FedMCP++: Integrating Modular Expert Heads with Prototype-Guided Contrastive Distillation for Wireless Personalized Federated Learning
Federated learning (FL) enables collaborative model training across decentralized clients while preserving data privacy, yet real-world deployments still face communication bottlenecks, performance degradation under heterogeneous client data, and limited personalization. In this study, we introduce FedMCP++, a modular and communication-efficient personalized FL framework in which every client owns a complete private model—a lightweight convolutional backbone with a private expert head—and collaboration is carried out entirely through knowledge exchange rather than parameter exchange. In each round, clients share only temperature-softened class predictions and class-wise feature prototypes computed on a small public proxy set; the server fuses them into an accuracy-weighted teacher and broadcasts the result, and clients realign their models through knowledge distillation, an instance-level contrastive objective, and prototype alignment. We evaluate FedMCP++, its ablations, and two knowledge-based baselines on six benchmark vision datasets with 10, 20, and 30 clients. The results indicate dataset-dependent trade-offs rather than uniform superiority: collaborative distillation improves average client-level accuracy over independent local training in twelve of eighteen configurations—most clearly under severe per-client data scarcity (e.g., up to +2.7 percentage points on KMNIST and +2.4 on STL-10 with 20–30 clients)—whereas independent training ensembles remain strongest on SVHN and CIFAR-10 at the studied budgets. Because no parameters are transmitted, the per-round uplink payload is a fixed-size 42.6 KB message, 9.9–12.8× smaller than full-model synchronization, and is invariant to model capacity. These properties make FedMCP++ a flexible framework for personalized FL in wireless edge and Internet of Things environments where bandwidth and privacy constraints are paramount.