Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks笔记

Code: https://github.com/mzweilin/EvadeML-Zoo

• Feature squeezing: reducing the color bit depth of each pixel and spatial smoothing.
• Framework:

• Adversarial examples attacks

  • $L_{p}norm$ attack
  • FGSM
  • BIM
  • DeepFool
  • JSMA
  • Carlini/Wagner attacks

• Defense:

  • Adversarial training
  • Gradient masking
  • Feature squeezing/input transformation

• Detecting adversarial examples

  • Sample statistics: maximum mean discrepancy
  • Training a detector
  • Prediction inconsistency: one adversarial example may not fool every DNN model.

• Color depth
  

• Spatial smoothing

  • Local smoothing
    
  • Non-local smoothing
    

• 这篇paper大篇幅都在survey adversarial attack and defense, 提出的方案很简单，并不effective

More Update：https://github.com/Billy1900/Backdoor-Learning