A Granularity-Centered Taxonomy of Personalized Federated Learning

Personalized Federated Learning (PFL) has emerged as a key approach to address performance degradation in FL systems under heterogeneous client data. While existing surveys typically categorize PFL methods based on optimization strategies or system-level mechanisms, they often overlook a fundamental question: where is personalization embedded within the model architecture? In this survey, we bridge this knowledge gap and introduce a granularity-centered taxonomy that organizes PFL approaches according to the structural depth of personalization, ranging from head-layer and layer-wise adaptation to model-wise and parameter-wise customization. This novel perspective helps practitioners select appropriate personalization strategies based on model architecture, data heterogeneity, and system constraints. By analyzing representative works published between 2016 and 2026, we identify recurring design bottlenecks and highlight key opportunities for improving personalization across different granularity levels. While existing surveys focus solely on algorithms, our taxonomy provides practitioners and researchers with the informed guidance needed to design scalable and effective PFL systems. Building upon this analysis, we also formulate and discuss several unaddressed open challenges currently present in the field, highlighting specific roadblocks and unresolved questions that must be addressed to drive the future evolution of PFL methodologies.