Pre-trained PubDef model for ImageNet
ZhengyuZhao opened this issue · comments
Hi Chawin,
Nice to see this impressive work!
It seems that the pre-trained PubDef model for ImageNet has not been released yet. Is it possible to get it?
Best,
Zhengyu
Thank you for taking a look at our work! As requested, I updated the evaluation script (https://github.com/wagner-group/pubdef?tab=readme-ov-file#evaluation-script) and put up the pretrained models and some example attacks (for CIFAR-10/100 and ImageNet) at https://www.kaggle.com/datasets/csitawarin/pubdef-defending-against-transfer-attacks/. Please let us know if you run into any trouble with the code, the files, or anything else!
Thank you for taking a look at our work! As requested, I updated the evaluation script (https://github.com/wagner-group/pubdef?tab=readme-ov-file#evaluation-script) and put up the pretrained models and some example attacks (for CIFAR-10/100 and ImageNet) at https://www.kaggle.com/datasets/csitawarin/pubdef-defending-against-transfer-attacks/. Please let us know if you run into any trouble with the code, the files, or anything else!
Many thanks for the updates!
Another small question:
Where can I find the details about the L_inf norms you have used for training (PubDef) and testing on ImageNet?
I can only find the one for testing from https://github.com/wagner-group/pubdef/blob/main/scripts/example_test.sh#L26
Thanks!
Ah, thank you for catching that!
The numbers in the script are accurate, i.e., 8/255 for CIFAR-10 and CIFAR-100 and 4/255 for ImageNet.