FedTrust-GNN: Federated Learning with Blockchain Trust and Graph Neural Networks for Decentralized Modeling

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Centralized user modeling systems inherently violate privacy regulations, with catastrophic failure points, opaque trust management mechanisms, and vulnerability to complex adversarial strategies such as poisoning attacks, model inversions, and membership inference. Federated learning (FL) approaches address data privacy issues by training models without centralized data storage. However, they suffer from Byzantine failures, lack of transparent reputation-based trust mechanisms, and inability to harness relational information implicit in user interactions. This paper proposes FedTrust-GNN, a secure and relationally aware decentralized user modeling framework that integrates federated learning, permissioned blockchain, and graph neural networks. The framework employs differentially private federated learning with secure multi-party computation (SMPC) to perform gradient aggregation while providing strong privacy guarantees. A permissioned blockchain using the Practical Byzantine Fault Tolerance (PBFT) consensus protocol maintains an immutable ledger of participant behaviors, model contributions, and reputation scores, eliminating reliance on a trusted central authority. To model participant relationships, a heterogeneous graph attention network (HGAT) infers dynamic trust scores by capturing multihop trust propagation, collusion patterns, and evolving reputation through attention-based message passing. These trust scores guide a trust-weighted robust aggregation (TWRA) mechanism that combines norm clipping with coordinate-wise median aggregation, achieving Byzantine fault tolerance against up to 30\% malicious participants. Experiments on Federated EMNIST, Stack Overflow, and synthetic datasets with 10,000--100,000 participants demonstrate that FedTrust-GNN achieves 94.2\% accuracy (within 1.3\% of centralized models), reduces successful label-flipping attacks by 82\% (34\% to 6.1\%), improves convergence stability by 41\%, and enables blockchain performance of 1,200 transactions per second with 2.3-second finality, demonstrating its practicality for large-scale deployment.