• Federated learning
• personalization

The thesis presents a novel framework to address core challenges in Federated Learning (FL), particularly scalability, communication efficiency, and personalization in non-IID data environments.

To tackle these issues, the work introduces CAMEO (Client Masking for Efficient Organization), a decentralized and scalable Personalized FL framework. CAMEO operates by embedding client data characteristics into a compact representation, then clustering clients based on these embeddings via Affinity Propagation. Within each cluster, CAMEO enforces a shared structured pruning mask, reducing model size while preserving the components most relevant to that cluster’s data distribution.

The thesis explores several core questions:
• Can clustering clients by data similarity improve both communication and model quality?
• Is it more efficient to use a single shared pruning mask per cluster, or should each client retain an individual mask?
• How does varying client participation impact model convergence, accuracy, and fairness?

Comprehensive experiments validate the proposed approach. CAMEO reduces communication overhead by up to 60% and maintains high performance, achieving comparable accuracy to more communication-intensive methods. While CAMEO does not lead to improvements in fairness or convergence speed, it establishes a practical and scalable approach to combining client clustering and pruning for personalized FL under non-IID settings.