jsikyoon / adv_attack_capsnet

Tensorflow Implementation of Adversarial Attack to Capsule Networks

Hi there 👋

I am Yoon, Jaesik (윤재식). I am currently working as a senior machine learning developer in SAP and usually researching a variety of ways to improve generality of Artificial Intelligence. This github is used to do that like sharing our proposed models or reproduced sourcecode of published papers. Hope you and your family be healthy who visits here and my logs be helpful! The details for works are in My HomePage / Google Scholar.

adv_attack_capsnet's Issues

bugs on max_epsilon setting

current implementation is using twice bigger max_epsilon setting than written one.

I will re-run the code and upload new result with bug fix.

Black or white box attacks?

Are the attacks reported on the frontpage black-box or white-box? In other words, what are the

(adversarial image producing model, testing model)

pairs for each attack?

Is this implementing Matrix Capsules with EM Routing?

I took a quick look at your code and it doesn't seem like it, but I wasn't sure.

I think it'd be interesting to try that, as quoting from its abstract: "Capsules also show far more resistance to white box adversarial attack than our baseline convolutional neural
network."

black-box attack generated by naiive model to CapsNets

RGB dataset (224*224)

How can we use ur code in other RGB dataset?
Suppose the structure of dataset is like that. it contains some sub-folder. Each sub-folder represents one class.

Class A:
0001.jpg 1
0002.jpg 1
Class B:
0001.jpg 2
0002.jpg 2