An Integrated Spatial–Temporal Deep Learning Approach for Detecting Fraud in Cloud Banking Ecosystems

The rapid expansion of cloud banking systems has led to a significant increase in transaction volume and complexity of interactions, thereby elevating the risk of sophisticated and hard-to-detect fraudulent activities. Traditional rule-based and statistical fraud detection methods are increasingly ineffective in such environments, as they fail to capture complex structural relationships among entities and cannot adapt to evolving user behavior patterns. To address these challenges, this study proposes an integrated spatio-temporal deep learning framework that combines Graph Convolutional Networks (GCNs) with a Temporal Attention mechanism. In the proposed approach, the GCN component models the spatial relationships among users, accounts, devices, and transactions, enabling the extraction of hidden and complex interaction patterns that conventional methods often overlook. Simultaneously, the Temporal Attention module analyzes the sequential and time-dependent nature of transaction data, allowing the system to focus on critical time periods where anomalous behavior is more likely to occur. This combination of spatial and temporal modeling enhances the detection of both explicit and subtle fraud patterns. The proposed framework is designed to be scalable, adaptive, and capable of real-time processing, making it well-suited for deployment in modern cloud banking infrastructures where efficient and proactive fraud detection is essential.