A Decentralized Blockchain-Anchored Federated Learning Framework for Real-Time Verification of Crisis-Related Social Media Content

Abstract

During humanitarian crises, natural disasters, and public emergencies, social media platforms serve as vital real-time information sources. However, they are also prone to the rapid propagation of misinformation, rumors, and malicious content. Traditional centralized verification systems present severe limitations, including single points of failure, algorithmic bias, censorship concerns, and data privacy issues. To address these vulnerabilities, this paper proposes DeFL-Crisis, a decentralized, blockchain-anchored federated learning framework designed for real-time verification of crisis-related social media content. DeFL-Crisis utilizes federated learning to allow edge client nodes (e.g., local emergency management centers, news agencies, and academic scrapers) to collaboratively train robust misinformation detection models on local data without sharing raw text, thus preserving user privacy. To secure the federated learning process against Byzantine model-poisoning and data-poisoning attacks, we anchor the model aggregation within a consortium blockchain ledger governed by a consensus-driven verification protocol. We implement a Proof-of-Accuracy (PoAC) consensus mechanism and smart-contract-based validation that evaluates client model updates against isolated, validated local validation sets, dynamically computing client reputation scores. Experimental evaluations using simulated benchmarks modeled after CrisisLexT26 and PHEME datasets show that DeFL-Crisis achieves a global verification accuracy of 94.2%, which is within 1.3% of the centralized training upper bound. Under hostile scenarios where 40% of clients perform aggressive model-poisoning attacks, DeFL-Crisis preserves high classification performance (F1-score of 0.88), whereas standard federated learning collapses to an F1-score of 0.45. Furthermore, we demonstrate that blockchain transaction latency is scalable (under 50 seconds for up to 50 active scaling nodes) and smart contract gas costs are highly optimized, indicating the realworld viability of our framework.