Federated learning for improved automatic modulation classification: data heterogeneity and low SNR accuracy

Abstract

This research introduces novel federated learning (FL)-based methods, namely FedAvgAMC, FedPerAMC, and FedAvgAMCAug, to enhance automatic modulation classification (AMC) in communication systems. By addressing data heterogeneity, incorporating higher-order modulation schemes, and exploring personalized models for clients, these methods overcome challenges associated with centralized learning and data leakage. Simulation results demonstrate significant improvements in accuracy, particularly at low and moderate signal-to-noise ratios (SNRs). FedAvgAMCAug, a privacy-preserving method, achieves superior performance, outperforming CentAMC by 18% in accuracy at an SNR of 5 dB and consistently surpassing existing learning techniques across all clients and SNR values. This research highlights the potential of federated learning to enhance AMC accuracy in real-world scenarios, paving the way for future optimizations and exploration of techniques to address data uncertainty for even more robust performance.