Network Intrusion Detection based on LDA for payload feature selection

Abstract

Anomaly Intrusion Detection System (IDS) is a statistical based network IDS which can detect attack variants and novel attacks without a priori knowledge. Current anomaly IDSs are inefficient for real-time detection because of their complex computation. This paper proposes a novel approach to reduce the heavy computational cost of an anomaly IDS. Linear Discriminant Analysis (LDA) and difference distance map are used for selection of significant features. This approach is able to transform high-dimensional feature vectors into a low-dimensional domain. The similarity between new incoming packets and a normal profile is determined using Euclidean distance on the simple, low-dimensional feature domain. The final decision will be made according to a pre-calculated threshold to differentiate normal and abnormal network packets. The proposed approach is evaluated using DARPA 1999 IDS dataset.