Detecting Rug Pull Risks in Cryptocurrency Projects on the Binance Smart Chain Using Machine Learning

Abstract

Cryptocurrency investments have grown exponentially, but the rapid expansion of decentralized finance (DeFi) ecosystems has been accompanied by the rise of sophisticated fraud schemes, particularly Rug Pulls. These scams occur when developers deliberately withdraw liquidity or sell large amounts of tokens, leaving investors with worthless assets. This research presents a machine learning-based framework for detecting rug-pull-prone projectson the Binance Smart Chain (BSC). A comprehensive dataset was constructed by aggregating transactional and smart contract features from reliable sources such as BscScan, TokenSniffer, DEXTools, and PeckShield Alerts. Data preprocessing included handling missing values, removing duplicates, detecting and mitigating outliers, and addressing severe class imbalance using Synthetic Minority Oversampling Technique (SMOTE). Seven machine learning algorithms were compared: Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGB), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). The top-performing models, Random Forest and XGBoost, were further validated using stratified holdout testing. Results demonstrate that XGBoost achieved the highest overall performance$(\mathrm{F1} = 0.82,\ \text{ROC-AUC} = 0.90,\ \text{PR-AUC} = 0.994)$confirming the model's robustness in identifying fraudulent patterns. This approach offers a scalable framework for blockchain fraud detection on BSC, with potential applicability to other networks such as Ethereum and Polygon.