A BLOCKCHAIN-ENHANCED RETRIEVAL-AUGMENTED GENERATION FRAMEWORK FOR DISINFORMATION DETECTION
DOI:
https://doi.org/10.26577/JMMCS201211426Keywords:
claim veracity assessment, disinformation, blockchain, retrieval-augmented classification, DistilBERT, FAISS, LIAR dataset, ablation study, audit trail, transformer models.Abstract
Automated assessment of claim veracity remains an open problem in natural language processing, particularly for political statements whose truthfulness is ambiguous, contextual and subject to expert disagreement. This paper presents VeraChain, a hybrid system that combines a fine-tuned DistilBERT classifier, a dual-index retrieval stage (retrieval-augmented classification in the sense of RAG; no text is generated), and a hash-chained ledger that records every verdict together with its evidence and, for borderline predictions, supplies the verdicts of similar past statements as a consensus signal. Both FAISS indices and the ledger are built from the LIAR training split only, and the ledger is frozen for the main test evaluation. A stepwise ablation over five random seeds on the 1,283 LIAR test statements shows that DistilBERT alone reaches 64.36 ±1.25 percent accuracy; adding a single index of REAL statements lowers accuracy by 0.75 percentage points, adding a second index of FAKE statements recovers this loss (+0.87 points, positive in four of five seeds), whereas the probability-ratio score and the ledger consensus step add no consistent gain (−0.05 and −0.11 points). With the original seed, VeraChain reaches 65.24 percent against 64.77 percent for the same system without the consensus step (exact McNemar p = 0.36; 95 percent bootstrap interval of the difference [−0.39,1.33] points). Retrieval fusion lowers the expected calibration error of DistilBERT from 0.112 to 0.045, although temperature scaling of DistilBERT alone gives 0.031. On a single laptop GPU, the median processing time is 17.5 milliseconds per statement (95th percentile 20.9 ms), including the write of the ledger entry to disk. The results show that a hash-chained ledger can provide tamper-evident audit records for automated veracity assessment at a small computational cost, whereas its use as a consensus signal does not improve accuracy on LIAR.











