Towards Adversarial Robustness of Bayesian Neural Network through Hierarchical Variational Inference [paper]

Baseline of this code is the official repository for this paper. We just replace the BNN regularizer from ELBO with enhanced Bayesian regularizer based on hierarchical-ELBO.

Citation

If you find this work helpful, please cite it as:

@misc{
lee2021towards,
title={Towards Adversarial Robustness of Bayesian Neural Network through Hierarchical Variational Inference},
author={Byung-Kwan Lee and Youngjoon Yu and Yong Man Ro},
year={2021},
url={https://openreview.net/forum?id=Cue2ZEBf12}
}

Hierarchical-Bayeisan-Defense

Dataset

CIFAR10
STL10
CIFAR100
Tiny-ImageNet

Network

VGG16 (for CIFAR-10/CIFAR-100/Tiny-ImageNet)
Aaron (for STL10)
WideResNet (for CIFAR-10/100)

Attack (by torchattack)

PGD attack
EOT-PGD attack

Defense methods

adv: Adversarial training
adv_vi: Adversarial training with Bayesian neural network
adv_hvi: Adversarial training with Enhanced Bayesian neural network based on hierarchical-ELBO

How to Train

1. Adversarial training

Run train_adv.sh

lr=0.01
steps=10
max_norm=0.03
data=tiny # or `cifar10`, `stl10`, `cifar100`
root=./datasets
model=vgg # vgg for `cifar10` `stl10` `cifar100`, aaron for `stl10`, wide for `cifar10` or `cifar100`
model_out=./checkpoint/${data}_${model}_${max_norm}_adv
echo "Loading: " ${model_out}
CUDA_VISIBLE_DEVICES=0 python ./main_adv.py \
                        --lr ${lr} \
                        --step ${steps} \
                        --max_norm ${max_norm} \
                        --data ${data} \
                        --model ${model} \
                        --root ${root} \
                        --model_out ${model_out}.pth \

2. Adversarial training with BNN

Run train_adv_vi.sh

lr=0.01
steps=10
max_norm=0.03
sigma_0=0.1
init_s=0.1
data=tiny # or `cifar10`, `stl10`, `cifar100`
root=./datasets
model=vgg # vgg for `cifar10` `stl10` `cifar100`, aaron for `stl10`, wide for `cifar10` or `cifar100`
model_out=./checkpoint/${data}_${model}_${max_norm}_adv_vi
echo "Loading: " ${model_out}
CUDA_VISIBLE_DEVICES=0 python3 ./main_adv_vi.py \
                        --lr ${lr} \
                        --step ${steps} \
                        --max_norm ${max_norm} \
                        --sigma_0 ${sigma_0} \
                        --init_s ${init_s} \
                        --data ${data} \
                        --model ${model} \
                        --root ${root} \
                        --model_out ${model_out}.pth \

3. Adversarial training with enhanced Bayesian regularizer based on hierarchical-ELBO

Run train_adv_hvi.sh

lr=0.01
steps=10
max_norm=0.03
sigma_0=0.1
init_s=0.1
data=tiny # or `cifar10`, `stl10`, `cifar100`
root=./datasets
model=vgg # vgg for `cifar10` `stl10` `cifar100`, aaron for `stl10`, wide for `cifar10` or `cifar100`
model_out=./checkpoint/${data}_${model}_${max_norm}_adv_hvi
echo "Loading: " ${model_out}
CUDA_VISIBLE_DEVICES=0 python3 ./main_adv_hvi.py \
                        --lr ${lr} \
                        --step ${steps} \
                        --max_norm ${max_norm} \
                        --sigma_0 ${sigma_0} \
                        --init_s ${init_s} \
                        --data ${data} \
                        --model ${model} \
                        --root ${root} \
                        --model_out ${model_out}.pth \

How to Test

Testing adversarial robustness

Run acc_under_attack.sh

model=vgg # vgg for `cifar10` `stl10` `cifar100`, aaron for `stl10`, wide for `cifar10` or `cifar100`
defense=adv_hvi # or `adv_vi`, `adv`
data=tiny-imagenet # or `cifar10`, `stl10`, `cifar100`
root=./datasets
n_ensemble=50
step=10
max_norm=0.03
echo "Loading" ./checkpoint/${data}_${model}_${max_norm}_${defense}.pth

CUDA_VISIBLE_DEVICES=0 python3 acc_under_attack.py \
    --model $model \
    --defense $defense \
    --data $data \
    --root $root \
    --n_ensemble $n_ensemble \
    --step $step \
    --max_norm $max_norm

How to check the learning parameters and KL divergence

Run check_parameters.sh

model=vgg # vgg for `cifar10` `stl10` `cifar100`, aaron for `stl10`, wide for `cifar10` or `cifar100`
defense=adv_hvi # or `adv_vi`
data=tiny-imagenet # or `cifar10`, `stl10`, `cifar100`
max_norm=0.03
echo "Loading" ./checkpoint/${data}_${model}_${max_norm}_${defense}.pth

CUDA_VISIBLE_DEVICES=0 python3 check_parameters.py \
    --model $model \
    --defense $defense \
    --data $data \
    --max_norm $max_norm \

How to check uncertainty by predictive entropy

Run uncertainty.sh

model=vgg # vgg for `cifar10` `stl10` `cifar100`, aaron for `stl10`, wide for `cifar10` or `cifar100`
defense=adv_hvi # or `adv_vi`
data=tiny-imagenet # or `cifar10`, `stl10`, `cifar100`
root=./datasets
n_ensemble=50
step=10
max_norm=0.03
echo "Loading" ./checkpoint/${data}_${model}_${max_norm}_${defense}.pth

CUDA_VISIBLE_DEVICES=0 python3 uncertainty.py \
    --model $model \
    --defense $defense \
    --data $data \
    --root $root \
    --n_ensemble $n_ensemble \
    --step $step \
    --max_norm $max_norm

Name		Name	Last commit message	Last commit date
Latest commit History 8 Commits
loader		loader
models		models
utils		utils
Figure.png		Figure.png
README.md		README.md
acc_under_attack.py		acc_under_attack.py
check_parameters.py		check_parameters.py
main_adv.py		main_adv.py
main_adv_hvi.py		main_adv_hvi.py
main_adv_mart.py		main_adv_mart.py
main_adv_mart_hvi.py		main_adv_mart_hvi.py
main_adv_mart_vi.py		main_adv_mart_vi.py
main_adv_trade.py		main_adv_trade.py
main_adv_trade_hvi.py		main_adv_trade_hvi.py
main_adv_trade_vi.py		main_adv_trade_vi.py
main_adv_vi.py		main_adv_vi.py
uncertainty.py		uncertainty.py

ByungKwanLee/Hierarchical-Bayesian-Defense

Folders and files

Latest commit

History

Repository files navigation