HYBRID FRAUD DETECTION IN DIGITAL FINANCIAL TRANSACTIONS USING XGBOOST AND CONVOLUTIONAL NEURAL NETWORKS

The rapid growth of digital financial services has been accompanied by increasingly sophisticated fraud schemes that exploit authentication weaknesses, sequential transaction patterns, and extreme class imbalance. Traditional rule-based systems and even strong tabular classifiers struggle to maintain high recall of rare fraud cases while controlling false-positive rates that damage customer experience. This paper presents a hybrid fraud-detection framework that couples a Convolutional Neural Network (CNN) for automated hierarchical feature extraction with an Extreme Gradient Boosting (XGBoost) classifier for final decision-making. Transaction feature vectors are reshaped into tensors, processed by a multi-layer CNN to produce discriminative embeddings, and then classified by a regularized, class-weighted XGBoost model. Experiments on the European Credit Card Fraud Detection dataset (284,807 transactions, 0.172% fraud) demonstrate that the hybrid architecture outperforms logistic regression, random forest, standalone XGBoost, and a standalone CNN across accuracy (0.995), precision (0.930), recall (0.920), F1-score (0.925), AUC-ROC (0.988), and AUPRC (0.930), while achieving the lowest false-positive rate (0.005). Feature-importance analysis confirms that CNN-derived embeddings rank among the most influential predictors. The results indicate that combining deep representation learning with gradient-boosted decision trees offers a practical, imbalance-aware, and relatively interpretable solution for near-real-time fraud detection in digital financial environments.