Effective feature selection based on manova

Abstract

Effectiveness in classifying malware is a critical issue which can overheat a classifier or reduce performance in real-time malware detection systems. However, the effectiveness in feature selection stage was not studied so far. As effectiveness should be taken into account at the earliest possible stages, in this paper, we focus on the effectiveness of feature selection. Firstly, we perform an analysis on instruction levels which consists of most frequencies mnemonics. Secondly, we propose new methods to select effective features by MANOVA statistical tests. Furthermore, we use those selected features fed to a classifier. Our approach reduces significantly the number of features from 390 to 4, which explains 99.4% variation of the data. With the selected features, we classify malware samples and have achieved 96.2% of accuracy and 0.6% of false positive. Copyright © 2020 Inderscience Enterprises Ltd.