TrafTrust‐Fed: Traffic‐Trust Federated Learning with Fuzzy Inference.

ABSTRACT Short‐term traffic forecasting is essential for intelligent transportation systems, but centralized learning requires collecting traffic data from geographically distributed sensors, which raises privacy, communication and scalability concerns. Federated learning offers a privacy‐preserving alternative by enabling collaborative model training without sharing raw data. However, under heterogeneous and non independent and identically distributed traffic conditions, conventional uniform aggregation becomes unreliable. This paper proposes TrafTrust Fed, an interpretable trust aware federated learning framework for short term traffic forecasting. The framework evaluates client reliability using four local indicators, namely validation performance, update drift, update alignment and update norm, and integrates them through fuzzy trust inference with temporal smoothing to stabilize trust across communication rounds. The main contribution is an interpretable and stability oriented aggregation mechanism for robust federated learning under heterogeneous urban traffic conditions. Experiments on the METR LA dataset with geographically partitioned clients show that TrafTrust Fed accepts a modest increase in average error, with a 5‐min denormalized MAE of 2.369, while achieving the best Macro F1 of 0.7841 and Recall of 0.7734 among the compared baselines. These results indicate that trust‐calibrated aggregation improves the robustness and interpretability of privacy‐preserving federated traffic forecasting.