Physical adversarial attacks against machine learning models for satellite imagery

Abstract

Machine learning (ML) models such as deep neural networks (DNNs) are increasingly employed to automatically process satellite imagery in various Earth observation (EO) applications. However, ML models are susceptible to adversarial attacks. The proposed research investigates physical adversarial attacks against ML models for satellite imagery, whereby a physical object with an optimized passive reflectance pattern is placed in the scene to manipulate the ML inference outcomes. Physical adversarial attacks in aerial imagery were first demonstrated by PI Chin’s team. There is a need to understand the types of adversarial vulnerability in ML models for aerial-satellite imagery, to inform the development of adversarial resilient ML models.

Document Details

Document Type
DoD Grant Award
Publication Date
Feb 16, 2024
Source ID
FA23862314082

Entities

People

  • Tat-Jun Chin

Organizations

  • Air Force Office of Scientific Research
  • United States Air Force
  • University of Adelaide

Tags

Fields of Study

  • Computer science

Readers

  • Computer Vision.
  • Irregular Warfare and Special Operations Cyberspace Operations against Adversarial Threats.
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks
  • Space