Privacy-Preserving Human-Centric Traffic Management: A Federated Multimodal Framework with Standardized Evaluation

Centralized ITS suffer from privacy leakage and poor cross-regional scalability, while human factors (e.g., driver cognitive state, personalized needs) are often overlooked. To solve these issues, this paper presents a privacy-preserving human-centric traffic framework based on federated learning (FL). The framework’s core innovations include: (1) dynamic sparse training to cut communication costs without performance loss; (2) few-shot adaptation to enable deployment in data-scarce regions; (3) multimodal modules (spoken language understanding, zero-shot recommendation) for intuitive human-ITS interaction; (4) driver state inference to adapt alerts based on cognitive status. A standardized evaluation using SBM-DEA measures technical efficiency, user experience, and adaptability. Results on two traffic datasets demonstrate 2.30 prediction MAE (18.7% better than FedAvg), 18.1 MB/round communication cost, and 0.89 DEA efficiency score. This framework advances ITS by unifying privacy protection, human-centric design, and rigorous performance assessment.