Robustness and Vulnerability Measurement of Deep Learning Methods for Cyber Defense
Abstract
The goal of this study is to investigate mathematical concepts and quantitative measures of robustness and vulnerability of machine learning systems to adversarial data and develop computational methods capable of quantitatively evaluating the robustness and vulnerability of deep learning tools that can be applied in cybersecurity settings. The first phase of the project is a literature review. The second phase of the study is focused on robustness analysis of infrastructure cyber security. Using a microgrid power system model and learning-based fault detection as the testbed, we investigate the robustness of neural networks subjected to noisy or poisoned data. Finally, the third phase of the project, explores distributional robustness. Neural networks may sometimes be used outside of the environment in which they were trained. If the distribution of the incoming data is significantly different from that of the training data, it could negatively impact the performance of the neural network. In addition to a quantitative analysis of robustness, the study reveals an underlying relationship between the robustness and the dynamical behavior the training data.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 2022
- Accession Number
- AD1194685
Entities
People
- Elana Kozak
- Phillip Smith
- Thor Martinsen
- Wei Kang
Organizations
- Naval Postgraduate School