https://github.com/ermongroup/ods
Code for "Diversity can be Transferred: Output Diversification for White- and Black-box Attacks"
https://github.com/ermongroup/ods
Last synced: 4 months ago
JSON representation
Code for "Diversity can be Transferred: Output Diversification for White- and Black-box Attacks"
- Host: GitHub
- URL: https://github.com/ermongroup/ods
- Owner: ermongroup
- Created: 2020-02-12T01:20:44.000Z (over 5 years ago)
- Default Branch: master
- Last Pushed: 2020-11-02T08:34:54.000Z (almost 5 years ago)
- Last Synced: 2025-06-20T03:03:22.075Z (4 months ago)
- Language: Python
- Homepage:
- Size: 1.73 MB
- Stars: 53
- Watchers: 8
- Forks: 5
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Output Diversified Sampling (ODS)
This is the github repository for the NeurIPS 2020 paper "[Diversity can be Transferred: Output Diversification for White- and Black-box Attacks](https://arxiv.org/abs/2003.06878)".
## Requirement
Please install PyTorch, pickle, argparse, and numpy
## Running experiments
### ODS for score-based black-box attacks
The following experiments combine ODS with [Simple Black-Box Attack (SimBA)](https://arxiv.org/abs/1905.07121).
#### Evaluation:
The evaluation is held for 5 sample images on ImageNet (images are already resized and cropped).
```shell
# untargeted settings with ODS:
python blackbox_simbaODS.py --num_sample 5 --ODS
# targeted settings with ODS:
python blackbox_simbaODS.py --num_sample 5 --num_step 30000 --ODS --targeted
```
___
### ODS for decision-based black-box attacks
The following experiments combine ODS with [Boundary Attack](https://arxiv.org/abs/1712.04248).
#### Additional Requirement
Please install Foolbox, Python>=3.6
#### Evaluation:
The evaluation is held for 5 sample images on ImageNet (images are already resized and cropped).
```shell
# untargeted settings with ODS:
python blackbox_boundaryODS.py --num_sample 5 --ODS
# targeted settings with ODS:
python blackbox_boundaryODS.py --num_sample 5 --ODS --targeted
# untargeted settings with random sampling:
python blackbox_boundaryODS.py --num_sample 5
# targeted settings with random sampling:
python blackbox_boundaryODS.py --num_sample 5 --targeted
```
#### Acknowledgement
Our codes for Boundary Attack are based on [Foolbox repo](https://github.com/bethgelab/foolbox).
___
### ODS for initialization of white-box attacks (ODI)
The following experiments combine ODI with PGD attack.
#### Training of target model (Adversarial Training):
```shell
python whitebox_train_cifar10.py --model-dir [PATH_TO_SAVE_FOLDER] --data-dir [PATH_TO_DATA_FOLDER]
```
#### Evaluation PGD attack with ODI:
```shell
# Evaluate PGD attack with ODI:
python whitebox_pgd_attack_cifar10_ODI.py --ODI-num-steps 2 --model-path [PATH_TO_THE_MODEL] --data-dir [PATH_TO_DATA_FOLDER]
# Evaluate PGD attack with naive random initialization (sampled from a uniform distribution):
python whitebox_pgd_attack_cifar10_ODI.py --ODI-num-steps 0 --model-path [PATH_TO_THE_MODEL] --data-dir [PATH_TO_DATA_FOLDER]
```
#### Acknowledgement
Our codes for white-box attacks are based on [TRADES official repo](https://github.com/yaodongyu/TRADES).
## Citation
If you use this code for your research, please cite our paper:
```
@inproceedings{tashiro2020ods,
title={Diversity can be Transferred: Output Diversification for White- and Black-box Attacks},
author={Tashiro, Yusuke and Song, Yang and Ermon, Stefano},
booktitle={Advances in Neural Information Processing Systems},
year={2020}
}
```