Research map / Large language models / Instruction and fine-tuning
Hey, That's My Model! Introducing Chain & Hash, An LLM Fingerprinting Technique
Mark Russinovich, Yanan Cai, Ahmed Salem
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 |
|---|---|---|---|
| add meta-prompt diversification to fingerprint training (ablation) | Chain & Hash (random questions) · Instruct models under adversarial meta-prompts | fingerprint strength: drops to 0 without -> >99% with | Improved |
| add meta-prompt diversification to fingerprint training (ablation) | Chain & Hash (natural questions) · Instruct models under adversarial meta-prompts | fingerprint strength: mean 82.8% | Improved |
| use random questions instead of natural questions (ablation) | Chain & Hash · Llama-3-8B-Instruct ChatDoc fine-tuning, no meta-prompt | required trials: 1 -> 25 | Hurt |
| use random questions instead of natural questions (ablation) | Chain & Hash · persistence under meta-prompts (Phi-3-Mini, Llama-3-8B-Instruct) | required trials: natural up to 21 -> random 1 | Improved |
| add random padding augmentation to fingerprint samples | Chain & Hash · post fine-tuning robustness | fingerprint robustness | Improved |
| expand fingerprint question set from 10 to 100 (ablation) | Chain & Hash · Llama-3-8B-Instruct meta-prompt persistence | fingerprint strength / convergence | Improved |
| reduce number of training meta-prompts to six (ablation) | Chain & Hash (random questions) · Llama-3-8B-Instruct | fingerprint strength | No clear effect |
| fingerprint model with Chain & Hash | Llama-3-8B · IFEval | IFEval change %: +46.7 | Improved |
| fingerprint model with Chain & Hash | Llama-3-8B and Phi-3-Mini-Instruct · GSM8K | GSM8K change %: -6.2 (Llama-3-8B natural), -3.21 (Phi-3 natural) | Hurt |
| apply INT8 quantization to fingerprinted model | Chain & Hash · fingerprinted models | fingerprint strength: <0.5% drop in most, <2.5% max | No clear effect |
| paraphrase output with GPT-4o before verification | Chain & Hash · stronger adversary | fingerprint detection: 99% -> 20% (2/10 fingerprints) | Hurt |
| paraphrase input with GPT-4o before querying | Chain & Hash · stronger adversary | fingerprint strength: 99% -> 79% | Hurt |
Closest papers by meaning
- Scalable Fingerprinting of Large Language Models (NeurIPS 2025 · similarity 0.89)
- SIF: Semantically In-Distribution Fingerprints for Large Vision-Language Models (CVPR 2026 · similarity 0.88)
- Queries, Representation & Detection: The Next 100 Model Fingerprinting Schemes (AAAI 2025 · similarity 0.87)
- HuRef: HUman-REadable Fingerprint for Large Language Models (NeurIPS 2024 · similarity 0.87)
- Fingerprinting Inference Systems of Large Language Models (NeurIPS 2026 · similarity 0.86)
- Intersecting-Boundary-Sensitive Fingerprinting for Tampering Detection of DNN Models (ICML 2024 · similarity 0.85)
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.
