Research map / Federated learning and privacy / Privacy attacks and secure learning
PISA: Privacy-Preserving Split Adaptation with Model IP Protection
Haocheng Yang, Xiang Cheng, ZONGDA HAN, Pengjie Wang, Changkang Chi, 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 MRP pre-training of the server head (ablation) | Naive-RS · GLUE average, eps=10 | accuracy (%): 45.4 -> 61.9 | Improved |
| add DSC to MRP (ablation) | PISA- · GLUE average, eps=10 | accuracy (%): 61.9 -> 67.9 | Improved |
| add DSC to MRP (ablation) | PISA- · SST-2, eps=10 | accuracy (%): 70.5 -> 80.4 | Improved |
| add UGR to MRP+DSC (ablation) | PISA-* · GLUE average, eps=10 | accuracy (%): 67.9 -> 71.3 | Improved |
| add UGR to MRP+DSC (ablation) | PISA-* · MNLI, eps=10 | accuracy (%): 62.6 -> 68.6 | Improved |
| add UGR to MRP+DSC (ablation) | PISA-* · GLUE average, eps=14 | accuracy (%): 75.7 -> 78.1 | Improved |
| use PISA instead of Hosted-LDP at strict privacy | Hosted-LDP · GLUE average, eps=10 | accuracy (%): 48.3 -> 71.3 | Improved |
| use PISA instead of Hosted-LDP at relaxed privacy | Hosted-LDP · SST-2, eps=14 | accuracy (%): 86.4 -> 87.1 | No clear effect |
| add MRP pre-training to Hosted-LDP (ablation) | Hosted-LDP · GLUE eps=12 | accuracy (%): 59.4 -> 75.4 | Improved |
| full PISA vs Hosted-LDP+MRP (ablation) | Hosted-LDP+MRP · GLUE eps=12 | accuracy (%): 75.4 -> 84.5 | Improved |
| increase Tail depth from L=1 to L=3 (ablation) | PISA · SST-2, eps=10 | accuracy (%): 80.7 -> 81.2 | Improved |
| increase Tail depth from L=3 to L=4 (ablation) | PISA · SST-2, eps=10 | accuracy (%): 81.2 -> 80.2 | Hurt |
Closest papers by meaning
- Split-and-Denoise: Protect large language model inference with local differential privacy (ICML 2024 · similarity 0.89)
- Synthesizing Privacy-Preserving Text Data via Finetuning *without* Finetuning Billion-Scale LLMs (ICML 2025 · similarity 0.88)
- Privacy-Preserving Instructions for Aligning Large Language Models (ICML 2024 · similarity 0.88)
- PrivAuditor: Benchmarking Data Protection Vulnerabilities in LLM Adaptation Techniques (NeurIPS 2024 · similarity 0.88)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.88)
- DP-MemArc: Differential Privacy Transfer Learning for Memory Efficient Language Models (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.
