A Novel Distributed Solution for Automatic Modulation Classification Based on Federated Learning and Modified LSTM

Abstract

Automatic modulation classification (AMC) is an indispensable technique in developing radio monitoring. It can automatically determine the modulation mode according to the collected radio signal. Due to the large amount of radio monitoring data being stored in the central server for AMC method, the risk of data leakage and insufficient communication bandwidth will arise. In this paper, an innovative learning framework - federated learning based on modified long short-term memory (FL-MoLSTM) is proposed for AMC. The framework of federated learning is adopted to save the limited communication bandwidth and improve data security. LSTM with attention mechanism is put forward, which can assign the weight of the learned features and reduce data redundancy. The federated averaging (FedAvg) algorithm is used for optimization. According to the characteristics of the modulated radio signals, the joint augmentation policy (JAP) combining rotation and flipping is drawn to improve the classification accuracy in FL-MoLSTM. Lastly, in FL-MoLSTM, the bandwidth shortage problem is addressed while protecting data privacy without causing severe performance loss. Our results show that the classification accuracy of FL-MoLSTM reaches more than 90% in AMC.