Federated Graph Learning for Privacy-Preserving Fraud Intelligence Across Distributed Financial Cloud Systems

Main Article Content

Mahatma Reddy Marri

Abstract

Financial fraud across distributed cloud ecosystems is a challenge that demands collaborative intelligence but is constrained by strict data-sovereignty laws that do not allow raw transaction data to be shared between institutions. In this paper, we propose a novel framework, Federated Graph Intelligence Layer (FedGIL), that integrates federated learning, graph attention networks, differential privacy and topology-aware gradient compression to facilitate privacy-preserving crossinstitutional fraud detection without sharing sensitive financial information with a central server. FedGIL represents each cloud node as a heterogeneous transaction graph, uses multi-head graph attention for local training, and only sends sanitized gradient updates, which are protected by a calibrated Gaussian mechanism with privacy budget (ε, δ), to a secure aggregator in the global cloud. Experiments on five real world and synthetic datasets with more than seven million transactions show that FedGIL is able to outperform the isolated per-node models by 85% on AUCROC and 80% on F1 score. A 2-way ANOVA test shows that data heterogeneity (η² = 0.312) and privacy budget (η² = 0.213) have large effect on model performance. Wilcoxon signed-rank tests show significant performance benefits over five competing baselines (p<0.05 and rank-biserial r ≥ 0.318). Through adaptive spectral gradient compression, the communication overhead can be reduced by 78% compared with the uncompressed federated baselines. FedGIL helps to push the envelope for collaborative financial intelligence in a privacy-preserving way and provides a deployable financial intelligence architecture for the real world regulatory environment.

Article Details

Section
Articles