Scholarlink

<p>As blockchain frameworks gain widespread adoption, malicious activities like phishing on networks such as Ethereum have surged, severely compromising the transactional integrity of the entire ecosystem and threatening the security of emerging multimedia digital assets (e.g., NFTs and Metaverse components). Despite various efforts, current detection approaches—which primarily depend on snapshot-based graphs or rudimentary temporal models—fail to adequately track the continuous evolution of transaction networks and synergize spatio-temporal characteristics to expose sophisticated deceptive behaviors. To address these critical gaps, we present <b>DyG-HSTA</b> (<b>Dy</b>namic <b>G</b>raph <b>H</b>ybrid <b>S</b>patio-<b>T</b>emporal <b>A</b>ttention), a dynamic graph framework for recognizing fraudulent operations. In particular, our framework initially models Ethereum trading histories as continuous-time dynamic graphs, a structure that accurately reflects the asynchronous, event-based properties of blockchain operations. Subsequently, we deploy a hybrid spatio-temporal attention module designed to dynamically extract intricate network relationships through the adaptive calibration of spatial, temporal, and content-based features. Ultimately, utilizing multi-layer Transformer encoders, we pool neighborhood data to produce highly expressive node embeddings, which are then applied to the downstream task of predicting phishing transaction links. Comprehensive evaluations utilizing four distinct blockchain datasets alongside four standard public networks demonstrate that DyG-HSTA achieves superior performance over eight representative baselines across both inductive and transductive configurations. Our codebase is publicly accessible at <a href="https://github.com/cmx12138/DyG-HSTA">https://github.com/cmx12138/DyG-HSTA</a>.</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

DyG-HSTA: a dynamic graph hybrid spatio-temporal attention model for Ethereum phishing detection

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Chao Li,
  • Mingxu Chen,
  • Chunmei Jiang,
  • Runshuo Liu,
  • Qianyu Song,
  • Ge Song,
  • Qingtian Zeng

摘要

As blockchain frameworks gain widespread adoption, malicious activities like phishing on networks such as Ethereum have surged, severely compromising the transactional integrity of the entire ecosystem and threatening the security of emerging multimedia digital assets (e.g., NFTs and Metaverse components). Despite various efforts, current detection approaches—which primarily depend on snapshot-based graphs or rudimentary temporal models—fail to adequately track the continuous evolution of transaction networks and synergize spatio-temporal characteristics to expose sophisticated deceptive behaviors. To address these critical gaps, we present DyG-HSTA (Dynamic Graph Hybrid Spatio-Temporal Attention), a dynamic graph framework for recognizing fraudulent operations. In particular, our framework initially models Ethereum trading histories as continuous-time dynamic graphs, a structure that accurately reflects the asynchronous, event-based properties of blockchain operations. Subsequently, we deploy a hybrid spatio-temporal attention module designed to dynamically extract intricate network relationships through the adaptive calibration of spatial, temporal, and content-based features. Ultimately, utilizing multi-layer Transformer encoders, we pool neighborhood data to produce highly expressive node embeddings, which are then applied to the downstream task of predicting phishing transaction links. Comprehensive evaluations utilizing four distinct blockchain datasets alongside four standard public networks demonstrate that DyG-HSTA achieves superior performance over eight representative baselines across both inductive and transductive configurations. Our codebase is publicly accessible at https://github.com/cmx12138/DyG-HSTA.