Develop an Intelligent Anomaly Intrusion Detection System in Computer Networks based on Resilient Back-propagation Neural Network

Abstract

Various anomaly attacks and disruptions to information networks are considered serious problems that affect the protection of information exchanged between these networks and affect the maintenance of reliability and confidentiality of information exchange. In the past decade, researchers around the world have faced many challenges and need to propose a set of systems with flexible architectures to accurately and automatically detect anomaly intrusion attacks to address their complexity. Related research has proposed many full-scale solutions based on machine learning ML techniques. Recent research has focused on building an anomaly intrusion detection system AIDS from a mathematical and architectural point of view, using sophisticated methods such as support vector machines (SVMs) and convolutional neural networks (CNNs). Many studies use moderate and low complexity AIDS based on the classical multilayer neural network MLNN. Therefore, the accuracy of MLNN classifiers in the testing phase is moderate or low. Based on relevant AIDS studies proposed in the literature and our detailed investigation, we find that the resilient backpropagation RBP algorithm is not used as a learning method for MLNN-based AIDS. In particular, RBP is an effective tool in many nonlinear binary classifiers. In this paper, we present an AIDS construction method based on MLNN trained by the RBP algorithm, using well-known related data NSL-KDD and CIC-DDoS2019. In this study, we carefully selected an appropriate AIDS architecture and made many attempts to avoid the above difficulties. Our AIDS was found to be stably trained without limitations in a reasonable amount of time, and subsequently tested on unprecedented data, with an accuracy of about

99%. We also compared the performance of our algorithm with other well-known MLNN learning algorithms (Levenberg Marquardt LM, Bayesian Regulated BR, and Quasi-Newton QN) using the same AIDS architecture and data set. The comparison results show that the RBP algorithm has the best performance among many algorithms.