FedGraph an Aggregation Method from Graph Perspective Zhifang Deng1 Xiaohong Huang1 Dandan Li1 and Xueguang Yuan2

2025-04-27 0 0 690.39KB 13 页 10玖币

侵权投诉

FedGraph: an Aggregation Method from Graph

Perspective ?

Zhifang Deng1, Xiaohong Huang1, Dandan Li1, and Xueguang Yuan2

1School of Computer Science (National Pilot Software Engineering School), Beijing

University of Posts and Telecommunications, Beijing, China

2School of Electronic Engineering, Beijing University of Posts and

Telecommunications, Beijing, China

Abstract. With the increasingly strengthened data privacy act and the

diﬃcult data centralization, Federated Learning (FL) has become an ef-

fective solution to collaboratively train the model while preserving each

client’s privacy. FedAvg is a standard aggregation algorithm that makes

the proportion of dataset size of each client as aggregation weight. How-

ever, it can’t deal with non-independent and identically distributed (non-

i.i.d) data well because of its ﬁxed aggregation weights and the neglect

of data distribution. In this paper, we propose an aggregation strategy

that can eﬀectively deal with non-i.i.d dataset, namely FedGraph, which

can adjust the aggregation weights adaptively according to the train-

ing condition of local models in whole training process. The FedGraph

takes three factors into account from coarse to ﬁne: the proportion of

each local dataset size, the topology factor of model graphs, and the

model weights. We calculate the gravitational force between local mod-

els by transforming the local models into topology graphs. The FedGraph

can explore the internal correlation between local models better through

the weighted combination of the proportion each local dataset, topology

structure, and model weights. The proposed FedGraph has been applied

to the MICCAI Federated Tumor Segmentation Challenge 2021 (FeTS)

datasets, and the validation results show that our method surpasses the

previous state-of-the-art by 2.76 mean Dice Similarity Score. The source

code will be available at Github.

Keywords: Federated Learning ·Brain Tumor Segmentation ·Fed-

Graph ·Coarse-to-ﬁne ·Topology.

1 Introduction

With the increasing demand for the precise medical data analysis, deep learning

methods are widely used in the medical image ﬁeld. However, with the promul-

gation of the data act and the strengthening of data privacy, especially in the

medical ﬁeld, it has become more diﬃcult to train models in large-scale central-

ized medical datasets. As one of the solutions, federated learning has attracted

a lot of attention from researchers.

?Supported by organization x.

arXiv:2210.02733v1 [cs.CV] 6 Oct 2022

2 Z. Deng et al.

Federated learning(FL)[1][2] is a distributed machine learning paradigm in

which all clients train a global model collaboratively while preserving their data

locally. The naive repeat steps of FL are: (i) each client trains its model with

local data; (ii) the server collects and aggregates the models from clients to get

a global model, then delivers the global model to clients. The data ﬂow between

clients and server is the trained models rather than the original data, which

avoids the leak of data privacy. As a crucial core of them, aggregation algorithm

plays an important role in releasing data potential and improving global model

performance. FedAvg[1], as pioneering work, is a simple and eﬀective aggregation

algorithm, which makes the proportions of local datasets size as the aggregation

weights of local models. [3] proposed FedProx to limit the updates between lo-

cal and global models by modifying the training loss of local models. FedMA[4]

matches and averages the hidden elements with similar feature extraction sig-

natures to construct the shared global model in a layer-wise manner. Federated

learning has attracted the attention of scholars in more research ﬁelds.

