PurduePAML / TrojanNN

Trojan Attack on Neural Network
https://purduepaml.github.io/TrojanNN/
181 stars 39 forks source link

Question about trigger generation #3

Open Karen0103 opened 5 years ago

Karen0103 commented 5 years ago

Hi,

This is a great work and thanks for sharing code.

In act_max.tvd.center_part.py, I think octaves hyperparams are used for denoise purpose. But I don't quite understand it. Could you please explain a little bit about it? Do I need to change these params when layers or neurons changes?

Looking forward to your reply. Thanks

Karen0103 commented 5 years ago

Hi @naiyeleo @shiqingma ,

Another question is about neuron selection.

The following code shows picking neuron that has the largest value of the sum of absolute weights connecting this neuron to the preceding layer. But the result neuron for layer 'fc6' is 2163 not 81 as mentioned in the paper. I am new to Caffe. Is there anything wrong here? Did I understand it correctly?
The original VGG model I used is http://www.robots.ox.ac.uk/~vgg/software/vgg_face/ as listed in README

Screen Shot 2019-05-03 at 9 24 55 PM Screen Shot 2019-05-03 at 9 24 45 PM