Performance Evaluation of Deep Learning Based Network Intrusion Detection System across Multiple Balanced and Imbalanced Datasets

Abstract

In the modern era of active network throughput and communication, the study of Intrusion Detection Systems (IDS) is a crucial role to ensure safe network resources and information from outside invasion. Recently, IDS has become a needful tool for improving flexibility and efficiency for unexpected and unpredictable invasions of the network. Deep learning (DL) is an essential and well-known tool to solve complex system problems and can learn rich features of enormous data. In this work, we aimed at a DL method for applying the effective and adaptive IDS by applying the architectures such as Convolutional Neural Network (CNN) and Long-Short Term Memory (LSTM), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU). CNN models have already proved an incredible performance in computer vision tasks. Moreover, the CNN can be applied to timesequence data. We implement the DL models such as CNN, LSTM, RNN, GRU by using sequential data in a prearranged time range as a malicious traffic record for developing the IDS. The benign and attack records of network activities are classified, and a label is given for the supervised-learning method. We applied our approaches to three different benchmark data sets which are UNSW NB15, KDDCup ’99, NSL-KDD to show the efficiency of DL approaches. For contrast in performance, we applied CNN and LSTM combination models with varied parameters and architectures. In each implementation, we trained the models until 100 epochs accompanied by a learning rate of 0.0001 for both balanced and imbalanced train data scenarios. The single CNN and combination of LSTM models have overcome compared to others. This is essentially because the CNN model can learn high-level features that characterize the abstract patterns from network traffic records data.