flower

By keeping data local and only sharing model updates (gradients or weights), Flower allows organizations to collaborate on model training while maintaining data sovereignty and regulatory compliance.

The framework is designed for both research experimentation and production deployment, supporting any ML framework (PyTorch, TensorFlow, JAX, scikit-learn, HuggingFace) as the local training backend through a simple client interface.

Flower handles the federated coordination layer server-side aggregation strategies (FedAvg, FedProx, FedAdam), client selection, communication compression, and differential privacy mechanisms independently of the local training implementation.

This architecture makes it straightforward to federate an existing centralized training pipeline by wrapping the training code in a Flower client.

Flower is open-source under the Apache 2.0 license and developed by Adap, with an active academic research community publishing federated learning algorithms as Flower strategies.

It scales from simulating federated learning on a single machine (using process-level parallelism to simulate multiple clients) to production deployments across thousands of mobile devices or hospitals.

Use cases include medical imaging models trained across hospital networks, mobile keyboard prediction trained on-device across millions of users, and financial fraud detection across banking institutions contexts where data pooling is prohibited but collaborative learning provides meaningful accuracy improvements.