Clustered Federated Learning via Embedding Distributions

Dekai Zhang, Matt Williams, Francesca Toni

Main Idea

EMD-CFL is a novel clustered federated learning method, which uses embedding space distribution distances to identify clusters of similar clients.

Project Structure

The main source code is in the src directory. Data should be placed in a newly created data directory. Scripts to run each of the CFL methods can be found in the root directory, along with a config.yaml. Model checkpoints can be evaluated using test.py and the test_config.yaml.