GPU-accelerating a deep learning anomaly detection platform

How can deep learning be employed to create a system that monitors network traffic, operations data, and system logs to reliably flag risk and unearth potential threats? Satish Dandu, Joshua Patterson, and Michael Balint explain how to bootstrap a deep learning framework to detect risk and threats in operational production systems, using best-of-breed GPU-accelerated open source tools.

How can deep learning be employed to create a system that monitors network traffic, operations data, and system logs to reliably flag risk and unearth potential threats? There are many challenges with developing such a system. With new types of behavior constantly emerging, a robust and exhaustively tagged dataset is extremely difficult to obtain. Training a model on such a large amount of data can often take longer than practicality might dictate. Once trained, it is often difficult to employ the model in an expeditious way and get actionable results in a production environment. Drawing on NVIDIA’s system for detecting anomalies on various NVIDIA platforms, Satish Dandu, Michael Balint, and Joshua Patterson explain how to bootstrap a deep learning framework to detect risk and threats in operational production systems, using best-of-breed GPU-accelerated open source tools. They then demonstrate how to speed up the training cycle by employing several algorithmic hacks and leveraging a cluster of GPUs, shortening the training time from days to hours, dramatically cutting the inferencing time, and generally making the entire system much more adaptive. Join Satish, Michael, and Josh as they walk you through how they built such a system, covering the architecture, the algorithms they implemented, how they sped up various parts of the data pipeline, and their future roadmap to incorporate more acceleration from the GPU Open Analytics Initiative, GOAI.