Research map / Large language models / Instruction and fine-tuning
Provably Protecting Fine-Tuned LLMs from Training Data Extraction while Preserving Utility
Tom Segal, Yuval Elovici, Asaf Shabtai
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 |
|---|---|---|---|
| aggregate over relative probability distributions instead of raw probabilities (CP-Delta to CP-Delta_r) | CP-Delta · Canary (MathAbstracts) | mean exposure: 8.8 -> 3.6 | Improved |
| add base-model smoothing (pre-fine-tuning base) to CP-Delta_r (ablation) | CP-Delta_r · Canary (MathAbstracts) | mean exposure: 3.6 -> 1.4 | Improved |
| add base-model smoothing (pre-fine-tuning base) to CP-Delta_r (ablation) | CP-Delta_r · PII extraction (Names2IDs) | full ID extraction rate (FER): .06 -> <.001 | Improved |
| add base-model smoothing (constant-logit base) to CP-Delta_r (ablation) | CP-Delta_r · PII extraction (Names2IDs) | full ID extraction rate (FER): .06 -> .05 | Improved |
| add base-model smoothing (pre-fine-tuning base) to CP-Delta_r (ablation) | CP-Delta_r · TTE (MathAbstracts) | attack AUC: .80 -> .76 | Improved |
| add base-model smoothing to CP-Delta_r (ablation) | CP-Delta_r · MathAbstracts | 99th percentile k_x: .98 -> .004 | Improved |
| use relative probabilities instead of raw probabilities (CP-Delta to CP-Delta_r) | CP-Delta · MathAbstracts | 99th percentile k_x: 11.6 -> .98 | Improved |
| replace RPD aggregation by raw-probability CP-Delta on the smoothed distributions (ablation) | SCP-Delta_r · Canary (MathAbstracts) | mean exposure: 1.4 -> 8.1 | Hurt |
| reduce smoothing parameter m from 10 to 1 (ablation) | SCP-Delta_r · MathAbstracts validation | next-token accuracy: 0.495 -> 0.39 | Hurt |
| increase smoothing parameter m from 10 to 1000 (ablation) | SCP-Delta_r · MathAbstracts validation | next-token accuracy: 0.495 | No clear effect |
| increase smoothing parameter m (ablation) | SCP-Delta_r · Canary (MathAbstracts) | protection | Hurt |
| insert canary 1 time versus 3 times (ablation) | CP-Delta_r · Canary (MathAbstracts) | mean exposure: 2.6 -> 3.6 | No clear effect |
Closest papers by meaning
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.89)
- Representation Noising: A Defence Mechanism Against Harmful Finetuning (NeurIPS 2024 · similarity 0.89)
- Scaling Laws for Differentially Private Language Models (ICML 2025 · similarity 0.89)
- Forget to Flourish: Leveraging Machine-Unlearning on Pretrained Language Models for Privacy Leakage (AAAI 2025 · similarity 0.88)
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.88)
- DR-Encoder: Encode Low-rank Gradients with Random Prior for Large Language Models Differentially Privately (AAAI 2025 · similarity 0.88)
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.
