Edge-deployed graph learning for real-time telecom and financial fraud detection: a systematic survey and reference architecture

Abstract Cross-domain telecom and financial fraud (attack chains that begin with a fraudulent SIM swap, intercept a one-time password, and clear an unauthorized transfer) now dominates high-value losses globally. Defenders restricted to a single layer see only half of the chain and routinely miss the relational coordination that characterizes modern fraud rings. Graph Neural Networks have closed the relational gap, but they are almost universally evaluated in centralized, offline settings, a regime structurally incompatible with the 100-millisecond latency budget of card-present payment authorization. This article systematically surveys 22 primary studies (2017 to 2025) at the intersection of graph-based fraud detection, multi-access edge computing, federated training, and blockchain audit, drawn from both the telecommunications venues and the broader machine-learning venues, and organized along a five-dimensional taxonomy. A coverage matrix exposes that no published system addresses all five dimensions simultaneously. We propose a reference architecture and identify five open research challenges. A two-experiment pilot implementation on the standard Sichuan Mobile CDR benchmark validates the spatial GAT backbone (F1 = 0.7737, AUC = 0.8039) and a FedAvg-based three-operator federated variant (F1 = 0.8622, AUC = 0.9263, Precision = 0.9364), with the federated configuration outperforming the centralized baseline on every metric, an outcome we attribute to implicit ensemble averaging and to cleaner per-silo graph structure. The result establishes that the regulatory and competitive obstacles to cross-operator data sharing need not be regarded as obstacles to high-accuracy cross-operator fraud detection.