Adaptive Model Aggregation for Decentralized Federated Learning in Vehicular Networks
Date
2023-12-20
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In vehicular networks, Decentralized Federated Learning (DFL) has emerged as a cutting-edge methodology for the collaborative development of machine learning models, offering a unique balance between collaborative learning and individual data privacy. This approach is particularly apt in scenarios where vehicles, equipped with their own data and computation capabilities, work together to build a comprehensive and shared global model. Each vehicle contributes to this goal by training local models and merging them through model aggregation with data from nearby vehicles. A notable challenge in this context arises from vehicular networks' characteristics - primarily, their high mobility and the often unreliable nature of wireless communications. These factors frequently result in vehicles receiving only fragments of models from their peers, presenting a complex challenge: how to effectively leverage these incomplete models to accelerate the overall training process while simultaneously managing to keep the computational and communication overhead to a minimum. This thesis introduces the Adaptive Model Aggregation (AMA) algorithm to meet this challenge. AMA functions asynchronously within each vehicle and utilizes a threshold-based mechanism to decide whether to include the partially received models in the aggregation process. Given the variable nature of vehicular networks, it becomes evident that a fixed threshold is insufficient. Consequently, this thesis innovates an algorithm inspired by Contextual Multi-Arm Bandit (CMAB) theory, adapting the threshold for each vehicle based on the current communication dynamics of the network. The efficacy of AMA is tested through a series of comprehensive evaluations conducted within simulated environments, which account for various vehicular mobility patterns. These evaluations provide robust evidence that AMA substantially reduces the overhead associated with DFL aggregation - by as much as 83% - compared to traditional, non-adaptive aggregation methods, without adversely affecting the accuracy of the training process.
Description
Keywords
Federated Learning, Vehicular Networks, Adaptive Learning
Citation
Movahedian Attar, M. (2023). Adaptive Model Aggregation for Decentralized Federated Learning in vehicular networks (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.