When using other than resnet18 as a proxy model, the accuracy seems wrong, especially in terms of white box about transferattack HOT 2 CLOSED

Hi, we have also noticed this issue! The reason for this problem is due to different versions of the model weights, the weights currently used are the latest version of the weights which are more robust, you can find this information at here.

You can load the old version of the weights so that it is possible to achieve a 100% attack success rate under the white box setting.
model = models.__dict__[model_name](weights="IMAGENET1K_V1") # Old weights, can achieve 100% attack success rates
model = models.__dict__[model_name](weights="DEFAULT") # New weights

from transferattack.

Thank you very much for your patient answer. I successfully solved my problem through the way you provided. Thank you again for your contribution to the Adversarial Attack direction

from transferattack.