Quantifying the Vulnerability of Credit Card Fraud Detection Systems to Gradient Based Adversarial Evasion

In financial cybersecurity, fraud detection is at the forefront of the battle against cybercriminals and deep learning models are the state-of-the-art method for real-time detection; yet their reliance on fixed feature distributions renders them inherently susceptible to adversarial attacks. With the rise of generative artificial intelligence (AI) in obfuscating fraudulent transactions, conventional performance indicators (accuracy and F1-score) do not reflect the cybersecurity resilience of these models. This research explores the Credit Card Fraud Detection dataset, comprising more than 550,000 anonymized records. The study uses a functional deep learning model as a benchmark and introduces the Fast Gradient Method (FGM) to simulate white box attacks on the latent features . The experiment results demonstrate a high detection accuracy of 99.85% under normal operating conditions, which dramatically reduces to 92.75% when card transactions are slightly modified by adversarial disturbances at an epsilon value of 0.1. Numerical evidence reveals a clear 7.10% "Security Gap," depicting a substantial degree of vulnerability, where transactions are mathematically reclassified from fraudulent to legitimate. Notably, our research suggests a non-linear "safety cliff" in the decrease of detection accuracy, where the model's reliability completely fails as the adversarial strength reaches 0.15 or higher. In addition, Principal Component Analysis demonstrates that gradient attacks have successfully distorted the latent features of fraudulent transactions, resulting in the movement of highly suspicious transactions towards the concentrated area of legitimate transactions, hence evading automated security screening while preserving the overall statistical distribution of the data.