Fake news detection methods

Fake news detection encompasses a range of computational approaches operating at different levels of analysis, from low-level linguistic features to high-level propagation patterns and user behavioral signals.

Taxonomy

Approaches broadly divide into four families based on what signal they exploit:

• Content-based detection: Text-only features (linguistic style, semantic coherence, readability), visual features, or multimodal combinations. Early work used hand-engineered features (LIWC, word n-grams); recent work uses neural representations (BERT, CNNs).
• Propagation-based detection: Temporal and structural patterns in how claims spread through social networks. Fake news often spreads faster, wider, and deeper; exhibits distinct bot participation patterns; shows characteristic engagement signatures.
• User-based detection: Demographic, network, and behavioral features of users who share false claims. Key finding: account age, follower count, and activity patterns predict sharing likelihood.
• External knowledge-based detection: Fact-checking claims against knowledge bases or retrieving related verified claims.
• AI-generated content detection: Identifying text produced by language models and other generative systems to detect synthetic disinformation and fraudulent content. Methods include statistical analysis, watermarking, rewriting-based detection, and neural classifiers.

Within these families, methods differ in scope: early detection (hours after posting) vs. full lifecycle, single-domain (trained and tested on one topic, e.g., politics) vs. cross-domain (generalization across topics), and supervised (requires labeled training data) vs. unsupervised (domain discovery, anomaly detection).

Key papers

Core references by family:

• On the Opportunities and Risks of Foundation Models — Discusses foundation models as powerful detectors of harmful content including misinformation and deepfakes; analyzes the arms race between generative and detection capabilities.
• Automated identification of media bias in news articles: an interdisciplinary literature review: Literature review mapping social science approaches to media bias analysis onto computational methods; covers content analysis, frame analysis, and automated NLP-based detection.
• Lazer et al. (2018) — Multidisciplinary overview framing detection as one of four complementary approaches (detection, inoculation, media literacy, platform design).
• Shu et al. (2017) — Survey of social media detection methods across content, user, network, and propagation.
• Sharma et al. (2018) — Complementary survey focusing on linguistic and stylistic approaches.
• Oshikawa et al. (2020) — Survey of NLP and deep learning methods.

Specific methodological innovations:

• Knowledge-based / Evidence-based: The Fact Extraction and VERification (FEVER) Shared Task — evidence retrieval from Wikipedia + natural language inference for claim verification; KAN: Knowledge-aware Attention Network for Fake News Detection — uses knowledge graph entities and contexts with attention mechanisms for detection
• Content-based / Transformer embeddings: FakeBERT: Fake News Detection in Social Media with a BERT-based Deep Learning Approach — BERT + parallel 1D CNNs achieving 98.90% accuracy, demonstrating state-of-the-art performance of contextualized embeddings for text classification
• Multimodal: TI-CNN: Convolutional Neural Networks for Fake News Detection, EANN: Event Adversarial Neural Networks for Multi-Modal Fake News Detection, SAFE: Similarity-Aware Multi-Modal Fake News Detection, Embracing Domain Differences in Fake News: Cross-domain Fake News Detection using Multimodal Data, Multimodal Emergent Fake News Detection via Meta Neural Process Networks
• Propagation networks: Information Credibility on Twitter, CSI: A Hybrid Deep Model for Fake News Detection, Hierarchical Propagation Networks for Fake News Detection: Investigation and Exploitation, Fake News Detection on Social Media using Geometric Deep Learning
• Early detection: Network-based Fake News Detection: A Pattern-driven Approach, Fake News Detection on Social Media using Geometric Deep Learning
• Transfer learning: EANN: Event Adversarial Neural Networks for Multi-Modal Fake News Detection
• Graph neural networks: Fake News Detection on Social Media using Geometric Deep Learning
• Ensemble methods with uncertainty: A Heuristic-driven Uncertainty based Ensemble Framework for Fake News Detection in Tweets and News Articles — soft-voting ensemble of pre-trained transformers (BERT, RoBERTa, XLNet, DeBERTa, ERNIE 2.0, ELECTRA) with Statistical Feature Fusion Network and uncertainty quantification via Monte Carlo Dropout; achieves F1=0.9892 on COVID-19 Fake News and F1=0.9156 on FakeNewsNet by combining text and metadata features
• Fake account detection: The Looming Threat of Fake and LLM-generated LinkedIn Profiles: Challenges and Opportunities for Detection and Prevention — Section and Subsection Tag Embedding (SSTE) method for detecting human-created and LLM-generated fake LinkedIn profiles; achieves 95% accuracy and 70%+ accuracy on unseen LLM-generated profiles
• AI-generated text detection: RAIDAR: Generative AI Detection via Rewriting — Raidar uses rewriting distance and structural properties (invariance, equivariance, output uncertainty) to distinguish human from machine-generated text; robust to adversarial prompts
• AI-generated text detection (comprehensive): Wu et al. (2023) — Systematic survey covering detection necessity (regulation, users, development, science, society), methods (watermarking, statistics-based, neural-based, human-assisted), datasets, evaluation metrics, and challenges like adversarial robustness and out-of-distribution generalization; Tang et al. (2023) — Survey of black-box and white-box detection approaches, including statistical analysis, linguistic features, fact verification, watermarking strategies, and adversarial attacks
• Neural-symbolic methods: TELLER: A Trustworthy Framework For Explainable, Generalizable and Controllable Fake News Detection — dual-system framework combining LLM-driven decomposition of claims into logical predicates (cognition) with interpretable rule learning via disjunctive normal form (decision); achieves 76% accuracy on GossipCop while maintaining explainability and human controllability

Connections

• Content-based detection is a major sub-category within the broader field.
• Multimodal fake news detection combines multiple modalities.
• Propagation-based fake news detection exploits network structure and temporal dynamics.