search

wgrathwohl / JEM

Project site for "Your Classifier is Secretly an Energy-Based Model and You Should Treat it Like One"
Apache License 2.0
415 stars 63 forks source link

How to generate Distal Adversarial Examples in paper? #6

Closed sndnyang closed 3 years ago

sndnyang commented 4 years ago

Hi, thank you for your great research. I have a problem about the distal adversaries.

Did you use the code like https://github.com/bethgelab/AnalysisBySynthesis/blob/441479b231fbd6a43615c10c7c68ccc86c31ae44/scripts/attacks.py for it?

if it is, can you share your configuration for distal adversaries?

opti = torch.optim.SGD([a_helper], lr=1, momentum=0.95)
confidence_level = model.confidence_level    # abs 0.0000031, CNN 1439000, madry 60, 1-NN 0.000000000004
logits_scale = model.logit_scale                      # ABS 430, madry 1, CNN 1, 1-NN 5

Thank you so much!