Usage

The code in this repository implements both the RMC ERMC and SRMC .

Curve Finding

Training the endpoints

To run the curve-finding procedure, you first need to train the two networks that will serve as the end-points of the curve. You can train the endpoints using the following command

python  train.py --dir=<DIR> \
                 --dataset=<DATASET> \
                 --data_path=<PATH> \
                 --transform=<TRANSFORM> \
                 --model=<MODEL> \
                 --epochs=<EPOCHS> \
                 --batch_size=<BS> \
                 --lr=<LR_INIT> \
                 --wd=<WD> \
                 --pgd=<PGD>
                 --gpus=<GPUS>
                 [--use_test]

Parameters:

DIR — path to training directory where checkpoints will be stored
DATASET — dataset name [CIFAR10/CIFAR100]
PATH — path to the data directory
TRANSFORM — type of data transformation [VGG/ResNet]
MODEL — DNN model name:
- VGG16/VGG16BN/VGG19/VGG19BN
- PreResNet110/PreResNet164
- SimpleViT
EPOCHS — number of training epochs
LR_INIT — initial learning rate
batch_size — batch size
WD — weight decay
PGD — type of pgd attack[1/2/inf/msd]
GPUS—id of gpus

Use the --use_test flag if you want to use the test set instead of validation set (formed from the last 5000 training objects) to evaluate performance.

For example, use the following commands to train PreResNet , WideResNet or ViT :

#PreResNet
python train.py --dir=<DIR> --dataset=[CIFAR10/CIFAR100] --data_path=<PATH>  --model=PreResNet110 --epochs=150 --batch_size=128  --lr=0.1 --wd=3e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd] 

python train.py --dir=<DIR> --dataset=ImageNet100 --data_path=<PATH>  --model=PreResNet110 --epochs=150 --batch_size=32  --lr=0.1 --wd=3e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd] --gpus=0,1,2,3

#WideResNet
train.py --dir=<DIR>  --dataset=CIFAR10 --data_path=<PATH> --model=WideResNet28x10 --epochs=200 --lr=0.1 --wd=5e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd]

#ViT
python train.py --dir=<DIR>  --dataset=CIFAR10 --data_path=<PATH> --model=ViT --epochs=150 --lr=0.001 --wd=1e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd]

Finetuning the endpoints from an existed endpoint

To finetune the endpoints from an existed endpoint, you can use the following command

python  train_.py    --dir=<DIR> \
                     --dataset=<DATASET> \
                     --data_path=<PATH> \
                     --transform=<TRANSFORM> \
                     --model=<MODEL> \
                     --epochs=<EPOCHS> \
                     --lr=<LR_INIT> \
                     --wd=<WD> \
                     --pgd=<PGD>
                     --origin_model=<CKPT>
                     [--use_test]

Parameters

CKPT — path to the existed endpoint merged from curves saved by merge.py

See the sections on [training the endpoints] for the description of other parameters.

#PreResNet
python train_.py --dir=<DIR> --dataset=[CIFAR10/CIFAR100] --data_path=<PATH>  --model=PreResNet110 --epochs=50 --batch_size=128  --lr=0.1 --wd=3e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd] --origin_model=<CKPT>

python train_.py --dir=<DIR> --dataset=ImageNet100 --data_path=<PATH>  --model=PreResNet110 --epochs=50 --batch_size=32  --lr=0.1 --wd=3e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd] --origin_model=<CKPT>

#WideResNet
train_.py --dir=<DIR>  --dataset=CIFAR10 --data_path=<PATH> --model=WideResNet28x10 --epochs=50 --lr=0.1 --wd=5e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd] --origin_model=<CKPT>

#ViT
python train_.py --dir=<DIR>  --dataset=CIFAR10 --data_path=<PATH> --model=ViT --epochs=50 --lr=0.001 --wd=1e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd]  --origin_model=<CKPT>

Training the curves

Once you have two checkpoints to use as the endpoints you can train the curve connecting them using the following command.

