Efficient Image-to-Image Diffusion Classifier for Adversarial Robustness

Hefei Mei, Minjing Dong, Chang Xu

Requirements

We evaluate the robustness of the IDC based on torchattacks (PGD, FGSM, MIFGSM, CW, AutoAttack) and DiffPure (only for BPDA+EOT). If you only need the classifier or the former robustness evaluation, you can follow the Requirements in BBDM and install the torchattacks. Otherwise, you can install the following combined environment.

CUDA=10.2
Python 3.9

Create environment:

conda create -n IDC python=3.9
conda activate IDC
pip install -r requirements

Put the file _pytree.py in anaconda/envs/IDC/lib/python3.9/site-packages/torch/utils

Train

cd IDC/

If you wish to train from the beginning

# For CIFAR-10
python main.py --config configs/Template-IDC-cifar10.yaml
--train
--sample_at_start
--save_top
--gpu_ids 0,1,2,3

# For CIFAR-100
python main.py --config configs/Template-IDC-cifar100.yaml
--train
--sample_at_start
--save_top
--gpu_ids 0,1,2,3

If you wish to continue training

# For CIFAR-10
python main.py --config configs/Template-IDC-cifar10.yaml
--train
--sample_at_start
--save_top
--gpu_ids 0,1,2,3
--resume_model path/to/model_ckpt
--resume_optim path/to/optim_ckpt

# For CIFAR-100
python main.py --config configs/Template-IDC-cifar100.yaml
--train
--sample_at_start
--save_top
--gpu_ids 0,1,2,3
--resume_model path/to/model_ckpt
--resume_optim path/to/optim_ckpt

We use auto_lr_scheduler=True with batch_size 256 (64*4), if you will train the model with other batch_size and lr, you can set use auto_lr_scheduler=False for similar accuracy.

Test

cd IDC/

# For CIFAR-10
bash test_torchattacks_cifar10.sh

# For CIFAR-100
bash test_torchattacks_cifar100.sh

cd IDC-BPDA/

# For CIFAR-10
bash test_bpda_cifar10.sh

# For CIFAR-100
bash test_bpda_cifar100.sh

Pre-trained Model

Dataset	Log	Checkpoint
CIFAR-10	log_cifar10 [Google, Quark]	ckpt_cifar10 [Google, Quark]
CIFAR-100	log_cifar100 [Google, Quark]	ckpt_cifar100 [Google, Quark]

Acknowledgement