A Reliable Vertical Federated Learning Framework for Traffic State Estimation with Data Selection and Incentive Mechanisms

TL;DR

This paper introduces a reliable vertical federated learning framework for traffic state estimation that incorporates data selection and incentive mechanisms to improve prediction accuracy and utility.

Contribution

It proposes a novel framework combining data provider selection and incentive design to enhance VFL-based traffic estimation reliability and performance.

Findings

Improved traffic flow prediction accuracy by 11.23%.

Enhanced density prediction accuracy by 23.15%.

Increased MA utility by 130 to 400 dollars.

Abstract

Vertical Federated Learning (VFL)-based Traffic State Estimation (TSE) offers a promising approach for integrating vertically distributed traffic data from municipal authorities (MA) and mobility providers (MP) while safeguarding privacy. However, given the variations in MPs' data collection capabilities and the potential for MPs to underperform in data provision, we propose a reliable VFL-based TSE framework that ensures model reliability during training and operation. The proposed framework comprises two components: data provider selection and incentive mechanism design. Data provider selection is conducted in three stages to identify the most qualified MPs for VFL model training with the MA. First, the MA partitions the transportation network into road segments. Then, a mutual information (MI) model is trained for each segment to capture the relationship between data and labels. Finally, using a sampling strategy and the MI model, the MA assesses each MP's competence in data provision and selects the most qualified MP for each segment. For the incentive mechanism design, given the MA can leverage the MI mode to inspect the data quality of MP, we formulate the interaction between MA and MP as a supervision game model. Upon this, we devise a penalty-based incentive mechanism to inhibit the lazy probability of MP, thereby guaranteeing the utility of MA. Numerical simulation on real-world datasets showcased that our proposed framework augments the traffic flow and density prediction accuracy by 11.23\% and 23.15\% and elevates the utility of MA by 130$\sim$400\$ compared to the benchmark.