Detecting abnormal behaviors by using intelligent system

Abstract

With the rising use of Internet technologies around the world, the number of network intruders and attackers has skyrocketed. For this reason, introducing systems for detecting attacks within the network security measures stops hackers from gaining access to data. In order to detect various forms of assaults, the development of intrusion detection systems is quite crucial. The election of features and elimination of the unrelated information can improve the classifier's accuracy execution because the network traffic dataset contains many useful and unhelpful features. Along these lines, in this work, three hybrid strategies for selecting the feature were implemented, which include the constant feature by standard deviation, the Quasi constant by variance threshold, and the information gain. These approaches were used to order and rank features, after which the best higher ranking was selected for classification and intrusion detection. These features were tested on three classifiers long shortـterm memory, convolutional neural network, and CNN-LSTM. From the acquired results, a high level of attacks was detected, and classification accuracy was achieved by cross-breeding the selection of the different best features. Furthermore, the convolutional neural network classifier achieved the highest accuracy rate, which exceeded the value of 99.5%.