Isolation and Induction: Training Robust Deep Neural Networks against Model Stealing Attacks

Introduction

This repository is the official PyTorch implementation of paper "Isolation and Induction: Training Robust Deep Neural Networks against Model Stealing Attacks"

Install

Requirements

• Python >= 3.6
• PyTorch >= 1.8

Data Preparation

This code supports following datasets:

• MNIST
• KMNIST
• EMNIST
• EMNISTLetters
• FashionMNIST
• CIFAR-10
• CIFAR-100
• SVHN
• Tiny ImageNet

Please download datasets from their official website, and put them in data/ directory.

If you want to customize your own datasets, please refer to src/datasets/__init__.py.

Usage

First you should generate an YAML file as the config of an experiment.

python yaml_all.py --path <exp_path> --defense [nd|ini] --attack [knockoff|jbda]

This will generate a config.yaml file at exp_path. You can manually modify the YAML file, or use commands to assign more parameters.

Then, you can run the experiments by running defense_entry.py and attack_entry.py. defense_entry.py is to train a defended model, and attack_entry.py is to perform model stealing attacks.

python defense_entry.py --config <exp_path>
python attack_entry.py --config <exp_path>

All results and checkpoints will save under <exp_path>.

Citation

If this work helps your research, please cite the following paper.

@article{guo2023isolation,
  title={Isolation and Induction: Training Robust Deep Neural Networks against Model Stealing Attacks},
  author={Guo, Jun and Liu, Aishan and Zheng, Xingyu and Liang, Siyuan and Xiao, Yisong and Wu, Yichao and Liu, Xianglong},
  journal={arXiv preprint arXiv:2308.00958},
  year={2023}
}