Scholarlink

The study addresses the challenge of detecting obfuscated malware, particularly memory-obfuscated variants, which evade conventional detection methods by targeting a system’s volatile memory. By leveraging transformer-based models, notably BERT, the research demonstrates promising advancements in malware detection. Through data augmentation and rigorous feature selection processes, the study enhances the CIC-MlMem-2022 dataset, improving its quality for training classification models. Comparative analysis with conventional machine learning techniques highlights the superior performance of BERT and DistilBERT, achieving approximately 74% accuracy in classifying malware families. Notably, BERT exhibits exceptional capability in generalizing to unseen malware, achieving a remarkable 100% success rate in categorizing new families. These findings underscore the potential of transformer-based models for effectively detecting obfuscated malware, emphasizing the need for intelligent detection mechanisms in cybersecurity.

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

Unveiling the Efficacy of BERT’s Attention in Memory Obfuscated Malware Detection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Md Mashrur Arifin,
  • Troy Suyehara Tolman,
  • Jyh-haw Yeh

摘要

The study addresses the challenge of detecting obfuscated malware, particularly memory-obfuscated variants, which evade conventional detection methods by targeting a system’s volatile memory. By leveraging transformer-based models, notably BERT, the research demonstrates promising advancements in malware detection. Through data augmentation and rigorous feature selection processes, the study enhances the CIC-MlMem-2022 dataset, improving its quality for training classification models. Comparative analysis with conventional machine learning techniques highlights the superior performance of BERT and DistilBERT, achieving approximately 74% accuracy in classifying malware families. Notably, BERT exhibits exceptional capability in generalizing to unseen malware, achieving a remarkable 100% success rate in categorizing new families. These findings underscore the potential of transformer-based models for effectively detecting obfuscated malware, emphasizing the need for intelligent detection mechanisms in cybersecurity.