A Deep Learning based Approach for Intrusion Detection System

Abstract

The internet has become a necessary tool in our life The convenience of online services has led to many applications However as they are increasingly popular the attacks on them are on the rise dramatically in recent years An intrusion detection system (IDS) has been considered as one of the main defense mechanisms Recently due to the popularity of machine learning more and more researches have applied machine learning techniques in intrusion detection systems to improve the detection efficacy The reason is that machine learning is a predictive model and can work very accurately on classification In this thesis we proposed an intelligent deep learning based approach for constructing an intrusion detection system The proposed IDS scheme consists of a pre-clustering method for discrete features a feature extraction scheme and a classification mechanism Applying a pre-clustering method on discrete features can reduce feature dimensions and effectively assist feature extraction Furthermore in the process of feature extraction we employed the Winner-take-all autoencoder which has been proven to be very effective in extracting data features and helpful on the subsequent classification work The proposed IDS scheme has been conducted experiments on the NSL-KDD and CICIDS2017 datasets The experimental results show that both Winner-take-all autoencoder and the pre-clustering method can effectively improve the accuracy of classification The proposed IDS scheme achieves the accuracy at 87 24% on the NSL-KDD dataset which is the highest accuracy among all the state-of-art researches Hence it also outperforms most researches on the CICIDS2017 dataset with excellent accuracy of 97 92% Therefore the proposed IDS improves the accuracy of intrusion detection and provides a novel method for intrusion detection

Date of Award	2019
Original language	English
Supervisor	Hui-Tang Lin (Supervisor)