Lightweight Federated Learning based White Space Detector for Cognitive Radios
Md Mehedi Hassan Galib, Mohamed Younis · 2024
In cognitive communication, detecting white space is critical for preventing interference with primary user’s transmissions. Rather than modifying the radio transceivers, several machine learning (ML)-based detection techniques have been proposed, where a model is trained offline and then employed to infer white spaces in real-time. Despite their viability for spectrum sensing, these techniques are computationally complex, especially when deep neural network (DNN) models are pursued. Moreover, training a centralized model requires transmitting massive data to the fusion center (FC), which ultimately results in congestion on the transmission channel rather than flagging opportunities for cognitive transmissions. This paper opts to fill the technical gap by proposing: (1) a federated learning model that enables effective design of spectrum monitors in a distributed manner, and (2) a spiking neural network (SNN)-based lightweight white space detector that solves the issue of computationally expensive DNN models and is suitable for resource-constrained devices. The SNN-based federated learning (SFL) model employs secondary users to train their respective SNN models using local data (spatial locality) and sends the gradient of SNN to FC. FC combines the individual SNN models and sends the aggregated model back to each edge node. Validation using live LTE data has demonstrated the effectiveness of SFL with a detection accuracy of 91.16%.