search

hfawaz / ijcnn19attacks

Adversarial Attacks on Deep Neural Networks for Time Series Classification
https://germain-forestier.info/src/ijcnn2019/
GNU General Public License v3.0
72 stars 28 forks source link

Sigmoid activation #2

Closed ulkursln closed 4 years ago

ulkursln commented 4 years ago

Hi, I am trying to use MSD function for custom ResNet trained with timeseries. It is a binary classification task, and the activation is sigmoid instead of softmax. Program throws following error: File "/home/../git/ijcnn19attacks/src/cleverhans_copy/utils_keras.py", line 132, in _get_softmax_name raise Exception("No softmax layers found") Exception: No softmax layers found

Is softmax activation mandatory, or just updating "softmax" with "sigmoid" in all functions of utils_keras.py enough?

ulkursln commented 4 years ago

By the way, since MDS function requires attacked dataset, I firstly run "python3 main.py attack fgsm " and that is when I got the error

hfawaz commented 4 years ago

Hi, Thanks for your interest. I believe that you should change the underlying code in order to use sigmoid. I would recommend to just have a softmax layer with two neurons in order to avoid modifying the code.