基于EfficientNetV2和特征融合的恶意软件分类方法

Malware Classification Based on EfficientNetV2 and Feature Fusion

针对现有恶意软件分类方法特征提取的单一性及对通道权重忽视的问题,本文提出了一种基于EfficientNetV2和特征融合的新型分类方法。该方法通过综合利用Byte和Asm文件从多角度提取特征图像,融合生成三通道图像以提供更全面的恶意软件特征表达,并采用EfficientNetV2深度学习模型进行分类,更精确地刻画恶意软件间的相似性,从而提高分类准确率。在BIG2015数据集上的实验结果表明,本文方法的分类准确率达到了99.14%,能够有效分类恶意软件家族,凸显了特征融合和深度学习模型在恶意软件分类领域的巨大潜力。Addressing the limitations of singularity of feature extraction and the neglect of channel weights in existing malware classification methods, this paper introduces a novel classification method based on EfficientNetV2 and feature fusion. This method combines Byte and Asm files to extract multi-dimensional feature images, creating three-channel images for a more comprehensive representation of malware features. Utilizing the EfficientNetV2 deep learning model, the approach enhances the accuracy of malware classification by capturing subtle similarities among malware more precisely. Experiments on the BIG2015 dataset demonstrate a classification accuracy of 99.14%, effectively categorizing malware families and highlighting the significant potential of feature fusion and deep learning models in the field of malware classification.