AdaBoost-based credit card Fraud data augmentation and adversarial learning analysis

Credit card fraud detection is a key technology for ensuring the security and stability of financial systems.However, the extremely low proportion of fraudulent transactions in real-world data causes severe class imbalance, degrading model performance.Traditional oversampling methods like SMOTE have been used to address this, but they rely on simple linear interpolation, leading to limited data diversity and vulnerability to adversarial attacks.This study applies various GAN-based data augmentation techniques-GAN, CGAN, WGAN, and WCGAN-to an AdaBoost model and evaluates detection performance under four adversarial attack scenarios: transaction volume manipulation, price spiking, transaction distortion, and trend manipulation.Experimental results show that GAN-based augmentation outperforms SMOTE in overall performance.In particular, WCGAN achieved the highest F1-score and maintained a stable precision-recall balance, demonstrating strong robustness.In contrast, models trained with SMOTE suffered significant performance degradation across all attacks, especially in price spiking and transaction distortion.These findings suggest that GAN-based augmentation is a more effective alternative to conventional oversampling.They also highlight the importance of evaluating models under adversarial conditions.Notably, WCGAN and WGAN exhibited strong resistance and general stability, offering practical value for the development of secure fraud detection systems.