PeerNets on PyTorch 1.0

A pytorch implementation of PeerNets: Exploiting Peer Wisdom Against Adversarial Attacks

Keywords

Adversarial Attack, Adversarial Example, Graph Convolutional Network(GCN)

Tutorails

• Adversarial Attack(Part 1): Adversarial Attack과 Defense Link
• Adversarial Attack(Part 2): PeerNets Link

Environement

• Docker 18.06 on Ubuntu 16.04
• PyTorch 1.0
• CUDA >= 9.1, CUDNN 7

Experiments

• LeNet5 on MNIST
• ResNet32 on CIFAR10
• ResNet110 on CIFAR100
• PR-LeNet5 on MNIST
• PR-ResNet32 on CIFAR10
• PR-ResNet110 on CIFAR100

Results

• Baseline: Accuracy(%)
• rho(0.02~1.0): Accuracy(%)/Fooling Rate(%)

MNIST

Method	Baseline	rho=0.2	0.4	0.6	0.8	1.0
LeNet-5	98.6	92.7/7.1	33.9/66.0	14.1/85.9	7.9/92.2	8.2/91.7
PR-LeNet-5	98.2	94.8/4.6	93.3/6.0	87.7/11.7	53.2/46.4	50.1/50.1
LeNet-5(ours)	98.5	94.2/4.0	75.8/32.0	64.5/42.0	57.1/50.0	42.0/72.0
PR-LeNet-5(ours)	-	-	-	-	-	-

CIFAR10

Method	Baseline	rho=0.04	0.08	0.10
ResNet-32	92.7	55.3/44.4	26.8/73.1	22.7/77.3
PR-ResNet-32	89.3	87.3/7.1	83.3/13.0	70.0/28.3
ResNet-32(ours)	93.3	84.7/20.0	75.2/32.0	69.5/46.0

CIFAR100

Method	Baseline	rho=0.02	0.04	0.06
ResNet-110	71.6	45.5/49.8	45.5/49.8	45.5/49.8
PR-ResNet-110	66.4	61.5/23.7	52.6/38.6	44.6/49.5
ResNet-110(ours)	68.8	-	-	-

How to run

nvidia-docker run -it tantara/peernets-pytorch /bin/bash # and CMD
nvidia-docker run -it tantara/peernets-pytorch CMD

• CMD can be Evaluation, Attack, Training as follows:

Evaluation

# bash eval.sh GPU_ID EXP
bash eval.sh 0 lenet5 # lenet5, pr-lenet5
bash eval.sh 0 resnet32 # resnet32, pr-resnet32
bash eval.sh 0 resnet110 # resnet110, pr-resnet110

Adversarial Attack

# bash attack.sh GPU_ID EXP NUM_SAMPLES RHO
bash attack.sh 0 lenet5 50 0.2 # lenet5, pr-lenet5
bash attack.sh 0 resnet32 50 0.04 # resnet32, pr-resnet32
bash attack.sh 0 resnet110 50 0.02 # resnet110, pr-resnet110

Training

# bash train.sh GPU_ID EXP
bash train.sh 0 lenet5 # lenet5, pr-lenet5
bash train.sh 0 resnet32 # resnet32, pr-resnet32
bash train.sh 0 resnet110 # resnet110, pr-resnet110

References

1. PeerNets: Exploiting Peer Wisdom Against Adversarial Attacks
   

   Jan Svoboda, Jonathan Masci, Federico Monti, Michael M. Bronstein, Leonidas Guibas. arXiv: 1806.00088.
   

   [link]. arXiv: 1806.00088, 2018.

2. aaron-xichen/pytorch-playground

3. akamaster/pytorch_resnet_cifar10

4. rusty1s/pytorch_geometric

Authors

• Taekmin Kim, @tantara

License

© Taekmin Kim, 2018. Licensed under the MIT License.