HyFed

Hybrid Federated Framework for Privacy-preserving Machine Learning

          

About

HyFed is a framework to develop federated, privacy-preserving machine learning algorithms. It is developed at the Technical University of Munich (TUM). HyFed consists of four main components:

1. WebApp to set up the (hyper-)parameters of the federated algorithm
2. Client to compute the local model and perturb it with noise
3. Compensator to calculate the aggregated noise by aggregating the noise values from the clients
4. Server to coordinate the training process and compute the global model by adding up the noisy local models from the clients and the negative of the aggregated noise from the compensator

HyFed provides enhanced privacy while preserving the utility of the global model, i.e. it hides the original values of the local parameters of a client from the server, compensator, or the other clients without adversely affecting the accuracy of the global model. HyFed supports the simulation mode, where all components are running in the same machine and communicate over the loopback network, and the federated mode in which the components are installed in separate machines and securely communicate with each other over the Internet. HyFed provides developers with the client and server Python API to implement their own federated, privacy-preserving algorithms.

To see how to install the HyFed framework, please see HyFed-Install.
To learn how to develop your own federated tool using the HyFed API, please see HyFed-Develop.
To run the HyFed framework, please see HyFed-Run.
To learn more about HyFed, please read the HyFed manuscript.

License

The HyFed source code is licensed under the Apache License Version 2.0. (C) 2021, the HyFed developers.

Citation

If you use the HyFed framework in your study, please cite it as follows:

@misc{nasirigerdeh2021hyfed,
       title={HyFed: A Hybrid Federated Framework for Privacy-preserving Machine Learning},
       author={Reza Nasirigerdeh and Reihaneh Torkzadehmahani and Julian Matschinske and Jan Baumbach and Daniel Rueckert and Georgios Kaissis},
       year={2021},
       eprint={2105.10545},
       archivePrefix={arXiv},
       primaryClass={cs.LG}
}

Contact

In case of questions or problems regarding HyFed, please contact the HyFed developers: Reza Nasirigerdeh, Reihaneh Torkzadehmahani, and Julian Matschinske.

Disclaimer

This framework is a research product and is provided as it is. We assume no liability for any user action or omission.