Weight embedding autoencoder as feature representation learning in an intrusion detection systems

The increasing need for safe internet access that withstands against various malicious attacks has gained much attention, especially in this abundant information age. Network intrusion detection systems have been proposed to tackle these kinds of attacks. Various machine learning algorithms, including deep learning, have been utilized in network intrusion detection systems. However, the limitations of current feature extraction methods for feature representation learning result in low detection accuracy. Our method utilizes an autoencoder to extract the low-level features in order to provide the necessary information for the classifier. We propose a weight embedding autoencoder to share feature representations between the autoencoder and the classifier. We apply our proposed method to improve two types of networks, a weight embedding autoencoder with a multi layers neural network (WE-AE DNN) and a weight embedding autoencoder with a convolutional neural network (WE-AE CNN) as the classifier. Experiments on two demanding benchmark datasets, such as NSL-KDD and UNSW-NB15 show the effectiveness and superiority of our proposed algorithm in terms of accuracy. The WE-AE DNN and WE-AE CNN improve the accuracy by 0.4% and 0.5% on NSL-KDD, respectively. Meanwhile, on the UNSWNB15 dataset, the WE-AE DNN and WE-AE CNN improve the accuracy by 2.8% and 0.5%, respectively.