Wednesday
Room 1
10:20 - 11:20
(UTC+01)
Talk (60 min)
Leveraging graph AI and GPUs to win the US Cyber Command AI challenge
While most tools show us events as tables, analyzing them as graphs and vector embeddings changes how we can approach investigation tooling and detection engineering. The core of our talk looks at the methods and tools behind our first-place finish in the US Cyber Command's Alert AI Challenge.
By combining some of our favorite graph, AI, and GPU techniques into a single pipeline, we were able reduce alert volumes in a fusion center covering ISPs and universities by 97% and create more useful situational awareness views for prioritization and interactive investigations. We will also touch on how we have been using LLMs & AI to further automate our graph workflows in more recent projects such as in zero-trust identity monitoring & investigations.
Key Points:
- Start thinking differently in how events can be used for investigations and detections
- Learn how modern graph AI techniques like hypergraphs, UMAP, graph neural networks, GPU ETL, and GPU feature engineering fit together
- Free & OSS: Get a feel for how tools like Jupyter Notebooks, Nvidia RAPIDS, PyGraphistry[AI], DGL/PyG, and cu_cat can boost your own workflows
Leo Meyerovich
Leo Meyerovich, founder and CEO of Graphistry, pioneered the visual graph AI platform, and created Louie.AI, the genAI-first analytics platform. They are leveraged by federal agencies, enterprises, and scientific organizations around the world for problems like threat hunting, anti-fraud measures, cancer genomics, and supply chain analysis. Before startups, Leo's research in high-performance computing, security, and program synthesis at UC Berkeley (PhD) and Brown (ScB) received the SIGPLAN 10 Year Test of Time award, multiple best paper awards, and features in prominent publications. His research has been adapted by popular browsers, web frameworks, and cloud providers. Projects include starting the GPU dataframe ecosystem, Apache Arrow's JavaScript implementation, the first functional reactive web framework, the first parallel web browser, and researching sociological foundations of programming languages.