DEEP LEARNING AND MACHINE LEARNING BASED ETHEREUM FRAUD FRAMEWORK

Abstract

This thesis presents a comprehensive framework for detecting fraudulent Ethereum addresses using machine learning and deep learning techniques. Leveraging a rich set of behavioral and transactional features, the framework incorporates robust preprocessing (including log transformation and scaling), feature selection via Recursive Feature Elimination (RFE), and classification using an FT-Transformer model tailored for tabular data. To address the challenge of class imbalance inherent in fraud detection, a weighted binary cross-entropy loss is employed. The proposed system achieves high performance, with an accuracy of 97.09% , F1-score of 93.55%, precision of 91.76%, recall of 95.41%, and AUC of 0.9961. Even after reducing the feature space, the model maintains comparable accuracy, demonstrating its efficiency and generalizability. This work introduces a scalable and interpretable deep learning-based approach for Ethereum fraud detection, contributing to the growing field of blockchain security.