Design and Implementation of Spectrum Sensing Algorithm Based on Federated Learning Training and Clustering Detection

In the scenario of crowd movement, Dynamic Spectrum Access (DSA) is crucial for improving spectrum utilization. However, traditional cooperative spectrum sensing technologies have certain limitations. For instance, they require extensive information exchange during data transmission, resulting in high communication overhead. Particularly, when users move dynamically, frequent position changes further reduce system efficiency. Moreover, in the face of large-scale nodes and complex mobile scenarios, problems such as excessive computational pressure on the central node and slow response speed easily arise, making it difficult to sense the spectrum in real time and accurately. Based on the aforementioned challenges, this paper proposes a spectrum sensing algorithm employing Federated Learning (FL) for model training and a clustered architecture for real-time detection. During the training phase, the Fusion Center (FC) initializes global model parameters and distributes them to all nodes. Each node independently performs multiple local training iterations and uploads its optimized parameters to the FC. The FC then aggregates these parameters through weighted averaging and redistributes the updated model. This process iterates until model convergence is achieved. In the detection phase, nodes are dynamically clustered. Each cluster performs autonomous spectrum sensing, with cluster heads making local decisions. The FC ultimately fuses these decisions for global inference. This architecture enhances parallel processing capability and responsiveness, demonstrating superior adaptability to dynamic scenarios with mobile user populations. Experimental results demonstrate that the proposed algorithm achieves higher accuracy and efficiency in dynamic spectrum sensing, significantly enhancing spectrum utilization.