Noisy Label, Weak Features, True Impacts on Login Risk Detection

10:30 am - 10:50 am

Objective: Guidance
Audience Level: Beginner/Intermediate
Session Type: Presentation

Attackers compromise accounts for malicious purposes. It is crucial to prevent these compromises by detecting malicious login events. Machine learning systems require reliable labels for training and evaluation. We present a method to improve labeling reliability through label correction based on history and future events. The resulting model significantly improves login risk detection accuracy.


, Senior Research Scientist, Yahoo