Tom Tom Festival 2019

We'll describe our work to achieve anomaly detection at network speed by combining probabilistic modeling, graph-based models, and more traditional machine learning techniques with the open source RAPIDS suite of software libraries. Traditional approaches to cybersecurity take a reactive approach, studying previous attacks to flag similar attacks in the future. This leaves systems vulnerable to day zero attacks in which adversaries use entirely new tactics to infiltrate a network. We'll explain how we address this issue by using multiple unsupervised models to alert cyber analysts about anomalous behavior, and then incorporate analysts' feedback to continuously update our models. The nature of anomaly detection in this low signal-to-noise space results in a high-false positive rate for most machine learning approaches. To mitigate this, we incorporate cyber analysts' feedback on alerts to continuously update our models. By combining the power of machine learning with experts' cyber knowledge in one integrated learning platform powered by GPUs, we improve the accuracy of future alerts, overall model performance, and reduce the time to detection for novel attacks. 

You need this ticket from Eventbrite to sign up: Applied Machine Learning Conference.