Scholarlink

<p>The research aims to tackle the problem of domain shift in malware image classification tasks where there is poor cross-domain generalization among different datasets. Two popular datasets for this task are considered: Malimg (grayscale images, 9,339 images) and Malevis (RGB images, approximately 14,200 images). The proposed method is based on a unified deep learning framework that incorporates Domain-Adversarial Neural Networks (DANN), semi-supervised learning, and a dual-branch Convolutional Neural Network (CNN) to tackle the domain shift problem. The proposed method is evaluated on three publicly available malware families: Agent, Allaple, and Autorun. The experimental results show that although poor cross-domain accuracy is observed for baseline and DANN models (&lt; 30%), the proposed method that incorporates semi-supervised learning and dual-branch CNN significantly improves cross-domain accuracy to 77.98% and 94.67% for Malimg to Malevis and Malevis to Malimg, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cross-domain malware image classification using unified deep learning and domain adaptation models

  • 期刊论文

  • 发表日期：

    2026年5月31日

查看全文
  • Bidhi Kataruka,
  • M. Gurupriya

摘要

The research aims to tackle the problem of domain shift in malware image classification tasks where there is poor cross-domain generalization among different datasets. Two popular datasets for this task are considered: Malimg (grayscale images, 9,339 images) and Malevis (RGB images, approximately 14,200 images). The proposed method is based on a unified deep learning framework that incorporates Domain-Adversarial Neural Networks (DANN), semi-supervised learning, and a dual-branch Convolutional Neural Network (CNN) to tackle the domain shift problem. The proposed method is evaluated on three publicly available malware families: Agent, Allaple, and Autorun. The experimental results show that although poor cross-domain accuracy is observed for baseline and DANN models (< 30%), the proposed method that incorporates semi-supervised learning and dual-branch CNN significantly improves cross-domain accuracy to 77.98% and 94.67% for Malimg to Malevis and Malevis to Malimg, respectively.