Research map / Federated learning and privacy / Differential privacy
Natural Identifiers for Privacy and Data Audits in Large Language Models
Lorenzo Rossi, Bartłomiej Marek, Franziska Boenisch, Adam Dziedzic
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 number of NIDs in the GSM8K suspect set from 50 to 100 (ablation) | NID dataset inference · GSM8K task-specific NIDs | p-value: 8.43\times10^{-4} -> 1.60\times10^{-6} | Improved |
| increase number of NIDs in the suspect set (ablation) | NID dataset inference · Pile and Dolma subsets | p-value | Improved |
| add stronger MIAs (CAMIA, SURP) to the DI features (ablation) | NID dataset inference · Pile DI | p-value | Improved |
| generate GIDs from a distribution mismatching the NIDs (e.g. changed casing) (ablation) | NID dataset inference · Pile DI | false positive rate | Hurt |
| use natural NIDs instead of injected canaries (ablation) | NID dataset inference · Pile DI | p-value | No clear effect |
| use longer or more structured NID formats (SHA-512, Java serialization) rather than MD5 (ablation) | NID dataset inference · DI by identifier type | p-value | Improved |
| increase candidate set cardinality c from 2 to 8 or 32 (ablation) | post-hoc DP auditing · Pythia DP-SGD finetuning, large epsilon | epsilon lower bound | Improved |
| increase candidate set cardinality c (ablation) | post-hoc DP auditing · Pythia DP-SGD finetuning, small epsilon | epsilon lower bound | Hurt |
| use Hinge instead of Loss as MIA (ablation) | NID vs GID MIA · Pile GitHub train | AUC: 71.8 -> 71.8 | No clear effect |
| use ReCALL instead of Loss as MIA (ablation) | NID vs GID MIA · Pile GitHub train | AUC: 71.8 -> 50.6 | Hurt |
Closest papers by meaning
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.91)
- The Canary’s Echo: Auditing Privacy Risks of LLM-Generated Synthetic Text (ICML 2025 · similarity 0.89)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.88)
- PrivAuditor: Benchmarking Data Protection Vulnerabilities in LLM Adaptation Techniques (NeurIPS 2024 · similarity 0.88)
- Unlocking Post-hoc Dataset Inference with Synthetic Data (ICML 2025 · similarity 0.87)
- Forget to Flourish: Leveraging Machine-Unlearning on Pretrained Language Models for Privacy Leakage (AAAI 2025 · 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.
