AUTOENCODER-BASED RISK ASSESSMENT FOR SUSPICIOUS ACTIVITY DETECTION IN FINANCIAL SYSTEMS

Abstract

The rapid digitization of financial services has led to a significant increase in transaction volumes, making traditional rule-based fraud detection systems inadequate in identifying complex and evolving suspicious activities. This study presents an autoencoder-based risk assessment framework for detecting anomalous and potentially fraudulent financial transactions. The proposed approach leverages unsupervised deep learning to learn normal transaction behavior from high-dimensional financial data and identify deviations using reconstruction error. A risk scoring mechanism is integrated with the autoencoder output to prioritize suspicious activities based on their severity and potential financial impact. The model is evaluated on large-scale transactional datasets using standard performance metrics such as precision, recall, F1-score, and area under the ROC curve. Experimental results demonstrate that the proposed framework effectively detects subtle and previously unseen fraud patterns while reducing false positives compared to conventional methods. The findings highlight the suitability of autoencoder-driven risk assessment systems for real-time deployment in modern financial environments, offering improved adaptability, scalability, and decision support for financial institutions.