Towards production-ready end-to-end federated learning for automotive applications

Abstract

The growing trend toward preserving user data privacy has embraced a new computing paradigm called Federated Learning (FL). FL enables edge devices to learn from a global shared model without cloud data storage. FL implements an efficient model training method for disseminating model training while maintaining user data privacy. Applying FL to the automobile industry solves numerous problems, including minimizing privacy concerns, establishing clients’ trust in data sharing, and constructing robust ML models. The time it takes to collect data is also mitigated because we do not have to wait for clients to share their data. Instead, the training at the edge node is enough. Meanwhile, the automotive industry is experiencing a dramatic increase in the popularity of Electric Vehicle (EV). In providing greener technologies, EVs offer substantial advantages. The batteries that power the EVs have a maximum energy consumption restriction. Consequently, it is vital to evaluate the battery consumption of EVs to resolve numerous issues, such as range anxiety and Charging Station (CS) usage, to name a few. Therefore, it is crucial to predict battery consumption under various traffic scenarios. Another critical issue to be tackled is the electric load forecasting of CSs in an area. This is especially important due to the frequency with which EVs utilize CSs. To end this, we implement an amalgamation of the FL in the automotive use cases. We implement an FL environment to embark on the battery capacity estimation of EVs and the energy demand of CSs in a specific region. We used the Simulation of Urban MObility (SUMO) simulator’s various toolkits to model the traffic, network simulation, and data acquisition. We use Flower, an open-source FL framework, to set the ground for our method. We implement and adapt the software stack in Flower to accommodate the two user-defined strategies, active client contribution (client picking strategy) and handling faulty clients (system fault tolerance). We eventually integrate the established FL into an experiment tracking tool called Weights and Biases (W&B). The implemented FL system runs a large client pool of about 1000 EVs, enabling a large-scale production-grade and scalable simulation in FL. The simulation environment is configured as resource-aware using the Virtual Client Engine (VCE) of Flower and Ray. Our experiments show significant convergence results preserving data privacy. The findings show that the FL model outperforms the local baseline model in many client cases. We also demonstrate the implementation of a fault-tolerant system using an extension to the Federated Averaging (FedAvg). Finally, we present the significance of adopting a differentially private model and the trade-off between end-to-end privacy protection and obtained accuracy.