Research map / Large language models / Language, safety and interpretability
FLIPS: Instance-Fingerprinting for LLMs via Pseudo-random Sequences
Richardeau Gurvan, Gohar Dashyan, Erwan Le Merrer, Gilles Tredan
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 |
|---|---|---|---|
| use instance-level fingerprinting (FLIPS) instead of adapted LLMmap | instance identification · 237 instances, closed-set | accuracy: 35% -> 96% | Improved |
| replace XGBoost classifier with other classifiers (ablation) | FLIPS · closed-set, Nt=8 | accuracy: 96.8% -> 68.1% (Gradient Boosting) | Hurt |
| increase verification queries Nt from 1 to 8 (ablation) | FLIPS · open-set | accuracy: 64% -> 90% | Improved |
| increase verification queries for LLMmap from 4 to 8 (ablation) | LLMmap-IF · closed-set | accuracy | No clear effect |
| use different token pairs across queries rather than the same pair repeatedly (ablation) | FLIPS · closed-set | accuracy | Improved |
| raise sampling temperature from 0.4 to 1.0 (ablation) | target instances · open-set, Nt=8 | recall: 99% -> 87% | Hurt |
| abliterated variant of a model (ablation) | FLIPS · open-set, Nt=8 | recall: 77% average vs 87%-99% for other variations | Hurt |
| truncate all bit sequences to fixed 100 bits (ablation) | FLIPS · closed-set and open-set | accuracy | No clear effect |
| use the 0-1 token pair instead of arbitrary token pairs (ablation) | FLIPS · closed-set, Nt=1 | fingerprinting accuracy | Hurt |
| use alpha-quantile threshold from training data instead of oracle test threshold (ablation) | FLIPS open-set · open-set, Nt=8 | AUROC: 0.947 -> 0.938 | No clear effect |
| better randomness of generated sequences (higher NIST score) | instances · closed-set, Nt=1 | fingerprinting accuracy | Hurt |
Closest papers by meaning
- Fingerprinting Inference Systems of Large Language Models (NeurIPS 2026 · similarity 0.90)
- HuRef: HUman-REadable Fingerprint for Large Language Models (NeurIPS 2024 · similarity 0.86)
- REEF: Representation Encoding Fingerprints for Large Language Models (ICLR 2025 · similarity 0.85)
- Scalable Fingerprinting of Large Language Models (NeurIPS 2025 · similarity 0.85)
- Language Models Can Predict Their Own Behavior (NeurIPS 2025 · similarity 0.85)
- Queries, Representation & Detection: The Next 100 Model Fingerprinting Schemes (AAAI 2025 · similarity 0.84)
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.
