A comparative study of deep transfer learning models for malware classification using image datasets

Abstract

This paper proposes deep convolution neural network-based malware classification approach. The proposed work is a transfer learning approach, where we have developed multiple deep learning classification models. The classification models are built by adapting multiple pre-trained convolutional neural networks, namely; Xception, VGG19, InceptionResNetV2, MobileNet, InceptionV3, DenseNet, and ResNet50. In the current work, weights of pre-trained models are embellished by adding three fully connected (FC) layers. The proposed models have been evaluated on two different malware datasets, Microsoft and MalImg, consisting of malware images. The focus of this paper is to analyse the performance of fine-tuned CNN models for malware classification. The results of our experiments show that InceptionResNetV2 and Xception models have performed considerably well for the Microsoft dataset with accuracy equal to 96% and 95%, respectively. In the case of the MalImg dataset, InceptionResNetV2, InceptionV3, and Xception models have achieved excellent performance with an accuracy of up to 96%.