Network Anomaly Detection With Convolutional Neural Network Based Auto Encoders

Abstract

As many devices can be accessed through the internet and sensitive information can be sent and copied over the internet, the importance of measures to be taken against cyber attacks are increasing. Today's applications for preventing cyber attacks are generally successful against attacks stored in databases, but not against the ones previously unknown, so called zero-day attacks. In this study, a deep learning based model has been devoloped in order to detect the known network attacks and increase the detection performance of the zero-day attacks. NSL-KDD data set which has been used to simulate the zeroday attacks and compare the performance with the previous studies. Our convolutional neural network based denoising, sparse stacked auto encoder (CNN-DSSAE) model, using the swish activation function in the last layer and SGD with decoupled weight decay (SGDW) as the optimization algorithm, has achieved higher performance than the studies done with different machine learning and deep learning models on the same dataset.