FedCORE: federated knowledge collaboration via orchestration and reciprocal enhancement for model-heterogeneous personalized federated learning

In personalized federated learning, the coexistence of model heterogeneity and non-independent and identically distributed (non-IID) data makes cross-client collaboration challenging. Client models may have inconsistent architectures and distinct local knowledge distributions, while conventional parameter aggregation or simple knowledge averaging is difficult to apply to heterogeneous model spaces and may introduce unreliable supervision. To address this problem, we propose FedCORE, a federated knowledge collaboration framework for model-heterogeneous personalized federated learning. FedCORE constructs a server-side Mixture-of-Experts (MoE) knowledge hub and uses a small class-balanced proxy dataset as a shared semantic anchor to support architecture-agnostic knowledge interaction. Each client first trains a personalized heterogeneous model on its local non-IID data. The server then queries these client models on the proxy data and performs Structured Knowledge Orchestration (SKO), which organizes heterogeneous client responses into a structured MoE global knowledge model through reliable teacher selection, dual-level supervision, teacher-expert responsibility assignment, and routing consistency optimization. Based on the structured server-side knowledge, Confidence-guided Personalized Knowledge Refinement (CGKR) further constructs lightweight client-specific refinement spaces by considering teacher reliability, client-specific class-wise knowledge states, and teacher-client confidence gaps. This enables each client to selectively absorb useful global knowledge while preserving its personalized discriminative structure. Experimental results under various model-heterogeneous and non-IID settings show that FedCORE improves average personalized test accuracy and constructs a server-side knowledge model with better cross-client generalization. Ablation studies and sensitivity analyses further validate the effectiveness and robustness of the proposed components.