Feature Reduction Based Intrusion Detection System Using Deep Learning Technique

Abstract

Attacks and intrusions have grown significantly due to the rapid development of networks.
These attacks' detection and prevention have risen in importance as security measures. One
of the key components of achieving high security in computer networks is the intrusion
detection system, which is used to repel various types of attacks. The curse of
dimensionality affects intrusion detection systems, resulting in less effective for
incorporating resources and more complicated to run over time. The consideration of
evaluating crucial features in intrusion detection system can significantly reduce the
dimensions. In order to identify the useful features, this study proposes an intelligent
system that first extracts features using kernel PCA (principal component analysis) as well
linear discriminant analysis (LDA). A convolutional neural network (CNN) is then fed these
reduced features for testing and training on NSL-KDD dataset. The extracted dataset, first
normalized in order to reduce irrelevant features for the sake of quality input which
ultimately classify the data into attacks and non-attacks. The standard evaluation
measures; precision recall and accuracy are employed to assess the behavior of proposed
system on with as well as without feature reduction. A comprehensive comparison with
state of art systems has also been presented to underline the proficiency of proposed work
along with the impact of feature reduction with improved QoS. Experimental evaluation
revealed that our proposed optimized CNN with LDA is an effective solution as it reached
to 96.1% of better accuracy with and high precision on NSL- KDD dataset, by making a
reasonable margin LDA reached over Kernel Principal Component Analysis (KPCA),
whereas optimized CNN with KPCA reduced the overall computational cost. We must
encourage future researchers to actively participate in this research for improving and
optimizing the IDSs.