Intrusion Detection System for Multiclass Detection based on a Convolutional Neural Network

Given the ever-increasing volume of traffic and the developing nature of internet communications, security has emerged as one of the primary considerations in the design and implementation of any system or network. The frequency of network attacks is increasing, so the need for protecting both the users’ and systems’ data is crucial. In recent years, there has been an increase in use of machine learning and deep learning techniques as an aid for detecting network attacks. Various deep learning models have been in focus of researchers for the past couple of years, having in mind their capabilities for processing large amounts of data. However, when it comes to handling multi-class imbalanced data, proper detection of multiple attacks is still a challenge. In this paper, a multiclass network intrusion detection system is implemented using a convolutional neural network design. The CICIDS-2017 dataset, a highly imbalanced dataset with 15 distinct traffic classes, is used to train and test the model. Two feature selection techniques are used, and depending on the input dataset dimensions and the hyperparameter variations, 24 different CNN model variations are trained. In order to evaluate the models objectively, macro average metric is used. A comparison is made with a DNN model, in order to show the CNN model capabilities. It is shown that the proposed model outperforms DNN and detects multiple classes that are represented by a limited number of inputs in the dataset with high precision.