Scholarlink

In the Internet of Things (IoT) domain, the heterogeneity of devices and the widespread deployment of resources significantly impact traditional malware classification methods. To effectively extract both local and global features of IoT malware and improve classification accuracy, a hybrid model, named SRViT-MCNet, is proposed. The model first visualizes IoT malware by applying the Occlusion Sensitivity method to capture regions where the model’s prediction confidence significantly drops. Topological Data Analysis (TDA) is then employed to generate Persistent Images (PI). The model utilizes a parallel feature extraction strategy, combining the SRViT module with the MCNet module to capture multi-dimensional features. The MC module integrated into the CNN enhances the model’s ability to capture fine-grained local features. Additionally, the SCFN design replaces the FFN in RepViT, further improving the model’s sensitivity and representation power for local features while retaining global context. Experimental results show that SRViT-MCNet achieves classification accuracies of 99.54% and 99.32% on the BIG2015 and Malimg datasets, respectively.

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

SRViT-MCNet: An IoT Malware Classification Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Changguang Wang,
  • Hongxuan Wang,
  • Xi Zhang,
  • Qingru Li,
  • Fangwei Wang

摘要

In the Internet of Things (IoT) domain, the heterogeneity of devices and the widespread deployment of resources significantly impact traditional malware classification methods. To effectively extract both local and global features of IoT malware and improve classification accuracy, a hybrid model, named SRViT-MCNet, is proposed. The model first visualizes IoT malware by applying the Occlusion Sensitivity method to capture regions where the model’s prediction confidence significantly drops. Topological Data Analysis (TDA) is then employed to generate Persistent Images (PI). The model utilizes a parallel feature extraction strategy, combining the SRViT module with the MCNet module to capture multi-dimensional features. The MC module integrated into the CNN enhances the model’s ability to capture fine-grained local features. Additionally, the SCFN design replaces the FFN in RepViT, further improving the model’s sensitivity and representation power for local features while retaining global context. Experimental results show that SRViT-MCNet achieves classification accuracies of 99.54% and 99.32% on the BIG2015 and Malimg datasets, respectively.