PENS: Leveraging Data Heterogeneity in Federated Learning

Examensarbete för masterexamen

Please use this identifier to cite or link to this item: https://hdl.handle.net/20.500.12380/302702
Download file(s):
File Description SizeFormat 
Master_thesis_Noa_Onozsko_och_Gustav Karlsson.pdf3.05 MBAdobe PDFThumbnail
View/Open
Bibliographical item details
FieldValue
Type: Examensarbete för masterexamen
Title: PENS: Leveraging Data Heterogeneity in Federated Learning
Authors: Onoszko, Noa
Karlsson, Gustav
Abstract: Federated learning (FL) is a decentralized machine learning technique where training is done cooperatively by exchanging model weights or gradients instead of sharing the raw data between the cooperating devices (clients). Classical FL algorithms such as federated averaging work best in the special case when the data is IID over clients. In this work, we address the problem of data heterogeneity in federated learning. We propose a decentralized federated learning (DFL) algorithm termed Performancebased Neighbour Selection Federated Learning Algorithm (PENS), that effectively leverages the data heterogeneity over clients. PENS is a cooperative communicationbased algorithm where clients communicate with other clients that have a similar data distribution. Specifically, model performance is used as a proxy for data similarity as no raw data is allowed to be shared among clients. Experiments on the CIFAR-10 dataset show that this communication scheme results in higher model accuracies than if clients communicate randomly with each other. The method is robust for different numbers of participating clients as long as the local datasets are sufficiently large.
Keywords: decentralized federated learning, federated learning, data heterogeneity, personalization, distributed machine learning, gossip learning, privacy, image classification
Issue Date: 2021
Publisher: Chalmers tekniska högskola / Institutionen för matematiska vetenskaper
URI: https://hdl.handle.net/20.500.12380/302702
Collection:Examensarbeten för masterexamen // Master Theses



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.