An Evaluation of Convolutional Neural Networks for Malware Family Classification

Abstract

There has been a rapid rise and diversification in the quantity and types of malware that are currently being propagated. Hence, the need for a proper mechanism to classify these different types of malware are of paramount importance. Academic researchers have been analysing malware samples to understand how they behave and they study the techniques used by malware developers to improve the security of the existing infrastructure. Malware analysis can be used for both the detection of malware and malware classification. In this work, modern convolutional neural networks (CNN) are evaluated for the task of malware classification using image data. The networks that are used for testing are, VGG, ResNet, Inception-V3, and Xception. These networks have proven to work with high performance on huge ImageNet dataset, but the possibility of using such CNN’s needs to be checked for the very specific task of malware classification. Comparing the results, Xception Network provided the best performance with an accuracy of 99% and proved to be the fastest network. In terms of training Inception Network was better. Furthermore, individual precision and recall values were calculated for each family.KeywordsMalware classificationConvolutional neural networkResNetInception-V3XceptionVGG