Research map / Large language models / Language, safety and interpretability
LLM Fingerprinting via Semantically Conditioned Watermarks
Thibaud Gloaguen, Robin Staab, Nikola Jovanović, Martin Vechev
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 |
|---|---|---|---|
| replace query-key memorization with semantically conditioned watermark fingerprint | LLM fingerprint · LoRA/full finetuning (Alpaca, Dolly, OMI, WildChatFr) | FSR: 0.0 -> 1.0 (Llama3.2-1B, Alpaca LoRA) | Improved |
| use statistical watermark signal instead of exact key matching | LLM fingerprint · pruning 50% (Wanda, SparseGPT) | FSR: 0.0 -> 1.0 | Improved |
| use statistical signal over 1000 queries vs SF 1024 query-key pairs | LLM fingerprint · Int4 quantization | FSR: 0.0/0.2 (SF) -> 1.0 | Improved |
| use broader semantic domain (French) instead of fixed queries | LLM fingerprint · system prompts | FSR: 0.0 (IF/SF, Advertise, Llama3.2-1B) -> 1.0 | Improved |
| embed fingerprint with watermark distillation plus regularization | Llama3.2-1B · HumanEval | accuracy: 0.37 -> 0.31 | Hurt |
| embed fingerprint with watermark distillation plus regularization | three instruct LLMs · average of 8 benchmarks | accuracy: 0.63 -> 0.62 (Llama3.1-8B) | No clear effect |
| embed fingerprint without regularization loss (ablation) | fingerprint embedding · benchmarks | accuracy | Hurt |
| increase number of detection queries (ablation) | fingerprint detection · detection | detectability | Improved |
| change sampling temperature (0.4-1.0) | all fingerprints · temperature | FSR: 1.0 for IF, SF, Ours | No clear effect |
Closest papers by meaning
- Can Watermarked LLMs be Identified by Users via Crafted Prompts? (ICLR 2025 · similarity 0.87)
- SIF: Semantically In-Distribution Fingerprints for Large Vision-Language Models (CVPR 2026 · similarity 0.86)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.86)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.86)
- Learning to Follow In-Context Watermark Instructions via Self-Distillation (NeurIPS 2026 · similarity 0.86)
- Fingerprinting Inference Systems of Large Language Models (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.
