ai4os/ai4os-federated-server

Federated Learning server for the AI4EOSC platform. Built using Flower

Federated Learning Server

Federated learning server with flower.

Summary: deploys a federated learning server to train machine/deep learning models on different clients following a federated learning architecture. Customization is allowed for the number of rounds the training is performed, the aggregation strategy used, the minimum number of clients, and the error metric to be used (aggregated) as validation.

Launching the federated learning server:

git clone https://github.com/ai4os/ai4os-federated-server
cd ai4os-federated-server
pip install -e .
python3 fedserver/server.py

Possible aggregation strategies introduced as FEDERATED_STRATEGY (see the implementation given in flower strategies):

• Federated average (FedAvg). [1]
• FedProx strategy (FedProx). [2]
• Adaptive Federated Optimization using Adam (FedAdam) [3]
• Federated Optim strategy (FedOpt) [3]
• Adaptive Federated Optimization using Yogi (FedYogi) [3]
• Federated Averaging with Momentum (FedAvgM) [4]
• Federated Median (FedMedian) [5]

The associated Docker image(s) for this module can be found in https://hub.docker.com/r/ai4oshub/ai4os-federated-server.

We provide some client examples as a guide for users. Users will need to adapt the uuid and the endpoint in those samples to point to their deployed Federated server.

🚀 Getting started

1. Deploy a federated server using the AI4EOSC dashboard (this is a tool inside the marketplace).
   • General configuration: Give a name and description to the deployment. In service to run select JupyterLab or VSCode if you need to start/stop the process and update the server configuration once started. If you only need to monitor the process you can run fedserver. In case you want to authenticate the client before connecting with the server, select the docker tag tokens, otherwise select latest. In case of deployong the server using the docker tag tokens, you can generate as many federated secrets for authenticating the clients with the server as needed (label them so that you can revoke them if necessary).
   • Hardware configuration: Select the number of CPUs you want, the GB of RAM and disk. Remember that you are deploying the server, which will not run ML/DL models.
   • Federated configuration: Set the number of rounds of the federated learning scheme, the minimum number of clients, the aggregation function, and the error/precision measurement metric (e.g. accuracy).
2. Once you have deployed the federated server, it will appear in your deployment list as a tool. In the tool's information, you can get the Deployment ID (uuid).
3. Each client of the scheme must enter the endpoint to the server as: fedserver-{uuid}.deployments.cloud.ai4eosc.eu. To create the client you can follow the example presented using MNIST.
4. Execute locally the code of each client to start the federated training. Each client must also introduce the token for connecting.

References

[1] McMahan, B., Moore, E., Ramage, D., Hampson, S., & y Arcas, B. A. (2017, April). Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics (pp. 1273-1282). PMLR.

[2] Li, T., Sahu, A. K., Zaheer, M., Sanjabi, M., Talwalkar, A., & Smith, V. (2020). Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2, 429-450.

[3] Reddi, S., Charles, Z., Zaheer, M., Garrett, Z., Rush, K., Konečný, J., ... & McMahan, H. B. (2020). Adaptive federated optimization. arXiv preprint arXiv:2003.00295.

[4] Hsu, T. M. H., Qi, H., & Brown, M. (2019). Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335.

[5] Yin, D., Chen, Y., Kannan, R., & Bartlett, P. (2018, July). Byzantine-robust distributed learning: Towards optimal statistical rates. In International conference on machine learning (pp. 5650-5659). Pmlr.