EAKD-CFND: continual fake news detection through adaptive knowledge distillation with external verification
Fake news spreads faster than detection models can adapt, and models updated on new data tend to forget what they learned before, an effect severe enough to make naive deployments impractical for continuous, streaming fake-news detection (FND). Continual-learning methods address this either by storing real past examples for replay, which is accurate but carries a memory and privacy cost many streaming deployments cannot accept, or by staying exemplar-free, which avoids that cost but can degrade badly under severe task drift; a deployer is typically forced to commit to one property or the other before knowing which their data will actually require. We introduce EAKD-CFND, a class-incremental continual-learning framework built around one mechanism, per-instance uncertainty-adaptive knowledge distillation with external fact-checking verification, offered in two configurations that share that mechanism: an exemplar-free default (EAKD-CFND) that never stores a past instance, and a rehearsal-augmented configuration (EAKD-CFND-R) that adds a small reservoir buffer for settings where storage is available and drift risk is high. We evaluate both configurations against five baselines (fine-tuning, EWC, LwF, the rehearsal-based DER, and the category-level adaptive-KD method LUD) under one identical protocol on three real class-incremental splits built from PHEME and FakeNewsNet, and measure calibration, hyperparameter sensitivity, and verification cost end to end. The exemplar-free configuration is statistically indistinguishable from every baseline tested, including rehearsal-based DER, on both FakeNewsNet splits, at zero storage cost. On PHEME’s more demanding five-task sequence, the exemplar-free configuration collapses to near-random accuracy alongside every other exemplar-free method tested, while DER’s stored exemplars carry it to 84.8%; switching to the rehearsal-augmented configuration closes that gap (85.5% on PHEME-Event, matching DER within noise), giving a working option for exactly the regime where the exemplar-free default falls short. The rehearsal-augmented configuration is not a single-benchmark fix: on FNN-Poli-Time it adds a further, measured gain over both the exemplar-free default and DER, while on FNN-Gossip-Time it adds nothing, a benchmark-specific split traced directly to how much genuine task drift each benchmark actually contains. The uncertainty signal itself is poorly calibrated in absolute terms (Expected Calibration Error 0.80–0.84 across four estimators), but the framework’s accuracy is robust to its two governing hyperparameters across a wide sweep, and its API cost is quantified rather than assumed. Together these results give practitioners a working choice between the two configurations, made on measured grounds rather than a qualitative argument: exemplar-free by default, rehearsal-augmented when storage is available and drift risk is severe.
Authors
- Aman Jolly (ORCID: https://orcid.org/0000-0003-2675-6722)
- Sankalp Paliwal (ORCID: https://orcid.org/0000-0002-3439-0102)
- Shailender Kumar
Institutions
- Manipal University Jaipur (IN)
- Delhi Technological University (IN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1038/s41598-026-69737-3
- Primary Topic
- Misinformation and Its Impacts
- Type
- article
- Field-Weighted Citation Impact
- 0.00