DEEPRISK: A HYBRID AUTOENCODER AND RISK-BASED LEARNING MODEL FOR FINANCIAL FRAUD DETECTION

Abstract

Financial institutions face a growing challenge in detecting fraudulent transactions within large-scale, high-velocity financial data. Traditional rule-based systems and shallow machine learning models often struggle to identify complex, evolving fraud patterns, leading to high false positives and delayed responses. This paper proposes DeepRisk, a hybrid model that integrates Autoencoder-based anomaly detection with a risk-based learning approach to identify suspicious financial transactions in real time. The Autoencoder component captures intricate, non-linear patterns in transaction behavior to detect anomalies, while the risk-based layer assigns weighted scores considering transaction amount, frequency, and behavioral history. By combining unsupervised feature extraction with interpretable risk profiling, DeepRisk enhances accuracy, scalability, and adaptability. Experimental evaluations on benchmark financial datasets demonstrate a substantial improvement in detection precision, recall, and false alarm reduction compared to traditional machine learning methods. The proposed system contributes to more secure and intelligent financial ecosystems through explainable deep learning driven fraud detection.