<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>审计方法学 on MessageDaily</title><link>https://inkeast.github.io/MessageDaily/tags/%E5%AE%A1%E8%AE%A1%E6%96%B9%E6%B3%95%E5%AD%A6/</link><description>Recent content in 审计方法学 on MessageDaily</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Wed, 09 Sep 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://inkeast.github.io/MessageDaily/tags/%E5%AE%A1%E8%AE%A1%E6%96%B9%E6%B3%95%E5%AD%A6/index.xml" rel="self" type="application/rss+xml"/><item><title>Split-LLM 隐私失效审计：返回梯度的零模式完美暴露真实数据 精读</title><link>https://inkeast.github.io/MessageDaily/posts/2026-09-09-split-llm-gradient-privacy-failure-paper-reading/</link><pubDate>Wed, 09 Sep 2026 00:00:00 +0000</pubDate><guid>https://inkeast.github.io/MessageDaily/posts/2026-09-09-split-llm-gradient-privacy-failure-paper-reading/</guid><description>独立研究者对两节点 split-LLM 训练的预注册审计：隐私损失忽略诱饵行→其返回梯度恰好为零→零模式逐帧完美暴露真实数据（9 种子 4096/4096 全中）。所有运行通过前向隐私检查与质量检查，加上返回梯度后同检查失败。逐行裁剪+加噪以 0.01 nats 代价封堵，并诚实声明五类未测攻击面。</description></item></channel></rss>