DMNAED: A Novel Framework Based on Dynamic Memory Network for Abnormal Event Detection in Enterprise Networks

Abstract

Abnormal event detection is a crucial step towards discovering insider threat in enterprise networks. However, most existing anomaly detection approaches fail to capture latent correlations between disparate events in different domains due to the lack of a panoramic view or the disability of iterative attention. In light of this, this paper presents DMNAED, a novel framework based on dynamic memory network for abnormal event detection in enterprise networks. Inspired by question answering systems in natural language processing, DMNAED considers the event to be inspected as a question, and a sequence of multi-domain historical events serve as a context. Through an iterative attention process, DMNAED captures the context-question interrelation and aggregates relevant historical events to make more accurate anomaly detection. The experimental results on the CERT insider threat dataset r4.2 demonstrate that DMNAED exhibits more stable and superior performance compared with three baseline methods in identifying aberrant events in multi-user and multi-domain environments.