Blockchain Fraud Detection Using Ensemble Graph Neural Networks

Abstract

Cryptocurrencies like Bitcoin promise secure, decentralized transactions, but their anonymity also attracts illicit activity, posing a challenge to regulators and exchanges in maintaining control. This study tackles fraud detection in Bitcoin's transaction network using the Elliptic dataset, a real-world collection of labeled transactions. We combine three powerful graph neural networks Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Graph Isomorphism Network (GIN) each capturing different patterns in the complex web of blockchain payments. By blending their predictions through ensemble techniques, such as tuned soft voting, we achieve a robust system that detects over 70% of illicit transactions while keeping false alarms below 1%. Our approach balances precision and coverage, making it practical for real-time anti-money laundering efforts. The modular framework adapts easily to new data, paving the way for scalable, reliable monitoring of cryptocurrency fraud.