It’s All Connected: Detecting Phishing Transaction Records on Ethereum Using Link Prediction

Abstract

Smart contracts are increasingly being used on platforms for virtual transactions, such as Ethereum, owing to new financial innovations. As these platforms are anonymous and easy to use, they are perfect places for phishing scams to grow. Unlike traditional phishing detection approaches that aim to distinguish phishing websites and emails using their HTML content and URL, phishing attacks on Ethereum focus on detecting phishing addresses by analysing the transaction relationships on the virtual transaction platform. This study proposes a link prediction framework for detecting phishing transactions on the Ethereum platform using 12 local network-based features extracted from the ether receiving (target) and initiating (source) addresses. The framework was trained and tested on over 280,000 verified phishing and legitimate transaction records. Experimental results indicate that the proposed framework with a LightGBM classifier provides a high recall of 89% and an AUC score of 93%.