A scalable and explainable framework for detecting Ponzi schemes in Ethereum smart contracts using a stacking model

Abstract

Blockchain technology has reshaped digital finance, enabling decentralized applications (DApps) on platforms like Ethereum. However, these innovations have also facilitated fraudulent schemes such as Ponzi schemes, which deceive users with false promises of high returns. These schemes cause financial losses and weaken trust in blockchain systems. Existing detection methods face key challenges, including limited labeled data, over-reliance on transaction history, and failure to identify scams early. To address these issues, we propose a framework that combines static and dynamic features of smart contracts for early Ponzi detection. Our feature set includes opcode patterns, developer behavior, temporal trends, and metadata, crafted to work independently of transaction data. We enhance feature representation using TF-IDF, CountVectorizer, and Word2Vec for deeper semantic understanding. These features are used to train multiple machine learning and deep learning models such as Random Forest, XGBoost, CNNs, and BiGRUs. A stacking ensemble with a neural meta-learner integrates predictions for improved performance. The model achieves 99% accuracy and an AUC of 0.9522 on a curated Ethereum dataset, handling class imbalance through oversampling and synthetic data generation. We also employ SHAP for model explainability, offering insights into feature importance and promoting transparency. Our framework is scalable and supports real-time monitoring of contracts, helping prevent financial damage by detecting fraud at deployment. This solution enhances the security and reliability of decentralized finance platforms.