本研究的主要貢獻包括：透過引入梯度懲罰與頻譜正規化來穩定對抗式生成模型的訓練過程，並建構一個結合多種模型優勢的堆疊式架構，以提升詐欺交易偵測的準確性與穩定性，並系統性分析增強倍率對詐欺偵測成效之影響。本架構可作為高度不平衡交易環境下詐欺偵測的穩健且具擴展性的解決方案。 ;The escalating reliance on online commerce has significantly heightened exposure to credit card fraud, posing critical challenges for financial institutions and consumers alike. Addressing this issue is complicated by the severe class imbalance inherent in real-world datasets. To mitigate this, we propose a hybrid detection framework that integrates a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) to generate diverse and realistic fraud samples, and a stacking ensemble classifier that combines conventional machine learning models (Random Forest, XGBoost, LightGBM, CatBoost) with deep learning architectures (DNN, CNN, LSTM). valuations are conducted on two publicly available Kaggle competition datasets—Credit Card Fraud Detection (with only 0.17% fraudulent cases) and IEEE-CIS Fraud Detection (3.52% fraud rate)—both of which are characterized by extreme class imbalance. Experimental results demonstrate that the proposed method consistently outperforms baseline oversampling techniques such as SMOTE and ADASYN across multiple evaluation metrics. The system achieves peak performance at optimized augmentation levels, with significant gains in recall and F1-score while maintaining high precision.Key contributions of this work include stabilizing adversarial training through gradient penalties and spectral normalization, constructing an effective stacking architecture that leverages heterogeneous model strengths, and systematically analyzing the impact of augmentation scale on detection outcomes. This framework provides a robust and scalable solution for enhancing fraud detection in highly imbalanced transactional environments.
