Abstract
Financial fraud across decentralized institutions presents a severe global challenge, complicated by stringent data privacy regulations that prevent direct cross-organizational data pooling. While Graph Neural Networks (GNNs) have emerged as highly effective tools for detecting complex topological patterns indicative of money laundering and fraud, applying GNNs in a decentralized, privacy-preserving setting remains difficult due to inter-institutional edge dependencies and structural leakage. In this work, we propose a novel Differentially Private Federated Graph Neural Network (DP-FGNN) framework specifically designed for cross-institutional financial fraud detection. DP-FGNN leverages local GNN architectures to capture relational dependencies within isolated institutional nodes while utilizing a differentially private aggregation scheme with Rényi Differential Privacy (RDP) guarantees to manage cross-boundary edge representations. By introducing a localized dynamic boundary alignment mechanism, our framework enables collaborative training without revealing sensitive transaction histories or underlying network topologies. Experimental evaluations conducted on semi-synthetic and real-world multi-institutional transaction graphs demonstrate that DP-FGNN achieves near-centralized detection performance (AUC-ROC of 0.942 under a tight privacy budget of $\epsilon = 2.0$), significantly outperforming baseline federated learning and differentially private models. Our findings highlight the viability of scalable, privacy-preserving graph intelligence for mitigating systemic financial risk.