Note that msd attacks in the current code contain all three attacks by default. If you need to reduce the number of attack types, manually adjust line 294 in the /attack/att.py file.

python  train.py --dir=<DIR> \
                 --dataset=<DATASET> \
                 --data_path=<PATH> \
                 --transform=<TRANSFORM>
                 --model=<MODEL> \
                 --epochs=<EPOCHS> \
                 --lr=<LR_INIT> \
                 --wd=<WD> \
                 --batch_size=<BS> \
                 --curve=<CURVE>[Bezier|PolyChain] \
                 --num_bends=<N_BENDS> \
                 --init_start=<CKPT1> \ 
                 --init_end=<CKPT2> \
                 --pgd=<PGD>
                 [--fix_start] \
                 [--fix_end] \
                 [--use_test]

Parameters:

CURVE — desired curve parametrization [Bezier|PolyChain]
N_BENDS — number of bends in the curve
CKPT1, CKPT2 — paths to the checkpoints to use as the endpoints of the curve

Use the flags --fix_end --fix_start if you want to fix the positions of the endpoints; otherwise the endpoints will be updated during training. See the section on [training the endpoints] for the description of the other parameters.

For example, use the following commands to train VGG16 or PreResNet :

#PreResNet
python  train.py --dir=<DIR> --dataset=[CIFAR10/CIFAR100] --use_test --transform=ResNet --data_path=<PATH> --model=PreResNet110 --curve=[Bezier|PolyChain] --num_bends=3  --init_start=<CKPT1> --init_end=<CKPT2> --fix_start --fix_end --epochs=50 --batch_size=128  --lr=0.03 --wd=3e-4 --pgd=[1/2/inf/msd]

python  train.py --dir=<DIR> --dataset=ImageNet100 --use_test --transform=ResNet --data_path=<PATH> --model=PreResNet110 --curve=[Bezier|PolyChain] --num_bends=3  --init_start=<CKPT1> --init_end=<CKPT2> --fix_start --fix_end --epochs=50 --batch_size=32  --lr=0.01 --wd=3e-4 --pgd=[1/2/inf/msd]

#WideResNet
python  train.py --dir=<DIR>  --dataset=CIFAR10 --use_test --transform=ResNet --data_path=./data --model=WideResNet28x10 --curve=Bezier --num_bends=3 --init_start=<CKPT1> --init_end=<CKPT2>--fix_start --fix_end --epochs=100 --lr=0.03 --wd=5e-4 --pgd=[1/2/inf/msd]

#ViT
python  train.py --dir=<DIR> --dataset=CIFAR10 --use_test --transform=ResNet --data_path=./data --model=ViT --curve=Bezier --num_bends=3  --init_start=<CKPT1> --init_end=<CKPT2> --fix_start --fix_end --epochs=50 --lr=0.001 --wd=3e-4 --pgd=[1/2/inf/msd]

Evaluating the curves

To evaluate the found curves, you can use the following command

python  eval_curve_pgd.py    --dir=<DIR> \
                             --dataset=<DATASET> \
                             --data_path=<PATH> \
                             --transform=<TRANSFORM>
                             --model=<MODEL> \
                             --batch_size=<BS> \
                             --curve=<CURVE>[Bezier|PolyChain] \
                             --num_bends=<N_BENDS> \
                             --ckpt=<CKPT> \ 
                             --num_points=<NUM_POINTS> \
                             --

Parameters

CKPT — path to the curves checkpoint saved by train.py
NUM_POINTS — number of points along the curve to use for evaluation (default: 61)

See the sections on [training the endpoints]and [training the curves] for the description of other parameters.

eval_curve.py outputs the statistics on train and test loss and accuracy along the curve.

python  eval_curve_pgd.py    --dir=<DIR> --dataset=[CIFAR10/CIFAR100/ImageNet100] --data_path=<PATH> --transform=PreResNet --model=[PreResNet110/WideResNet28x10/ViT] --batch_size=128  --curve=Bezier--num_bends=3 --ckpt=<CKPT>  --num_points=61

Merging the curves

To merge the trained curves into an endpoint at t=T, you can use the following command

python merge.py --ckpt=<CKPT> --t=[0.0<=T<=1.0] --dir=<DIR>

Parameters

CKPT — path to the curves checkpoint saved by train.py
T — where the point you want to get from the curves

See the sections on [training the endpoints]and [training the curves] for the description of other parameters.

