Blockchain-Based Decentralized Federated Learning for Privacy-Preserving Traffic Flow Prediction: A Case Study with PeMS-8 Data

Abstract

A timely traffic flow information is essential for traffic management, traffic prediction has become a key component of intelligent transportation systems. However, current centralized machine learning-based traffic flow prediction algorithms need the collection of raw data for the train model, which poses significant privacy breach hazards. Federated learning, a recent innovation that effectively protects privacy by sharing model changes without transferring raw data, has been developed to solve these issues. The current federated learning frameworks, however, are built on a centralized model coordinator that continues to experience serious security issues, such as a single point of failure. In this paper, we proposed BDFL a block-chained decentralized federated-learning (DFL) architecture for traffic flow prediction using PeMS-8 data. The suggested technique provides decentralized model training on PeMS-8 while ensuring the privacy of the underlying data. The long short-term memory (LSTM) model, which is often employed for signal and time-series data, was used in this study. A total of 3035520 traffic locations were covered by the PeMS-8 dataset, which was obtained from the Caltrans Performance Measurement System (PeMS). These locations' data included timestep and junction details as well as traffic flow, occupancy, and speed. The total accuracy of our long-short-term memory (LSTM) model is 99.0, with a loss of 3.296.