Malware Classification and Visualization Using EfficientNet and B2IMG Algorithm

Abstract

The growth of malware has been significantly high in the last few years. Signature-based and heuristic methods have already been used in malware classification for a long time, but since the growth of polymorphic, both methods have become irrelevant. This problem makes machine learning and deep learning popular to overcome this problem. EfficientNet is one of the transfer learning models, a deep learning subfield. It said that this model could beat the state of the art of deep learning classification in ImageNet. In this paper, the researcher already implemented several EffiicientNet models into two type of Malware BIG 2015 that had been visualized into grayscale and RGB format. From the experiment, we found that EfficientNetB7 implemented into RGB dataset got 99.63% of accuracy, 98.36% of precision, 98.35% of recall, 98.34% of F1-Score, and 98.30% of AUC, with only takes 10 epochs in training process. It could outperformed other pretrained model within a few epochs.