BXAI-Audit: A Blockchain-Enabled Explainable AI Framework for Tamper-Proof Fraud Detection
TL;DR
A Blockchain-Ass Explainable AI Audit system employing gradient boosting models and explainability techniques along with the benefits of blockchain technology in the context of fraud detection systems demonstrates 100% accuracy with the Hyperledger and Ethereum blockchain platforms with low latency of 1-3 seconds.
Abstract
Financial fraud detection systems employing machine learning techniques are known to demonstrate strong predictive accuracy but are opaque and non-auditable in nature. This study proposes a Blockchain-Assisted Explainable AI Audit (BXAI-Audit) system employing gradient boosting models and explainability techniques along with the benefits of blockchain technology in the context of fraud detection systems. The IEEE-CIS dataset containing more than 590,000 transactions with a significant level of class imbalance between the classes (\~{}3.5% fraud) is used in the study. Stratified sampling is implemented to generate a large subset of the dataset containing around 288,000 transactions to assess the stability of the model's performance. From the experimental results obtained with the proposed system employing the LightGBM algorithm, the model demonstrates strong accuracy of 98.21%, precision of 92.87%, and recall of 80.12%, with an ROC-AUC value of 0.957 compared to the XGBoost model. SHAP values are obtained to determine the important contributing factors towards the prediction of fraud at the global and local levels. To ensure the integrity of the system, the output of the explainability model is hashed using SHA-256 and recorded on the blockchain platform. The proposed system demonstrates 100% accuracy with the Hyperledger and Ethereum blockchain platforms with low latency of 1-3 seconds.