The rapid increase of mobile banking services over 5G networks has enabled instant, real-time financial transactions but also facilitates possibilities for new sophisticated fraud attacks. This paper presents a novel, end-to-end framework that combines lightweight neural inference at the network edge with a permissioned blockchain ledger to detect, record, and mitigate fraudulent transactions within a single-digit-millisecond budget. First, we design and deploy compact anomaly-detection models on mobile devices and local bank gateways, leveraging transaction metadata (e.g., velocity patterns, geolocation consistency) to flag suspicious activity before it leaves the edge. Second, every alert and model update is immutably recorded on a permissioned blockchain layer shared among trusted financial institutions, ensuring auditability, accountability, and regulatory compliance. We evaluate our framework on a synthetic, federated transaction dataset under emulated 5G conditions, measuring inference latency, detection accuracy, and network overhead. Results demonstrate that edge-based detection reduces average response time by over 80% compared to cloud-only approaches and completes end-to-end decisions in approximately 7 ms (95th < 10 ms). While recall is high ( \(\ge \) 0.90), precision remains limited because of the class imbalance and the relatively small set of features. We discuss threshold calibration, two-stage filtering, and human-in-the-loop review as practical remedies for production deployment.