BEaCON: Benchmarking Ensemble and Classical Models for Ethereum Fraud Detection

Abstract

The detection of Ethereum fraud is still one of the biggest problems due to the highly unbalanced nature of blockchain transactions and the sheer volume of today's cryptocurrency networks. In this paper, we compared six supervised learning models on the XBlock-ETH transaction graph with 2.97 million nodes. Using a scalable preprocessing pipeline, dimensionality reduction, and class balancing with the SMOTE technique, the training samples were extracted and processed to obtain graph-derived transaction features. The experimental results show that the ensemble-based methods always perform better than the linear based and single-tree-based methods, with SVM Nyströem achieving the highest fraud recall (0.944) and Gradient Boosting achieving the strongest ROC-AUC (0.983), indicating complementary strengths in detection sensitivity and overall discrimination. Additionally, the study shows that in the case of severe class imbalance, the overall accuracy can be misleading, as the Random Forest model can reach 0.991 accuracy but only identify 52.8% of fraudulent accounts. XGBoost has the best performance on both detection and computational costs among the evaluated models and can be a viable choice for large-scale fraud detection systems on the Ethereum network.