Energy Efficient User Scheduling for Hybrid Split and Federated Learning in Wireless UAV Networks

The use of unmanned aerial vehicles (UAVs) as flying users provides various applications by exploiting machine learning (ML) algorithms. Recently, distributed learning algorithms, federated learning (FL) and split learning (SL), have been exploited to train ML models distributedly via sharing model parameters rather than large raw datasets in the conventional centralized learning algorithms. Considering the diversity of users with heterogeneous resources, computation capabilities, and data distributions, we propose a hybrid split and federated learning (HSFL) framework that allows users to select either split training (ST) or federated training (FT) method based on the characteristics of the users in wireless UAV networks. Due to unreliable wireless channels and the limited energy of the users, we further formulate a user scheduling and training method selection problem within HSFL framework as a Multiple-Choice Knapsack Problem (MCKP) and propose an energy-efficient user scheduling algorithm to select a subset of users in each round and schedule each user with either ST or FT method. The simulations demonstrate that our proposed HSFL framework consumes less energy while having the same good test accuracy performance compared to the currently distributed learning algorithms, and the proposed user scheduling algorithm achieves energy-efficient selection of ST or FT method under different distributions.

I. Introduction

Wireless unmanned aerial vehicles (UAVs) network is foreseen to be an integral component in the upcoming sixth-generation (6G) networks [1], [2], which has the potential to support various applications, such as video streaming and disaster surveillance. In these scenarios, the UAVs (users) fly over the target area under the control of the base stations (BSs) to collect data (e.g., images and videos) and then transmit them to the BS for data processing. Each user can observe its local environment and collects a sub-dataset that only contains partial environment information, so all the users should transmit their local sub-datasets to the BS for integrated data processing (e.g., training machine learning (ML) models, including deep neural network (DNN) and convolutional neural network (CNN)). However, the transmission of large datasets causes high communication overhead and large energy consumption for the users, and may potentially reveal user privacy [3]. Thanks to the increased computational capability brought by GPUs, distributed learning becomes more attractive by enabling local learning at users and model aggregation at the BS via only sharing model parameters rather than raw data.