ETHIAD: A novel explainable model for detecting illicit accounts on Ethereum

Area of Science:

• * Blockchain Technology
• * Cybersecurity
• * Machine Learning

Background:

• * Ethereum is a major platform for decentralized applications (Dapps), initial coin offerings (ICOs), and decentralized finance (DeFi).
• * The platform is increasingly targeted for illicit activities, including fraud, money laundering, and illegal fundraising.
• * Existing fraud detection models struggle with imbalanced datasets and lack interpretability.

Purpose of the Study:

• * To propose a novel, explainable model for detecting illicit accounts on the Ethereum network.
• * To address the challenges of class imbalance and model interpretability in fraud detection.

Main Methods:

• * Pre-processing the dataset using ADASYN oversampling and Lasso feature selection for effective transaction structure modeling.
• * Training the Ethereum Illicit Account Detection (ETHIAD) model using the XGBoost algorithm.
• * Employing the SHAP framework for multi-perspective analysis of key factors influencing illicit accounts.

Main Results:

• * ETHIAD achieved high performance metrics: 99.70% accuracy, 99.51% precision, 99.02% recall, 99.26% F1 score, and 99.45% AUC.
• * The model demonstrated superior performance compared to existing state-of-the-art (SOTA) models, outperforming them by 0.05%-1.1%.
• * The SHAP framework provided strong explainability, identifying key factors contributing to illicit account detection.

Conclusions:

• * ETHIAD offers a robust and explainable solution for identifying illicit activities on the Ethereum blockchain.
• * The model's high accuracy and interpretability enhance the security and trustworthiness of the Ethereum ecosystem.
• * The findings contribute to advancing machine learning applications in cybersecurity for blockchain platforms.