Federated Learning for Spectrum Sensing in THz Bands

Spectrum sensing at terahertz (THz) frequencies presents significant challenges due to extreme signal attenuation, device heterogeneity, and real-time processing constraints. In this work, we propose a lightweight federated learning (FL) framework for cooperative spectrum sensing, tailored to decentralized secondary users operating under non-identically distributed signal conditions. We adopt a lightweight convolutional neural network (CNN) with a multi-head attention refinement and soft attention-based pooling, enabling efficient processing of magnitude-domain MIMO-OFDM signals while preserving key spectral features. We evaluate our attention-enhanced CNN model within a federated learning setup, comparing it against a standard FL aggregation baseline, a centralized model trained on aggregated data, and a local-only model trained on a single device. Results show that our model with FedProx consistently outperforms all alternatives, highlighting the benefits of proximal regularization under non-IID conditions, and demonstrating that federated learning can match centralized performance while preserving privacy and scalability in heterogeneous THz environments.