A Hybrid CNN-LSTM Approach for Detecting Anomalous Bank Transactions: Enhancing Financial Fraud Detection Accuracy

Аннотация

Detecting fraudulent bank transactions is crucial for maintaining the integrity of financial institutions and preserving customer trust. This study introduces a hybrid Convolutional Neural Network– The American Journal of Management and Economics Innovations 116 https://www.theamericanjournals.com/index.php/tajmei Long Short-Term Memory (CNN-LSTM) model designed transactions, addressing existing limitations of to enhance the accuracy and efficiency of fraud traditional fraud detection techniques. detection systems. Utilizing the European Credit Card Literature Review Fraud Detection dataset comprising 284,807 The detection of fraudulent transactions has been transactions with significant class imbalance, extensive extensively studied, employing various machine preprocessing techniques were applied, including Min- learning and deep learning techniques. Logistic Max scaling and Synthetic Minority Over-sampling regression, decision trees, random forests, and Technique (SMOTE). Recursive Feature Elimination gradient boosting models such as XGBoost have (RFE) identified the top 20 impactful features, historically been popular for this task due to their ease optimizing model performance. The proposed hybrid of implementation and interpretability (Dal Pozzolo et model demonstrated remarkable effectiveness, al., 2014; Bhattacharyya et al., 2011). However, these achieving superior accuracy (99.5%), precision (93.1%), models often struggle with imbalanced data and may recall (92.1%), F1-score (92.6%), and an Area Under the fail to accurately detect intricate fraud patterns. Receiver Operating Characteristic Curve (AUC-ROC) of Convolutional Neural Networks (CNN) have shown 97.5%. Comparative analyses revealed that the hybrid promising results by effectively identifying spatial CNN-LSTM model significantly outperformed correlations within data (Jurgovsky et al., 2018). traditional machine learning algorithms such as Logistic Nevertheless, CNNs alone do not effectively capture Regression, Random Forest, and XGBoost. These the temporal sequences that characterize fraudulent findings underscore the potential of CNN-LSTM hybrid transaction patterns. To address this limitation, Long models in addressing complex fraud detection Short-Term Memory (LSTM) networks, known for their scenarios, providing financial institutions with a robust ability to capture long-term dependencies in sequential and reliable tool for transaction anomaly detection. data, have been employed with significant success in Keywords: Fraud Detection, CNN-LSTM Hybrid Model, various sequential data problems, including financial Credit Card Fraud, SMOTE, Feature Selection, Machine fraud detection (Roy et al., 2018). Learning, Financial Security Hybrid models integrating CNN and LSTM architectures Introduction: Financial institutions increasingly rely on have been introduced to leverage both spatial and digital transactions, significantly raising concerns sequential data processing capabilities. For instance, related to fraudulent activities. Fraudulent transactions Zhang and Chen (2019) employed a CNN-LSTM hybrid not only result in direct financial losses but also severely architecture to enhance the detection of complex impact customer trust and institutional credibility. patterns, significantly outperforming traditional Consequently, the development of robust and reliable models in terms of accuracy and precision. Similarly, systems to detect and prevent fraud is paramount. recent studies have applied CNN-LSTM models to Traditionally, manual detection methods have been various domains, such as intrusion detection and employed; however, due to the enormous volume and healthcare analytics, further supporting the complexity of transaction data, such methods have effectiveness of these hybrid approaches in anomaly become ineffective and inefficient. Machine learning detection (Yin et al., 2017; Pham et al., 2020). techniques have emerged as powerful tools, Despite these advances, challenges remain, particularly significantly enhancing the accuracy and efficiency of regarding the imbalance of transaction datasets, fraud detection systems (Awoyemi et al., 2017). interpretability of models, and the computational In recent years, hybrid machine learning models, efficiency of deep learning architectures. Hence, this specifically combining Convolutional Neural Networks study aims to build upon existing literature by (CNN) and Long Short-Term Memory (LSTM) networks, specifically focusing on mitigating these challenges have demonstrated substantial potential in improving through sophisticated data preprocessing techniques detection capabilities by effectively capturing both and optimized hybrid CNN-LSTM architectures. spatial and sequential patterns in transaction data (Zhang & Chen, 2019). This study proposes a hybrid