Research On Ethereum Phishing Node Detection Technology Based On One-class Graph Attention Network

As a representative of blockchain 2.0,Ethereum has attracted attention due to its smart contract functionality.The Ethereum transaction network is a decentralized,freely circulating,and tamper-evident trading platform built on a peer-to-peer network.However,with the continuous increase in transaction volume,attacks on the transaction network have become increasingly common,and network security issues are becoming more and more severe.Among them,phishing scams account for more than half of all network crimes,causing users significant losses.In Ethereum phishing detection tasks,traditional detection methods based on manual features only consider Ethereum’s transaction history,with simple feature extraction and a lack of consideration of Ethereum transaction network characteristics.As a result,the detection results have certain limitations.Phishing detection methods based on random walks focus more on the local characteristics of target nodes,with limited feature selection and limited application space.Furthermore,most current methods are two-stage,which decouples upstream and downstream and cannot learn features related to the task.To address these problems,this thesis further explores phishing fraud detection methods and proposes corresponding detection schemes based on Ethereum transaction network characteristics.The main work carried out in this thesis is as follows:(1)A binary model for Ethereum phishing detection based on random walk with restart is proposed.The core idea of this model is to construct a large-scale Ethereum transaction graph,dividing accounts into phishing nodes and normal nodes,and representing transaction relationships between accounts through edges.To obtain the embedding features of nodes,the random walk with restart algorithm is employed in this thesis.Additionally,considering the characteristics of the Ethereum transaction network,a biased random walk algorithm based on transaction amount and transaction time is designed.This algorithm takes into account the weighting factors of transaction amount and transaction time,making the walking process more in line with real transaction behavior.(2)Furthermore,an end-to-end detection framework based on graph neural networks is proposed.This framework converts the classification task into an anomaly detection task based on graph attention networks and classical hypersphere learning,addressing the issue of data imbalance and the drawbacks of two-stage detection.At the same time,to fully utilize the high-order information of graph attention networks,according to Ethereum’s transaction network characteristics,this thesis injects the structure information of the transaction network into the attention layer during the training process,improving the calculation of traditional attention coefficients to fully explore the superiority of the attention network and improve the detection effectiveness of the model.