pip install the following:

torch
torchvision
numpy
timm
Pillow
Pandas

have the images of validation set from 2012 Imagenet (ILSVRC 2012) under the directory /dataset/imagenet/val
anonymously download validation images from: https://drive.google.com/file/d/10FQEQ1yOT9Jtd7hESfDbYcT2cqa-Eg5t/view?usp=sharing

download and put trained model under /lr-models
anonymously download trained model from: https://drive.google.com/file/d/1zSIYrBdqvDlcmQAnAsbli0CaCkoU7mvO/view?usp=sharing

to reproduce, run the following scripts. First script is to get clean score, second is for adv scores.
available attack_type options = [pgd, bim, pif, apgd, vmi, vni, anda]

Script 1:
python pipeline_inception --attack False

Script 2:
python pipeline_inception --attack True --attack_type pgd

after acquire the score files under results directory,
use auc.py to get AUC, by default it calculates inception-pgd pair.
