An Improved FEDformer–Dilated Residual CNN Framework with Cross-Attentive Fusion for Cloud-Based Cyber Threat Detection and Financial Fraud Prediction

The fast growth in cloud computing has presented serious cybersecurity and anti-fraud challenges because enormous volumes and diverse patterns of data have to be processed. Conventional security strategies and traditional Machine Learning (ML) algorithms usually find it difficult to process such complexity, hence introducing delay and inefficiency in identifying and containing cyber threats and fraud. In order to resolve the issues mentioned above, the paper here describes a cutting-edge hybrid model of cyber threat detection using cloud and financial fraud forecasting. FEDformer, an end-to-end Transformer-based long-term time-series forecasting model, and Dilated Residual Convolutional Neural Networks (DRCNNs) are merged, with the fusion of temporal and spatial information made through a Cross-Attentive Fusion module. This novel method enhances the model’s ability to recognize sophisticated patterns from multi-source information, improving both scalability and accuracy. The framework is cloud-optimized, such that it is highly appropriate for real-time execution without compromising on performance. The proposed model in this research achieves an accuracy rate of 97.99%, a sensitivity/recall rate of 97.96%, and an F1-Score rate of 98.21%. These rates beat current models remarkably, indicating the model performs much better than present solutions to advanced cloud-based security and fraud prevention problems.