Blockchain-Enabled Maximum Support Deep Belief Neural Network Technique for Cyber Threat Detection in the Financial Sector

Abstract

The increasing digitisation of financial transactions has led to a growing vulnerability to cyber threats, including fraud, unauthorised access, and data breaches. Existing approaches for detecting financial threats suffer from several issues, including low detection rates due to data imbalance, insufficient feature selection methods, and the inability to extract subtle transactional patterns. However, the lack of integrated frameworks that combine smart detection with encryption and access control in existing security solutions makes these solutions susceptible to advanced cyberattacks. To overcome these problems, the proposed Blockchain-enabled Maximum Support Deep Belief Neural Network (MSDBN2) technique is used to detect cyber threats in the financial sector. The Malicious Activity Prone Factor (MAPF) scheme is used to analyze the behaviour of processes from a collective dataset. Then, the Optimized Dragon Fly with Decision Tree (ODF-DT) approach is used to select the essential features of financial cyber threats. After that, the proposed MSDBN2 approach learns the complex patterns and correlations of the selected features. Here, the Maximum support weight vectors determine the optimal hyperplane that separates normal from threat classes. Based on threat detection, the Circular Shift Elliptic Curve Data Encryption (CSECDE) method encrypts the financial information. Next, the Blockchain-based Dynamic Chain Link Policy (BDCLP) scheme is used to prevent unauthorized access. Finally, the Quantum Key Authentication (QKA) technique is employed to verify the authorized users in the financial sector. Therefore, the proposed method achieves higher security performance and threat-detection accuracy than other methods. The proposed model is assessed using several performance metrics, including accuracy, precision, recall, F1-score, time complexity, encryption and decryption efficiency, authentication performance, and security performance. The experimental results reveal that the proposed approach achieves 96.72% accuracy and 96.10% improved security performance.