Research on Ethereum Phishing Scam Detection Technology Based on Multi-transaction-view Graph Neural Networks

Abstract

As the cryptocurrency market continues to evolve, phishing scams are considered one of the most deceptive forms of fraud. Currently, most existing Ethereum phishing detection methods rely on traditional machine learning or graph representation learning, mainly depending on local statistical and structural features. This can lead to insufficient utilization of transaction graph data across different scales. To address this challenge, we propose Multi-transaction-view Graph Attention Network (MTvGAT), which fully leverages edge features between nodes at different scales and discovers relationships between nodes. Two types of graphs are used to model Ethereum transactions: global views and local views. Global views are constructed by partitioning the complete transaction graph using graph clustering algorithms and inputting them into MTvGAT to obtain global view representations. For each target node, a local view is constructed by sampling K-hop neighbors from the transaction network. Importantly, attention coefficients are calculated between nodes, and edge coefficients are obtained by fusing edge features and attention coefficients, utilizing spatial structure and edge coefficients to enable the phishing detection model to access multi-view sources of information. Experimental results demonstrate that the multi-view graph attention network outperforms existing algorithms in detecting Ethereum phishing scams datasets.