jiaxiaojunQAQ / Comdefend

The code for ComDefend: An Efficient Image Compression Model to Defend Adversarial Examples (CVPR2019)
112 stars 25 forks source link

Paper Fig.2 是否有誤? #11

Open mangoerya opened 5 years ago

mangoerya commented 5 years ago

请问paper Fig.2的Gaussian noise是否应加在ComCNN最后一层conv layer的sigmoid input, 而不是加在3 b​​its map? According to paragraph 3.5, "In particular, we add the random Gaussian noise ϕ (the mean of the gaussian noise is 0 and the variance of the gaussian noise is ϕ) to the output before the sigmoid function. "

jiaxiaojunQAQ commented 5 years ago

请问paper Fig.2的Gaussian noise是否应加在ComCNN最后一层conv layer的sigmoid input, 而不是加在3 b​​its map? According to paragraph 3.5, "In particular, we add the random Gaussian noise ϕ (the mean of the gaussian noise is 0 and the variance of the gaussian noise is ϕ) to the output before the sigmoid function. " 加在ComCNN的输出上