Flower

An open-source Python framework for federated learning: you train one model across several sites while patient data stays where it is.

What it is¶

With Flower you write a client app that trains on local data and a server app that combines the model updates, using a built-in strategy such as federated averaging or one you write yourself. It works with common machine learning libraries, including PyTorch, TensorFlow, scikit-learn, XGBoost and the medical imaging library MONAI. You can simulate a federation on one machine first and then deploy it across real sites. It also offers add-ons for differential privacy and secure aggregation, which reduce what the server can learn from a single site’s updates. Flower started as a research project at the University of Oxford and is now maintained by Flower Labs with an open community.

In healthcare¶

Federated learning suits multi-hospital projects where data-sharing agreements or the GDPR prevent pooling patient records. Note that federated training alone does not make a model private: model updates can still leak information, so combine it with differential privacy or secure aggregation and a security review.

Where to use it in The ELSA Way¶

• Privacy and security › Privacy-preserving techniques: Flower implements federated learning with optional differential privacy and secure aggregation, which lets you measure the performance trade-offs this passage warns about.

References¶

• Beutel DJ, Topal T, Mathur A, Qiu X, Fernandez-Marques J, Gao Y, et al. Flower: A Friendly Federated Learning Research Framework. arXiv preprint arXiv:2007.14390. 2020. Beutel et al. (2020)

Related tools¶

• ARX: Free desktop software and Java library for anonymising structured personal data, with built-in checks of re-identification risk and of how much useful information remains.

• MLflow: An open-source platform that records what went into each model run and keeps a versioned registry of trained models, which lets you trace any deployed model back to its code, data and settings.