Federated Learning: Biased Client Selection Reproduction

Developed a comprehensive tutorial for the Flautim Platform, focusing on the reproducibility of the paper “Towards understanding biased client selection in federated learning” [Jee Cho et. al, 2022].

Code available at https://github.com/carlabferreira/flautim_tutoriais/blob/main/TUTORIAL_4.ipynb

Implementation

Utilized Python and the Flower Framework to architect a federated environment.

Simulation

Modeled a network of 100 clients, implementing a probabilistic algorithm to select 3 clients per round (from 30 candidates) based on dataset size and local loss.

Objective

To empirically demonstrate the impact of non-uniform client selection on global model convergence and bias.