Scholarlink

<p>In this paper, we introduce the GRT (Graphical Representation of Tabular data) framework that transforms individual tabular sample into relational intelligence graph. Our model can learn relationships between data instances by using the cosine distance to capture both individual sample features similarities and the overall relational patterns within the dataset. We use Graph Convolutional Networks (GCNs) to detect ransomware and leverage their ability to capture high-order relationships to identify and classify it. Our method was tested on three various and current datasets: Ransomware Behavioral Analysis (RBA 2024), Ransomware PE Header Feature Dataset (RPHF 2023) and Windows PE Malware Detection (WPD). The findings indicate that the proposed GCN achieves competitive detection performance relative to state-of-the-art ML and DL baselines across all three datasets, with accuracy reaching up to 99%. The two layer GCN is more resistant to obfuscation than the traditional single-instance analyses since it learns higher-order patterns of relationships. A comparison against GAT and GraphSAGE on the same GRT graph, under both transductive and inductive evaluation protocols, confirms that detection performance is robust to the choice of message-passing operator across all datasets, establishing similarity-based relational graph learning as a scalable and practical direction for next-generation ransomware detection.</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

Similarity-driven graph convolutional networks for ransomware detection

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Sana Aroussi,
  • Fatima Boumahdi,
  • Abdelmoumen Khaldi,
  • Mohamed Abdelkarim Remmide,
  • Narhimane Boustia,
  • Fatma Zohra Souaci,
  • Noor Amalou

摘要

In this paper, we introduce the GRT (Graphical Representation of Tabular data) framework that transforms individual tabular sample into relational intelligence graph. Our model can learn relationships between data instances by using the cosine distance to capture both individual sample features similarities and the overall relational patterns within the dataset. We use Graph Convolutional Networks (GCNs) to detect ransomware and leverage their ability to capture high-order relationships to identify and classify it. Our method was tested on three various and current datasets: Ransomware Behavioral Analysis (RBA 2024), Ransomware PE Header Feature Dataset (RPHF 2023) and Windows PE Malware Detection (WPD). The findings indicate that the proposed GCN achieves competitive detection performance relative to state-of-the-art ML and DL baselines across all three datasets, with accuracy reaching up to 99%. The two layer GCN is more resistant to obfuscation than the traditional single-instance analyses since it learns higher-order patterns of relationships. A comparison against GAT and GraphSAGE on the same GRT graph, under both transductive and inductive evaluation protocols, confirms that detection performance is robust to the choice of message-passing operator across all datasets, establishing similarity-based relational graph learning as a scalable and practical direction for next-generation ransomware detection.