python merge.py --ckpt=<CKPT>  --t=0.1 --dir=./mergesave/rmc3/0.1

Training the endpoints from an existed endpoint merged from curves

To train the endpoints from an existed endpoint merged from curves,you can use the following command

python  train_.py    --dir=<DIR> \
                     --dataset=<DATASET> \
                     --data_path=<PATH> \
                     --transform=<TRANSFORM> \
                     --model=<MODEL> \
                     --epochs=<EPOCHS> \
                     --lr=<LR_INIT> \
                     --wd=<WD> \
                     --pgd=<PGD>
                     --origin_model=<CKPT>
                     [--use_test]

Parameters

CKPT — path to the existed endpoint merged from curves saved by merge.py

See the sections on [training the endpoints] for the description of other parameters.

#PreResNet
python train_.py --dir=<DIR> --dataset=[CIFAR10/CIFAR100] --data_path=<PATH>  --model=PreResNet110 --epochs=50 --batch_size=128  --lr=0.1 --wd=3e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd] --origin_model=<CKPT>

python train_.py --dir=<DIR> --dataset=ImageNet100 --data_path=<PATH>  --model=PreResNet110 --epochs=50 --batch_size=32  --lr=0.1 --wd=3e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd] --origin_model=<CKPT>

#WideResNet
train_.py --dir=<DIR>  --dataset=CIFAR10 --data_path=<PATH> --model=WideResNet28x10 --epochs=50 --lr=0.1 --wd=5e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd] --origin_model=<CKPT>

#ViT
python train_.py --dir=<DIR>  --dataset=CIFAR10 --data_path=<PATH> --model=ViT --epochs=50 --lr=0.001 --wd=1e-4 --use_test --transform=ResNet --pgd=[1/2/inf/msd]  --origin_model=<CKPT>

Citation

@article{wang2023robust,
  title={Bridging Models to Defend: A Population-Based Strategy for Robust Adversarial Defense},
  author={Wang, Ren and Li, Yuxuan and Chen, Can and Wang, Dakuo and Chen, Pin-Yu and Liu, Sijia and Shahidehpour, Mohammad and Hero, Alfred},
  journal={arXiv preprint arXiv:2303.10225},
  year={2025}
}

@inproceedings{wang2024deep,
  title={Deep Adversarial Defense Against Multilevel $-$\backslash$ell\_ $\{$P$\}$ $ Attacks},
  author={Wang, Ren and Li, Yuxuan and Hero, Alfred},
  booktitle={2024 IEEE 34th International Workshop on Machine Learning for Signal Processing (MLSP)},
  pages={1--6},
  year={2024},
  organization={IEEE}
}

Acknowledgement

We would like to express our gratitude for the support provided by NSF grants 2246157 and 2319243, as well as the ORAU Ralph E. Powe Junior Faculty Enhancement Award. Additionally, we are thankful for the computational resources made available through ACCESS and CloudBank.

Name		Name	Last commit message	Last commit date
Latest commit History 63 Commits
CKPT		CKPT
attack		attack
models		models
README.md		README.md
connect.py		connect.py
curves.py		curves.py
d.py		d.py
data.py		data.py
eval_curve.py		eval_curve.py
eval_curve_pgd.py		eval_curve_pgd.py
eval_ensemble.py		eval_ensemble.py
fge.py		fge.py
merge.py		merge.py
other_utils.py		other_utils.py
pgdtest.py		pgdtest.py
plane.py		plane.py
plane_plot.py		plane_plot.py
requirements.txt		requirements.txt
test.py		test.py
train.py		train.py
train_.py		train_.py
utils.py		utils.py

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Usage

Curve Finding

Training the endpoints

Finetuning the endpoints from an existed endpoint

Training the curves

Evaluating the curves

Merging the curves

Training the endpoints from an existed endpoint merged from curves

Citation

Acknowledgement

About

Uh oh!

Releases

Packages

Uh oh!

Contributors 2

Uh oh!

Languages

wangren09/MCGR

Folders and files

Latest commit

History

Repository files navigation

Usage

Curve Finding

Training the endpoints

Finetuning the endpoints from an existed endpoint

Training the curves

Evaluating the curves

Merging the curves

Training the endpoints from an existed endpoint merged from curves

Citation

Acknowledgement

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors 2

Uh oh!

Languages

Packages