DEVELOPMENT OF HYBRID INTELLIGENT FAKE ALERT DETECTION SYSTEM

CHIAHA, CHIDERA STEPHEN (2026) DEVELOPMENT OF HYBRID INTELLIGENT FAKE ALERT DETECTION SYSTEM. Other thesis, Godfrey Okoye University, Enugu Nigeria.

Text GODFREY OKOYE UNIVERSITY (2).pdf Download (854kB)

Abstract

Electronic banking fake alert poses a severe and escalating threat to financial institutions globally, with Nigerian banks experiencing growing losses attributable to increasingly sophisticated fakeulent transactions that overwhelm traditional rule-based monitoring systems. This project develops hybrid intelligent fake alert detection system that employs a hybrid ensemble model combining Decision Tree, XGBoost, LightGBM, and Random Forest algorithms with SMOTE (Synthetic Minority Oversampling Technique) resampling to address the critical challenge of extreme class imbalance inherent in real-world financial transaction data. The system was trained and evaluated on three industry-standard benchmark datasets — the Kaggle Credit Card fake Detection dataset (284,807 transactions, 0.17% fake rate), the PaySim Synthetic African Mobile Money dataset (6,362,620 transactions, 0.13% fake rate), and the IEEE-CIS fake Detection dataset (590,540 transactions, 3.5% fake rate) — collectively forming a unified training corpus of 7,237,967 samples with 29,368 confirmed fake cases, split 70% for training and 30% for testing using stratified sampling. The hybrid model architecture uses a two-step resampling strategy of RandomUnderSampler followed by SMOTE to achieve class balance, and employs RobustScaler for feature normalization. Individual model outputs are combined through a weighted ensemble scheme with weights automatically computed from AUC-ROC performance, and the final decision threshold is optimized using the Precision-Recall curve to balance sensitivity and specificity. Experimental results demonstrate an accuracy of 99.7%, precision of 84.39%, recall of 30.57%, F1-score of 44.88%, and AUC-ROC of 98.24%, with the high AUC confirming strong discriminative capability despite the challenge of extreme imbalance at 0.41% fake 7 prevalence. The system is deployed as a full-stack web application comprising a React.js frontend and Flask REST API backend, featuring real-time single-transaction analysis, batch CSV processing, SHAP-based explainability, role-based access control, an in-app training pipeline with live progress monitoring, and comprehensive model evaluation dashboards. This research contributes a production-ready, explainable, and institutionally deployable fake detection platform relevant to Nigerian and African electronic banking contexts.

Item Type:	Thesis (Other)
Subjects:	Q Science > Q Science (General)
Divisions:	Faculty of Computing And Information Technology (FACIT)
Depositing User:	Cynthia Ugwuoti
Date Deposited:	24 Jul 2026 12:46
Last Modified:	24 Jul 2026 12:46
URI:	http://eprints.gouni.edu.ng/id/eprint/5968

Actions (login required)

View Item