Research map / Adversarial robustness and security / Attacks on text and graphs
AliMark: Enhancing Robustness of Sentence-Level Watermarking Against Text Paraphrasing
Yuexin Li, Wenjie Qu, Linyu Wu, Yulin Chen, Yufei He, Tri Cao, Bryan Hooi, Jiaheng Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase next-sentence candidate budget Q from 8 to 64 (ablation) | AliMark · Booksum, GPT-3.5 paraphrase | TPR@5%: 29.6 -> 66.6 | Improved |
| increase next-sentence candidate budget Q from 8 to 64 (ablation) | AliMark · C4, GPT-3.5 paraphrase | TPR@5%: 15.0 -> 51.6 | Improved |
| increase candidate budget Q from 32 to 64 (ablation) | AliMark · C4, DIPPER paraphrase | TPR@5%: 43.8 -> 49.8 | Improved |
| use sentence embedder all-mpnet-base-v2 instead of multi-qa-mpnet-base-dot-v1 (ablation) | AliMark · Booksum, GPT-3.5 paraphrase | TPR@5%: 55.2 -> 66.6 | Improved |
| use sentence embedder all-mpnet-base-v2 instead of all-distilroberta-v1 (ablation) | AliMark · C4, DIPPER paraphrase | TPR@5%: 41.6 -> 49.8 | Improved |
| use simple one-step re-structurer instead of a learned LLM re-structurer (ablation) | AliMark · Booksum, GPT-3.5 paraphrase | TPR@5%: 56.0 -> 66.6 | Improved |
| use multi-step re-structurer instead of one-step (ablation) | AliMark · Booksum, GPT-3.5 paraphrase | TPR@5%: 66.4 -> 66.6 | No clear effect |
| add the Re-Structurer (RS) module (ablation) | AliMark · Booksum, strong paraphrasers (DIPPER, GPT-3.5) | TPR@5% | Improved |
| add adaptive bit sequence alignment (ablation) | AliMark · Booksum, strong paraphrasers (DIPPER, GPT-3.5) | TPR@5% | Improved |
| add the Re-Structurer module (ablation) | AliMark · Detection runtime, N=128 sentences | time (s): 0.07 -> 0.34 | Hurt |
| block size M=8 instead of M=2 (ablation) | AliMark · Booksum and C4 | TPR@5% | Improved |
| increase block size M from 2 to 16 (ablation) | AliMark · Booksum and C4 text quality | perplexity: 12.9 -> 12.8 | No clear effect |
Closest papers by meaning
- Robust Multi-bit Text Watermark with LLM-based Paraphrasers (ICML 2025 · similarity 0.90)
- BiMark: Unbiased Multilayer Watermarking for Large Language Models (ICML 2025 · similarity 0.88)
- Learning to Watermark: A Selective Watermarking Framework for Large Language Models via Multi-Objective Optimization (NeurIPS 2025 · similarity 0.87)
- GaussMark: A Practical Approach for Structural Watermarking of Language Models (ICML 2025 · similarity 0.87)
- On the Reliability of Watermarks for Large Language Models (ICLR 2024 · similarity 0.87)
- ArcMark: Distortion-Free Multi-Byte LLM Watermark via Optimal Transport (NeurIPS 2026 · similarity 0.86)
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.
