Federated split GANs for collaborative training with heterogeneous devices

Applications based on machine learning (ML) are greatly facilitated by mobile devices and their enormous volume and variety of data. To better safeguard the privacy of user data, traditional ML techniques have transitioned toward new paradigms like federated learning (FL) and split learning (SL). Ho...

Full description

Saved in:
Bibliographic Details
Published inSoftware impacts Vol. 14; p. 100436
Main Authors Liang, Yilei, Kortoçi, Pranvera, Zhou, Pengyuan, Lee, Lik-Hang, Mehrabi, Abbas, Hui, Pan, Tarkoma, Sasu, Crowcroft, Jon
Format Journal Article
LanguageEnglish
Published Elsevier B.V 01.12.2022
Subjects
Federated learning
GAN
Hardware heterogeneous
Privacy preservation
Split learning
Hardware heterogeneous
Privacy preservation
Federated learning
GAN
Split learning
Online AccessGet full text

Cover

More Information
Summary:Applications based on machine learning (ML) are greatly facilitated by mobile devices and their enormous volume and variety of data. To better safeguard the privacy of user data, traditional ML techniques have transitioned toward new paradigms like federated learning (FL) and split learning (SL). However, existing frameworks have overlooked device heterogeneity, greatly hindering their applicability in practice. In order to address such limitations, we developed a framework based on both FL and SL to share the training load of the discriminative part of a GAN to different client devices. We make our framework available as open-source software11https://github.com/YukariSonz/FSL-GAN.. •Novel federated split GAN frameworks.•Flexible software allowing any type of GAN networks to be plugged in.•Flexible software allowing any type of client selection algorithms to be plugged in.•Feasibility for deploying complex GAN networks on resource-constrained mobile devices.
ISSN:2665-9638
2665-9638
DOI:10.1016/j.simpa.2022.100436