Explaining and harnessing adversarial examples (2015)
- 1. Explaining and Harnessing Adversarial Examples (2015) Ian J. Goodfellow, Jonathon Shlens, Christian Szegedy @mikibear_ 朱 襴 170118
- 4. 蟇一 手覺
- 5. Intriguing properties of neural networks (2013, Christian Szegedy at el.)
- 6. 覿襯襯 企 ろ語 覈 覓願碓 襯 螳企 ex) Alexnet, VGG, ResNet, Inception...
- 7. 朱語 れ螻 螳 螳レ (Examples)襯 覿襯 覈 : "願 'ろ' 螳" ( 螳) 覿襯 覈 : "願 'ろ' 螳螳 "
- 8. ?????????
- 10. 伎 語伎 襦 覿襯 覲 讌 confidence襦 る襯 讌
- 11. , 讌 るジ 讌. '朱 覿襯 讌' ' 語伎'襯 覃 '覈覦燕蟆 る襯 讌' 襷 も 蟆 覲 朱語 讌.
- 12. 蠏朱 讌 螳螳 襷 朱慨企? 螳 (99%) "譟" (99%)
- 13. 蠏朱 讌 螳螳 襷 朱慨企? 螳 (99%) "譟" (99%)
- 14. 蠏碁 企 語伎襯 企至 蟆 谿場 ?
- 15. 覓企 Optimization Problem...
- 16. 覓企 Optimization Problem... 覿襯 覈語 Black-box 覲 企語 語伎 朱襖
- 17. 1) 企語 語伎襯 る襯襯 殊狩る 語伎襯 谿城, 2) Norm 螳 蟆 谿場 襴覃...
- 18. 蠏 れ Optimization 蠍磯 蟇碁 . 朱語 L-BFGS襯 蟇瑚 給. 蠏朱 覓語 Non-convex朱れ
- 19. Explaining and Harnessing Adversarial Examples (2015, Ian J. Goodfellow, Jonathon Shlens & Christian Szegedy) れ 豌朱 ...
- 20. 譬 ｸ蟆 企 Adversarial Example襯 谿場 蟾?
- 21. Linear Model... 襯 螳ロ 蟆 れ Decision Boundary襯 蟆 蠍磯 term 谿場朱癌 (磯殊 覈語 レ high-dimension殊襦 企 襯 谿剰鍵 讌.)
- 22. Non-linear model linear 語伎 る Adversarial Example 襷り鍵 "The linear view of adversarial examples suggests a fast way of generating them. We hypothesize that neural networks are too linear to resist linear adversarial perturbation."
- 23. "neural networks are too linear"
- 25. 蠏碁蟾 Non-linearり 覦れ讌 覈語 襷 linear perturbation ｌ伎 蠏 覈語 蟾讌る, 蠏 覈語 豢覿 linearり 覲 も 企 襷.
- 26. Non-linear model linear 語伎 る Adversarial Example 襷り鍵 Backpropagation朱 覓企 所 蟲 Gradient
- 30. LeNet, 覦
- 31. 豌覯讌 覓, '企蟆 詞伎 Adversarial Example 螳り れ 覈語 旧る 企蟾?'
- 33. 蟆磯, 'Adversarial Example 豌伎 螻手 企 訖襷 , Model Generalization 螻朱 . 讌 蠏 焔レ Dropout覲企る .'
- 34. (蟆) 讌襷, 覈覈 るジ 朱碁れ 覲企 螻手 螻, 讌 るジ Adversarial Example 語り 譬 蟆所伎 覿覿 螳給.
- 35. 覯讌 覓, '蠏碁覃 譬 Non-linear RBF network 企り?'
- 36. , 譬 覲 一危一 訣蟆 谿願 . 讀, Adversarial Example 譬 robust譯.
- 37. 碁讌 覓, 'Ensembel 蠍磯 磯 譬 讌 蟾?' -> るれ. る讌 覓, '誤 手朱 distortion 蟇碁伎 旧 る 譬 讌 蟾?' -> るれ.
- 38. 朱語 蟆磯, 1) Universal approximation theorem 譟危 覈 覈語 Adversarial Example 襷蠍一 覓 Linear 2) 蠏朱 Adversarial Example襦 覈語 旧る 譬
- 39. References, 1) Intriguing properties of neural networks https://arxiv.org/abs/1312.6199 2) Explaining and Harnessing Adversarial Examples https://arxiv.org/abs/1412.6572 3) Adversarial Examples http://www.iro.umontreal.ca/~memisevr/dlss2015/goo dfellow_adv.pdf
- 40. 襴 伎 蟇磯 譴 觜語 蟆曙 れ殊語! @mikibear