Learning Based Ethereum Phishing Detection: Evaluation, Robustness, and Improvement

Abstract

Phishing attacks continue to pose a significant threat to the Ethereum ecosystem, accounting for a major share of Ethereum-related cybercrimes. To enhance the detection of such fraudulent transactions, this dissertation develops a comprehensive framework for machine learning-based phishing detection in Ethereum transactions. The framework addresses critical aspects such as feature selection, class imbalance, model robustness, and the vulnerability of detection models to adversarial attacks. By systematically evaluating these key practices, this work contributes to the development of more effective detection methods. The first part of the dissertation assesses the current state of phishing detection methods, identifying gaps in feature selection, dataset composition, and model optimization. We propose a systematic framework that evaluates these factors, providing a foundation for improving the overall performance and reliability of detection models. The second part explores the vulnerability of machine learning models, including Random Forest, Decision Tree, and K-Nearest Neighbors, to single-feature adversarial attacks. Through extensive experimentation, we analyze the impact of various adversarial strategies on model performance and uncover alarming weaknesses in existing models. However, the varied effects of these attacks across different algorithms present opportunities for mitigation through adversarial training and improved feature selection. Finally, the dissertation investigates how phishing detection models generalize across datasets, focusing on the role of preprocessing techniques such as feature engineering and class balancing. Our findings show that optimizing these techniques enhances model accuracy and robustness, making detection methods more adaptable to evolving threats. Overall, this work presents a comprehensive framework that addresses the critical elements of phishing detection in Ethereum transactions, offering valuable insights for the development of more robust and generalizable machine learning-based security models. The proposed framework has broad implications for improving blockchain security and advancing the field of phishing detection.