Research map / Large language models / Instruction and fine-tuning
SecP-Tuning: Efficient Privacy-Preserving Prompt Tuning for Large Language Models via MPC
Jinglong Luo, Zhuo Zhang, Yehong Zhang, Shiyu Liu, Ye Dong, Hui Wang, Yue Yu, Xun Zhou, Zenglin Xu
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 |
|---|---|---|---|
| replace gradient-based tuning with forward-only tuning (no secure backward pass or optimizer) (ablation) | MPC prompt tuning · RoBERTa-large LAN, seq len 512 | total time (s): 882.075 -> 174.138 (Prompt Tuning vs FoT) | Improved |
| add privacy-preserving Random Feature Attention in place of softmax attention (ablation) | SecP-Tuning (FoT) · RoBERTa-large LAN, seq len 512 | total time (s) and communication (GB): 174.138 -> 55.172 s; 205.359 -> 56.545 GB | Improved |
| use SecP-Tuning instead of gradient-based prompt tuning in MPC | Prompt Tuning · RoBERTa-large LAN | speedup and communication: 16x speedup, 20x less communication | Improved |
| use gradient-based prompt tuning instead of full SFT under MPC | SFT · RoBERTa-large LAN | total time (s): 651.598 -> 882.075 | Hurt |
| use the proposed cosine protocol inside RFA (ablation) | privacy-preserving RFA · sequence lengths 64 to 512 | time and communication | Improved |
| lower bandwidth and higher latency (WAN 100Mbps/80ms vs LAN) | SecP-Tuning · RoBERTa-large WAN | speedup over Prompt Tuning: 16x (LAN) -> 34x (WAN) | Improved |
| initialize FoT with pre-trained prompt embeddings (ablation) | FoT · RTE and MRPC, 16 shots | accuracy / F1: RTE 52.59 -> 77.62; MRPC 61.56 -> 75.51 | Improved |
| use SecP-Tuning (FoT+RFA) instead of plaintext FoT without pre-trained prompt | FoT · Yelp Polarity, 16 shots | accuracy: 91.50 -> 85.23 | Hurt |
| use SecP-Tuning instead of gradient-based prompt tuning without pre-trained prompt | Prompt Tuning · SST-2 and Yelp P., 16 shots | accuracy: SST-2 68.23 -> 88.11; Yelp 61.02 -> 85.23 | Improved |
| use SecP-Tuning instead of SFT on harder tasks | SFT · AG's News, MRPC, RTE, Yelp P. 16 shots | accuracy / F1: AG's News 86.36 -> 81.27; Yelp 91.82 -> 85.23; MRPC 77.35 -> 75.33 | Mixed |
Closest papers by meaning
- DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer (ICLR 2024 · similarity 0.91)
- Privacy-Preserving Instructions for Aligning Large Language Models (ICML 2024 · similarity 0.89)
- ScaleOT: Privacy-utility-scalable Offsite-tuning with Dynamic LayerReplace and Selective Rank Compression (AAAI 2025 · similarity 0.88)
- DPZero: Private Fine-Tuning of Language Models without Backpropagation (ICML 2024 · similarity 0.87)
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.87)
- Differentially Private Bias-Term Fine-tuning of Foundation Models (ICML 2024 · 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.
