This study presents the design and evaluation of a quantum-resistant federated learning protocol (QFLP) for secure, privacy-preserving financial fraud detection across globally distributed institutions. Traditional federated learning frameworks suffer from exposure to gradient inference attacks, model poisoning, and vulnerability to future quantum adversaries. To overcome these challenges, QFLP integrates lattice-based post-quantum cryptographic primitives (CRYSTALS-Kyber, Dilithium), secure multi-party aggregation, and anomaly-aware trust scoring into a unified federated architecture. Each participating node locally trains a fraud detection model using non-identically distributed transactional data and transmits encrypted, masked updates to a coordinating server for robust aggregation. Simulation experiments under adversarial, non-IID, and bandwidth-constrained conditions demonstrate that QFLP achieves faster convergence, superior fraud recall, and up to 42% reduction in false positive rates compared to baseline FL methods. Despite minor increases in bandwidth and encryption latency, the protocol sustains low computational overhead while delivering strong cryptographic guarantees against both classical and quantum attacks. The QFLP framework enables compliant, cross-border collaboration without raw data exchange, offering a scalable foundation for modernizing decentralized financial security. This work contributes a resilient software architecture aligned with emerging regulatory, cryptographic, and operational demands in post-quantum secure fintech environments.