Research map / Large language models / Instruction and fine-tuning
From Prompts to Responses: Dual-Sided Data Leakage and Defense in Split Large Language Models
Zixuan GU, Xiaojun Ye, Yang Liu
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 Dual-Sided Initialization (DSI) in place of vanilla init (ablation) | PIDI on Llama3-8B · Fin | AP(alpha=0.5): 0.697 -> 0.883 | Improved |
| add Patched Model Inversion (PMI) on top of DSI (ablation) | DSI+VMI · Fin Llama3-8B | AP(alpha=0.5): 0.716 -> 0.883 | Improved |
| add PMI to DSI on Qwen2.5-7B (ablation) | DSI+VMI · Dolly | AP(alpha=0.5): 0.814 -> 0.985 | Improved |
| add PMI to DSI on Llama3.2-3B Med (ablation) | DSI+VMI · Med | AP(alpha=0.5): 0.925 -> 0.901 | Hurt |
| use DSI initialization instead of MC initialization with VMI (ablation) | VMI · Fin Llama3-8B | AP(alpha=0.5): 0.423 -> 0.716 | Improved |
| increase patch length l (ablation) | PMI · Fin Llama3.2-3B | AP_input | Hurt |
| deepen split point from 3-3 to 6-6 without defense (ablation) | PIDI attack · Fin Llama3.2-3B | AP(alpha=0.5): 0.950 -> 0.726 | Hurt |
| reduce attacker auxiliary dataset size (ablation) | PIDI · Fin | AP | No clear effect |
| apply ADMI defense | PIDI attack on Llama3.2-3B · Fin | AP(alpha=0.5): 0.868 -> 0.017 | Improved |
| apply ADMI defense | Llama3.2-3B · Fin | MP (METEOR): 0.56 -> 0.516 | Hurt |
| use ADMI instead of DualGuard | PIDI defense on Qwen2.5-7B · Fin | AP(alpha=0.5): 0.572 -> 0.004 | Improved |
| apply ADMI defense | Qwen2.5-7B · Dolly | inference throughput (token/s/GPU): 24.892 -> 20.418 | Hurt |
Closest papers by meaning
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.91)
- Split-and-Denoise: Protect large language model inference with local differential privacy (ICML 2024 · similarity 0.91)
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.89)
- Better Language Model Inversion by Compactly Representing Next-Token Distributions (NeurIPS 2025 · similarity 0.89)
- DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer (ICLR 2024 · similarity 0.88)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · 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.
