Research map / Vision-language and multimodal / Human and embodied modeling
Promising Multi-Granularity Linguistic Steganography by Jointing Syntactic and Lexical Manipulations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MMLS, a modification-based linguistic steganography framework that embeds secret bits in both syntactic space (distance-aware bins coding over paraphrase templates) and symbolic space (BERT-based lexical substitution with Huffman coding), raising embedding capacity while preserving semantics.
Problem
Modification-based linguistic steganography built on language models is limited to one embedding space (syntactic or symbolic), which constrains embedding capacity while preserving semantics and security.
Approach
Jointly embeds secret information at two granularities: a syntax-controlled paraphrase generator with spectral-clustering-based distance-aware syntactic bins for the syntactic level, followed by BERT masked-word substitution with autoregressive Huffman encoding for the lexical level, plus a consistency checker.
Key results
- MMLS reaches 0.6305 bpw with SIM 0.7377 and detection Acc 0.7205 (k=4), roughly doubling the capacity of LSCD (0.3327 bpw) at similar quality.
- MMLS achieves the highest SIM among compared methods, about 5% higher than other baselines, and the lowest Acc in the k=4 setting per the paper's claims.
- LSCD (cosine-distance bins) gives best BLEU/MAUVE and lowest detection among syntactic-only variants, e.g. BLEU 0.4765 and MAUVE 0.9323 at H=4, k=4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add lexical substitution to syntactic-space steganography | MMLS · QQP-Pos test, k=4 | bpw: 0.3327 -> 0.6305 | Helped |
| Add lexical substitution to syntactic-space steganography | MMLS · QQP-Pos test, k=4 | SIM: 0.7904 -> 0.7377 | Hurt |
| Use DBC-CD instead of DBC-LD (ablation) | Syntactic-space steganography · QQP-Pos test, k=4, H=4 | SIM: 0.7576 -> 0.7904 | Helped |
| Use DBC-CD instead of DBC-TED (ablation) | Syntactic-space steganography · QQP-Pos test, k=4, H=4 | SIM: 0.7698 -> 0.7904 | Helped |
| Increase syntactic payload k from 1 to 4 (ablation) | LSCD · QQP-Pos test, H=4 | BLEU: 0.7887 -> 0.4765 | Hurt |
| Increase syntactic payload k from 1 to 4 (ablation) | LSCD · QQP-Pos test, H=4 | SIM: 0.9187 -> 0.7904 | Hurt |
| Increase masking interval s from 2 to 3 (ablation) | MMLS, k=3, τ=0.1 · QQP-Pos test | bpw: 0.5561 -> 0.4277 | Hurt |
| Increase masking interval s from 2 to 3 (ablation) | MMLS, k=3, τ=0.1 · QQP-Pos test | BLEU: 0.3900 -> 0.4586 | Helped |
| Increase candidate-word threshold τ from 0.1 to 0.3 (ablation) | MMLS, k=3, s=2 · QQP-Pos test | bpw: 0.5561 -> 0.3975 | Hurt |
| Increase candidate-word threshold τ from 0.1 to 0.3 (ablation) | MMLS, k=3, s=2 · QQP-Pos test | BLEU: 0.3900 -> 0.4527 | Helped |
| Increase syntactic payload k from 3 to 4 (ablation) | MMLS, s=2, τ=0.1 · QQP-Pos test | bpw: 0.5561 -> 0.6305 | Helped |
| Increase syntactic payload k from 3 to 4 (ablation) | MMLS, s=2, τ=0.1 · QQP-Pos test | BLEU: 0.3900 -> 0.3299 | Hurt |
Surprising findings
- Table 1 shows MMLS has lower BLEU (0.3299) and MAUVE (0.8585) than LSCD, so combining lexical substitution trades text quality for capacity.
Limitations
- Text quality and anti-steganalysis security are lower than the syntactic-only LSCD variant.
- Evaluation uses a single steganalysis detector (Peng et al. 2021) and a single dataset (QQP-Pos).
Experimental setup
Datasets: QQP-Pos
Benchmarks: QQP-Pos test set
Metrics: bpw BLEU MAUVE SIM Acc F1
Baselines: PhraseLS SPLS HISS LSCD
Builds on: HISS SI-SCP Ueoka et al. 2021 BERT substitution steganography SPLS spectral clustering
Closest papers by meaning
- STEAD: Robust Provably Secure Linguistic Steganography with Diffusion Language Model (NeurIPS 2025 · similarity 0.87)
- STLC-KG:A Social Text Steganalysis Method Combining Large-Scale Language Models and Common-Sense Knowledge Graphs (AAAI 2025 · similarity 0.84)
- Robust Multi-bit Text Watermark with LLM-based Paraphrasers (ICML 2025 · similarity 0.83)
- GaussMark: A Practical Approach for Structural Watermarking of Language Models (ICML 2025 · similarity 0.82)
- A Semantic Invariant Robust Watermark for Large Language Models (ICLR 2024 · similarity 0.82)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.