Next Generation Machine Learning Platform for Advanced Vulnerability Analysis and Reverse Engineering

Next Generation Machine Learning Platform for Advanced Vulnerability Analysis and Reverse Engineering Next Generation Machine Learning Platform for Advanced Vulnerability Analysis and Reverse Engineering Existing Department of Defense projects at Arizona State University, described in Section 3, pursue a wide range of areas relating to the automated and manual analysis of software for security vulnerabilities. In recent years, program analysis researchers have increasingly incorporated Machine Learning into their proposed approaches [2, 3, 5]. However, our attempts to use these techniques in real-world scenarios to reason about the security of real code and systems has resulted in lower than expected efficacy. Digging into the reason for this, we realized that, though modern approaches to complex, capable networks in fields such as NLP have tended toward the use of massive models requiring hundreds of gigabytes of simultaneously-addressable GPU memory [7], models currently used in cybersecurity applications instead reside in tens of gigabytes [5]. This reduction in model parameters is associated with a reduced applicability of emerging ML-driven program analysis techniques on real-world software. We propose to build an integrated GPU cluster, connecting 12 GPUs using Nvidias NVLink and NVSwitch [4] technology, to build a cluster with an aggregate GPU containing 960 gigabytes of GPU memory. This would allow us to apply cutting-edge techniques in core ML research, in combination with results in the intersection of cybersecurity and ML, supercharging our existing DoD projects with advanced AI capabilities as described in Section 3. Additionally, our existing DoD work has both implicit and explicit priorities for education and training, for which the proposed hardware would enable additional capabilities as well. These capabilities will allow us to deliver above and beyond planned milestones on our existing DoD project portfolio.