跳到主要内容Skip to main content

差分隐私Differential privacy

Differential privacy

  • 数据与检索Data & retrieval
  • 第四站 · 看懂Station 4 · Under the Hood
  • 引用核验 2026-11-19refs checked 2026-11-19
解决什么问题What problem it solves

在统计/训练结果里加可控噪声,让单个样本的存在与否无法被反推。Add controlled noise to statistics or training so one sample’s presence cannot be inferred.

这是什么What it is

用隐私预算 ε 衡量「加入一条数据会改变输出多少」。

常用机制是加高斯噪声或拉普拉斯噪声。

ε 越小越隐私,但模型效用通常越差。

Privacy budget ε measures how much one record can change the output.

Common mechanisms add Gaussian or Laplacian noise.

Smaller ε means more privacy but usually worse utility.

为什么需要知道Why it matters

训练数据里可能有病历、聊天记录。差分隐私提供可量化的隐私保证,而不是「删掉姓名」这种伪匿名。Training data may contain medical or chat records. Differential privacy gives a quantifiable guarantee, unlike name removal.

出处Sources
  • Deep Learning with Differential Privacy(arXiv)Deep Learning with Differential Privacy (arXiv 1607.00133) ↗