跳到主要内容Skip to main content

对抗样本Adversarial example

Adversarial example

  • 核心概念Core concepts
  • 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

在输入上加人几乎看不出的扰动,让模型高置信度地答错。Add a perturbation humans barely notice and make the model answer wrong with high confidence.

这是什么What it is

扰动通常很小,肉眼或语义上几乎不变。

但它能让分类器/模型输出完全不同的结果,且置信度很高。

对抗训练是常见防御,但会降低干净样本上的精度,也挡不住新攻击。

The perturbation is small and often semantically invisible.

It can flip a classifier or model output while confidence stays high.

Adversarial training is a common defence, but it hurts clean accuracy and does not stop new attacks.

为什么需要知道Why it matters

图像识别、自动驾驶、内容审核都怕这种「看起来正常、模型却判错」的输入。Vision, driving and content moderation all fear inputs that look normal but are misclassified.

出处Sources
  • Intriguing properties of neural networks(arXiv)Intriguing properties of neural networks (arXiv 1312.6199) ↗