Federated Multi-Agent Reinforcement Learning for Real-Time Risk Scoring in Cross-Border Payment Networks

As financial ecosystems become increasingly digitized and globalized, cross-border payments present escalating challenges in real-time risk assessment, fraud detection, and regulatory compliance. Traditional centralized models for risk scoring often fall short in scalability, privacy preservation, and adaptability. This research proposes an innovative architecture that fuses Federated Learning (FL) with Multi-Agent Reinforcement Learning (MARL) to enable decentralized, privacy-preserving, and adaptive real-time risk scoring across international payment gateways. Each participating financial entity acts as an autonomous agent, trained on local data without compromising user privacy, and collaborates through a shared policy gradient to enhance the collective risk intelligence. The system is implemented as a microservices-based framework, ensuring resilience, modularity, and interoperability. Simulation results demonstrate improved detection accuracy, reduced false positives, and compliance alignment with data protection regulations such as GDPR and PCI-DSS. This paper advances the frontier of intelligent financial systems by aligning federated AI with secure, scalable fintech architectures.