Exploring Vulnerability and Robustness of Deep Learning Systems

Abstract

Deep neural networks have enabled modern computer vision systems to reach new heights of performance on a variety of challenging tasks. Despite the accuracy and efficiency benefits, the highly parameterized non-linear nature of deep networks makes them very difficult to interpret and prone to failure in the presence of an adversary or anomalous data. This vulnerability makes the integration of these models into our real-world systems troubling. This project has two broad threads: (1) we explore the vulnerability of deep neural networks by developing state-of-the-art adversarial attacks, and (2) we improve the robustness of models in challenging operating environments such as in open-world target recognition and federating learning scenarios.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
May 01, 2022
Accession Number
AD1170105

Entities

People

  • Jingyang Zhang
  • Mathew Inkawhich
  • Randy Linderman
  • Yiran Chen

Organizations

  • Duke University

Tags

Communities of Interest

  • Autonomy
  • Energy and Power Technologies
  • Ground and Sea Platforms
  • Materials and Manufacturing Processes
  • Sensors
  • Space

DTIC Thesaurus Topics

  • Air Force
  • Artificial Intelligence
  • Artificial Intelligence Software
  • Birds
  • Computer Languages
  • Computer Vision
  • Computers
  • Deep Learning
  • Detection
  • Detectors
  • Dimensionality Reduction
  • Distance Learning
  • Information Processing
  • Information Science
  • Information Systems
  • Machine Learning
  • Neural Networks
  • Standards
  • Supervised Machine Learning
  • Synthetic Aperture Radar
  • Target Recognition

Fields of Study

  • Computer science

Readers

  • Cybersecurity.
  • Neural Network Machine Learning.
  • Systems Analysis and Design

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks