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, Sejun Park, Minkyu Kim, Heung-Chang Lee, Doguk Kim and Jinwoo Shin.

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); Salman et al (2019); Jeong and Shin (2020)). 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.

Testing Scripts

All the testing scripts is originally from https://github.com/locuslab/smoothing:

• The script 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 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 contains some useful classes and functions to analyze the result data from certify.py or predict.py.