GPT-3 and Me: How Supercomputer-scale Neural Network Models Apply to Defensive Cybersecurity Problems presented at Black Hat USA 2022

by Joshua Lee,

Tags: Network Security AI ML & Data Science


Summary : A key lesson of recent deep learning successes is that as we scale neural networks, they get better, and sometimes in game-changing ways.In this talk, we'll demonstrate and explain how supercomputer-scale neural networks open new vistas for security, qualitatively changing the horizons for machine learning security applications in surprising and powerful ways. Specifically, we'll demonstrate two applications of large neural networks to security problems that wouldn't have been tractable with smaller models: generating custom, human-readable explanations of difficult-to-parse attacker behavior, and detecting malicious behaviors even when we have very few examples of the kind of behaviors we're looking for.We'll describe each example application in transparent and reproducible detail, and then show you how you can use our work, or do your own large neural network experimentation, using publicly available models like OpenAI's GPT-3 series of models.