Entropy-based Adaptive Gradient Quantization in Federated Learning for Internet of Vehicles

Abstract

Federated learning for internet of vehicles builds an intelligent transportation system with real-time responsiveness and intelligent collaborative training of high-quality models by integrating traffic data between vehicle nodes, roadside units, and infrastructure. As the internet of vehicles architecture continues to expand, frequent gradient data interactions between roadside units and vehicle nodes lead to increased uplink channel load and communication delay in federated learning systems. To alleviate the communication delay problem, existing works propose gradient quantization algorithms to reduce the communication bandwidth overhead by reducing the transmission of redundant data. However, the existing gradient quantization algorithms' undifferentiated discarding of gradient data leads to a reduction in the accuracy of the aggregation model. To balance model accuracy and communication overhead, we propose an entropy-based adaptive gradient quantization for federated learning (eaqfed). The eaqfed dynamically adjusts the gradient quantization level through the entropy property during model updating to maintain model accuracy while reducing communication cost.