Adaptive Edge-Cloud Intelligence for Low-Latency Anomaly Detection in Digital Payment Ecosystems
Main Article Content
Abstract
Digital payment systems process billions of transactions every day on a set of diverse infrastructure that ranges from resource-constrained point-of-sale terminals, near-edge gateways, regional cloud nodes, to central orchestration clusters. Round-trip cloud-based fraud detection systems are 100 to 200 milliseconds, which isn't feasible for on-demand payment authorisation, and pure on-device fraud detection systems are simply insufficient to represent the growing complexity and sophistication of fraud patterns. We introduce a novel framework called AEC-AD (Adaptive Edge-Cloud Anomaly Detector), which dynamically splits the inference between edge and cloud by leveraging a learned offloading policy across the edge-cloud continuum, includes lightweight streaming models on the edge, periodically synchronised teacher networks on cloud using knowledge distillation, and integrates an online concept drift detector to refresh the model proactively when the payment behaviour changes. AEC-AD achieves an AUC-ROC of 0.953 and an F1 score of 0.924 with an inference latency of 23.4ms at the 99th percentile performance, a 4.9× latency improvement over the cloud-only baselines and an 87% accuracy gain towards the prohibitively-high accuracy ceiling. A twoway ANOVA for bandwidth and system-load conditions shows that both bandwidth and load have significant main effects (η² = 0.341 and 0.202, respectively). Pairwise Wilcoxon signed-rank tests are used to determine the statistically significant advantage over six competing baselines (p < 0.05, r ≥ 0.442, rank-biserial). The six incremental configurations are used in an ablation study to determine the contribution of each architectural component. The results of these findings make AEC-AD a deployable, regulation aware, intelligence architecture for the next generation payment security infrastructures.