Adaptive Cloud-Native Fraud Detection via Uncertainty-Aware Dynamic Model Routing
Real-time payment fraud detection in cloud-native platforms is constrained by two forces that are usually optimized separately: the statistical difficulty of highly imbalanced and drifting transaction streams, and the operational latency budget of synchronous microservice inference. Prior cloud-native payment intelligence work has shown that lightweight models can offer attractive latency and footprint properties, while tree ensembles can improve ranking quality at substantially higher inference cost. This paper extends that line of work by proposing UDMR, an uncertainty-aware dynamic model routing framework that invokes a fast-linear scorer for every transaction and routes only high-risk, uncertain, or drift-exposed transactions to a heavier expert model. The router combines fast-model risk, decision-margin uncertainty, and a training-window drift distance into a fixed single-row routing score, then selects the routing threshold and operating threshold on a validation stream under a minimum precision floor. Because the public ULB/Worldline fraud data were not available inside the execution sandbox, the empirical artifact uses a deterministic synthetic transaction stream with 60,000 chronological transactions, a 0.91% fraud rate, and controlled concept drift; the released code automatically uses a real ULB-style creditcard.csv file when supplied. On the synthetic drift benchmark, the random-forest expert achieves the strongest average precision (0.859) and F1 (0.803) but requires serving every request through the heavy model. UDMR routes 39.6% of test transactions to the expert, reaches average precision of 0.557 and F1 of 0.528, and reduces median single-row latency from 2429 microseconds for the expert-only service to 345 microseconds. The results show that uncertainty-aware routing is a practical cloud-native design pattern when median capacity, alert precision, and graceful degradation matter, while also revealing that high routing rates still expose p95/p99 latency to expert-model cost.