Research map / Federated learning and privacy / Privacy attacks and secure learning
OSNIP: Balancing the Privacy-Utility-Efficiency Trilemma in LLM Inference via Obfuscated Semantic Null Space
Zhiyuan Cao, Zeyu Ma, Chenhao Yang, HAN ZHENG, Mingang Chen
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 the key-conditioned perturbation (ablation) | OSNIP · Adaptive random-key attack on Qwen3-32B | ASR: 100.00% -> 0.00% | Improved |
| add the diversity loss L_div (ablation) | OSNIP · Adaptive random-key attack on Qwen3-32B | ASR: 100.00% -> 0.00% | Improved |
| add oracle key knowledge for the attacker (ablation) | OSNIP · Adaptive oracle-key attack on Qwen3-32B | ASR: 0.00% -> 100.00% | Hurt |
| use OSNIP instead of Cape | privacy-preserving inference · KNN-Top10 attack on Qwen3-32B | ASR: 0.578 -> 0.000 | Improved |
| use OSNIP instead of DYNText | privacy-preserving inference · 10-task average on Qwen3-32B | accuracy (Avg): 0.428 -> 0.748 | Improved |
| use OSNIP instead of InferDPT | privacy-preserving inference · WikiText-2 on Qwen3-32B | perplexity: 1155.74 -> 6.94 | Improved |
| increase model scale from Llama-3.2-1B to Qwen3-32B | OSNIP · KNN-Top10 attack | ASR: 0.066 -> 0.000 | Improved |
| increase model scale from Llama-3.2-1B to Qwen3-32B | OSNIP · WikiText-2 | perplexity increase over non-private: +31% -> +3% | Improved |
| use OSNIP instead of InferDPT | summarization · CNN/DailyMail ROUGE-L on Qwen3-32B | ROUGE-L: 0.197 -> 0.182 | Hurt |
| use OSNIP instead of InferDPT | summarization · CNN/DailyMail on Llama-3.2-1B | BERTScore: 0.868 -> 0.840 | Hurt |
| exclude stop words from vocabulary-matching attack ASR | OSNIP · Vocabulary-matching attack, Qwen3-32B L10 | ASR: 18.76 -> 7.23 | Improved |
Closest papers by meaning
- Encryption-Friendly LLM Architecture (ICLR 2025 · similarity 0.86)
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.86)
- Portcullis: A Scalable and Verifiable Privacy Gateway for Third-Party LLM Inference (AAAI 2025 · similarity 0.86)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.85)
- CipherPrune: Efficient and Scalable Private Transformer Inference (ICLR 2025 · similarity 0.85)
- Split-and-Denoise: Protect large language model inference with local differential privacy (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.
