README.md

Package required

install python 3.6.9, pytorch 1.7.0, torchvision 0.8.1, tensorboardX 2.5, matplotlib 3.3.4, tqdm 4.60.0, pandas 1.1.5, numpy 1.18.5.

Train Models

In each dataset and model, first train the base model, than the enhanced model.

CIFAR10-resne32

cd cifar10/resnet32
mkdir data
sh train_CIFAR.sh
sh train_finetune_branch.sh

CIFAR10-vgg16

cd cifar10/vgg16
mkdir data
sh train_CIFAR.sh
sh train_finetune_branch.sh

CIFAR100-resne32

cd cifar100/resnet32
mkdir data
sh train_CIFAR.sh
sh train_finetune_branch.sh

CIFAR100-vgg16

cd cifar100/vgg16
mkdir data
sh train_CIFAR.sh
sh train_finetune_branch.sh

stl10-resne32

cd stl10/resnet32
mkdir data
sh train_STL.sh
sh train_finetune_branch.sh

stl10-vgg16

cd stl10/vgg16
mkdir data
sh train_STL.sh
sh train_finetune_branch.sh

tinyimagenet-resne32

cd tinyimagenet/resnet32
mkdir data
sh train_tinyimagenet.sh
sh train_finetune_branch.sh

tinyimagenet-vgg16

cd tinyimagenet/vgg16
mkdir data
sh train_tinyimagenet.sh
sh train_finetune_branch.sh

TBT Attacks

First enter a folder to attack the target model, e.g., cd ./Aegis/TBT/resnet32-cifar10/

Run non-adaptove attacks

python3 TBT_noadaptive.py

Run adaptive attacks

python3 TBT_adaptive.py

Results

The attack success rate under our defense are shown on the screen.

TA-LBF Attacks

DATA

./utils.py modify line23 to specify positions to save data. When you first run attack scripts. We will automatically download CIFAR-10, CIFAR-100, STL-10. But you need to download tiny-ImageNet by yourself.

Run non-adaptove attacks:

cd TA-LBF/non-adaptive
Then run attacks. for example, sh ./attack_reproduce_k=50_vgg16_cifar10.sh
All sh files are listed as follows.
cifar10 resnet32: attack_reproduce_k=50_resnet32_cifar10.sh
cifar100 resnet32: attack_reproduce_k=50_resnet32_cifar100.sh
stl10 resnet32: attack_reproduce_k=50_resnet32_stl10.sh
tinyimagenet resnet32: attack_reproduce_k=50_resnet32_tinyimagenet.sh
cifar10 vgg16: ./attack_reproduce_k=50_vgg16_cifar10.sh
cifar100 vgg16: ./attack_reproduce_k=50_vgg16_cifar100.sh
stl10 vgg16: ./attack_reproduce_k=50_vgg16_stl10.sh
tinyimagenet vgg16: ./attack_reproduce_k=50_vgg16_tinyimagenet.sh

Run adaptove attacks:

cd TA-LBF/adaptive
Then run attacks. for example, sh ./attack_reproduce_k=50_vgg16_cifar10.sh
All sh files are listed as follows.
cifar10 resnet32: attack_reproduce_k=50_resnet32_cifar10.sh
cifar100 resnet32: attack_reproduce_k=50_resnet32_cifar100.sh
stl10 resnet32: attack_reproduce_k=50_resnet32_stl10.sh
tinyimagenet resnet32: attack_reproduce_k=50_resnet32_tinyimagenet.sh
cifar10 vgg16: ./attack_reproduce_k=50_vgg16_cifar10.sh
cifar100 vgg16: ./attack_reproduce_k=50_vgg16_cifar100.sh
stl10 vgg16: ./attack_reproduce_k=50_vgg16_stl10.sh
tinyimagenet vgg16: ./attack_reproduce_k=50_vgg16_tinyimagenet.sh

Results

The attack success rate under our defense are shown on the screen.

ProFlip Attacks

First enter a folder to attack the target model, e.g., cd ./Aegis/ProFlip/resnet32-cifar10/

Run non-adaptove attacks

Run the instruction to generate a trigger: python3 trigger_nonadaptive.py
Then, attack: python3 CSB_nonadaptive.py

Run adaptive attacks

Run the instruction to generate a trigger: python3 trigger_adaptive.py
Then, attack: python3 CSB_adaptive.py

Results

The attack success rate under our defense are shown on the screen.