Federated Learning — Client Selection Techniques and Model Pruning

Machine Learning & Deep Learning — Politecnico di Torino

Overview

This project explores advanced strategies for Federated Learning (FL) — a machine learning paradigm where a model is trained across multiple decentralized clients without sharing raw data. The focus is on three key challenges in FL:

1. Client Selection — which clients participate in each training round
2. Domain Generalisation — making models robust across heterogeneous data distributions
3. Model Pruning — reducing model size and communication cost without sacrificing performance

The goal was to improve the efficiency and generalisability of federated models under realistic heterogeneous data conditions.

Techniques Implemented

Client Selection Strategies

• Random selection (baseline)
• Loss-based selection — prioritising clients with higher local loss
• Power-of-Choice — biased client selection to improve convergence on heterogeneous data

Domain Generalisation

• FedAvg (baseline federation algorithm)
• Experiments with data heterogeneity (non-IID distributions across clients)

Model Pruning

• Structured pruning — removing entire filters/neurons to reduce model size
• Evaluated trade-off between model compression and accuracy retention

Dataset

Experiments conducted on standard FL benchmarks with non-IID data partitioning to simulate realistic heterogeneous client distributions.

Tech Stack

Python, PyTorch, NumPy etc.

Authors

Silva Bashllari & collaborators — Politecnico di Torino, MLDL January 2024

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)