- The code is the official implementation of paper Self-Ensemble Protection: Training Checkpoints Are Good Data Protectors
- This repository supports data protection on CIFAR-10, CIFAR-100, ImageNet subset
- protecting DNN and appropriator DNN: ResNet18, SENet18, VGG16, DenseNet121, GoogLeNet
- The experiments are run in an NVIDIA A100 GPU, but could modify the batch size to run on small GPUs
- Install dependencies
conda env create -f pt.yaml
- train the protecting DNN for CIFAR-10, CIFAR-100, ImageNet subset
- crafting protective samples (CIFAR-10, SEP)
python ens.py --num_model=30 --eps=2 --target_batch=0
- crafting protective samples (CIFAR-10, SEP-FA)
python ens_feature.py --num_model=30 --eps=2 --target_batch=0
- crafting protective samples (CIFAR-10, SEP-FA-VR)
python ens_feature_svre.py --num_model=15 --eps=2 --target_batch=0
- crafting protective samples (CIFAR-100, SEP-FA-VR)
python ens_feature_svre100.py --num_model=15 --eps=2 --target_batch=0
- crafting protective samples (ImageNet subset, SEP-FA-VR)
python ens_feature_svreimg.py --num_model=15 --eps=2 --target_batch=0
- train the appropriator DNN
python vanilla.py --uledir=samples/XX --eps=2
python vanilla100.py --uledir=samples/XX --eps=2
python vanillaimg.py --uledir=samples/XX --eps=2
├── ens_feature.py
├── ens_feature_svre100.py
├── ens_feature_svreimg.py
├── ens_feature_svre.py
├── ens.py
├── models
│ ├── densenet.py
│ ├── dpn.py
│ ├── efficientnet.py
│ ├── googlenet.py
│ ├── __init__.py
│ ├── lenet.py
│ ├── mobilenet.py
│ ├── mobilenetv2.py
│ ├── pnasnet.py
│ ├── preact_resnet.py
│ ├── regnet.py
│ ├── resnet.py
│ ├── resnext.py
│ ├── senet.py
│ ├── shufflenet.py
│ ├── shufflenetv2.py
│ └── vgg.py
├── pt.yaml
├── README.md
├── utils
│ ├── data.py
│ ├── __init__.py
│ ├── output.py
│ └── tmp.py
├── vanilla100.py
├── vanillaimg.py
└── vanilla.py