Privacy Failure in Split-LLM Training, The Returned Gradient Nullifies the Decoys

  • 类型:arxiv
  • 标识:2609.04382
  • 链接:https://arxiv.org/abs/2609.04382
  • 主分类:engineering
  • 形态:method
  • 被引:2
  • 被引来源:Semantic Scholar
  • S2被引:2
  • OpenAlex被引:0
  • 影响力被引:0
  • TLDR:A systems-security case study of a two-node split-LLM training system whose privacy evaluation passed while leaving an observable channel untested, but the system is not thereby safe: five classes of attack, including those accumulating observations across training steps, were never measured.
  • OpenAlex ID:W7211870434
  • OpenAlex DOI:10.48550/arxiv.2609.04382
  • DOI:10.48550/arxiv.2609.04382
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://arxiv.org/pdf/2609.04382
  • OpenAlex更新:2026-09-10
  • 待LLM分类:否
  • 标题中文:Split-LLM 训练中的隐私失败:返回的梯度使诱饵失效
  • TLDR中文:本文对一个双节点 split-LLM 训练系统进行系统安全案例研究:其隐私评估通过,却遗留一条未被测试的可观测信道;系统因此并不安全——包括跨训练步骤累积观测在内的五类攻击从未被测量。
  • 来源文件:
  • /inbox/tom/_candidates/2026-09-09-agent-rag-longcontext-candidates.json
  • [OpenAlex backfill]
  • [S2 enrich]