Authors: Roman V. Yampolskiy; Leon Reznik; Mike Adams; Joshua Harlow; Dima Novikov
Addresses: Duthie Center for Engineering – 215, Speed School of Engineering, University of Louisville, Louisville, KY 40292, USA. ' Computer Science, Rochester Institute of Technology, Room 3-521, Building 70, Rochester, NY 14623, USA. ' Acme Packet, Inc., 100 Crosby Drive, Bedford, MA 01730, USA. ' Yahoo!, Inc., 701 First Avenue Sunnyvale, #A1334, San Jose, CA 94089, USA. ' VirtualScopics, Inc., 500 Linden Oaks, Rochester, NY 14625, USA
Abstract: High learning and adaptation ability of intelligent agents based on artificial neural networks (ANNs) has made them a popular tool in design and implementation of intrusion detection systems (IDS). However, ANN might consume significant resources during their retraining because of network changes. The paper investigates the design of ANN structures that may reduce the resource consumption without a substantial performance degradation. It describes the results of empirical studies examining a variety of design solutions, such as the choice of the ANN architecture and its parameters, the choice of an ANN fully connected topology versus a partial connectivity and the IDS design in a form of a hierarchical system of heterogeneous ANN-based agents. The results are analysed and design recommendations are provided. The fully connected ANN structure optimised with genetic algorithms has been found to achieve the best performance, while partial connectivity might save resources without a significant sacrifice of possible accomplishments.
Keywords: intrusion detection; design; genetic algorithms; artificial neural networks; ANNs; resource awareness; computational intelligence; intelligent agents.
International Journal of Advanced Intelligence Paradigms, 2011 Vol.3 No.3/4, pp.305 - 322
Published online: 26 Mar 2015 *Full-text access for editors Access for subscribers Purchase this article Comment on this article