Leveraging graph AI and GPUs to win the US Cyber Command AI challenge
summary
First-place system for the US Cyber Command Alert AI Challenge combines graph correlations, UMAP embeddings and graph neural nets with GPU ETL and a GPU reimplementation of Dirty Cat to turn 100M+ raw events into incident-level graphs. Pipeline uses Jupyter, Nvidia RAPIDS, PyGraphistry[AI], DGL/PyG and cu_cat to vectorize, reduce and link alerts, producing 97% alert volume reduction for an ISP/university fusion center and enabling UMAP embeddings and GPU feature engineering for interactive investigations.
AI-written summary · about
chapters (18)
- 0:00Intro
- 8:26Running the query
- 13:10Drawing a graph
- 20:10The challenge
- 21:31Natural language
- 22:14Multiphase analysis
- 24:55Network scans
- 27:34Umap
- 28:43Rapid prototyping
- 34:28Automatic feature engineering
- 36:53Anomaly detection
- 38:19Connecting entities
- 39:46Umap view
- 42:04Graph Neural Networks
- 43:35Security
- 45:39Data wrangling
- 48:09Speed
- 49:23Takeaways
related talks
40mHow to Map Networked Devices Using Graph-Based AI to Preempt Cyber Threats
48mHunting Stealth Adversaries with Graphs & AI
27mChagu: Graph-Powered LLM Shielding and Anomaly Detection for Secure AI Pipelines
20mAttack of the Graph: Visual Tools for Cyber Analysis
56mGraph Analytic Systems
12mMake ML outcomes roar with graph neural networks on TigerGraph and AWS (PRT097)
17mElevating security investigations with generative AI
25mBeyond Human Scale: AI and Automation as Force Multipliers in Cyber Threat Intelligence