Introduction to FL frameworks

First, we introduce the FL frameworks and platforms to be used in the subsequent implementation-focused sections.

Flower

Flower (https://flower.dev/) is an open source and ML framework-agnostic FL framework that aims to be accessible to users. Flower follows a standard client-server architecture, in which the clients are set up to receive the model parameters from the server, train on local data, and send the new local model parameters back to the server.

The high-level orchestration of the federated learning process is dictated by what Flower calls strategies, used by the server for aspects such as client selection and parameter aggregation.

Flower uses Remote Procedure Calls (RPCs) in order to perform said orchestration through client-side execution from messages sent by the server. The extensibility of the framework allows researchers to experiment with novel approaches such as new aggregation algorithms and communication methods (such...