GeoFed: A Geography-Aware Federated Learning Approach for Vehicular Visual Crowdsensing

Internet of Things (IoT) technology enables enhanced connectivity and information sharing among various devices and platforms. In the context of vehicular crowds ensing, this connectivity has opened up new way to collect environmental data via Vehicle-based Visual Crowdsensing. However, the heterogeneity of data sources and the presence of vehicle outliers pose challenges of ensuring the reliability and accuracy of the machine learning (ML) models. We propose GeoFed, a geography-aware federated learning (FL) approach for vehicular visual crowdsensing. Here, geographically similar vehicular fog nodes (VFNs) collaborate to train a cluster model unlike the traditional FL approaches where vehicles participate to train a model. To further improve GeoFed's performance, we employ the deep Q-Network (DQN) algorithm to intelligently determine the participation of vehicles in the FL process. Through extensive experiments on our own collected real-world dataset, we find that our proposed GeoFed not only outperforms the state-of-art FedAvg with higher F1 score (1.18 x) and mAP (1.14 x), but also achieves a faster convergence rate with less loss (80%).