---
id: 20260829-T0-06
title: "无需重训：差分隐私对齐实现LLM推理时隐私保护"
title_en: "Differentially Private Alignment Protects Privacy at LLM Inference Time"
url: https://ai.daily.yangsir.net/daily/20260829-T0-06
issue_date: 2026-08-29
publish_date: 2026-08-28T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2608.26324
---

# 无需重训：差分隐私对齐实现LLM推理时隐私保护

针对带有隐私敏感数据的推理任务，arXiv新论文提出了一种差分隐私的推理时对齐方法。该方法在Best-of-N采样基础上增加隐私预算控制，防止通过多次采样逆推训练数据。实验结果表明，与标准BoN相比，该方法在保护隐私的同时，回复质量下降不到5%，且无需重新训练模型。这为医疗、金融等需严格数据保护的领域提供了安全的大模型应用方案。

## English Version

**Differentially Private Alignment Protects Privacy at LLM Inference Time**

Addressing privacy in sensitive inference tasks, a new arXiv paper introduces differential privacy to Best-of-N sampling for inference-time alignment. It adds a privacy budget to prevent training data leakage via repeated sampling. Compared to standard BoN, quality drops under 5% while offering strong privacy guarantees, without retraining. This enables safer LLM deployment in healthcare and finance.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2608.26324)

**详情页**：https://ai.daily.yangsir.net/daily/20260829-T0-06

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*