Forecasting Anomaly score for Financial Fraud Detection using Convolution Neural Networks

Abstract

Financial statement fraud has become a critical concern, causing substantial economic losses to a wide range of stakeholders, including investors, governments, and financial institutions. The growing complexity and digitization of global financial systems have opened new avenues for fraudulent activities, challenging the effectiveness of traditional detection methods. To address this issue, the study proposes a Financial Fraud Detection Model (FFDM) leveraging Graph Neural Networks (GNNs) and Convolutional Neural Networks (CNNs). The model integrates both network-based and feature-based analysis to uncover hidden patterns and suspicious behaviours within dynamic transaction networks. Emphasis is placed on predicting anomaly scores using CNNs to proactively identify potential fraud before it occurs. The methodology incorporates mobile money platforms, credit card fraud, and fraudulent phone call detection as case studies. Experimental evaluation using various machine learning models demonstrates that CNNs outperform traditional methods like Decision Trees and Logistic Regression, achieving a significantly lower Mean Squared Error (MSE) in probability score prediction. This research highlights the potential of advanced AI techniques in enhancing the accuracy and speed of financial fraud detection systems.