WAFFLE: Watermarking in Federated Learning

This repo contains the Pytorch implementation of the WAFFLE: Watermarking in Federated Learning (arXiv link), and allows you to reproduce experiments presented in the paper. The paper will appear in the Proceedings of SRDS 2021.

Setup

You need to have a working conda setup (Conda 4.8.3 and Python 3.7.7 would be the best for the complete reproduction). We also assume that you have access to a GPU with CUDA >=9.2. After meeting hardware and softwarre requirements, clone this repository and install necessary software packages using conda environment.

1. Create the environment from the pysyft.yml file: conda env create -f pysyft.yml
2. Activate the new environment: conda activate pysyft

We use PySyft 0.2.x, an ecosystem that supports federated learning, differential privacy and homomorphic encryption. Its codebase is in its own branch here, but OpenMined announced that they will not offer official support for this version range. You can visit here to check the latest updates.

Training federated learning models

1. Train an FL model without watermarking: (Files under the folder "baseline_experiments" contains configuration files for this purpose)

python main.py --config_file configurations/baseline_experiments/mnist/1.ini --experiment training

2. Train an FL model with watermarking: (Files under the folder "waffle_experiments" contains configuration files for this purpose. Please check confuguration files for correctly setting the watermark set type)

python main.py --config_file configurations/waffle_experiments/cifar10/1.ini --experiment training

Watermark removal attacks

In order to run an attack (fine-tuning, pruning, Neural Cleanse), set the --experiment argument with the desired attack. It will run the chosen attack with different number of adversaries.

python main.py --config_file configurations/waffle_experiments/cifar10/1.ini --experiment fine-tuning/pruning/neural-cleanse/evasion

You can run evasion the verification experiments by seting the --experiment argument to evasion. However, make sure that the tiny-imagenet dataset is downloaded into "data/datasets/tiny-imagenet-200/" folder. You can download the dataset from here.

Citation

If you find our work useful in your research, you can cite the paper as follows:

@article{atli2020waffle,
  title={WAFFLE: Watermarking in Federated Learning},
  author={Tekgul, Buse G.A. and Xia, Yuxi and Marchal, Samuel and Asokan, N},
  eprint={2008.07298},
  archivePrefix={arXiv},
  year={2020},
  primaryClass={cs.LG}
}