The Impact of PCA-Scale Improving GRU Performance for Intrusion Detection

Abstract

A device or software appliance monitors a network or systems for malicious activity is an Intrusion Detection System (IDS). Conventional IDS does not detect elaborate cyber-attacks such as a low-rate DoS attack as well as unknown attacks. Machine Learning has attracted more and more interests in recent years to overcome these limitations. In this paper, we propose a novel method to improve intrusion detection accuracy of Gated Recurrent Unit (GRU) by embedding the proposed PCA-Scale with two options including PCA-Standardized and PCA-MinMax into the layer of GRU. Both optional methods explicitly enforce the learned object feature maps by affecting the direction of maximum variance with positive covariance. This approach can be applied to GRU model with negligible additional computation cost. We present experimental results on two real-world datasets such as KDD Cup 99 and NSL-KDD demonstrate that GRU model trained with PCA-Scaled method achieves remarkable performance improvements.