信息网络安全 ›› 2022, Vol. 22 ›› Issue (10): 31-38.doi: 10.3969/j.issn.1671-1122.2022.10.005

• 入选论文 • 上一篇    下一篇

基于多尺度卷积神经网络的恶意代码分类方法

刘家银1,2,3, 李馥娟1,2,3,4(), 马卓1,2,3, 夏玲玲1,2,3   

  1. 1.江苏警官学院计算机信息与网络安全系,南京 210031
    2.江苏省电子数据取证分析工程研究中心,南京 210031
    3.江苏省公安厅数字取证重点实验室,南京 210031
    4.南京大学计算机软件新技术国家重点实验室,南京 210093
  • 收稿日期:2022-08-12 出版日期:2022-10-10 发布日期:2022-11-15
  • 通讯作者: 李馥娟 E-mail:lifujuan@jspi.cn
  • 作者简介:刘家银(1986—),男,重庆,讲师,博士,主要研究方向为信息安全、机器学习|李馥娟(1974—),女,陕西,教授,硕士,主要研究方向为信息安全|马卓(1993—),女,山西,讲师,博士,主要研究方向为隐私保护、时间序列分析|夏玲玲(1988—),女,江苏,副教授,博士,主要研究方向为网络安全技术、网络传播动力学
  • 基金资助:
    国家自然科学基金(62272203);江苏省市场监督管理局科技计划项目(KJ21125027);江苏省公安厅科技研究项目(2020KX008);江苏省公安厅科技研究项目(2021KX011);江苏省高等学校自然科学基金(21KJD520003);计算机软件新技术国家重点实验室(南京大学)开放课题(KFKT2022B23)

Malware Classification Method Based on Multi-Scale Convolutional Neural Network

LIU Jiayin1,2,3, LI Fujuan1,2,3,4(), MA Zhuo1,2,3, XIA Lingling1,2,3   

  1. 1. Department of Computer Information and Cyber Security, Jiangsu Police Institute, Nanjing 210031, China
    2. Jiangsu Electronic Data Forensics and Analysis Engineering Research Center, Nanjing 210031, China
    3. Key Laboratory of Digital Forensics of Jiangsu Provincial Public Security Department, Nanjing 210031,China
    4. State Key Laboratory for Novel Software Technology at Nanjing University, Nanjing 210093, China
  • Received:2022-08-12 Online:2022-10-10 Published:2022-11-15
  • Contact: LI Fujuan E-mail:lifujuan@jspi.cn

摘要:

恶意代码文件大小差异巨大,使用传统卷积神经网络对其可视化图像进行训练时会因分辨率调整导致大量信息丢失。为此,文章提出一种基于多尺度卷积神经网络的恶意代码分类方法。该方法首先将不同大小的恶意代码生成为多种特定分辨率的图像;然后利用DenseNet网络提取特征,避免因调整至同一分辨率导致信息损失;最后通过空间金字塔模型处理多尺度特征,进而训练分类模型。实验结果表明,该方法有效提高了恶意代码分类性能。

关键词: 恶意代码分类, 空间金字塔, 多尺度, 卷积神经网络

Abstract:

Because of the huge difference in size between different malware, one has to manually unify the resolution of their visualization images while training deep neural networks for malware classification, which may in turn cause severe information loss due to resolution adjustments. To this regard, this paper proposed a novel malware classification method based on the merits of multi-scale convolutional neural networks. Specifically, this method first visualized malware of different sizes into images of various specific resolutions, and then adopted the DenseNet network for feature extraction to avoid information loss in resolution unification. Finally, multi-scale features were processed through the spatial pyramid model to train the classification model. Extensive experimental results show that the proposed method could effectively improve the performance of malware classification.

Key words: malware classification, spatial pyramid, multi-scales, convolutional neural network

中图分类号: