이더리움의 피싱노드 탐지 연구: GNN을 활용한

초록

Purpose This study aims to compare and analyze the performance of Graph Neural Networks (GNN) with traditional graph embedding techniques for scam detection in the Ethereum network. Specifically, we evaluate the effectiveness of LINE, Node2Vec, Metapath2Vec, and GNN across various network scales to provide comprehensive insights into their strengths and limitations. Design/methodology/approach Our methodology consists of constructing a large-scale Ethereum transaction network from 74,577 initial scam addresses, expanded using Breadth-First Search (BFS) to include 805,327 nodes and 17,104,978 edges. We evaluate the performance of different methods on three subgraph scales (30,000, 50,000, and 70,000 nodes) using metrics such as AUC and MAE. The implementation involves a two-stage approach: graph embedding/neural network processing followed by SVM classification. Findings The experimental results demonstrate that GNN consistently outperforms traditional embedding methods across all network scales, achieving the highest performance (AUC 0.96, MAE 0.01) in the largest network configuration. While Metapath2Vec approaches GNN's performance in large-scale networks, LINE and Node2Vec show relatively lower performance. The results validate the effectiveness of GNN's message-passing mechanism in capturing the structural characteristics of Ethereum transaction networks.

키워드