DEFENSE-BANK: Sea Horse Optimized Deep Learning Framework for Fraud Detection in the Banking Sector

The rapid digital transformation of the banking sector has increased reliance on online transactions, making financial systems increasingly vulnerable to sophisticated cyberattacks and large-scale fraud attempts. Due to the high volume of transactions and the complexity of networks, banks require sophisticated, intelligent, and adaptive security mechanisms that are capable of analyzing massive, evolving, and imbalanced streams of data. However, existing intrusion detection frameworks still suffer from major shortcomings, including high false alarm rates (FAR), severely imbalanced datasets, limited temporal relationship learning, and growing data privacy concerns. To overcome these challenges, a novel Deep learning Enabled Fraud dEtectioN via sea horSE optimization in the banking sector (DEFENSE-BANK) method has been proposed in this paper to enhance accurate and efficient DDoS and cyberattack detection. Stacked Contractive Autoencoder (SCAE) is used to extract compact and noise-resistant features, reducing the impact of imbalance and enhancing privacy. The Temporal Graph-based Multi-Head Convolutional Neural Network (TGMH-CNN) module captures essential temporal and relational patterns to improve cyberattack classification and lower FAR. Additionally, Sea Horse Optimization (SHO) performs adaptive hyperparameter tuning, enabling faster convergence and higher accuracy. The DEFENSE-BANK method has been implemented employing a Python simulator. The proposed DEFENSE-BANK method achieves a higher accuracy of 99.73%, whereas the previous approaches, including SCARFF, STA-GT, and FraudGNN-RL, achieve the accuracy of 98.16%, 97.65% and 98.74% respectively. Additionally, the DEFENSE-BANK method improves the security by 3.38%, 5.85%, and 3.81% better than the existing SCARFF, STA-GT, and FraudGNN-RL methods.