FedPML: Federated Prototype Meta-Learning for Adaptation and Generalization in Distributed Environments

Federated learning (FL) holds significant promise for enabling automatic modulation classification (AMC) across distributed wireless Internet-of-Things (IoT) systems while preserving data privacy. Yet, its deployment in real-world scenarios is hindered by two fundamental limitations: the inability to generalize across non-independent and identically distributed (non-IID) clients without sacrificing personalization, and the lack of fast, data-efficient local adaptation mechanisms under non-IID, resource-constrained conditions. In this work, we present FedPML, a novel federated prototype-based meta-learning framework that directly addresses these challenges through a unified and scalable approach. FedPML leverages a lightweight Reptile-based meta-optimization strategy and class-level prototypes to jointly optimize for cross-client consistency and local task adaptability. This design enables rapid personalization with minimal data and computation, while preserving global knowledge alignment across heterogeneous clients. Extensive experiments on standard AMC benchmarks under a wide range of IID and non-IID conditions show that FedPML consistently outperforms strong FL baselines, achieving higher accuracy, faster convergence, and reduced communication overhead. The results demonstrate that FedPML is a powerful and practical solution for realizing resilient, low-latency AMC in decentralized IoT networks.