Robustness and Vulnerability Measures of Deep Learning Methods for Cyber Defense

Abstract

Navy networks and infrastructures are under frequent cyberattack. One developing area of application of Artificial Intelligence (AI) and Machine Learning (ML) is cybersecurity. However, some weakness of machine learning, such as the lack of interpretability and the susceptibility to adversarial data, are important issues that must be studied for reliable and safe applications of AI tools. The robustness of deep learning (DL) techniques used in computer vision and language processing have been extensively studied. However, less is currently known about the vulnerabilities and robustness of DL methods suitable in cybersecurity applications. The goal of this research is to investigate mathematical concepts and quantitative measures of robustness and vulnerability to adversarial data for cybersecurity DL and to create computational algorithms capable of quantitatively evaluating the robustness and vulnerability of DL tools. The tasks of the project include literature review, an innovative study of mathematical concepts, the development of computational algorithms, the validation of the concepts and algorithms through examples. The deliverables of the project include technical reports, student thesis, and technical papers for publication. This work will enhance understanding of vulnerabilities of deep learning systems that could be incorporated in future DoN networks, and provide the US Navy with computational tools capable of measuring the robustness of the AI enabled systems.