QST-Fed: Quantum-Secure Spatiotemporal Federated Learning for Cross-Region 5G Traffic Forecasting

Accurate cross-regional 5G traffic forecasting is critical for network operations but is hampered by data silos and privacy vulnerabilities of conventional federated learning (FL). Moreover, existing models struggle to effectively capture the complex spatiotemporal dependencies in base station traffic (BST). To address these challenges, this paper proposes QST-Fed, a Quantum-Secure Spatiotemporal FL framework. We design a hybrid spatiotemporal model at client level: a 1D-convolutional neural network captures local temporal patterns, a service-aware encoding informed by Prophet-decomposed periodicity adapts to daily traffic cycles, and a multi-dimensional graph convolutional network models complex spatial correlations by fusing geographic proximity, regional type, and traffic similarity. For privacy preservation, the framework integrates quantum key distribution(QKD) to establish one-time pad keys between the server and clients, enabling block-level encryption of model parameters to achieve information-theoretic security—a solution that eliminates the three key pain points of accuracy loss, prohibitive computational overhead and scalability bottlenecks in large-scale 5G scenarios, compared with traditional privacy-enhancing technologies. Experiments on real-world multi-region 5G dataset demonstrate that QST-Fed significantly outperforms baseline models in short, medium, and long-term forecasting tasks. Furthermore, with INT8 model quantization, the framework improves quantum encryption/decryption efficiency by over 3-fold and reduces model storage and bandwidth requirements by 75%, all with negligible loss in prediction accuracy.