question

[sakura-king]

I would like to know how many percentage points the forgetting scores you mentioned are worse than DU, and whether there is a picture of the experimental results.

[BjoernNieth]

Hey, here is a picture with the experimental results. I hope that helps answer your question.

[sakura-king]

Thank you for providing the images. However, this should be a test of the accuracy of the clean dataset, and the accuracy of adversarial samples is the subject of your paper. Should the accuracy of adversarial samples on the CIFAR-10 test set be used as a judging method?

[BjoernNieth]

These are the results of the preliminary experiments, where we decided on a data importance metric based on the performance in standard training. We cannot test for adversarial accuracy because the networks were not trained with adversarial training. Here is the part of the paper describing our experiment with forgetting scores:

[sakura-king]

I got it, thank you very much