Federated learning for radio environment mapping

Abstract

My work extends previous work on proactive spatial prediction of radio environments for spectrum sharing by integrating it into a distributed Federated Learning (FL) framework. In this enhanced version of the Federated ProSpire system, clients locally train their models and communicate after several rounds of local training. This decentralized approach allows for the secure and efficient sharing of learned models without requiring centralized data collection. We investigate the impact of hyperparameters such as local epochs, communication rounds, and data augmentation on model performance. Additionally, we explore the effects of data distribution under both IID (Independent and Identi cally Distributed) and non-IID conditions, with the latter simulating location-based variations and RSS strength differences across clients. Experiments show that the Federated ProSpire system maintains comparable accuracy in signal strength prediction, with a mean absolute error of around 5 dB, while enhancing scalability, privacy, and robustness against interference under some conditions. These findings demonstrate the suitability of our framework for emerging spectrum-sharing paradigms in next-generation wireless networks, offering a flexible and scalable solution for collaborative, privacy preserving radio environment prediction.