A Malware Classification Algorithm Based on Multiple Data Augmentation and Transfer Learning Using Pretrained Image Classification Models

In order to solve the problem of limited number of training data and insufficient feature representation for malware classification, in this paper, the data augmentation and pretrained image classification models based algorithm is proposed. First, two malware image construction and augmentation approaches are proposed. The malware data is preprocessed and converted into RGB image by mapping. Then, the deep convolutional generative adversarial network is utilized for image augmentation of each channel. And the structural similarity index is proposed for generated malware image filtering. For another, the denoising diffusion implicit model is utilized for malware data augmentation of each class. After the cosine similarity and Jensen-Shannon divergence based data filtering, the Gramian angular summation field method is used for malware image construction. Second, the pretrained VGG16 and ResNet50 models are proposed for feature extraction by transfer learning. Moreover, three feature fusion strategies are designed. At last, by the off-line training, the malware classification model is obtained. Experiment results demonstrate that the proposed algorithm has better malware classification performance than some existing methods.