Temporal Behavioural Modelling for Malicious Entity Detection in Ethereum Blockchain

Abstract

The quick proliferation of decentralized finance and smart platforms on Ethereum blockchain has greatly exposed people to fraudulent practices like scams, phishing accounts, Ponzi schemes, and transaction mixers. Such harmful entities are dynamic in their behaviors and have large volumes of transactions making it hard to detect them early and accurately. The given study introduces a time-based behavioral modeling framework, which uses machine learning algorithms to identify malicious Ethereum accounts through the analysis of transactions based on time intervals. It is a hybrid of behavioral attributes at a particular point of time and interval based transactional attributes that are pulled over several time windows, 1 month, 3 months, 6 months, 1 year, 2 years, and 3 years that allow a detailed temporal profiling of account activity. Preprocessing of data entails the removal of non-informative identifiers, the encoding of labels in binary and multiclass classification, scaling of features through standardization and stratified train test separation to ensure that the classes are balanced. Three models are used to compare the traditional and deep learning methods, namely Support Vector Machine (SVM), 1D Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM). The experimental findings indicate that predictive performance is higher with longer observation windows and SVM has a maximum accuracy of 0.84 in binary classification and CNN has a maximum accuracy of 0.74 in multiclass classification. The findings prove the importance of incorporating a temporal transaction pattern to improve the strength of detection and proactive detection of the ill-intentional blockchain entities.