alinlab / consistency-adversarial Goto Github PK

Hi, thanks for your nice job! I have some questions about the paper:

1、I try to reproduce the results of the Table 3 in paper:
I train the model PreAct-ResNet-18 generates the seen adversarial samples with l∞ of ε = 8/255 in training time for training a robust model and also generates the unseen adversarial samples with different sized l∞ balls and other types of norm ball, e.g., l1, l2 for testing the robustness of the model with "unseen attacks". However, I find that the defense model trained with l∞ of ε = 8/255 achieves the better performance on the adversarial samples (generated by the trained defense model ) with l2 of ε = 300/255 than the results in paper, e.g. the accuracy on adversarial samples with l2 of ε = 300/255 is 38.48% (36.87% in Table 3 of the paper) only in 5th epoch. I want to know whether there is a problem in my generation of the unseen adversarial samples with l2 of ε = 300/255 and lead to the fake better results than the paper?

2、Whether the defense model in the Table 3 is trained on PGD 100 with l∞ of ε = 8/255? It seems that there is no related descriptions about it.

I would be grateful if you can help me with the above puzzles. Thank you!

This seems like a good model, and it'd be great if it was submitted to robustbench.