FedBA Non-IID Federated Learning Framework in UA V Networks Pei Li

2025-04-27 0 0 470.68KB 5 页 10玖币

侵权投诉

FedBA: Non-IID Federated Learning Framework in

UAV Networks

Pei Li

School of Data Science and Technology

Heilongjiang University

Harbin, China

2212623@s.hlju.edu.cn

Zhijun Liu

School of Data Science and Technology

Heilongjiang University

Harbin, China

2222708@s.hlju.edu.cn

Luyi Chang

School of Data Science and Technology

Heilongjiang University

Harbin, China

2202518@s.hlju.edu.cn

Jialiang Peng

School of Data Science and Technology

Heilongjiang University

Harbin, China

jialiangpeng@hlju.edu.cn

Yi Wu*

School of Data Science and Technology

Heilongjiang University

Harbin, China

1995050@hlju.edu.cn

Abstract—Advances in artiﬁcial intelligence technology and the

popularity of the Internet of Things (IoT) devices have brought

great convenience to people’s lives and signiﬁcantly improved pro-

ductivity. As a new type of Internet of Things device, Unmanned

Aerial Vehicles (UAVs) have broad development prospects and

have become a hot research ﬁeld. However, due to privacy

concerns and the limited communication resources of UAVs, it is

impractical for UAV devices to transmit their raw data via the

air link. Compared with centralized machine learning, Federated

Learning (FL) necessitates the exchange of gradient instead of

local data among participating clients and servers, effectively

protecting user privacy and reducing the communication cost

of participating devices, which is especially suitable for UAV

networks. Nevertheless, there are signiﬁcant differences in the

images captured by different types of drones carrying cameras

to different areas (i.e., the problem of statistical heterogeneity),

which is still challenging for training FL models. To this end, we

propose an aggregation rule based on the distance between local

and global models, named FedBA, to alleviate the problem of

data heterogeneity in UAV-assisted FL. Results from experiments

demonstrate that, on three real-world data sets (i.e., CIFAR-100,

MNIST, and Fashion-MNIST), our proposed approach performs

noticeably better than the conventional FL algorithm.

Index Terms—federated learning, statistical heterogeneity, un-

manned aerial vehicle, aerial computing

I. INTRODUCTION

Due to the rapid development of network and commu-

nication technologies in recent years, drones have shown

the great commercial value and are widely used in trafﬁc

monitoring, site management, aerial tourism photography and

commercial performances [1]. Especially with the maturity and

perfection of 5G technology and the continuous development

of 6G technology, the computing speed as well as data

The Heilongjiang Provincial Natural Science Foundation of China (Grant

No. LH2020F044), the 2019-“Chunhui Plan” Cooperative Scientiﬁc Research

Project of the Ministry of Education of China (Grant No. HLJ2019015), and

the Fundamental Research Funds for Heilongjiang University of China (Grant

No. 2020-KYYWF-1014) have all contributed to the funding of this work.

*Yi Wu is the corresponding author.

transmission rate has been greatly improved [2]. Unmanned

Aerial Vehicle (UAV) is gradually being used as an aerial

service device with its excellent mobility, high computing

power, and high perception capability [3]. Drones are uniquely

suited for remote and dangerous areas (e.g., wildlife photo

collection, atmospheric cloud weather prediction) as well as

for scenarios that require extreme real-time performance (e.g.,

trafﬁc management, mountain power maintenance, disaster

prediction). In addition, the UAV cluster, as an aerial edge

device, can collect data through its sensors and then transmit

the data to the central server for processing and modeling [4].

However, this process can cause privacy leakage of users and

increase communication costs [5].

The two main reasons for the above problems are as follows.

First, when users upload their private data to a third-party

central server if the third-party central server is attacked by

hackers or the third-party central server is untrustworthy, it will

cause the risk of user private data leakage. Second, traditional

centralized machine learning requires that a large amount of

data be uploaded to a third-party server for computing. Due

to the limited communication resources of UAVs, they cannot

afford high communication costs [6].

In response to these challenges, Federated Learning (FL)

[7] is gradually becoming the main focus of popular research.

By merely uploading model parameters, federated learning

addresses the issue of communication costs and lowers the

danger of data leaking. Nevertheless, when multiple UAVs

work jointly, the difference in the monitoring area of each

UAV causes heterogeneity in the collected sensing data. This

eventually causes the local device’s ability to recognize data

and the precision of the global model to decline [8] [9]. To

this end, we propose a brand-new calculation method for an

aggregate based on the distance between training models to

palliate the non-IID problem caused by local data. The system

overview of this paper is shown in Fig. 1. Consequently, the

following are the contributions to this paper:

arXiv:2210.04699v2 [cs.LG] 27 Dec 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

FedBA:Non-IIDFederatedLearningFrameworkinUAVNetworksPeiLiSchoolofDataScienceandTechnologyHeilongjiangUniversityHarbin,China2212623@s.hlju.edu.cnZhijunLiuSchoolofDataScienceandTechnologyHeilongjiangUniversityHarbin,China2222708@s.hlju.edu.cnLuyiChangSchoolofDataScienceandTechnologyHeilongjiangUnivers...

展开>> 收起<<

FedBA Non-IID Federated Learning Framework in UA V Networks Pei Li.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

FedBA Non-IID Federated Learning Framework in UA V Networks Pei Li

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: