Improving Financial Security: A Hybrid AI-Based Credit Card Fraud Detection Framework

The explosive expansion of digital financial transactions demand for robust solutions to address the increasing number of fraudulent activities, perpetrated against credit cards. Although traditional ML models, such as XGBoost, can produce a high level of accuracy, problems such as the class skew, incompatibility with computational resources, and difficulty in interpretability still exist. The proposed method is a new hybrid approach that integrates Temporal Convolutional Network (TCN) with CatBoost, and it is supported by Generative Adversarial Network (GAN)-based synthetic data generation and Bayesian hyperparameter optimization. Class imbalance in the context of SHAP may be mitigated using a hybrid SMOTE-Tomek resampling approach which enables a better interpretability of the selected features. Tested on the anonymized set of $\mathbf{2 8 4, 8 0 7}$ transactions, our model yields an accuracy of $\mathbf{9 9. 1 2 \%}$ and an AUC-ROC score of $\mathbf{0. 9 9 3}$, quite comparable to the achievements of existing benchmarks. We show the possibility of real-time deployment by showing model quantization and edge computing incorporation. This study highlights the success of deep learning hybrids and the combination of advanced optimization in fraud detection.