pardef

Requirements

pip install -r requirements.txt

For Tiny-ImageNet, download the dataset and place it at tinyimagenet/resnet32/data/tiny-imagenet-200 and tinyimagenet/vgg16/data/tiny-imagenet-200.

Training

Train and apply the defense for each dataset and architecture:

cd cifar10
python train.py --arch resnet32
python train.py --arch vgg16

cd ../cifar100
python train.py --arch resnet32
python train.py --arch vgg16

cd ../tinyimagenet
python train.py --arch resnet32
python train.py --arch vgg16

Checkpoints are saved to {arch}/checkpoint/. The defended model is saved as defended.pth.

Validation

Evaluate the defended model and compute the ARI threshold:

cd cifar10
python validate.py --arch resnet32
python validate.py --arch vgg16

cd ../cifar100
python validate.py --arch resnet32
python validate.py --arch vgg16

cd ../tinyimagenet
python validate.py --arch resnet32
python validate.py --arch vgg16

Attacks

Each attack directory is named {attack}/{arch}-{dataset}/ and contains five scripts. Run from inside the directory.

ProFlip

cd ProFlip/resnet32-cifar10
python nonadaptive.py
python adaptive.py
python trigger_nonadaptive.py
python trigger_adaptive.py

P3A

cd P3A/resnet32-cifar10
python nonadaptive.py
python adaptive.py
python trigger_nonadaptive.py
python trigger_adaptive.py

APA

cd APA/resnet32-cifar10
python nonadaptive.py
python adaptive.py
python trigger_nonadaptive.py
python trigger_adaptive.py

Replace resnet32-cifar10 with any of:

• resnet32-cifar100
• resnet32-tinyimagenet
• vgg16-cifar10
• vgg16-cifar100
• vgg16-tinyimagenet

Results are saved to ./result/ inside each attack directory.