AI-Assisted Financial Fraud Detection: Exploration of Model Architecture in Performance and Accuracy

Abstract

Financial fraud involves the intentional use of deception to obtain money or assets and poses a significant threat to individuals, businesses, and governments worldwide. As financial systems increasingly rely on digital platforms, fraud has become more frequent and complex, resulting in hundreds of billions of dollars in global losses each year and undermining public trust in financial institutions. Traditional fraud prevention methods, such as manual audits and rule-based monitoring, are often costly, slow, and difficult to scale, making them less effective against evolving fraud strategies. To address these limitations, organizations have begun adopting automated, data-driven approaches that leverage analytics, anomaly detection, and machine learning to identify suspicious activity in real time. This project investigates how automated fraud detection techniques can enhance accuracy and efficiency while reducing dependence on manual review processes. To facilitate automated and scalable fraud detection, several machine learning (ML)/artificial intelligence (AI) based approaches were leveraged for fraud classification based on a synthetic dataset. Among the tested models, Random Forest achieved the strongest overall performance (precision 0.74, recall 0.93, F1-score 0.82), followed closely by Logistic Regression (F1-score 0.80), while Decision Tree performed moderately (F1-score 0.79) and both Neural Network and SVM showed lower F1-scores (0.69), reflecting challenges with overfitting and scalability. Feature weight analysis in the Logistic Regression model indicated that transaction metadata variables, particularly payment channel and device-related attributes, contributed most strongly to fraud classification. These findings suggest that ensemble and well-regularized linear models provide reliable and scalable solutions for financial fraud detection in large transactional datasets, with meaningful implications for improving real-world forensic auditing and automated compliance monitoring.