Research map / Federated learning and privacy / Differential privacy
PrivSV: Differentially Private Steering Vector for Large Language Models
Haocheng Yang, Xiang Cheng, Chenhao Sun, Pengfei Zhang, Sen Su
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 LNR2 compression before noise instead of naive noise on the full SV (ablation) | Mistral-7B SV · Refusal (eps=2) | accuracy: .521 -> .854 | Improved |
| add DPC reference-vector interpolation (alpha 0.2) to LNR2 (ablation) | PrivSV- · Refusal and Corrigibility (eps=2) | accuracy: Refusal .854 -> .959; Corrigibility .874 -> .963 | Improved |
| add POPD adaptive alpha on top of LNR2 + DPC (ablation) | PrivSV-* · Refusal and Corrigibility (eps=2) | accuracy: Refusal .959 -> .979; Corrigibility .963 -> .980 | Improved |
| add DPC with fixed alpha 0.2 to LNR2 (ablation) | PrivSV- on Survival · Survival (eps=2, Mistral-7B) | accuracy: .803 -> .756 | Hurt |
| use PrivSV instead of PSA | steering vector with DP · Refusal (eps=2, Mistral-7B) | accuracy: .477 -> .979 | Improved |
| use PrivSV instead of PSA | steering vector with DP · Sycophancy (eps=2, Mistral-7B) | accuracy: .635 -> .556 | Hurt |
| reduce latent dimension m' from 64 to 4 (ablation) | PrivSV HCC · Survival (eps=2, Mistral) | accuracy: .979 -> .827 | Hurt |
| increase latent dimension m' from 64 to 1024 (ablation) | PrivSV HCC · Coordination (eps=2, Mistral) | accuracy: .878 -> .518 | Hurt |
| increase privacy budget epsilon (ablation) | PrivSV · Coordination and Survival (4 LLMs) | accuracy | Improved |
| use PrivSV instead of PSA for generation | steering vector with DP · Survival GPT-4 score | GPT-4 score: 6.47 -> 6.86 | Improved |
| use PrivSV instead of PSA for membership inference defense | steering vector with DP · Hallucination, Qwen-2.5 7B | empirical epsilon: 1.6 -> 1.0 | Improved |
Closest papers by meaning
- Split-and-Denoise: Protect large language model inference with local differential privacy (ICML 2024 · similarity 0.89)
- DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer (ICLR 2024 · similarity 0.88)
- DPZero: Private Fine-Tuning of Language Models without Backpropagation (ICML 2024 · similarity 0.88)
- Privacy-Preserving Instructions for Aligning Large Language Models (ICML 2024 · similarity 0.88)
- Differentially Private Steering for Large Language Model Alignment (ICLR 2025 · similarity 0.88)
- Scaling Laws for Differentially Private Language Models (ICML 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.
