Detecting Fraudulent Financial Transactions using Deep Learning and Transaction Log Analysis

Abstract

The financial sector has seen a significant rise in fraudulent activities, costing businesses billions of naira annually. Fraudulent transactions often involve unauthorized account access, manipulation of transaction data, or theft, leaving businesses and consumers vulnerable. This study aims to enhance financial fraud detection using a deep learning approach combining Bidirectional Long Short-Term Memory (BiLSTM) and Multilayer Perceptron (MLP). The goal is to develop a more efficient model, especially for analyzing large datasets. Transaction log data from Kaggle, containing diverse fraudulent and non-fraudulent transaction records worldwide, was used for this experiment. Data preprocessing involved imputation to fill empty entries. The architectures of BiLSTM, MLP, and the hybrid BiLSTM-MLP were designed for smooth model integration and data pipeline processes. After preprocessing, the data was input into the BiLSTM model, which reads data in forward and backward directions. The output was passed to the MLP model layers for predicting the authenticity of transactions. The combined model was evaluated using accuracy, precision, recall, F1 score, AUC, ROC, and confusion matrix. The standalone BiLSTM model achieved 89% accuracy, while the MLP model scored 92%. The hybrid BiLSTM-MLP also scored 92%, similar to the MLP model. Upon critical examination, the combined model demonstrated better outcome in making accurate predictions compared to standalone models. Keywords: Multiple Layer perceptron; Bidirectional Long-short Term Memory; Deep Learning; Machine Learning; Fraudulent