Exploring the Removal of Bit-Planes for Increased Adversarial Robustness (Preprint)
Abstract
Machine Learning has been found to be a very valuable and powerful tool, that will almost certainly see an increase in use in the future. However, it has also been found to have some vulnerabilities that could be exploited. This is concerning overall, but particularly important if the machine learning model is being used for safety-critical, but even for non-safety-critical applications. Many defenses have been worked on to combat this issue, but most defenses seem to be broken shortly after being proposed. These defenses are also typically very resource hungry as well as cause normal performance to drop. In this work, we present several ideas as to how to make robust models using fewer resources and without impacting clean performance. We explore several different combinations of defenses as well as metrics to determine the impact of the attacks on metrics besides just accuracy.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 10, 2023
- Accession Number
- AD1200986
Entities
People
- Alex Hildenbrandt
- Ashley Diehl
- Christopher Menart
- Hannah Richards
- Melissa Robertson
- Robert Canady
Organizations
- Air Force Research Laboratory
- Brigham Young University
- Vanderbilt University