Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/jh-jeong/smoothmix
Code for the paper "SmoothMix: Training Confidence-calibrated Smoothed Classifiers for Certified Robustness" (NeurIPS 2021)
https://github.com/jh-jeong/smoothmix
adversarial-defense deep-learning machine-learning neurips-2021 paper randomized-smoothing robustness
Last synced: 3 months ago
JSON representation
Code for the paper "SmoothMix: Training Confidence-calibrated Smoothed Classifiers for Certified Robustness" (NeurIPS 2021)
- Host: GitHub
- URL: https://github.com/jh-jeong/smoothmix
- Owner: jh-jeong
- License: mit
- Created: 2021-10-26T05:22:58.000Z (about 3 years ago)
- Default Branch: main
- Last Pushed: 2022-09-27T02:20:10.000Z (about 2 years ago)
- Last Synced: 2023-10-20T20:44:55.476Z (about 1 year ago)
- Topics: adversarial-defense, deep-learning, machine-learning, neurips-2021, paper, randomized-smoothing, robustness
- Language: Roff
- Homepage: https://arxiv.org/abs/2111.09277
- Size: 16.8 MB
- Stars: 20
- Watchers: 1
- Forks: 3
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- Awesome-Mixup - [Code
README
# SmoothMix: Training Confidence-calibrated Smoothed Classifiers for Certified Robustness (NeurIPS2021)
This repository contains code for the paper
**"SmoothMix: Training Confidence-calibrated Smoothed Classifiers for Certified Robustness"**
by [Jongheon Jeong](https://sites.google.com/view/jongheonj), [Sejun Park](https://sites.google.com/site/sejunparksite/), Minkyu Kim, Heung-Chang Lee, Doguk Kim and [Jinwoo Shin](http://alinlab.kaist.ac.kr/shin.html).
## Dependencies
```
conda create -n smoothmix python=3.7
conda activate smoothmix
# Below is for linux, with CUDA 11.1; see https://pytorch.org/ for the correct command for your system
conda install pytorch torchvision cudatoolkit=11.1 -c pytorch -c nvidia
conda install scipy pandas statsmodels matplotlib seaborn
pip install tensorboardX
```
## Scripts
### Training Scripts
Our code is built upon a previous codebase from several baselines considered in the paper
([Cohen et al (2019)](https://github.com/locuslab/smoothing);
[Salman et al (2019)](https://github.com/Hadisalman/smoothing-adversarial);
[Jeong and Shin (2020)](https://github.com/jh-jeong/smoothing-consistency)).
The main script is `code/train.py`, and the sample scripts below demonstrate how to run `code/train.py`.
One can modify `CUDA_VISIBLE_DEVICES` to further specify GPU number(s) to work on.
```
# SmoothMix (Ours): MNIST, w/ one-step adversary, eta=5.0
CUDA_VISIBLE_DEVICES=0 python code/train.py mnist lenet --lr 0.01 --lr_step_size 30 --epochs 90 --noise 1.0 \
--num-noise-vec 4 --eta 5.0 --num-steps 8 --alpha 1.0 --mix_step 1 --id 0
```
For a more detailed instruction to reproduce our experiments, see [`EXPERIMENTS.MD`](EXPERIMENTS.MD).
### Testing Scripts
All the testing scripts is originally from https://github.com/locuslab/smoothing:
* The script [certify.py](code/certify.py) certifies the robustness of a smoothed classifier. For example,
```python code/certify.py mnist model_output_dir/checkpoint.pth.tar 0.50 certification_output --alpha 0.001 --N0 100 --N 100000```
will load the base classifier saved at `model_output_dir/checkpoint.pth.tar`, smooth it using noise level σ=0.50,
and certify the MNIST test set with parameters `N0=100`, `N=100000`, and `alpha=0.001`.
* The script [predict.py](code/predict.py) makes predictions using a smoothed classifier. For example,
```python code/predict.py mnist model_output_dir/checkpoint.pth.tar 0.50 prediction_outupt --alpha 0.001 --N 1000```
will load the base classifier saved at `model_output_dir/checkpoint.pth.tar`, smooth it using noise level σ=0.50,
and classify the MNIST test set with parameters `N=1000` and `alpha=0.001`.
* The script [analyze.py](code/analyze.py) contains some useful classes and functions to analyze the result data
from [certify.py](code/certify.py) or [predict.py](code/predict.py).
### Pre-trained models
We release our ImageNet models trained with SmoothMix in the following GDrive link: [ImageNet](https://drive.google.com/drive/folders/1M6Garbgg201tjxPosGIATZ7kLeB8HJQO?usp=sharing)
### Citation
```
@inproceedings{jeong2021smoothmix,
author = {Jeong, Jongheon and Park, Sejun and Kim, Minkyu and Lee, Heung-Chang and Kim, Do-Guk and Shin, Jinwoo},
booktitle = {Advances in Neural Information Processing Systems},
editor = {M. Ranzato and A. Beygelzimer and Y. Dauphin and P.S. Liang and J. Wortman Vaughan},
pages = {30153--30168},
publisher = {Curran Associates, Inc.},
title = {SmoothMix: Training Confidence-calibrated Smoothed Classifiers for Certified Robustness},
url = {https://proceedings.neurips.cc/paper/2021/file/fd45ebc1e1d76bc1fe0ba933e60e9957-Paper.pdf},
volume = {34},
year = {2021}
}
```