FedSup: A communication-efficient federated learning fatigue driving behaviors supervision approach

Abstract

Listen

Translate article

With the proliferation of smart devices and the Internet of Vehicles (IoV) technologies, intelligent fatigue detection has become one of the most-used methods in our daily driving. Data sharing among vehicles can be used to optimize fatigue detection models and ensure driving safety. However, data privacy issues hinder the sharing process. Besides, due to the limitation of communication and computing resources, it is difficult to carry out training and data transmission on vehicles. To tackle these challenges, we propose FedSup, a communication-efficient federated learning method for fatigue driving behaviors supervision. Inspired by the resources allocation mechanism in edge intelligence, FedSup dynamically optimizes the sharing model with tailored client-edge-cloud architecture and reduces communication overhead by a Bayesian Convolutional Neural Network (BCNN) data selection strategy. To improve the sharing model optimize efficiency, we further propose an asynchronous parameters aggregation algorithm to automatically adjust the mixing weight of each edge model parameter. Extensive experiments demonstrate that the FedSup method is suitable for IoV scenarios and outperforms related federated learning methods in terms of communication overhead and model accuracy. • We design cloud–edge-client fatigue driving supervision approach FedSup, which is communication-efficient, accurate, and privacy-preserving. • In FedSup, we propose an efficient BCNN-based method to quantify the uncertainty of the image, which minimizes the amount of uploaded data and requirements of clients’ computing resources. • In FedSup, we propose an asynchronous parameter aggregation algorithm based on uncertainty weight to more effectively integrate local models trained in edge nodes, which improves the accuracy of the central model and reduces communication rounds in the training process. • We show that our method, FedSup, significantly outperforms other related FL-based methods in communication overhead and accuracy, across multiple clients with both batch-IID and NonIID data.