Privacy-preserving Machine Learning Models for Traffic Forecasting

Traffic congestion in the transportation system has a substantial negative impact on productivity, living standards, the economy, and the environment. Mitigating these effects requires accurate traffic forecasting, a task crucial for effective transportation management. However, conventional approaches often entail privacy concerns and suboptimal forecasting accuracy. Leveraging drivers' data in its raw form introduces vulnerabilities in privacy and may not yield optimal results due to potential inaccuracies or inconsistencies. To address these challenges, this research investigates the domain of privacy-preserving traffic forecasting, aiming to revolutionize the way we approach this critical aspect of transportation systems. The project seeks to develop privacy-preserving techniques that enable the prediction of future traffic conditions without compromising the location privacy of individual drivers. The endeavor involves the utilization of advanced mathematical operations on encrypted data, ensuring that sensitive location information remains concealed from external entities. Traditional approaches, such as homomorphic encryption, while effective in some contexts, prove less efficient for this application due to computational constraints and inflexibility in accommodating diverse deep learning models. Consequently, this project adopts a cutting-edge approach centered on quadratic functional encryption. In practical terms, this entails the construction of a semi- private neural network. This network is designed with two distinct components: a private section that handles encrypted user location reports, aggregates them, and computes outputs for subsequent plaintext layers. These layers play a crucial role in augmenting the overall forecasting performance, with a particular focus on reducing storage, computation, and bandwidth overhead. One pivotal aspect of this research is the fine-tuning of the privacy-forecasting tradeoff. Recognizing that privacy preferences vary among drivers, the schemes empower individuals to customize their level of privacy protection. For instance, certain segments of a route may be disclosed with minimal privacy impact, resulting in enhanced forecasting accuracy. To achieve these objectives, the research team adopts functional encryption as a cryptographic primitive, affording us a robust foundation to pursue various critical aims. First and foremost, the team establishes a comprehensive route format that offers a spatiotemporal perspective on traffic congestion. Drivers are encouraged to furnish detailed routes for their journeys, with each segment bearing a unique encrypted identifier. Additionally, the team analyzes the influence of privacy levels on accuracy, recognizing that while privacy-preserving measures bolster trust in location sharing, they can potentially lead to a loss in forecasting precision. Through the implementation of advanced encryption schemes that support element-wise computations, the team aims to strike an optimal balance between privacy protection and accuracy. Furthermore, the project encompasses an adaptable framework to account for dynamic changes in road and traffic conditions. By integrating models developed in tandem with other project tasks, the team anticipates a robust solution capable of swiftly adapting to unforeseen circumstances such as accidents, social events, and weather variations. Finally, recognizing that privacy-preserving schemes entail heightened computation and communication overhead for users, the team investigates the tradeoff between enhanced accuracy and resource allocation. Strategic selection of the d cells in a driver's route aims to optimize both privacy and performance, ensuring an effective and sustainable traffic forecasting solution.

• Record URL:
  https://utc.engineering.osu.edu/privacy-preserving-machine-learning-models-traffic-forecasting

Language

• English

Project

• Status: Active
• Funding: $225000
• Contract Numbers:

  69A3552348327

• Sponsor Organizations:

  Office of the Assistant Secretary for Research and Technology

  University Transportation Centers Program
  Department of Transportation
  Washington, DC  United States  20590
• Managing Organizations:

  Center for Automated Vehicle Research with Multimodal Assured Navigation

  Ohio State University
  Columbus, OH  United States  43210
• Project Managers:

  Kline, Robin

• Performing Organizations:

  North Carolina A&T State University

  1601 E. Market Street
  Greensboro, NC  United States  27411
• Principal Investigators:

  Mahmoud, Mahmoud

• Start Date: 20231030
• Expected Completion Date: 20240830
• Actual Completion Date: 0
• USDOT Program: University Transportation Centers Program

Subject/Index Terms

• TRT Terms: Data privacy; Drivers; Information management; Machine learning; Neural networks; Traffic forecasting
• Subject Areas: Data and Information Technology; Highways; Planning and Forecasting; Vehicles and Equipment;

Filing Info

• Accession Number: 01901374
• Record Type: Research project
• Source Agency: Center for Automated Vehicle Research with Multimodal Assured Navigation
• Contract Numbers: 69A3552348327
• Files: UTC, RIP
• Created Date: Dec 4 2023 5:03PM