Agents and Neural Networks for Intrusion Detection

Abstract

Up to now, several Artificial Intelligence (AI) techniques and paradigms have been successfully applied to the field of Intrusion Detection in Computer Networks. Most of them were proposed to work in isolation. On the contrary, the new approach of hybrid artificial intelligent systems, which is based on the combination of AI techniques and paradigms, is probing to successfully address complex problems. In keeping with this idea, we propose a hybrid use of three widely probed paradigms of computational intelligence, namely Multi-Agent Systems, Case Based Reasoning and Neural Networks for Intrusion Detection. Some neural models based on different statistics (such as the distance, the variance, the kurtosis or the skewness) have been tested to detect anomalies in packet-based network traffic. The projection method of Curvilinear Component Analysis has been applied for the first time in this study to perform packet-based intrusion detection. The proposed framework has been probed through anomalous situations related to the Simple Network Management Protocol and normal traffic.