Image-Based Malware Classification Powered by Deep Learning based DenseNet-121 for Enhancing Cybersecurity

Cybersecurity is threatened by the increased number of malware and the growing emergence of sophisticated malware types. Traditional anti-malware software, commonly binarily dependent, cannot effectively cope with new and disguised threats. In this paper, authors investigate the possibility of rank-based image classification of malware using DenseNet-121 within the context of the Malimg dataset. The malware binaries introduced here are instead transformed into grayscale images: this way, the task becomes an image classification, where the model is capable of distinguishing fine features and structures which are distinctive of families of malware. DenseNet-121 uses a densely connection which enables feature propagation, reduces vanishing gradient and improves the representation learning since the features are reused at different levels. Numerous experiments on the Malimg dataset show that DenseNet-121 performs with high accuracy and is less sensitive to dataset bias than the other models in the field. The work further explores the effects of several preprocessing methods, data augmentation, and hypers parameter optimization on the model results. The results further emphasize the effectiveness of image-based analysis as a supplementary approach to malware detection, which is extendable and malleable against the emerging malicious software threats.