In medical image segmentation, Since [5] and [6] explored the feasibility of

FL in brain tumor segmentation(BraTS), FL on medical image segmentation is

in full swing. Liu et al. [7] proposed FedDG to make the model generalize to

unseen target domains via episodic learning in continuous frequency space in

retinal fundus image segmentation. Xia et al. [8] proposed Auto-FedAvg, where

the aggregation weights are dynamically adjusted according to the data distribu-

tion, to accelerate the training process and get better performance in COVID-19

lesion segmentation. Zhang et al. [9] proposed SplitAVG to overcome the perfor-

mance drops from data heterogeneity in FL by network split and feature map

concatenation strategies in the BraTS task. More than this, the ﬁrst computa-

tional competition on federated learning, Federated Tumor Segmentation(FeTS)

Challenge1[10] is held to measure the performance of diﬀerent aggregation algo-

rithms on glioma segmentation[11,12,13,14]. Leon et al. [15] proposed FedCost-

WAvg to get a notable improvement compared to FedAvg by including the cost

function decreased during the last round and won the challenge. However, most

of these methods only study the single granularity or add other regular terms to

the aggregation method, without considering the ﬁner granularity factors, which

limit the performance of global model.

Diﬀerent from the above methods, in this paper, we propose a novel aggrega-

tion strategy, FedGraph, which attempts to explore the aggregation algorithm

of FL from the topological perspective of neural networks. After the server col-

lects the local models, the FedGraph explores the internal correlations between

local models by three aspects from coarse to ﬁne: the proportion of each local

dataset size, the topology structure of model graphs, and the model weights.

The proportion of local dataset size factor is similar to FedAvg. We compute

the topological correlation by mapping the local models into topological graphs.

Meanwhile, the ﬁner grain model weights correlations are taken into account.

Through the weighted combination of three diﬀerent granularity factors from

1https://fets-ai.github.io/Challenge/

FedGraph: an Aggregation Method from Graph Perspective 3

client 1 client 2 client K

∙∙∙

server

global model

FedTopAvg

Sample size Topology Model weights

Coarse to fine







G1G2G3

_(,)

kki

top sim top sim G G







top k

top sim







ttt

kkg

www



 



wk t













Weighted sum and aggregation









∙∙∙

Fig. 1. Overview of proposed FedGraph. The FedGraph aggregates local models and

explore the correlations by considering coarse-to-ﬁne three factors: the sample size of

local datasets, network topology and model weights.

coarse to ﬁne, the proposed method promotes the more eﬀective aggregation of

local models.

The primary contributions of this paper can be summarized as: (1) We pro-

pose FedGraph, a novel aggregation strategy which takes coarse-to-ﬁne three

factors: the sample size, the topology of model graphs and the model weights,

especially from the topological perspective of neural network; (2) We propose an

aggregation method which introduce the concept of graph into federated learn-

ing, and the aggregation weights can be adjusted adaptively; (3) The superior

performance is achieved by the proposed method, on the public FeTS challenge

datasets.

2 Related Work

Federated Learning The FedAvg[1] plays a cornerstone role in federated learn-

ing tasks because of its eﬃcient algorithm design, and the proportion of each

client’s dataset is taken as weight during the aggregation process. Recently, Some

methods[19-24] based on FedAvg are proposed because of the poor performance

caused by heterogeneous data distribution across clients in the real world. For

example, FedProx[3] limits the bias between global model and local model by im-

posing additional regularization terms. FedDyn introduces a regularization term

for local training based on the global model to address the problem. FedNova

improves the aggregation stage by normalizing and scaling the local updates

of each client according to their local steps before updating the global model.

SCAFFOLD addresses non-i.i.d problem by introducing variance among clients

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：

FedGraph:anAggregationMethodfromGraphPerspective?ZhifangDeng1,XiaohongHuang1,DandanLi1,andXueguangYuan21SchoolofComputerScience(NationalPilotSoftwareEngineeringSchool),BeijingUniversityofPostsandTelecommunications,Beijing,China2SchoolofElectronicEngineering,BeijingUniversityofPostsandTelecommunicati...

展开>> 收起<<

FedGraph an Aggregation Method from Graph Perspective Zhifang Deng1 Xiaohong Huang1 Dandan Li1 and Xueguang Yuan2.pdf

共13页,预览3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！

FedGraph an Aggregation Method from Graph Perspective Zhifang Deng1 Xiaohong Huang1 Dandan Li1 and Xueguang Yuan2

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: