Fraudulent Transaction Detection and Prevention Using Deep Neural Networks

The worldwide explosion of information technology has significantly streamlined financial transactions via electronic devices like computers, mobile phones, and tablets, all made possible by the ubiquity of the Internet. While this has provided significant convenience in settling financial transactions, it has also led to a severe threat, as evidenced by the increasing incidents of fraudulent transactions. These fraudulent activities result in financial losses for both financial institutions and customers. Card fraud incidents increased by 14.34% to $566 million in 2022, compared to $495 million in 2021 in Australia. To combat these fraudulent transactions, researchers have been actively exploring various methods, including but not limited to statistical, machine learning, and deep learning techniques. Given the effectiveness of deep learning in various artificial intelligence applications, this has demonstrated better performance than traditional machine learning in detecting fraudulent transactions. 

A multi-pronged approach was adopted for the conduct of this study. This project explores deep learning techniques for identifying and preventing fraudulent transactions. Deep learning is a machine learning approach utilising artificial neural networks with multiple layers to automatically extract hierarchical representations from data, enabling complex pattern recognition. This research involves generating features from financial transactions to train the deep learning models. Additionally, it emphasises the optimisation of the deep learning model and proposes an efficient approach for classifying fraudulent transactions. After reviewing the existing literature, it was found that fraud detection faces several challenges. Firstly, the data is imbalanced, meaning fraudulent transactions are considerably less frequent than non-fraudulent transactions. Secondly, fraudulent data is scarce due to organisations' reluctance to disclose such data for privacy and reputational concerns. We have also explored several classical and deep machine learning to detect fraudulent transactions, such as Logistic Regression (LR), Conventional Neural Networks, K-Nearest Neighbour (KNN), Decision Trees, Random Forests, Convolutional Neural Networks (CNN), Autoencoders, Recurrent Neural Networks (RNN), Long Short Term Memory (LSTM), and Deep Belief Network (DBN), among others. Although these methods perform well in detecting fraud, there is room for improvement by employing techniques such as hyperparameter tuning, feature engineering, and mitigating overfitting issues. 

Two novel methods were thus proposed for improving the efficiency of deep learning models for fraud detection. Firstly, a feature engineering technique called Deep Feature Synthesis (DFS) was deployed with the CNN model to develop a new model. This feature engineering technique automatically generates features from existing features by applying mathematical functions based on the relationship among the datasets. Following this, an optimised CNN model was developed by fine-tuning the hyperparameters. Two algorithms, random search and grid search were employed for hyperparameters optimisation. A step-by-step process was followed in tuning the hyperparameters, where each hyperparameter was adjusted within the defined search space, tested with the test data, and subsequently updated the model with the tuned hyperparameters. This process continued until the optimised hyperparameters were identified. A research methodology was designed to validate and test the efficacy of the proposed improvements, including data collection, pre-processing, data splitting, model training, optimisation, and model testing and evaluation. An analysis of DFS's impact on CNN's performance regarding the accuracy, precision, recall, F1 score, and AUC was conducted. The results demonstrated a significant improvement, with performance increasing by 11% in all four indices after incorporating DFS in the pre-processing step of the CNN deep learning model training compared to the standalone CNN model. On the other hand, the CNN model optimised through random search showed a performance enhancement of 7% over the baseline and a 10% improvement compared to the existing model in the current literature, and the optimised CNN using the grid search Algorithm exhibited a 5.5% enhancement over the baseline and an 8.9% improvement over the existing model. 

Overall, this study focuses on detecting fraudulent transactions using deep learning techniques, specifically Convolutional Neural Network (CNN) models. A systematic literature review is conducted to address research questions, exploring six experimental steps for fraudulent transaction detection using deep learning. The integration of DFS improves fraud detection performance, resulting in a notable enhancement in recall, F1 score, and accuracy. Hyperparameter tuning, utilising grid search and random search algorithms yield remarkable improvements across precision, recall, and F1 score for the optimised CNN models compared to the baseline. These optimised CNN models surpass the baseline and outperform a state-of-the-art method.