An Application of CNN-based Models with Fine-tuning Techniques on Malevis Malware Dataset

Abstract

In recent years, the increasing numbers of cyber-attacks has marked a concerning trend in community in both frequency and severity. The malware could exist in several forms, so that it is much more complicated to detect. The previous works on applying machine learning models in malware classification mostly focusing on the MalImg dataset which contains 9339 malware byteplot images from 25 different families. However, this dataset, dating back to 2011, lacks updates to accommodate evolving malware families. Moreover, prior studies solely addressed malware classification using grayscale images from MalImg, neglecting both malware detection and color photo analysis. This paper aims to review and apply convolutional neural networks (CNNs) model for detecting and classifying malware across grayscale and color image datasets. By deploying CNNs, this research undertakes a comparative analysis to assess their efficacy in addressing these dual challenges. The evaluation utilizes the Malevis dataset, chosen for its contemporary nature and reliability, offering diverse representations of malware types through color images. The study anticipates that training CNN models on the Malevis dataset will yield insights into their accuracy in malware detection.