Federated Spatiotemporal Deep Learning for Privacy-Preserving Urban Traffic Flow Prediction

Abstract

This work explores the use of federated spatiotemporal deep learning to achieve accurate predictions of urban traffic flow in practice under real-world privacy restrictions. The primary objective is to create a private, scalable system that can deliver accurate forecasts throughout the entire city. This method uses adaptive attention mechanisms and graph-based neural networks to extract complex spatiotemporal properties from the dispersed sensor data without exchanging raw data. To guarantee generality and model robustness, a customized aggregation algorithm dynamically adjusts to local pattern variance and node dependability. The approach outperforms conventional centralized and federated baselines, with an increase of up to 12.7% in Root Mean Squared Error and up to 14.1% in rare congestion event F1 scores, according to experiments done on three large-scale metropolitan datasets. If the budget for privacy is high, the communication overhead can be lowered by up to 22%. The aforementioned findings demonstrate that the suggested framework can currently accomplish comparatively high forecast accuracy and operating efficiency in a variety of complex urban situations. In summary, by satisfying the requirements of technical viability, expandability, and the new data protection standard, this article offers a strong basis for the future development of smart city traffic management.