Mathematical Modelling of Fraud Detection in Mobile Financial Transactions Using Deep Learning

Abstract

The rapid growth of mobile financial services has introduced complex vulnerabilities, making fraud detection a critical priority for digital financial systems. This study presents a mathematically grounded deep learning framework for detecting fraudulent mobile transactions by modeling them as multivariate time-series classification problems. The methodology employs Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and a hybrid CNN-LSTM architecture to capture both spatial feature patterns and temporal behavioral dependencies. The dataset, comprising anonymized and synthetic mobile transaction records, was preprocessed through normalization, categorical encoding, and class imbalance correction using SMOTE. Experimental evaluations reveal that the CNN-LSTM model outperformed baseline architectures, achieving an F1-score of 0.955 and AUC-ROC of 0.97, indicating superior detection capability and generalizability. Misclassification analysis highlighted threshold-sensitive trade-offs between false positives and false negatives, while explainability and robustness assessments demonstrated the model’s transparency and resistance to adversarial input manipulation. Conclusively, the proposed framework offers a scalable, interpretable, and high-performing solution for fraud mitigation in mobile financial platforms, contributing to enhanced cybersecurity and regulatory compliance in real-time transaction systems.