K-Means Clustering for Identifying Suspicious Patterns in Ethereum Networks (International Journal of Industrial Engineering)

Summary

• The experimental results show that the proposed K-Means clustering model achieved a classification accuracy of 98.07% and an F1-score of 98.08%.p.1
• Ethereum daily block creation exceeds 6,000 blocks, with total accounts surpassing 207 million.p.1
• Analysis of an Ethereum dataset with 9840 transactions revealed that the XGBoost classifier achieved an accuracy of 99.65%.p.3
• The study examined 184 Ponzi scheme addresses from an open-source dataset, successfully recovering bytecode for 172 of them.p.4
• Evaluation of the confusion matrix showed 1930 out of 1968 total transactions were correctly classified, including 970 legitimate and 960 fraudulent transactions.p.6
• Testing accuracy of the model increased from 88.75% to 98.07%, while testing loss decreased from 0.371 to 0.052.p.7
• The average period between transactions involving Ponzi smart contracts was measured at 526K days, compared to 79K days for non-Ponzi smart contracts.p.8

Key insights

1. The K-Means clustering model differentiates fraudulent transactions from genuine activities with a 98.07% accuracy using an 80:20 train-test split.p.1
2. Machine learning models trained on 30 million Bitcoin transactions achieved a recall and accuracy of 0.987.p.1
3. XGBoost outperformed Decision Tree and Random Forest classifiers on Ethereum fraud detection with a 99.65% accuracy rate.p.3
4. Ether-scan source code validation was utilized to collect 3,203 non-Ponzi smart contracts engaged in at least 10 transactions.p.4
5. Code-based features provide immediate 0-day categorization capabilities upon uploading a smart contract to the blockchain.p.5
6. Cluster evaluation using the Silhouette coefficient identified k = 3 as the optimal number of clusters for transaction partitioning.p.7
7. Ponzi smart contracts exhibit a significantly lower first quartile transaction interval under 6 minutes compared to non-Ponzi contracts at 70 minutes.p.8
8. Initial centroid selection and class imbalance present key operational limitations for K-Means clustering in cryptocurrency fraud detection.p.9