MH-DRNN: An intelligent approach in financial fraud detection and prevention

Financial fraud detection requires models capable of identifying subtle and evolving anomalies within large-scale transaction streams. Existing deep learning and anomaly-detection techniques struggle with sparse fraud patterns, nonlinear feature interactions, and limited memory for rare events. To address these gaps, this study proposes a hybrid MH-DRNN framework that integrates Kernel PCA–based nonlinear feature extraction, LOF-based anomaly pre-screening, Modern Hopfield Networks for associative memory retrieval, and Deep Recurrent Networks for temporal sequence modelling. Using a benchmark Kaggle financial-transaction dataset containing 10,000 records, the model achieved 98.5% accuracy, 98% recall, 97% precision, and a 97% F1-score. These improvements exceed the performance of IF-AE, CNN, and CNN-GRU-Attention models. The results demonstrate that the combined associative-memory and sequential-learning architecture enhances the model’s ability to capture sparse, complex, and evolving fraud behaviour. The study also discusses data-privacy implications of AI-based fraud detection and highlights the importance of reliable, high-quality datasets for secure financial analytics.