Image-based Malware Classification using Deep Convolutional Neural Network and Transfer Learning

Abstract

Malware classification is a major challenge as they have multiple families and its type has been ever increasing. With the involvement of deep learning and the availability of massive data, neural networks can easily address this problem. This experimental work focuses on classifying the malware that are in the form of grayscale images into their respective families with high accuracy and low loss. We used transfer learning in a pretrained VGG16 model obtaining an accuracy of 88.40% of accuracy. Additionally, upon experimenting with the ResNet-18, InceptionV3 model to classify the malware images into their families didn't give better results as compared to our custom model. The custom model based on convolution neural network achieved better accuracy and was able to classify malware with 98.7% validation accuracy. Upon comparing our custom model with VGG16, ResNet-18, InceptionV3 the custom model gave better accuracy with a better f1 score of 0.99. But improper tuning of VGG16 yielded low accuracy and high parameter loss. In overall the approach of classifying malware by converting them into images and classifying thus obtained images makes the malware classification problem easier.