Adaptive Financial Decision-Making in DeFi: A Comprehensive Approach Using MARL and GNN

Abstract

Decentralized Finance (DeFi) faces critical security challenges due to its pseudonymous and permissionless nature, which exposes it to fraud and market instability. Existing approaches, such as single-agent reinforcement learning (RL) and static graph-based fraud detection, struggle to capture dynamic multi-agent interactions and evolving financial risks. This study proposes an integrated framework combining Multi-Agent Reinforcement Learning (MARL) and Graph Neural Networks (GNNs) to address adaptive decision-making and real-time fraud detection in DeFi. MARL agents, trained in DeepMind’s Melting Pot environment, optimize trading, liquidity provisioning, and arbitrage strategies, while GNNs analyze transaction graphs to detect anomalous patterns. Experimental results demonstrate that MARL agents achieve a 210% increase in average profit per trade and a 57% improvement in market adaptation, alongside a 120% rise in liquidity utilization. The GNN model attains a converged loss below 0.10, reducing false positives by 29%. The integrated system enhances market stability, achieving a stability impact score of 175 within 10 training episodes. This work establishes a scalable, intelligent framework for fraud-resistant trading, cross-chain compliance, and decentralized risk management, advancing the security and efficiency of DeFi ecosystems.