| 俞山青,唐政,彭松涛.基于深度图信息增强的以太坊异常检测算法研究[J].高技术通讯(中文),2025,35(8):837~846 |
| 基于深度图信息增强的以太坊异常检测算法研究 |
| Research on Ethereum anomaly detection algorithm based on depth graph information enhancement |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 08. 003 |
| 中文关键词: 钓鱼检测； 子图增强； 对比学习； 小样本学习 |
| 英文关键词: phishing detection, subgraph augmentation, comparative learning, small-sample learning |
| 基金项目: |
| 作者 | 单位 | | 俞山青 | （浙江工业大学信息工程学院杭州 310023） | | 唐政 | | | 彭松涛 | |
|
| 摘要点击次数: 88 |
| 全文下载次数: 66 |
| 中文摘要: |
| 随着区块链技术的普及应用，以太坊已发展成为去中心化交易生态的核心基础设施。与此同时，钓鱼节点的存在导致异常交易行为频发，因此针对以太坊的异常检测问题变得尤为紧迫。但是，以太坊的庞大数据及正、异常样本比例的极不均衡，使得现有方法缺乏足够的可扩展性，检测成本高昂。针对此问题，本文提出了一个基于深度图信息增强策略的自监督对比学习框架（residual graph infomax contrastive learning，ResGI-CL）。首先，利用交易信息构建交易图网络，根据用户自身的资金能力与用户同邻居之间的互动能力提出节点邻居置信度（neighbor confidence，NC）策略，以获取增强子图。然后，对子图数据进行深度增强，生成图信息差异化的正向样本和负向样本。最后，模型引入了残差图神经网络来对比高正负数据差异以实现钓鱼节点检测。实验结果表明，本文的异常检测模型在小样本数据上比多种代表性方法的性能提升了7.4%，模型中提出的子图采样策略对其他方法有普遍的增强效果，同时该模型在均衡数据集上表现出稳定的检测性能，为钓鱼节点检测提供了新的研究思路和理论支持。 |
| 英文摘要: |
| As the adoption and application of blockchain technology have expanded, the Ethereum platform has emerged as the core infrastructure of decentralized transaction ecosystems. However, the presence of phishing nodes has led to an increase in anomalous transaction behaviors, thus highlighting the urgency of addressing Ethereum’s anomaly detection issues. Nonetheless, the substantial volume of Ethereum’s data, coupled with the highly imbalanced ratio of normal to anomalous samples, has rendered existing methods lacking in scalability and incurring high detection costs. To tackle this challenge, this paper proposes a self-supervised contrastive learning framework based on a deep graph information augmentation strategy, termed ResGI-CL. The framework starts by constructing a transaction graph network using transaction information and devises a node confidence (NC) strategy based on users’ financial capacities and interactions with neighboring nodes, thus obtaining augmented subgraphs. Subsequently, deep augmentation is applied to the subgraph data, generating positively and negatively differentiated samples of graph information. Finally, the model introduces a residual graph neural network to contrast high positive-negative data disparities for phishing detection. Experimental results demonstrate that the proposed anomaly detection model outperforms various representative methods by 7.4% on small-scale data. Additionally, the introduced subgraph sampling strategy generally enhances the performance over conventional methods, while the model exhibits stable detection performance on balanced datasets. This research offers novel insights and theoretical support for phishing node detection by pioneering a novel research direction and introducing ResGI-CL, a self-supervised framework with deep graph information augmentation. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
