foolbox
foolbox copied to clipboard
Top-k misclassification on ImageNet
Running an untargeted attack on an ImageNet classifier results in a church being classified as a monastery instead. Do you maybe have a suggestion or a workaround to introduce top-k misclassification?
We did support top-k misclassification in Foolbox 1 and 2: https://github.com/bethgelab/foolbox/blob/v2.4.0/foolbox/criteria.py#L209 It has not yet been ported to Foolbox 3. Adding the criterion should be easy, making sure it works with certain attacks probably requires some more testing.
@jonasrauber thanks for the hint! I am mainly running experiments with C&W. Do you have any insights regarding this? Otherwise I will report again whether it worked.
@ines21 Is there an update from your side?