Blockchain Based E-Invoicing Platform for Global Trade: An Explainable Hybrid CatBoost–Isolation Forest Framework for Security Compliance and Fraud-Resilient Invoice Processing

Abstract

While electronic invoicing has helped streamline global trade, existing blockchain-based platforms mostly have focused on immutability, traceability and auditing of invoice disputes, while failing to automatically detect fraudulent invoice behaviour. The objective of this study is to create a methodology that is explainable, predictive and also anomaly detection that can complement the blockchain-based e-invoicing. A new Explainable Hybrid Invoice Fraud Detection framework (EHIFD) is proposed, which is based on the Isolation Forest algorithm and a CatBoost classifier tuned by Optuna. The invoice records are cleaned and mixed numerical and categorical columns are processed and possible target-leakage variables are eliminated and the data is split into stratified training, validation and testing sets. CatBoost becomes class-balanced and is trained to recognize known fraud patterns, its hyperparameters are optimized by maximizing validation PR-AUC with the Tree-structured Parzen Estimator algorithm. Isolation Forest is typically trained on legitimate invoices, and then used to detect potentially suspicious or out-of-sight transactions. The normalized anomaly score is fused with the CatBoost fraud probability in a weighted way (80% CatBoost probability, 20% Isolation Forest anomaly score). A threshold for fraud decisions is then chosen from the validation set to maximize the F1 score. SHAP explanations are able to determine the attributes that lead to a prediction, and invoice hashes, model versions, risk scores and decisions can be stored on the blockchain without revealing the private details of the invoice. The framework was tested with 25 000 synthetic invoice records with 5 000 test records having a fraud prevalence of 4.02%. EHIFD achieved 93.84% accuracy, 27.80% precision, 33.33% recall, a 30.32% F1-score, 80.44% ROC-AUC, and 23.55% PR-AUC. The hybrid algorithm outperformed tuned CatBoost by 5.97 percentage points in terms of recall and 3.22 percentage points in terms of F1-score. But the highest F1-Score was obtained by the baseline CatBoost with 33.90% and PR-AUC was 24.33%. These results demonstrate that hybrid anomaly integration boosts the sensitivity of the tuned classifier, but cross-validated optimization and evaluation on real-world invoice data must be done before deployment.