A dynamic spectrum access scheme for Internet of Things with improved federated learning

Abstract

The traditional spectrum management paradigm is no longer sufficient to meet the increasingly urgent demand for efficient utilization of spectrum resources by Internet of Things (IoT) devices. Dynamic spectrum access, as an emerging solution, allows devices to intelligently select appropriate spectrum resources based on real-time demands and environmental changes. In this paper, we propose a dynamic spectrum access scheme based on a federated deep reinforcement learning framework, incorporating federated learning, graph neural networks (GNN), and deep Q networks (DQN). In the method, the GNN undertakes the Q-value prediction task, giving full play to its ability to capture inter-device relationships and environmental features. Meanwhile, the DQN learns by interacting with the environment and continuously adapts its strategy to maximize long-term cumulative rewards. To enhance the stability and learning efficiency of the model, we also apply techniques such as empirical playback buffering and updating the target network at fixed intervals. In particular, the use of the FedAge algorithm in federated learning helps to coordinate knowledge sharing and model updates across multiple devices, further enhancing the performance and operational efficiency of the entire system. After several simulation training, the results show that the system model of this paper’s scheme is close to or even better than the traditional federated deep reinforcement learning model in terms of convergence effect and stability while maintaining the privacy-preserving advantages of federated learning. Particularly noteworthy is that in terms of operational efficiency, this paper’s scheme significantly outperforms traditional federated deep learning models.