The process of fraud detection on digital financial systems exists within two long-term constraints: confirmed fraud symbols are scarce and frequently latent, whereas fraudsters adapt to implemented protection against countermeasures. These factors diminish the performance of just discriminative models which are learned on distributions of by far skewed data which change over time. Generative artificial intelligence, with Generative Adversarial Networks (GANs) being a more specific format, provides an alternative, attaining structured representations of transaction data and producing synthetic samples that cover more frequency of rare patterns of fraud without replicating sensitive records. This paper presents GAN-based fraud detection using a system-level design approach instead of an independent modeling measure. We explain how conditional tabular GANs may be integrated into an end-to-end pipeline that includes data preprocessing, stabilized training of GANs, generative fraud, retraining of models, probabilistic calibration of probabilities and drift checks. Precision, recall, F1-score, and AUC are used to analyze model effectiveness in the case of severe class imbalance, but mostly the behavior of the precision and recall and the operational decision thresholds are considered. Possible constraints and ethical issues are also addressed, not to mention the privacy risks, low-fidelity synthetics, and dual-use.
Latha et al. (Thu,) studied this question.