SemanticDFL: Similarity-Aware Pull-based Personalized Decentralized Federated Learning

Personalized decentralized federated learning (PDFL) seeks to tailor models to heterogeneous clients without a central coordinator, yet gossip-style mixing on large graphs dilutes minority signals and assumes any-to-any connectivity. We present SemanticDFL, a fully decentralized, pull-based personalization layer that organizes peers into a hierarchical semantic overlay network (SON). Each client publishes a compact top- P model signature; proximity-bounded discovery forms zones that are clustered using affinity propagation and stewarded by replica-backed super-peers that route bounded-fanout similarity queries. Clients then pull only the K most similar models for personalized aggregation, concentrating communication and computation where they matter most. We prove a lower bound that links spectral mixing and data heterogeneity to an irreducible mis-aggregation penalty for graph-oblivious, push-based overlays, thereby motivating the proposed similarity-aware pull method. A prototype and large-scale evaluation on FMNIST, Tiny ImageNet, Google Speech Commands, and 20 Newsgroups under Dirichlet and pathological splits (50--400 peers on the EU SLICES testbed) show that SemanticDFL improves final accuracy by 3--12% over strong decentralized personalized baselines, reaches target accuracy with ≈2.5× fewer rounds than FedAvg, and requires ≈1.3× fewer rounds than the best DPFL alternative. It adds only ≈1.7--12.6% per-round overhead across all settings while maintaining Recall@K ≈ 0.88--1.00, positioning similarity-aware pull over semantic overlays as a scalable path to high-quality personalization in decentralized FL.