ADVERSARIAL ATTACKS ON FRAUD DETECTION MODELS: A REVIEW OF EVASION, DATA POISONING, AND CAMOUFLAGED FRAUDSTERS

Graph Neural Networks (GNNs) have become a leading approach to fraud detection because fraud is a networked activity, but the same graph structure gives adversaries ways to mislead them. This narrative review examines adversarial attacks on fraud detection models, with a focus on GNN-based systems. It groups attacks into evasion attacks at inference time, data poisoning attacks at training time, and camouflage, in which fraudsters imitate legitimate users or attach themselves to reputable nodes. It then reviews defenses, including generative graph representations, adversarial training, robust aggregation, domain constraints, explainability, and real-time monitoring. The reviewed studies indicate that small changes to a few nodes, edges, or features can degrade graph models, that dynamic settings such as blockchain networks widen the attack surface, and that camouflage is hard to separate from unusual but legitimate behavior. The review adds summary tables of attacks and defenses, a pipeline diagram, and a toy example that shows how neighbor-smearing lowers a node's aggregated risk score. Most evidence comes from single studies of narrow settings, and comparable benchmarks for attacks and defenses are still missing. The article follows no formal search protocol. Future work should test defenses under attack and report results in a standard format.