Can you share the Hyperparameters of evaluated attacks in the leaderborad?

[vtddggg]

Where can I get the Hyperparameters (such as steps, n_restarts) of evaluated methods in leaderboard ?.

I test Deepfool to attack TRADES and got acc of 54%, but in your leaderboard, Deepfool linf only got 61.38%. It seems strange. Maybe some of hyperparameters are different. I want to use some of your result in my paper if you can share more informantion.

[yaodongyu]

Thanks for your interest in our work.

For evaluating the Deepfool attack, we directly applied the default hyperparameters in Foolbox (https://github.com/bethgelab/foolbox). It is possible that you could achieve better attack results by tuning the parameters (as long as satisfies the perturbation constraint).

Let me know if you have any other questions. Yaodong

[vtddggg]

Thanks for your interest in our work.

For evaluating the Deepfool attack, we directly applied the default hyperparameters in Foolbox (https://github.com/bethgelab/foolbox). It is possible that you could achieve better attack results by tuning the parameters (as long as satisfies the perturbation constraint).

Let me know if you have any other questions. Yaodong

Thanks for your response. Sorry that there's one more question for me: The top1 attack ODI-PGD gets 53.01% robust accuracy, but I cant reproduce the result. If you know how many steps and num of restarts here? I try 100 steps and 20 restarts but only get 54.5%

[hongyanz]

Hi vtddggg,

The ODI-PGD adversarial example file on our GitHub leaderboard is submitted by Yusuke Tashiro and we do not know exactly the hyper-parameter details on how to generate these adversarial examples. Perhaps a best way is to ask the authors of ODI-PGD for your concern.