Cross-cluster precision-guided knowledge fusion for fair and personalized federated learning

We propose a Personalized Precision-Centric Clustered Federated Learning (PPCFL) framework to address statistical heterogeneity and model personalization challenges in Federated Learning. PPCFL introduces a layer-wise, angular dissimilarity-based clustering mechanism that groups nodes with similar optimization trajectories, guiding both intra- and inter-cluster weight aggregation. A class precision-centric aggregation strategy assigns weights to model updates based on each node’s class-wise predictive performance, ensuring a more representative and personalized global model. To preserve local model fidelity after aggregation, we adopt a bias-aware model consolidation approach using functional parameter decoupling, aggregating feature extraction and classification layers separately based on dataset size and class-wise precision. We further introduce a cross-cluster knowledge fusion strategy that leverages shared hidden layer weights to integrate generalizable representations across clusters while retaining domain-specific adaptations. To support scalability and real-time coordination, we incorporate a secure Digital-Twin based orchestration layer that virtualizes edge nodes for efficient management. Unlike traditional simulation-driven digital twins, our implementation provides a web-based control interface enabling real-time node registration, training configuration, and performance monitoring. Empirical evaluations across heterogeneous data modalities demonstrate improved convergence and enhanced generalization under non-IID conditions. This work lays a practical foundation for scalable, secure, and adaptive federated learning systems in real-world decentralized environments.