FraudSentinel: Federated Multi-Agent Reinforcement Learning for Privacy-Preserving Cross-Marketplace Fraud Detection in Distributed E-Commerce Ecosystems

E-commerce fraud costs the global economy $48 billion annually, with sophisticated fraud rings operating across multiple online marketplaces. While individual platforms deploy fraud detection systems, fraudsters exploit the lack of cross-platform intelligence sharing. Traditional centralized fraud databases violate privacy regulations and create single points of failure. We present FRAUDSENTINEL, a federated multi-agent reinforcement learning framework that enables privacy-preserving fraud pattern sharing across distributed e-commerce marketplaces without exposing sensitive customer data. FRAUDSENTINEL introduces three key innovations: (1) a hierarchical federated multi-agent architecture where each marketplace operates independent RL agents while contributing to a global fraud detection policy through secure gradient aggregation, (2) a privacy-preserving protocol combining homomorphic encryption, differential privacy, and secure multi-party computation to ensure zero-knowledge fraud pattern sharing, and (3) an adaptive cross-marketplace learning mechanism that enables rapid fraud pattern transfer across platforms with different business models, inventory types, and customer demographics. We evaluate FRAUDSENTINEL on the IEEE-CIS fraud detection dataset (590K real-world e-commerce transactions) extended with synthetic multi-marketplace scenarios across five verticals (general retail, fashion, electronics, groceries, digital goods). Our framework achieves 96.8% fraud detection accuracy with 0.31% false positive rate—a 9.1% relative improvement over isolated single-marketplace systems—while maintaining