulnerable to word- and sentence-level perturbations, while variable-length schemes can suffer bit-slippage from local extraction-length mismatches.
Method
The framework encodes messages through hierarchical semantic subspaces in sentence embedding space and uses GRM with LT codes to recover from local symbol errors.
Results
The method’s extraction error remains below 2.7% under word-level attacks, while experiments report effective capacity, linguistic quality, and security against steganalysis.
Takeaways & Limitations
Sentence-space embedding and global resynchronization support robust coverless steganography under word- and sentence-level textual attacks.
Takeaways & Limitations
Increasing LT-code redundancy improves robustness but decreases the amount of actual secret message embedded, creating a robustness-capacity trade-off.
Abstract
from arXiv · showhide
Linguistic steganography enables covert communication through natural language. Existing methods heavily rely on token-level operations and struggle to maintain reliability under word- and sentence-level textual perturbations. Moreover, variable-length coding-based schemes are highly susceptible to bit-slippage under minor disturbances, as perturbations cause desynchronization between embedded and extracted bit sequences. To address these issues, we propose a robust coverless steganographic framework that operates in the sentence embedding space rather than the token space. Specifically, secret messages are encoded as hierarchical clustering paths in the sentence embedding space, which enhances decoding stability against word- and sentence-level textual perturbations. To tackle the bit-slippage problem, we introduce a Global Resynchronization Mechanism (GRM) that reframes variable-length bitstreams as discrete symbols anchored to semantic subspaces, decoupling local embedding failures from global message recovery. Experimental results demonstrate that under word- and sentence-level perturbations, our approach achieves substantial improvements in robustness, while maintaining effective embedding capacity and exhibiting strong resistance to statistical analysis.
1 Introduction
Text steganography supports covert communication, but token-oriented methods remain vulnerable to word- and sentence-level rewriting. This paper shifts embedding into sentence embedding space and adds mechanisms for efficient partitioning and global recovery.
- Text steganography embeds secret information in natural-language carriers, making it an important research area for covert communication.
- Existing modification- and retrieval-based methods are fragile because deletions or replacements can break shared dictionary or codebook correspondence.
- Word- and sentence-level attacks remain an unresolved challenge because token-level defenses mainly address perturbations aligned with token boundaries.
- The proposed framework embeds information in a shared sentence embedding space partitioned into hierarchical, codable semantic subspaces.
- HCM reduces hierarchical clustering complexity from exponential to linear, while GRM uses symbol erasures and LT codes for global recovery from local embedding failures.
2 Related Work
Related work spans retrieval-based, generative, and attack-resistant linguistic steganography. Existing approaches address naturalness, segmentation ambiguity, or local robustness, but retrieval-based methods have not systematically investigated robustness.
- Retrieval steganography selects unmodified carriers through shared datasets and coding rules, preserving the naturalness of transmitted text.
- Generative methods embed messages through language-model token probabilities, random prefixes, or sampling procedures designed to resemble ordinary text generation.
- SECC adds an LT-code-like error-correction layer for character-level erasures, while Syncpool addresses segmentation ambiguity by merging tokens with shared prefixes.
- GTSD, Winstega, and STEAD study active textual attacks using diffusion-based candidates, discontinuous embedding, or duplicated codes.
- Sentence embeddings represent semantically similar sentences with higher similarity and smaller distance, with vector components used to compute similarity.
- Encoder choices trade speed and resource demands against the dimensionality and text-processing capabilities of the resulting semantic space.
3 Proposed Method
The proposed framework constructs a shared sentence-embedding semantic space, encodes messages through hierarchical subspace paths, and uses global resynchronization to limit error propagation under perturbations. It combines semantic-space robustness with efficient hierarchical clustering and symbol-level recovery.
- The system assumes sender and receiver share a dataset, sentence encoder, LT-code identifiers, and a pseudorandom key for consistent semantic-space construction and extraction.
- Semantic-space construction: A shared sentence encoder maps sentences into a semantic space partitioned into disjoint subspaces, where subspace identity rather than exact sentence coordinates carries secret information.Perturbed sentences can remain decodable when they stay within the original cluster boundary.
- Hierarchical Clustering Mechanism: Hierarchical clustering distributes b secret bits across l layers, reducing the large-space clustering cost associated with creating 2^b distinguishable states.HCM recursively selects clusters along the steganographic path instead of clustering the entire space at once.
- Hierarchical Clustering Mechanism: With b̄ set to 2 or 3, the hierarchical procedure has overall time complexity O(N · l) rather than exponential dependence on the full branching space.The stated per-layer complexity is l · O(N · 2^b̄), with computation concentrated on the selected path.
- Subspace Sampling: Message-driven pseudorandom subspace sampling reduces distributional bias caused by unequal subspace sizes while preserving reproducible selection from a shared key.Sampling partitions [0, 1) into 2^b equal intervals and uses a key-derived pseudorandom position.
- Global Resynchronization: GRM converts variable-length local outcomes into fixed-size symbol blocks, reorganizes recovered blocks globally, and uses LT-code redundancy to recover messages despite missing blocks.This prevents a local decoding error from shifting all subsequent bit positions and isolates deviations at the sentence level.
4 Experiment
Experiments evaluate robustness, quality, embedding capacity, security, and ablations under word- and sentence-level attacks. The proposed method remains highly robust, while redundancy improves recovery at the cost of embedding capacity.
- Evaluation setup: The evaluation measures extraction error under deletion, replacement, insertion, swap, and paraphrase attacks, with attack intensity defined by the attacked-word ratio for word-level attacks.Paraphrase is treated as sentence-level, while the other listed attacks are word-level.
- Robustness results: Under word-level attacks, our method keeps P_E below 2.7% with three-times LT-code redundancy, while baseline error rates approach 50% as attacks intensify.The baseline methods become increasingly vulnerable as attack strength rises.
- Error analysis: For α=5, extraction errors arise from different mechanisms: BER dominates AC and ADG, while BLR dominates Discop and STEAD; resynchronization and semantic stability keep our method's BER and P_E low.Discop's zero-padding artificially lowers its apparent error rate because padded sequences are excluded from BLR calculations.
- Robustness results: Under paraphrase attacks, baseline methods remain vulnerable, with Discop and STEAD completely unable to decode, whereas our method retains strong robustness.Paraphrase provides the sentence-level robustness test.
- Quality, capacity, and security: Our method achieves lower PPL than generative methods, outperforms other methods in attack-scenario EER except FSBTS, and produces steganalysis accuracy close to 50%.The quality comparison uses PersonaChat, C4, and IMDB; the embedding-capacity comparison uses mixed attacks at α=5.
- Ablation and trade-offs: Increasing cluster count raises P_E while increasing capacity, whereas increasing LT-code redundancy reduces BER and improves recovery but decreases the amount of embedded secret message.Two-times redundancy keeps BER below 10% but leaves an error floor; three-times redundancy achieves near-perfect reconstruction across attack types and intensities.
5 Conclusion
The framework operates in sentence embedding space and combines HCM with GRM to improve robustness while preserving embedding capacity, linguistic quality, and security against steganalysis.
- The framework shifts steganographic operations from fragile token sequences to sentence embedding space.
- HCM efficiently partitions semantic space, while GRM addresses bit-slippage through LT codes.
- Experiments show substantial robustness improvements under textual perturbations while maintaining effective embedding capacity, superior linguistic quality, and strong security against steganalysis.