Robust aggregation function in federated learning

Rahim Taheri*, Farzad Arabikhan, Alexander Gegov, Negar Akbari

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Maintaining user data privacy is a crucial challenge for machine learning techniques. Federated learning is a solution that enables machine learning models to be trained using data residing on different devices without centralizing the data. This training method offers several advantages: Firstly, federated learning helps preserve user privacy by storing data on separate devices rather than transferring it to a central location for training. Secondly, training machine learning models on a diverse set of devices through federated learning improves their robustness, ensuring optimal performance in a wide range of real-world scenarios. Finally, federated learning can promote scalability by enabling simultaneous training on a vast number of devices. So, it can increase the scale of training and enable the development of more sophisticated models. The aggregation function is one of the main steps in federated learning and is used on the server side to aggregate local models sent from the client side. The most widely used aggregation function is Krum, which, despite the research done to improve its robustness, is still vulnerable to adversarial samples. In this paper, a method is proposed to improve the robustness of the Krum aggregation function. The results confirm that the proposed method is more robust against adversarial samples than the original version of the Krum aggregation function.
Original languageEnglish
Title of host publicationAdvances in Information Systems, Artificial Intelligence and Knowledge Management
Subtitle of host publication6th International Conference on Information and Knowledge Systems, ICIKS 2023, Portsmouth, UK, June 22–23, 2023, Proceedings
EditorsInès Saad, Camille Rosenthal-Sabroux, Faiez Gargouri, Salem Chakhar, Nigel Williams, Ella Haig
PublisherSpringer Nature
Pages168-175
Number of pages8
ISBN (Electronic)9783031516641
ISBN (Print)9783031516634
DOIs
Publication statusPublished - 20 Jan 2024
EventThe 6th International Conference on Information and Knowledge Systems (ICIKS 2023), June 22-23, 2023 - Portsmouth, Portsmouth, United Kingdom
Duration: 22 Jun 202323 Jun 2023

Publication series

NameLecture Notes in Business Information Processing
PublisherSpringer Nature
Volume486
ISSN (Print)1865-1348
ISSN (Electronic)1865-1356

Conference

ConferenceThe 6th International Conference on Information and Knowledge Systems (ICIKS 2023), June 22-23, 2023
Abbreviated titleICIKS 2023
Country/TerritoryUnited Kingdom
CityPortsmouth
Period22/06/2323/06/23

Keywords

  • Federated Learning
  • Robustness
  • Krum Aggregation

Cite this