Deriving Optimal Deep Learning Models for Image-based Malware Classification

Analyzing a huge amount of malware is a major burden for security analysts. Since emerging malware is often a variant of existing ones, automatically classifying malware into known families greatly reduces their burden. Image-based malware classification with deep learning is an attractive approach for its simplicity, versatility, and affinity with existing technologies. However, the impact of different deep learning models and the degree of transfer learning on the classification accuracy has not been fully investigated. In this paper, we conducted an exhaustive study of deep learning models using 24 models pre-trained with ImageNet and 5 fine-tuning parameters, 120 models in total, for malware on two platforms. As a result, we derived the optimal deep learning models by fine-tuning the pre-trained models and achieved the cross-validation accuracy on the Malimg (98.96%) and Drebin (91.03%) datasets.