GNN-Augmented RL for Fraud Detection in Decentralized Finance

Decentralized Finance (DeFi) has revolutionized financial transactions by enabling open, permissionless access to financial services. However, its lack of centralized oversight and pseudonymous architecture have also brought by fraudulent activities. This study presents a novel framework for fraud detection in DeFi that integrates graph neural networks (GNNs) with multi-agent reinforcement learning (MARL). Leveraging a directed transaction graph comprising 50,000 Ethereum addresses and over 120,000 token transfers, this paper evaluates four detection pipelines: extreme gradient-boosted decision trees (XGBoost), a GNN-only model (GCN), a standalone reinforcement learning agent (PPO), and a proposed GNN+RL hybrid model. The hybrid system combines graph-based embeddings with adversarial policy learning, where a fraudster and a detector co-evolve through a multi-agent PPO setup using PettingZoos ParallelEnv. Synthetic fraud strategies are generated using a GAN and projected into the GCN embedding space to simulate adaptive threats. Experimental results show that while GCNs outperform flat-feature models, the GNN+RL hybrid achieves superior balance across accuracy (84.58%), AUC (0.8176), and F1 score (0.7493), capturing both structural and behavioral fraud signals. Reward convergence curves further illustrate emergent adversarial dynamics. The proposed framework demonstrates the effectiveness of combining relational inductive biases, dynamic decision-making, and adversarial augmentation for resilient fraud detection. Future work includes extending to cross-chain analytics and enriching contextual understanding through integration with large language models.