PODE: privacy-enhanced distributed federated learning approach for origin-destination estimation

TL;DR

This paper introduces PODE, a privacy-focused method using federated learning to estimate travel destinations while keeping sensitive data local.

Contribution

A novel federated learning approach for origin-destination estimation that preserves privacy by training models locally.

Findings

PODE achieves 93.20% accuracy on the server side for predicting truck destination states.

The method reduces computational burden on the server by using a two-client architecture.

Label reduction from 51 to 11 improves learning efficiency in the model.

Abstract

The statewide consumer transportation demand model analyzes consumers’ transportation needs and preferences within a particular state. It involves collecting and analyzing data on travel behavior, such as trip purpose, mode choice, and travel patterns, and using this information to create models that predict future travel demand. Naturalistic research, crash databases, and driving simulations have all contributed to our knowledge of how modifications to vehicle design affect road safety. This study proposes an approach named PODE that utilizes federated learning (FL) to train the deep neural network to predict the truck destination state, and in the context of origin-destination (OD) estimation, sensitive individual location information is preserved as the model is trained locally on each device. FL allows the training of our DL model across decentralized devices or servers without exchanging raw data. The primary components of this study are a customized deep neural network based on federated learning, with two clients and a server, and the key preprocessing procedures. We reduce the number of target labels from 51 to 11 for efficient learning. The proposed methodology employs two clients and one-server architecture, where the two clients train their local models using their respective data and send the model updates to the server. The server aggregates the updates and returns the global model to the clients. This architecture helps reduce the server’s computational burden and allows for distributed training. Results reveal that the PODE achieves an accuracy of 93.20% on the server side.

Figures