Spectrum occupancy detection supported by centralized or distributed federated learning

• Federated learning improves detection and robustness for distributed sensing. • Larger models boost accuracy but raise costs, lowering spectral efficiency. • Centralized FL limits messages; distributed FL scales better in large networks. Spectrum occupancy detection is a key issue in enabling dynamic spectrum access. Nowadays, to improve detection effectiveness, the application of machine learning (ML) solutions is popular, and the issue of federated learning (FL) deserves special attention. A key challenge in spectrum occupancy detection is the limited access to labeled training data at the sensors. The sensor network studied in this work comprises a few sensors. These sensors have only partial access to labeled data during the ML model training process. The approach presented in this work uses the exchange of ML model coefficients to overcome the issue of limited training data at individual sensors. Both centralized and distributed FL variants were considered. The results of a hardware experiment involving the detection of the DVB-T signal by several sensors are discussed. In the presence of the sensor without access to labeled data probability of detection in the system increases from 52.06 % to 68.96 % in distributed variant, and to 78.87 % in centralized one. Moreover, the influence of neural network size on additional data transfer was analyzed due to the ML model coefficients exchange.