FedTT: A Federated Spatio-Temporal Learning Framework for Cross-City Traffic Knowledge Transfer

Abstract

Traffic prediction aims to forecast future traffic conditions from historical spatio-temporal data. In practice, many cities face limited traffic data due to incomplete sensing infrastructures, making accurate prediction difficult. Federated traffic knowledge transfer offers a promising solution by allowing data-rich source cities to assist data-scarce target cities while keeping local data private. However, existing methods still face key challenges, including cross-city distribution discrepancies, incomplete or noisy observations, and potential privacy leakage from shared gradients or model parameters during federated optimization. To this end, we propose FedTT, a federated spatio-temporal learning framework for cross-city traffic knowledge transfer. FedTT features three co-designed components: (i) a Traffic Domain Adapter (TDA) that explicitly aligns heterogeneous traffic data distributions across cities to improve transfer effectiveness, (ii) a Traffic View Imputation (TVI) module that enhances data quality by completing missing traffic observations through spatio-temporal dependency modeling, and (iii) a lightweight Traffic Secret Aggregation (TSA) protocol that enables attack-resistant knowledge aggregation with formal privacy analysis, without relying on heavy cryptographic primitives. Extensive experiments on four real-world traffic datasets demonstrate that FedTT consistently outperforms 18 state-of-the-art baselines, achieving improved prediction accuracy while maintaining strong empirical resistance to attacks.