Privacy-Preserving Cross-Domain Federated Learning for Real-Time Traffic Anomaly Detection

Smart city traffic operations require learning from data generated by multiple cities agencies and sensing platforms under strict privacy and latency constraints. Centralized training is impractical due to regulatory limits bandwidth cost and operational risk while existing federated learning methods struggle with cross domain shift real time data streams and static privacy control. This paper proposes a privacy preserving cross domain federated learning framework for real time traffic anomaly detection that integrates federated transfer learning adaptive differential privacy and streaming model updates within a unified architecture. Regional clients train locally on UAV and UGV sensor or video data and share privacy controlled model updates for global aggregation without exposing raw data. A domain adaptation mechanism enables zero shot generalization to new cities with unseen traffic patterns while an adaptive privacy budget dynamically balances data sensitivity trust level and update frequency against model utility. Online inference and incremental learning support early detection of congestion incidents and abnormal traffic events under live traffic streams. Experimental evaluation against standard federated learning cross silo federated learning and FedBCD demonstrates improved cross city generalization faster adaptation under non IID data reduced communication overhead and lower privacy cost at comparable anomaly detection accuracy.