Data-Driven Federated Autonomous Driving

Intelligent vehicles optimize road traveling through their reliance on autonomous driving applications to navigate. These applications integrate machine learning to extract statistical patterns and sets of rules for the vehicles to follow when facing decision-making scenarios. The immaturity of such systems, caused by the lack of a diverse dataset, can lead to inaccurate on-road decisions that could affect road safety. In this paper, we devise a decentralized scheme based on federating autonomous driving companies in order to expand their access to data and resources during the learning phase. Our scheme federates companies in an optimal manner by studying the compatibility of the federations' dataset in the federations formation process, without exposing private data to rivalries. We implement our scheme for evaluation against other formation mechanisms. Experiments show that our approach can achieve higher model accuracy, reduce model loss, and increase the utility of the individuals on average when compared to other techniques.