Investigating Bio-Inspired Feature Selection for an Enhanced Ensemble-Based Intrusion Detection System

Resumo

An intrusion in a network traffic is a serious attack on the security of systems, which can sometimes collapse if their data is not protected enough or if the intrusion is not detected at the right time. The network traffic state is commonly represented in a detection intrusion system (IDS) by datasets with a wide range of parameters, which can considerably slow down the system’s response time, and thus preventing detection at the right moment to deal with the attack. Hence the need to integrate a feature selection engineering module into any effective IDS. The current work revolves around the selection of the most relevant and representative characteristics of the network to be able to detect the intrusion with the best precision while minimizing the system computation time. An investigation of four bio-inspired models is made to extract the best subset according to each one, i.e. genetic algorithm, gray wolf algorithm, bat algorithm, and particle swarm optimizer. Results are gathered in a pool where two strategies are developed to decide on the final structure of the subset, which is validated by a classification process. Two strong classifiers namely logistic regression and decision tree, are used in an individual way, and then in fusion approach to benefit from the strengths of several machine learning models. The ensemble model showed that we can achieve highly accurate and reliable decisions (99,8%) while working only on reduced subsets.