Official Codes for Extracting Robust Models with Uncertain Examples (ICLR2023) [pdf]

Requirements

1. pytorch >= 1.9.0
2. torchvision
3. numpy
4. tqdm
5. torchattacks

Prepare Dataset

python split_set.py --dataset [cifar10, cifar100] --num 5000 --class_num [10, 100]

Victim Model and Pretrained Model

put your victim model into folder ./models/[cifar10, cifar100]/

put and rename your pretrained model into folder ./pretrained/tiny/[mobilenet, resnet, vgg, wrn]_pretrained/[mobilenet, resnet, vgg, wrn].pkl

We provide some pre-trained model on Tiny-ImageNet, you can find them in this [repo].

Run Model Extraction Attack

python extraction_attack.py --arch the architecture of the victim model -- ext the architecture of the pretrained model --dataset [cifar10, cifar100] --num 5000 --class_num [10, 100] --save1 the checkpoint name for the victim model --save the name you want to save your extracted model --exp experiment name --method BEST --aug [0,1]

Citation

If you find our work useful, please cite it:

@inproceedings{
li2023extracting,
title={Extracting Robust Models with Uncertain Examples},
author={Guanlin Li and Guowen Xu and Shangwei Guo and Han Qiu and Jiwei Li and Tianwei Zhang},
booktitle={International Conference on Learning Representations},
year={2023},
url={https://openreview.net/forum?id=cMAjKYftNwx}
}