Trust-Aware Blockchain Framework for Secure, Explainable Federated Artificial Intelligence

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).

Federated Learning (FL) allows organizations to train shared models without pooling raw data, but it leaves open questions of adversarial robustness, accountability, and decision transparency. This paper presents FedPP-XAI, a Federated and Privacy-Preserving AI framework that integrates blockchain-based trust management, differential privacy, and SHAP-driven explainability within a single architecture, rather than treating them as independent add-ons: blockchain immutability anchors the audit trail that explainability outputs reference, while differential privacy bounds the leakage that blockchain records expose, and smart contracts remove the need for a trusted central aggregator. At the core of the framework is TEWA-Fed (Trust- and Explanation-Weighted Adaptive Aggregation), an aggregation rule that replaces sample-count-only weighting with a composite weight derived from each client's blockchain-verified trust score and its round-over-round SHAP attribution consistency, coupled to a convex privacy-budget allocation that gives well-behaved clients tighter noise without weakening the framework's overall differential-privacy guarantee. Across benchmark datasets spanning network intrusion detection and cardiovascular diagnostics, FedPP-XAI outperformed centralized, standard-federated, and blockchain-only baselines on accuracy, privacy leakage, and poisoning-attack resistance (97.3% accuracy, 11.2% membership-inference success rate, 100% poisoning detection, cross-client SHAP ρ = 0.91). A controlled proof-of-concept with real poisoning attacks further shows TEWA-Fed's trust term holding accuracy stable where FedAvg, FedProx, SCAFFOLD, and FedDyn collapse.