Detecting Suspicious Transactions in Blockchain Networks Using Temporal Graph AI

Abstract

Bitcoin and Ethereum record every transaction on a public ledger, yet criminals conceal illicit activity by routing funds through long chains of intermediate addresses. Simple rule-based tools miss these patterns and produce high volumes of false alerts. This paper presents an explainable AI framework that models blockchain transactions as a time-evolving network and uses this structure to prioritise suspicious activity for human analyst review. Two complementary models are proposed and evaluated against published benchmarks: XGBoost, a fast and interpretable classifier, and a Temporal Graph Attention Network (TGAT) that captures the shape and timing of suspicious flows. Results drawn from the Elliptic Bitcoin benchmark show that XGBoost (F1 = 0.85, ROC-AUC = 0.97) substantially outperforms a static Graph Convolutional Network (F1 = 0.63) while remaining fully explainable via SHAP values. Temporal graph models project further improvement (TGAT F1 = 0.75, TGN F1 = 0.77), consistent with temporal attention gains reported in Ethereum phishing detection studies. Both models generate explanations alongside risk scores so that analysts understand each alert. The framework, evaluation design, and explainability pipeline are described in full.