HAE-HRL: A network intrusion detection system utilizing a novel autoencoder and a hybrid enhanced LSTM-CNN-based residual network

Abstract

As networks evolve, their attacks become ever more varied - which creates an increasing variety of features-rich information which models must incorporate during training. However, this data often includes redundant and irrelevant features that impede its effectiveness as an intrusion detection system. Hybrid Autoencoder- Hybird ResNet-LSTM, an advanced hybrid residual network which combines an innovative hybrid Autoencoder with an enhanced LSTM-CNN architecture, was introduced here to enhance detection capabilities of models and identify pertinent feature subsets within datasets more quickly and efficiently. Initial feature selection within the dataset is performed using a modified self-encoder that incorporates CNN and GRU components, in order to reduce data dimensionality while pinpointing an optimal subset. This paper assesses a proposed intrusion detection model against three datasets commonly used for intrusion detection studies: UNSW-NB15, NSL-KDD, and CICIDS-2018. Experimental findings demonstrate high accuracy rates of 95.7%, 94.9% and 96.7% in intrusion detection for NSL-KDD, UNSW-NB15, and CICIDS-2018 datasets respectively. A comparative analysis with methods proposed by other researchers illustrates how effective our method presented here can be at significantly enhancing intrusion detection accuracy.