With the growing growth of cyberspace, reliable fraud detection in the banking industry has become more and more crucial. While successful, traditional centralized handling of data infrastructures has drawbacks regarding scalability, privacy, and single points of breakdown. Decentralized techniques, while promoting privacy and fault tolerance, can confront complexity and high interacting costs. To overcome these issues, this study presents a novel architecture for fraud detection that employs federated learning inside a semi-decentralized structure. This architecture incorporates regional hubs to ad-minister clusters of nearby banks, allowing for rapid, confidential model training and aggregation. The suggested schema improves scalability, privacy, and fault tolerance by integrating the advantages of centralized and de-centralized techniques, resulting in a robust and safe method for detecting bank fraud.