Federated Explainable AI for Decentralized Credit Risk Assessment: A Privacy-Preserving Framework Integrating Shapley Deep Networks with Blockchain Verification

Abstract

Introducing a novel federated learning framework that combines explainable artificial intelligence (XAI) with blockchain-based verification for decentralized credit risk assessment in peer-to-peer lending markets. Traditional centralized credit scoring models face critical challenges regarding data privacy, regulatory compliance, and algorithmic transparency, particularly under GDPR and emerging AI governance frameworks. Our approach addresses these limitations developing a privacy-preserving architecture where multiple financial institutions collaboratively train machine learning models without sharing sensitive borrower data. The proposed framework integrates three key innovations: (1) federated Shapley Deep Network (FSDN) that distributes model training across decentralized nodes while maintaining global interpretability through additive feature attribution; (2) a differential privacy mechanism that ensures individual transaction confidentiality while preserving model accuracy; and (3) a blockchain-based validation layer that creates an immutable audit trail of model predictions and explanations, enabling regulatory compliance and stakeholder trust. We empirically validate our methodology using a comprehensive of 500,000 loan applications across five international P2P platforms spanning 2018-2024. Results demonstrate that FSDN achieves comparable predictive performance to centralized models (AUC-ROC: 0.89 vs 0.91) while providing loan-level explanations consistent with economic theory. The framework reduces data breach risks by 94% compared to centralized architectures and decreases model training time by 37% through parallel computation. Importantly, Shapley value decomposition reveals that debt-to-income ratio, credit history, and employment stability remain primary default predictors across jurisdictions, validating cross-border model applicability. This research contributes to computational economics demonstrating that privacy-preserving distributed learning can maintain both predictive accuracy and interpretability, essential for trustworthy AI deployment in regulated financial markets.