Model Aggregation Federated Learning Approach for Vehicular Traffic Forecasting

Abstract

To make the Intelligent Transportation System (ITS) more efficient and robust, researchers are working hard. Analysis of traffic data helps ITS to be more helpful. Mobile phones are the prime source of traffic data. The vast availability of data and increased processing speed of mobile phones is making ITS more robust. Presently for traffic prediction, the entire mobile user’s data is accumulated at the central server. The information is then aggregated together to make predictions. In this approach, sensitive user data have the risk of privacy and security—massive user data uploading on the server results in latency.This paper proposes a decentralized approach for vehicular traffic prediction that allows ‘selected’ local mobiles/ organizations (clients) to train the model and share the trained model securely to the server. The selection of organizations to participate in the training process is made by clustering algorithms. The server then aggregates the locally trained model and shares the aggregated model to all the clients again