Defending Object Detectors in Autonomous Vehicles Against Adversarial Attacks with Diffusion Models

Object detection stands as a cornerstone task in computer vision, serving as the foundation for autonomous vehicles. Although machine learning-based object detectors achieve remarkable accuracy and efficiency, they are vulnerable to adversarial attacks, which exploit the inherent weaknesses of machine learning models to mislead them into producing incorrect outputs. In particular, physical adversarial patch attacks (e.g., stickers to be placed on real-world objects) have attracted significant attention from the security community as their real-world implications are severe for the safety and functionality of object detection systems. In this research project, the research team aims to utilize the latest advancements in generative models, particularly diffusion models, to preprocess input images before feeding them into object detection systems. The goal is to develop a defense mechanism that can address different physical adversarial patch attacks, regardless of their shape or format. Therefore, the proposed method is both patch-agnostic and attack-agnostic. Leveraging the generative power of diffusion models, the system will automatically detect and replace adversarial patches with contextually consistent content drawn from surrounding areas.

• Record URL:
  • https://www.clemson.edu/cecas/tracr/research/projects/#defendingobjectdetectorsinautonomousvehiclesagainstadversarialattackswithdiffusionmodels

Language

• English

Project

• Status: Active
• Funding: $209,512.00
• Contract Numbers:

  69A3552344812

  69A3552348317

• Sponsor Organizations:

  Office of the Assistant Secretary for Research and Technology

  University Transportation Centers Program
  Department of Transportation
  Washington, DC  United States  20590

  Clemson University

  216 Lowry Hall
  Clemson, SC, SC  United States  29634

  Benedict College

  1600 Harden Street
  Columbia, South Carolina  United States  29204
• Managing Organizations:

  National Center for Transportation Cybersecurity and Resiliency (TraCR)

  Clemson University
  Clemson, SC  United States 

  Clemson University

  216 Lowry Hall
  Clemson, SC, SC  United States  29634
• Project Managers:

  Chowdhury, Mashrur

• Performing Organizations:

  Clemson University

  216 Lowry Hall
  Clemson, SC, SC  United States  29634

  Benedict College

  1600 Harden Street
  Columbia, South Carolina  United States  29204
• Principal Investigators:

  Cheng, Long

  Luo, Feng

  Iyangar, Balaji

• Start Date: 20250101
• Expected Completion Date: 20251231
• Actual Completion Date: 0
• USDOT Program: University Transportation Centers

Subject/Index Terms

• TRT Terms: Autonomous vehicles; Computer security; Connected vehicles; Image processing; Machine learning; Object detection; Simulation
• Subject Areas: Data and Information Technology; Highways; Security and Emergencies; Vehicles and Equipment;

Filing Info

• Accession Number: 01950244
• Record Type: Research project
• Source Agency: National Center for Transportation Cybersecurity and Resiliency (TraCR)
• Contract Numbers: 69A3552344812, 69A3552348317
• Files: UTC, RIP
• Created Date: Mar 28 2025 2:10PM