This paper proposes FedKWAZ, a framework that enhances heterogeneous federated learning by bridging semantic and decision gaps via Knowledge Weak-Aware Zones.