FedTT: Cross-City Federated Traffic Knowledge Transfer with Privacy Preservation

Zhihao Zeng, Ziquan Fang, Yuting Huang, Yujie Fang, Lu Chen, Yunjun Gao

Keywords: Traffic Prediction, Transfer Learning, Federated Learning

Abstract: Traffic prediction (TP) is a core task in urban computing, aiming to forecast future traffic conditions from historical observations. To overcome the scarcity of traffic data in emerging cities, recent studies have explored Federated Traffic Knowledge Transfer (FTT), which leverages data-rich source cities to assist data-scarce target cities without raw data sharing. However, existing FTT approaches are limited by three unresolved challenges: (i) potential *privacy leakage* since gradients or parameters generated during federated computing can still be inverted, (ii) severe *cross-city distribution discrepancies* that reduce transfer effectiveness, and (iii) \textit{low data quality} caused by missing or unreliable sensor readings. To address these challenges, we propose **FedTT**, a novel federated framework for cross-city traffic knowledge transfer with privacy-preserving. FedTT introduces three innovations: (i) a lightweight **Traffic Secret Aggregation (TSA)** protocol that achieves secure knowledge aggregation without sacrificing efficiency or accuracy; (ii) a **Traffic Domain Adapter (TDA)** that explicitly aligns heterogeneous source–target distributions for more effective transfer, and (iii) a **Traffic View Imputation (TVI)** method that leverages spatio-temporal dependencies to complete missing traffic data robustly. Extensive experiments on four real-world datasets show that FedTT achieves significant improvements over 14 state-of-the-art baselines, consistently reducing prediction error while maintaining strong privacy protection.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 23780