FediTT: A Federated Learning Approach for Privacy-Preserving Traffic Flow Prediction

Traffic flow prediction (TFP) is critical to the development of intelligent transportation systems, yet current deep learning models often rely on centralized training with sensitive location and mobility data, leading to significant privacy concerns. To address this, we propose FediTT, a novel federated learning framework that integrates the iTransformer architecture to enable accurate and privacy-preserving traffic forecasting. By training models locally on distributed clients and aggregating updates without sharing raw data, FediTT ensures data privacy while maintaining strong modeling capacity. Unlike traditional federated baselines such as FedGRU, FediTT employs a Transformer-based backbone to better capture long-range dependencies and multivariate correlations in time series data. We conduct extensive experiments on real-world traffic datasets, evaluating performance across multiple prediction horizons (96, 192, and 336 time points). Results show that FediTT consistently achieves lower MSE and MAE than centralized deep learning models (Transformer, PatchTST, Crossformer) and federated baselines, demonstrating both superior accuracy and enhanced privacy protection. These findings validate the effectiveness of combining a high-capacity forecasting model with federated learning and position FediTT as a practical and scalable solution for privacy-sensitive applications in smart transportation systems.