Robust machine learning for fraud detection under imbalance, rarity, and adversarial behavior: Evidence across financial reporting, payments, and graph ecosystems

Fraud detection is a high-stakes machine learning (ML) problem characterized by extreme class imbalance, rare-event dynamics, temporal distribution shift, and adaptive adversarial behavior. Existing reviews typically emphasize a single fraud domain, model family, or performance comparison, leaving limited synthesis of how robustness mechanisms transfer across financial reporting, payment systems, and relational graph ecosystems. This systematic literature review develops a robustness-by-design framework organized around four dimensions: rarity and asymmetric-cost robustness, evaluation integrity, dynamic/adversarial robustness, and structural/evidential robustness. Guided by PRISMA logic, the review synthesizes 38 studies using a common failure-mode/design-response lens rather than treating the three application ecosystems as separate literatures. The synthesis indicates that reliable fraud detection depends on integrating rarity-aware data design, cost-sensitive objectives, temporal or entity-aware validation, explicit stress testing under drift and manipulation, and interpretable relational evidence. Persistent weaknesses include inconsistent metric choice, limited external and adversarial validation, heterogeneous reporting, and insufficiently standardized robustness tests. The review contributes a cross-ecosystem framework and research agenda for fraud-detection systems that are auditable, transferable, and operationally resilient.