Research map / Large language models / Language, safety and interpretability
Reconstruction Attack-Resistant Inference Paradigm for LLM Cloud Services
Zipeng Ye, Wenjian Luo, Qi Zhou, Yubo Tang
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 |
|---|---|---|---|
| apply scaling defense to context | Llama-3-8B · HellaSwag (0-shot) | accuracy: 66.7 -> 65.4 | Hurt |
| apply scaling defense to context | Phi-3-14B · GSM8K (0-shot CoT) | accuracy: 91.1 -> 85.2 | Hurt |
| apply scaling defense to context | Llama-3-8B · HumanEval (pass@1) | pass@1: 55.5 -> 51.2 | Hurt |
| add 8-bit quantization of user-side layers (ablation) | defense, Mistral-7B · GSM8K (0-shot CoT) | accuracy: 46.8 -> 46.6 | No clear effect |
| add 4-bit quantization of user-side layers (ablation) | defense, Phi-3-14B · GSM8K (0-shot CoT) | accuracy: 85.2 -> 78.1 | Hurt |
| add 4-bit quantization of user-side layers (ablation) | defense, Mistral-7B · BoolQ (0-shot) | accuracy: 82.8 -> 42.6 | Hurt |
| perturb embeddings instead of scaling defense (ablation) | Mistral-7B · GSM8K (0-shot CoT) | accuracy: 46.8 -> 2.1 | Hurt |
| replace tokens with nearest neighbors instead of scaling defense (ablation) | Llama-3-8B · HumanEval (pass@1) | pass@1: 51.2 -> 0 | Hurt |
| add defense to 3-shot examples vs 1-shot undefended | Llama-3-8B · BBH CoT (selected tasks) | accuracy: 57.4 -> 67.5 | Improved |
| apply defense to MMLU few-shot examples | Llama-3-70B-AWQ · MMLU | accuracy (STEM): 77.7 -> 74.0 | Hurt |
Closest papers by meaning
- Split-and-Denoise: Protect large language model inference with local differential privacy (ICML 2024 · similarity 0.91)
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.91)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.89)
- Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory (ICLR 2024 · similarity 0.88)
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.88)
- Privacy Auditing of Large Language Models (ICLR 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.
