Distributed Optimal-Transport Clustering for Malicious User Rejection in Federated-Learning VANETs.

Pacheco, Lucas; Braun, Torsten (6 October 2022). Distributed Optimal-Transport Clustering for Malicious User Rejection in Federated-Learning VANETs. In: 3rd KuVS Fachgespräch "Machine Learning & Networking". Hasso-Plattner-Institut

[img] Text (Abstract)
kuvs_fachgespraeche.pdf - Submitted Version
Restricted to registered users only
Available under License Publisher holds Copyright.

Download (140kB)

Federated Learning (FL) has rapidly become a crucial paradigm for training Machine Learning models when datasets are spread across several devices without compromising the privacy of the data owners. In vehicular networks, we can use FL to train driving models and object detection and classification over sensitive datasets to improve the experience of users and driving safety. However, the majority of FL implementations cannot efficiently filter malicious vehicular users and low-quality contributions. We propose DOTFL, an aggregation mechanism based on the clustering of the received trained Neural Networks (NN) at the vehicular devices and on outlier detection. The proposed mechanism can detect malicious contributions by comparing them to previously received contributions for clustering. Furthermore, the convergence time of the FL process is improved by distributing trained NN weights directly through vehicle-to-vehicle links. Experimental analysis shows an improvement of up to 19% in terms of convergence time compared to state-of-the-art approaches. This is achieved by clustering models and removing outliers, enabling a significantly lower presence of malicious contributions in aggregated models.

Item Type:

Conference or Workshop Item (Abstract)

Division/Institute:

08 Faculty of Science > Institute of Computer Science (INF) > Communication and Distributed Systems (CDS)
08 Faculty of Science > Institute of Computer Science (INF)

UniBE Contributor:

De Sousa Pacheco, Lucas, Braun, Torsten

Subjects:

000 Computer science, knowledge & systems
500 Science > 510 Mathematics

Publisher:

Hasso-Plattner-Institut

Language:

English

Submitter:

Dimitrios Xenakis

Date Deposited:

14 Jul 2023 13:18

Last Modified:

28 Aug 2023 15:31

Related URLs:

BORIS DOI:

10.48350/184752

URI:

https://boris.unibe.ch/id/eprint/184752

Actions (login required)

Edit item Edit item
Provide Feedback