Engineered Statistical Features Dominate Network Anomaly Detection in Complex Deep Architectures

Written by

in

Hina Naaz and Reema Ajmera evaluated an anomaly detection pipeline based on a customized LeNet and found that removing time-window statistical features reduced the area under the precision-recall curve by as much as 0.22, while multi-scale pyramid pooling, adaptive activation, and transfer learning offered only minor gains or reduced diagnostic accuracy. The experiments used 528,690 enterprise log events covering 151 users across seven departments; 126 features were retained after information-gain-based selection.

Source:BIOENGINEER.ORG · Read Original

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *