Federated Learning for Spatiotemporal Traffic Prediction: Privacy, Efficiency, Security and Emerging Extensions

Abstract

Spatiotemporal traffic prediction is very important for improving traffic flow and reducing congestion. But standard centralized prediction methods often cause big worries about data privacy and security, because they need to collect and store lots of sensitive traffic data on central servers. To solve this problem, this survey looks at using distributed federated learning for spatiotemporal traffic prediction. This paper reviews and sorts recent studies in this area, grouping them into a few main directions. These include privacy-preserving Federated Learning (FL) frameworks, efficiency-oriented FL methods, security and trust-enhanced FL systems, emerging optimization and heterogeneity-aware extensions. The survey shows that distributed federated learning can reach good prediction results, and it protects raw traffic data by keeping it on local devices. It also shows that recent studies make better models using graph-based spatial modeling, time learning modules, adaptive aggregation, blockchain-assisted checking, and personal collaboration. But, the literature also points out some big limits. These are mainly high communication costs, non-independent and non-identically distributed data, and the hard work of building secure, scalable, and steady real-world systems. So, this survey finds that distributed federated learning is a good way for privacy-preserving traffic prediction. Even so, further study still needs more research on adaptive optimization, strong collaboration under mixed data, and combined privacy and security tools for real large-scale uses.