Data preprocessing for non-image data: using convolution neural networks for network intrusion detection systems
2021
Application of machine learning models in network intrusion detection systems has been the subject of extensive investigation and testing.Modern day networks produce data in quantities that put more emphasis on the accuracy and precision of the intrusion detection systems.This produces the drive for more accurate and time-efficient i ntrusion d etection s ystems, a nd m achine l earning w as i nvestigated a s a v iable solution.Research into machine learning models in other fields has yielded several different algorithms and approaches, highly specialised to those particular data types.Testing for intrusion detection has found that the models that process the network data best tend to yield higher accuracy and lower false-positive rates, whereas those models that perform best on their original data have struggled.One such model that under performed in intrusion detection when compared to the original field i s c onvolution n eural n etwork.T his p aper a ims to investigate preprocessing methods for network data to increase the effectiveness of using a convolution neural network model as part of a network intrusion detection system.Specifically, the paper will analyse the use of the DeepInsight architecture,using a modified t-distributed stochastic neighbour embedding technique, the positioning of features in isolation and a control class of simple reshaping data from vector to matrix form.