A Secure Wireless Traffic Prediction With Federated Learning

Feng Li, Xiaona Chen, Kwok-Yan Lam, Li Wang, Xin Liu

Abstract: Accurate traffic prediction is crucial for enhancing the performance of intelligent cellular networks, as it directly impacts the effective allocation of network resources and user satisfaction. The instability and dynamics of traffic data pose challenges to centralized prediction methods. These methods are not only limited in prediction accuracy but may also lead to issues concerning data privacy breaches and response delays. Furthermore, traditional Federated Learning (FL) typically employs a simple averaging strategy during model aggregation, which does not account for the heterogeneity of client data, thereby affecting the generalization performance of the global model. Therefore, this study proposes a wireless traffic prediction model that combines FL, Long Short-Term Memory (LSTM), and Kolmogorov-Arnold Networks (KAN). The aim is to enhance prediction accuracy while ensuring data privacy. The model leverages edge computing and FL techniques, enabling multiple edge devices to locally train LSTM models. By integrating KAN to explore the inherent complexity of traffic data, it enhances the model's adaptability. Furthermore, the Federated Normalized Averaging (FedNova) model is introduced, which thoroughly considers the heterogeneity of data among devices and optimizes the process of updating model parameters. The extensive experimental results conducted on the Milano real dataset demonstrate that the FedNova-LSTM-KAN model exhibits significant advantages compared to existing centralized and FL methods when dealing with uneven client traffic data.

External IDs:doi:10.1109/tvt.2025.3594326