跳到主要内容Skip to main content

机器遗忘Machine unlearning

Machine unlearning

  • 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

让模型「忘掉」指定训练数据或知识,而不必从零重训。Make a model forget specified training data or knowledge without retraining from scratch.

这是什么What it is

目标是在删除某个数据点后,模型行为接近「从未见过它」。

常见做法包括梯度上升、影响函数定位、教师蒸馏替代。

要同时检查遗忘是否彻底,以及模型整体能力是否被破坏。

The target is behaviour close to “never saw that data point”.

Common methods include gradient ascent, influence-function localisation and teacher distillation.

Check both forgetting completeness and collateral damage to general ability.

为什么需要知道Why it matters

用户行使删除权、版权方撤回授权、敏感数据误入训练集,都需要可验证的遗忘。Deletion rights, withdrawn licences and leaked sensitive data all need verifiable forgetting.

出处Sources
  • Machine Unlearning: A Survey(arXiv)Machine Unlearning: A Survey (arXiv 2303.04211) ↗