基于AGNN-GBDT的链上欺诈账户检测模型  

AGNN-GBDT-based model for on-chain fraudulent account detection

作　　者：冮君泽 李海明 Gang Junze;Li Haiming(College of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 200090,China)

机构地区：[1]上海电力大学计算机科学与技术学院,上海200090

出　　处：《国外电子测量技术》2023年第8期102-110,共9页Foreign Electronic Measurement Technology

摘　　要：针对虚拟数字货币的市场逐渐升温,大量非法交易和行为难以追踪溯源的问题,提出了基于BGNN的链上欺诈账户检测模型——AGNN-GBDT。通过分析真实账户交易数据和以太坊官方提供的欺诈账户数据的特点,使用GReaT进行数据增强,并在GNN网络中设计了基于节点通道和语义通道的双通道注意力机制来学习节点自身和图网络结构的特征信息,同时保留GBDT处理异质特征数据优势,引入SHAP值来判断特征的重要性。实验结果表明,该模型在准确率上达到84.2%,F1-score为84.2%,其实验效率和结果相较于以前学者提出的模型方法都有一定程度的提升,能够较为准确地识别链上的欺诈账户,对于改善区块链的交易环境有积极作用。In response to the increasing market for virtual digital currencies,there is a problem with a large number of illegal transactions and behaviours that are difficult to trace and track.Therefore,a fraud account detection model on the chain based on BGNN called AGNN-GBDT is proposed.By analysing the characteristics of real account transaction data and fraud account data provided by Ethereum,data enhancement is done using GReaT.A dual-channel attention mechanism based on the node channel and the semantic channel is designed in the GNN network to learn the feature information of the nodes themselves and the graph network structure while retaining the advantage of GBDT in processing heterogeneous feature data.The SHAP value is introduced to judge the importance of features.The experimental results show that the model achieves an accuracy rate of 84.2%and an F1-score of 84.2%;its experimental efficiency and results have been improved to a certain extent compared with the model methods proposed by previous scholars;and it can accurately identify fraud accounts on the chain,which can promote the improvement of the blockchain transaction environment.

关 键 词：欺诈账户检测 以太坊 GREAT 双通道注意力机制 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]