Federated Spiking Network Framework for Spectrum Access Prediction in Interweave Cognitive Communications

This paper presents FeSAP, a Federated Spiking Adaptive Prediction framework for opportunistic spectrum access in distributed Cognitive Radio Networks (CRNs). FeSAP integrates a Recurrent Temporal Network (RTN), trained using unsupervised Spike-Timing Dependent Plasticity (STDP), with a Feedforward Spectral Network (FSN), trained locally at each Secondary User (SU) via supervised learning using surrogate gradients with Batch Normalization Through Time (BNTT) and globally aggregated at the Fusion Center (FC) using Federated Averaging (FedAvg). This hybrid spiking architecture enables anticipatory prediction of Primary User (PU) activity while addressing key challenges in federated learning, including data heterogeneity, intermittent SU participation, and privacy constraints. Evaluations on 5G Non-Standalone (NSA) downlink transmission data with Long-Term Evolution (LTE) resource reservations, as well as synthetically constructed non-Independent and Identically Distributed (non-IID) scenarios, demonstrate that FeSAP achieves an F1 score of 87.95%, full utilization of all 13 symbols per subframe, and convergence within 25–30 global rounds with only 60–80% SU participation. Temporal offset analysis further confirms its robustness when operating with delayed subframe inputs, while its low parameter count and modest computational footprint underscore its suitability for deployment in edge-based CRNs with limited processing resources.