FedXChain: Explainable Federated Learning with Adaptive Trust Scoring and Blockchain-Based Audit Trails

Federated learning faces challenges in explainability and trust when aggregating models from heterogeneous nodes with non-IID data distributions. This study presents FedXChain, a framework that combines privacy-preserving Federated-SHapley Additive exPlanations (SHAP) aggregation with Node-Specific Divergence Scores (NSDS) to quantify local explanation fidelity, adaptive trust-based aggregation, and blockchain-verified audit trails for transparent and verifiable collaboration. It validates FedXChain across three fundamentally different model architectures (Logistic Regression, Multi-Layer Perceptron, and Random Forest) on real-world medical data from the Wisconsin Breast Cancer dataset (569 clinical breast tissue samples). The experimental results show that FedXChain achieves 96.50% accuracy with excellent statistical reproducibility (CV < 2% across 5 independent runs). FedXChain also provides NSDS-based interpretability tracking, with observed NSDS values ranging from 0.1926 to 0.5768 across the evaluated architectures, supporting the analysis of explanation divergence under heterogeneous clients. In the final-round comparison, FedXChain reaches 96.5% accuracy under non-IID settings (α = 0.3), outperforming FedProx (89.5%, non-IID α = 0.5) and remaining competitive with FedAvg under IID conditions (96.0%).