Federated Learning-Driven GRU for Modulation Recognition in Cognitive Radio Networks

In the advancing landscape of$\mathbf{6 G}$communication, efficient spectrum utilization and data privacy are crucial. This study presents a distributed spectrum sensing framework that employs Gated Recurrent Units (GRUs) within a Federated Learning (FL) environment to classify modulation types in Cognitive Radio Networks (CRNs). Departing from traditional centralized architectures that raise privacy concerns through a Fusion Center (FC), this approach enables secondary users (SUs) to collaboratively train a global GRU model without exchanging raw data. The architecture promotes privacy compliance while minimizing data leakage risks. To improve model robustness, an outlier filtering mechanism is applied locally at each device, discarding anomalous or misleading samples before model updates. GRUs effectively capture temporal variations in spectrum signals, making the system adaptive to real-time 6 G environments. Experimental results confirm improved classification accuracy, reduced latency, and faster convergence. The proposed GRU-FL system achieved a peak accuracy of$\mathbf{9 2. 6 \%}$, outperforming both centralized and existing federated models.