A Scalable Ensemble ML Framwork for Phishing Detection in Ethereum

Abstract

Blockchain enables secure transactions through asymmetric encryption and decentralized storage, making cryptocurrencies such as Ethereum increasingly popular and attracting greater attention to decentralized finance (DeFi). This increased attention has also made Ethereum a target for phishing attacks, which account for 26% of cryptocurrency-related fraud. These attacks pose a significant threat because blockchain transactions are irreversible. Many fraud detection methods rely on features manually designed by experts. This can make them less adaptable to changing fraud patterns and difficult to scale to large and evolving datasets. This study proposes an automated feature extraction and ensemble learning framework that utilizes seven machine learning models: Decision Tree, Random Forest, Extra Trees, Gradient Boosting, XGBoost, LightGBM, and AdaBoost. These models are integrated using a soft-voting classifier for phishing detection on the Ethereum network, leveraging its transparent transaction history. The proposed framework achieves an accuracy of 96.2%. The aim of this research is to develop a phishing detection classification system with dynamic scalability.