A CNN-BASED FRAMEWORK FOR PAYROLL FRAUD DETECTION IN SIMULATED NIGERIAN PUBLIC FINANCIAL DATA

Abstract

 This study develops an artificial intelligence–driven framework for detecting payroll fraud in Nigeria’s Public Financial Management (PFM) system using Convolutional Neural Networks (CNNs). The model targets common fraud patterns such as ghost workers, salary inflation, duplicate payments, and payments to retired or overage employees. A dual-stage validation strategy was adopted. First, a publicly available IBM Human Resources dataset containing 1,470 records was used to benchmark the model, achieving an F1-score of 0.87. Second, a realistically simulated Nigerian payroll dataset comprising 12,110 records with 31 features was constructed to reflect local administrative structures and fraud typologies. To address severe class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied during model training. The proposed model demonstrated strong generalization performance on the 
simulated dataset, achieving an accuracy of 0.958, precision of 0.910, recall of 0.841, F1-score of 0.874, and PR-AUC of 0.857. Furthermore, the trained model was integrated into a real-time fraud alert dashboard to support proactive monitoring and decision-making. The study demonstrates the potential of deep learning approaches for fraud detection using simulated data.