FedWS: Dealing with Heterogeneous Data on Federated Learning

Unmanned Aerial Vehicles (UAVs) significantly deliver safe, sustainable, and autonomous services to Intelligent Transportation Systems and urban air mobility applications. In this context, image classification tasks have an important role in these applications, and Federated Learning (FL) can empower these services with a decentralized approach to machine learning tasks taking advantage of UAV's broad access to distributed data and the edge computing capacity of these devices. However, data can have a heterogeneous distribution, negatively affecting federated training and impacting the accuracy of results. To address this issue, we propose a method called Federated Learning Algorithm with Weight Standardization on Convolutional Layers (FedWS) that locally normalizes the weights of a neural network's convolutional layers. The results show that FedWS provided more smoothness on losses in the training process at the aggregation level, allowed 3 % to 6 % higher accuracy in different levels of heterogeneity, and reduced the communication cost in a margin of 25% to 50% in a more linear manner compared to others methods on image classification tasks in the EuroSAT dataset.