论文阅读笔记：Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey

论文阅读笔记：Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey

要点

1. (universal adversarial perturbation)[1]
2. 3D物理对抗样本

对抗样本生成

1. Box-constrained L-BFGS
   minρ min ρ

2. Fast Gradient Sign Method (FGSM)
   ρ=ϵsign(∇J(θ,Ic,l)) ρ = ϵ s i g n ( ∇ J ( θ , I c , l ) )
   one-shot 生成方式要求 ϵ ϵ 不能太小

3. Basic Iterative Method (BIM)
   Ik+1=clip(Ik+αsign(∇J(θ,Ik,l)) I k + 1 = c l i p ( I k + α s i g n ( ∇ J ( θ , I k , l ) )

4. Iterative Least-likely Class Method (ILCM)
   选取测试概率最低的那个类作为target class进行BIM targeted adversarial example generation

5. Jacobian-based Saliency Map Attack (JSMA)
6. One Pixel Attack
7. Carlini and Wagner Attacks (C&W)
8. DeepFool
9. Universal Adversarial Perturbations
   对几乎所有输入均有效、不依赖于输入的对抗扰动
10. UPSET(Universal Perturbations for Steering to Exact Targets)
11. ANGRI(Antagonistic Network for Generating Rogue Images)

对抗训练

参考文献

[1]