EXPLAINABLE AI FOR DEFI FRAUD DETECTION: A COMPARATIVE STUDY WITH LARGE-SCALE TRANSACTION DATA

Abstract

Fraud in blockchain-based financial applications is becoming more and more sophisticated. This can negatively impact transaction trust and security. It also has implications for the global uptake of blockchain financial services. This research proposes a comparative machine learning model for fraud detection. It leverages the BCCC-DeFiFraudTrans-2025 dataset of 177,586 balanced Ethereum transactions. The transactions are represented by 78 predictive features. We compare the performance of five classification models, trained on a stratified 80/20 train test split. These models include Logistic Regression, Random Forest, XGBoost, LightGBM and CatBoost. LightGBM exhibits the best overall performance in terms of all the evaluation metrics. It delivers accuracy, precision, recall and F1 scores greater than 99.9%. Explainability is evaluated using SHAP values of the XGBoost model. It reveals the most important features are those related to transaction value. This finding supports robust model performance while providing insights into model predictions. The results highlight the need for explainable artificial intelligence (XAI) in financial fraud detection. In conclusion, the use of ensemble learning models is successful in studying complex high-dimensional DeFi data. These can enhance trust, reliability and security in practical blockchain financial applications.