Hybrid approaches for optimal network intrusion detection using deep learning classifiers

Abstract

Intrusion detection has become one of the challenging fields of research in every networked environment. The IDS models should be able to handle the huge volume and velocity associated with network communications. Current researches in areas of intrusion detection have tended towards network-based systems and how to improve on their intrusion detection. However, both host-based and network-based systems should be involved to effectively detect attacks from insider as well as outsider users. The extraction of spatial-temporal features is very rare in the IDS. Likewise, the packets-based network IDS are the traditional method. Since network link speeds and traffic volume grew over the last years, packet-based analysis became difficult, leading to the development of alternative approaches for flow-based analysis which are still in their earlier stage of real-time implementation. This thesis presents three contributions, OCNN-HMLSTM, MR-BWO-ConvLSTM, and MSHIDS that can be used for fast and efficient detection of intrusions in networked environments by overcoming these described limitations. The first contribution presents the Optimized Convolutional Neural Networks-Hierarchical Multi-scale Long Short-Term Memory (OCNN-HMLSTM) based IDS model that is used for improving intrusion detection by integrating the spatial and temporal aspects in the intrusion datasets. This model has combined both the optimized CNN and HMLSTM models. In this model, OCNN is used to learn the spatial aspects while HMLSTM learns the temporal aspects and classifies the data effectively. The OCNN model is developed integrating the LSO algorithm with the standard CNN classifier to optimally select the values of the CNN weights and hyper-parameters to tune its performanc newline