TrustGraph: Federated Graph Neural Networks for Cross-Platform Trust and Fraud Propagation Analysis
Tejas Patel, Arun Kumar, Madhushree Kumari, Rajesh Purushothaman, Rakesh Keshava, Milan Parikh
Abstract
Centralized fraud detection systems in e-commerce ecosystems face significant limitations due to stringent data privacy regulations, platform heterogeneity, and the distributed nature of sophisticated fraud rings operating across multiple marketplaces. Existing approaches predominantly rely on isolated platform-specific models or centrally aggregated data, fundamentally limiting their ability to capture cross-platform trust relationships and fraud propagation dynamics that characterize modern coordinated fraud campaigns. We propose TrustGraph, a federated graph neural network framework designed for privacy-preserving cross-platform fraud detection in distributed e-commerce environments. TrustGraph models entities and their interactions as graph-structured data and enables collaborative learning across independent marketplaces without requiring raw data sharing. By synergistically combining federated optimization with graph-based representation learning, TrustGraph captures complex structural fraud propagation patterns while maintaining strict data locality and regulatory compliance. Through extensive experiments on the YelpChi benchmark dataset under realistic non-IID data distributions with K = 10 federated clients and severe heterogeneity (Dirichlet α = 0.5), we demonstrate that TrustGraph achieves 0.93 AUC, approaching within 0.02 of centralized performance while significantly outperforming localonly models by 0.06 AUC and federated non-graph baselines by 0.08 AUC, with statistical significance confirmed across multiple architectural configurations and heterogeneity levels.