Personalized Federated Learning With Mixture-of-Experts for Automatic Modulation Classification Under Data Heterogeneity

Automatic Modulation Classification (AMC) is a pivotal task for cognitive and intelligent communication systems. Existing AMC methods predominantly rely on centralized deep learning, necessitating the collection and centralized training of modulation signals from various terminals to recognize modulation types. However, in practical applications, there are still two main challenges: (i) signal data are often distributed across different devices and data privacy and security constraints often prohibit centralized data aggregation; (ii) Complex and heterogeneous wireless environments lead to Non-Independent and Identically Distributed (Non-IID) data and missing-class phenomena across terminals, rendering a unified global model ineffective for personalized recognition. To address these limitations, this paper proposes a heterogeneous personalized federated learning framework based on a Mixture-of-Experts (MoE) model, termed pFedMixMoE, for the AMC task. Specifically, the proposed framework employs a dynamic gating mechanism to adaptively determine the fusion ratio between a globally shared expert for generalization and client-specific local experts for personalization, thus effectively balancing global knowledge transfer and tailored local adaptation. Furthermore, we design a QKV-based frequency-domain aggregation strategy to improve robustness under heterogeneous environments. Extensive Experimental results demonstrate that pFedMixMoE enhances recognition accuracy by 17.43% and 28.32% compared to mainstream baseline methods under various data distribution and sample type scenarios. These results validate the superior effectiveness and robustness of pFedMixMoE and provide new ideas for constructing personalized models in wireless networks.