SADCNN: Supervised anomaly detection based on convolutional neural network models

As the importance of the Internet increases and computer networks get more extensive, network security becomes one of the most important aspects that must be considered. Since the network attacks evolve at an astounding pace and the network data is high-volume, anomaly detection approaches based on Deep Learning (DL) are introduced as a solution to detect unknown attacks in voluminous network data. In this paper, we focus on the problem of supervised anomaly detection. Since Convolutional Neural Networks (CNN) are widely used in data processing with voluminous data, we proposed a model based on CNN named SADCNN. SADCNN is a CNN with two convolutional layers, the Glorot uniform weight initializer, the Tanh activation function, 20 epochs, 32 filters, and the Adam optimizer. To evaluate our approach on real network traffic, the KDDCUP'99, NSL-KDD-Train+, and UNSW-NB15 datasets are attempted. The simulation results demonstrate that SADCNN is a proper feature extractor for network data. It gives better results for F1-score and AUC1 in comparison to the previous models.