zlh-thu / stealingverification

Defending against Model Stealing via Verifying Embedded External Features

Defending against Model Stealing Attacks via Verifying Embedded External Features

This is the official implementation of our paper Defending against Model Stealing Attacks via Verifying Embedded External Features, accepted by the AAAI Conference on Artificial Intelligence (AAAI), 2022. This research project is developed based on Python 3 and Pytorch, created by Yiming Li and Linghui Zhu.

Reference

If our work or this repo is useful for your research, please cite our paper as follows:

@inproceedings{li2022defending,
  title={Defending against Model Stealing via Verifying Embedded External Features},
  author={Li, Yiming and Zhu, Linghui and Jia, Xiaojun and Jiang, Yong and Xia, Shu-Tao and Cao, Xiaochun},
  booktitle={AAAI},
  year={2022}
}

Pipeline

Requirements

To install requirements:

pip install -r requirements.txt

Make sure the directory follows:

stealingverification
├── data
│   ├── cifar10
│   └── ...
├── gradients_set 
│   
├── prob
│   
├── network
│   
├── model
│   ├── victim
│   └── ...
|

Dataset Preparation

Make sure the directory data follows:

data
├── cifar10_seurat_10%
|   ├── train
│   └── test
├── cifar10  
│   ├── train
│   └── test
├── subimage_seurat_10%
│   ├── train
|   ├── val
│   └── test
├── sub-imagenet-20
│   ├── train
|   ├── val
│   └── test

📋 Data Download Link:
data

Model Preparation

Make sure the directory model follows:

model
├── victim
│   ├── vict-wrn28-10.pt
│   └── ...
├── benign
│   ├── benign-wrn28-10.pt
│   └── ...
├── attack
│   ├── atta-label-wrn16-1.pt
│   └── ...
└── clf

📋 Model Download Link:
model

Collecting Gradient Vectors

Collect gradient vectors of victim and benign model with respect to transformed images.

CIFAR-10:

python gradientset.py --model=wrn16-1 --m=./model/victim/vict-wrn16-1.pt --dataset=cifar10 --gpu=0
python gradientset.py --model=wrn28-10 --m=./model/victim/vict-wrn28-10.pt --dataset=cifar10 --gpu=0
python gradientset.py --model=wrn16-1 --m=./model/benign/benign-wrn16-1.pt --dataset=cifar10 --gpu=0
python gradientset.py --model=wrn28-10 --m=./model/benign/benign-wrn28-10.pt --dataset=cifar10 --gpu=0

ImageNet:

python gradientset.py --model=resnet34-imgnet --m=./model/victim/vict-imgnet-resnet34.pt --dataset=imagenet --gpu=0
python gradientset.py --model=resnet18-imgnet --m=./model/victim/vict-imgnet-resnet18.pt --dataset=imagenet --gpu=0
python gradientset.py --model=resnet34-imgnet --m=./model/benign/benign-imgnet-resnet34.pt --dataset=imagenet --gpu=0
python gradientset.py --model=resnet18-imgnet --m=./model/benign/benign-imgnet-resnet18.pt --dataset=imagenet --gpu=0

Training Ownership Meta-Classifier

To train the ownership meta-classifier in the paper, run these commands:

CIFAR-10:

python train_clf.py --type=wrn28-10 --dataset=cifar10 --gpu=0
python train_clf.py --type=wrn16-1 --dataset=cifar10 --gpu=0

ImageNet:

python train_clf.py --type=resnet34-imgnet --dataset=imagenet --gpu=0
python train_clf.py --type=resnet18-imgnet --dataset=imagenet --gpu=0

Ownership Verification

To verify the ownership of the suspicious models, run this command:

CIFAR-10:

python ownership_verification.py --mode=source --dataset=cifar10 --gpu=0 

#mode: ['source','distillation','zero-shot','fine-tune','label-query','logit-query','benign']

ImageNet:

python ownership_verification.py --mode=logit-query --dataset=imagenet --gpu=0 

#mode: ['source','distillation','zero-shot','fine-tune','label-query','logit-query','benign']

An Example of the Result

python ownership_verification.py --mode=fine-tune --dataset=cifar10 --gpu=0 

result:  p-val: 1.9594572166549425e-08 mu: 0.47074130177497864