Combination of Traditional and Deep Learning based Architectures to Overcome Class Imbalance and its Application to Malware Classification

Classifying malware programs is useful to improve our understanding of their malicious features. Several studies that use pattern recognition, machine learning, and deep learning based algorithms have been proposed in this area. However, class imbalance remains a challenge to these learning algorithms. In this paper, we hybridize existing ones to effectively overcome that challenge. We propose an architecture that uses 1) a novel feature extraction technique followed by non-parametric classification method for cases with limited data and 2) a deep learning-based grouping algorithm for cases with sufficient data thereby creating a new effective approach for classifying imbalanced data. We make use of the publicly available database (originally provided for the Microsoft Malware Classification Challenge BIG2015) to evaluate the proposed approach. More specifically, we show that for the case where there are very few exemplars, our approach yields improved results over an existing state of the art system (existing benchmark).