---
id: 20260910-T0-10
title: "隐私保护新思路：PAC生成让模型输出也安全"
title_en: "PAC Privacy: Calibrating Noise to Output Disagreement in Private LLMs"
url: https://ai.daily.yangsir.net/daily/20260910-T0-10
issue_date: 2026-09-10
publish_date: 2026-09-09T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2609.05676
---

# 隐私保护新思路：PAC生成让模型输出也安全

针对私密文本训练的模型通过API服务时，隐私泄露主要发生在生成输出而非模型权重。新研究提出PAC（概率近似正确）私有自回归生成方法，通过将噪声校准到集成模型的预测分歧程度，在不显著降低生成质量的前提下保护每次输出的隐私。该方法比现有PMixED等方法更高效，为隐私敏感场景下的模型部署提供了新选项。

## English Version

**PAC Privacy: Calibrating Noise to Output Disagreement in Private LLMs**

When language models are trained on private text, privacy leaks through generated outputs, not exposed weights. This paper introduces PAC-Private Autoregressive Generation, which calibrates noise to ensemble disagreement, offering a more efficient privacy protection alternative to methods like PMixED. The approach enables safer deployment of private models via API without drastically sacrificing output quality.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2609.05676)

**详情页**：https://ai.daily.yangsir.net/daily/20260910-T0-10

---

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