Multi-Instance Coupled Modular Neural Network for Attack Detection in IoT-Enabled Banking Systems

Abstract

Most modern banks have incorporated IoT into their operations to offer customers better services, while the IoT nature of banking systems poses several security threats. There is nothing more irreplaceable than the banking sector, which deals with sensitive financial information; hence, the need for more protective measures against cybercriminals. In response, this research work puts forward a new Multi-instance Coupled Modular Neural Network with Newton-Raphson Optimizer (M-CMNNet-NRO) to achieve efficient attack detection within an IoT-based banking system. The model begins with text processing of raw data gathered from various IoT devices, utilizing Zero-shot Text Normalization (Z-STN), which aims to normalize and standardize textual data without requiring any prior sample data. Then, the Geometric Algebra Transformer (GAT) further delves into the data to incorporate additional features while navigating complex structures and patterns. The M-CMNNet effectively analyzes multiple instances of data simultaneously and can therefore identify new and complex attack patterns as well. To enhance the model's efficiency, the Newton-Raphson-based Optimizer (NRO) further adjusts the weights to improve convergence and performance. The findings show that the integrated model outperforms existing systems, achieving an accuracy of 99.45%, a recall of 99.60%, and an AUC of 99.90%, with the lowest error rate of 0.55%. These are valuable results because they emphasize its stability and fast running, which would make it applicable to the purpose of real-time fraud detection.