Standard federated learning often operates under simplifying assumptions, such as clients having identically distributed data and similar computational resources. Real-world deployments frequently violate these assumptions. Client datasets typically exhibit statistical heterogeneity (Non-IID data), meaning the local data distribution $P_k(x, y)$ for a client $k$ can differ substantially from the global distribution $P(x, y)$. Concurrently, variations in client hardware, network speed, and availability introduce systems heterogeneity. Both types can hinder convergence, degrade model accuracy, and lead to fairness issues.

This chapter provides techniques to address these common difficulties. We will investigate methods designed to mitigate the negative effects of Non-IID data and system variability. Recognizing that heterogeneity implies a single global model may not serve all clients optimally, we will also study personalization in federated learning. You will learn about strategies including:

• Handling different data distributions across clients.
• Grouping similar clients using clustered federated learning.
• Applying meta-learning principles for efficient client-specific model adaptation.
• Structuring the problem using multi-task learning frameworks.
• Adjusting models to fit diverse device capabilities.