On Addressing Heterogeneity in Federated Learning for Autonomous Vehicles Connected to a Drone Orchestrator

TL;DR

This paper proposes a dynamic wireless resource allocation method in federated learning for autonomous vehicles, improving the learning speed and accuracy of critical object detection amid data heterogeneity and non-IID distributions.

Contribution

It introduces a reactive, dynamic resource allocation approach based on learner contribution, enhancing federated learning performance in autonomous vehicle scenarios.

Findings

Proactive resource allocation improves model accuracy and learning speed.

FedProx with adjusted local optimizer enhances performance with deeper neural networks.

Dynamic allocation outperforms static methods in non-IID data environments.

Abstract

In this paper we envision a federated learning (FL) scenario in service of amending the performance of autonomous road vehicles, through a drone traffic monitor (DTM), that also acts as an orchestrator. Expecting non-IID data distribution, we focus on the issue of accelerating the learning of a particular class of critical object (CO), that may harm the nominal operation of an autonomous vehicle. This can be done through proper allocation of the wireless resources for addressing learner and data heterogeneity. Thus, we propose a reactive method for the allocation of wireless resources, that happens dynamically each FL round, and is based on each learner's contribution to the general model. In addition to this, we explore the use of static methods that remain constant across all rounds. Since we expect partial work from each learner, we use the FedProx FL algorithm, in the task of computer vision. For testing, we construct a non-IID data distribution of the MNIST and FMNIST datasets among four types of learners, in scenarios that represent the quickly changing environment. The results show that proactive measures are effective and versatile at improving system accuracy, and quickly learning the CO class when underrepresented in the network. Furthermore, the experiments show a tradeoff between FedProx intensity and resource allocation efforts. Nonetheless, a well adjusted FedProx local optimizer allows for an even better overall accuracy, particularly when using deeper neural network (NN) implementations.