Research map / Large language models / Language, safety and interpretability
Perturb Your Data: Paraphrase-Guided Training Data Watermarking
Pranav Shetty, Mirazul Haque, Petr Babkin, Zhiqiang Ma, Xiaomo Liu, Manuela Veloso
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 |
|---|---|---|---|
| sample paraphrases with score-matched strategy instead of maximum Min-K%++ (ablation) | SPECTRA · Wiki (member) | p-value: 1.00 -> 4E-19 | Improved |
| sample paraphrases with score-matched strategy instead of maximum Min-K%++ (ablation) | SPECTRA · HN (member) | p-value: 3E-6 -> 3E-60 | Improved |
| sample paraphrases with score-matched strategy instead of random (ablation) | SPECTRA · PeS2o (member) | p-value: 1E-3 -> 2E-12 | Improved |
| sample paraphrases with score-matched strategy instead of random (ablation) | SPECTRA · Wiki (non-member) | p-value: 2E-5 -> 0.02 | Improved |
| use SPECTRA instead of STAMP watermarking | SPECTRA · PeS2o (member) | p-value: 0.15 -> 2E-12 | Improved |
| use SPECTRA instead of LLM-DI | SPECTRA · Wiki (member) | p-value: 0.02 -> 4E-19 | Improved |
| replace Min-K%++ with Loss in SPECTRA (ablation) | SPECTRA · Wiki (non-member) | p-value: 0.02 -> 6.2E-20 | Hurt |
| replace Min-K%++ with DC-PDD in SPECTRA (ablation) | SPECTRA · Wiki (member) | p-value: 4E-19 -> 1.0 | Hurt |
| use smaller scoring model Pythia-160m for ranking instead of OLMo-7b (ablation) | SPECTRA · PeS2o paraphrase ranking | Spearman rho: 0.826 -> 0.699 | Hurt |
| vary scoring model in the statistical test to OLMo-7b (ablation) | SPECTRA · PeS2o | p-value: 2E-12 -> 3E-5 | No clear effect |
| increase number of samples to 100-150 (ablation) | SPECTRA · all datasets | p-value: 100-150 samples | Improved |
| increase pretraining tokens beyond 3.5 billion (ablation) | SPECTRA · all datasets | p-value: 3.5 billion tokens | No clear effect |
Closest papers by meaning
- DE-COP: Detecting Copyrighted Content in Language Models Training Data (ICML 2024 · similarity 0.88)
- On the Reliability of Watermarks for Large Language Models (ICLR 2024 · similarity 0.87)
- STAMP Your Content: Proving Dataset Membership via Watermarked Rephrasings (ICML 2025 · similarity 0.87)
- On the Learnability of Watermarks for Language Models (ICLR 2024 · similarity 0.87)
- Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · similarity 0.87)
- Watermarking Makes Language Models Radioactive (NeurIPS 2024 · similarity 0.87)
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.
