Fusion of linear and non-linear dimensionality reduction techniques for feature reduction in LSTM-based Intrusion Detection System

Abstract

Securing networks is becoming increasingly crucial due to the widespread use of information technology. Intrusion Detection System (IDS) plays a crucial role in network security by detecting potential security threats in real-time. In order to create effective IDS, Deep Learning (DL) techniques like Auto Encoder (AE) and Long Short-Term Memory (LSTM) have been widely used. However, the high dimensionality and complexity of network traffic data make it challenging to extract meaningful information. The paper proposes a fusion-based approach that combines AE and Principal Component Analysis (PCA) techniques for dimensionality reduction in IDS. The proposed approach aims to capture both linear and non-linear relationships between features while reducing the dimensionality of the input data. NSL-KDD, UNSW-NB15, CIC-IDS-2017, and MSCAD datasets are used to evaluate this proposed method. The proposed approach has improved accuracy compared to the existing AE+LSTM-based approach by 3% for NSL-KDD and around 1% for UNSW-NB15 and CIC-IDS-2017, while the proposed approach gives comparable results for the MSCAD dataset. The Wilcoxon signed-rank test has been applied to confirm the statistical significance of the result.