Building Visual Malware Dataset using VirusShare Data and Comparing Machine Learning Baseline Model to CoAtNet for Malware Classification

The present work takes inspiration from the work of Zihang Dai, Hanxiao Liu, Quoc V. Le, Mingxing Tan at Google Research, Brain Team about CoAtNet. In that work it was showed that it is possible to combine the strengths from both convolution and transformer architectures, by unifying convnets and self-attention into a machine learning model. We want to apply the CoAtNet to a visual dataset of malware images and compare its performances to a baseline CNN model. For this reason we need a data set of appropriate size and format. From these needs triggers the requirement to find or generate a visual dataset of the malware images capable to measure the accuracy of the constructed model. As will be seen, the creation of a new dataset will be preferred to the search for an existing dataset. Although the visual approach has already been extensively tested in recent years, there is still a need for more customised data for the model under examination. The work described in this paper can serve as a guide to a balanced and dimensioned construction of an optimal malware visual image dataset.