Robust Network Intrusion Detection Systems for Outlier Detection

Abstract

Machine Learning Algorithms have become a crucial tool for designing Intrusion Detection Systems(IDS). The research community has identified deep learning architectures like Convolutional Neural Networks(CNN) as the go-to solution for IDS. However, these deep learning models are not immune to new outliers. We propose a Robust Network intrusion Detection system (RNIDS) model, which combines a CNN architecture followed by K Nearest Neighbors method. The proposed RNIDS model can classify different known attacks, and then predict if a new arriving traffic is an outlier with very high accuracy. We train and evaluate a CNN-based model which can classify attacks with an accuracy of 98.3% using up only 70,252 training parameters.