Research map / Vision-language and multimodal / Human and embodied modeling
Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds AMG, a multimodal fake news dataset with five fine-grained attribution categories plus real news, and proposes MGCA, a multi-granularity clue alignment model for joint detection and attribution.
Problem
Existing multimodal fake news datasets provide only binary real/fake labels, lacking fine-grained attribution of how a fake post is fake (e.g., image fabrication, entity/event/time inconsistency), so detectors cannot explain error patterns.
Approach
Attribute each fake post to a fine-grained cause using expert annotation guided by fact-checking rulings, and detect/attribute using multi-view clues (entity, event, temporal, manipulation features) aligned via Compare-Net-style consistency modules fused with gated per-feature probabilities.
Key results
- MGCA achieves 0.8323 detection accuracy and 0.7385 attribution accuracy on AMG, versus 0.8079 and 0.6687 for the next best baseline BMR.
- On public datasets MGCA reaches F1 of 0.905 (Twitter), 0.899 (Weibo), and 0.913 (Weibo21 per its table), outperforming the compared baselines.
- Removing the event-level module reduces detection accuracy from 0.8323 to 0.7917.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MGCA instead of BMR | AMG detection · AMG | accuracy: 0.8079 -> 0.8323 | Helped |
| use MGCA instead of BMR | AMG detection · AMG | F1: 0.8057 -> 0.8310 | Helped |
| use MGCA instead of BMR | AMG attribution · AMG | accuracy: 0.6687 -> 0.7385 | Helped |
| use MGCA instead of BMR | AMG attribution · AMG | F1: 0.5193 -> 0.5666 | Helped |
| add PSCC-NET manipulation features (ablation) | MGCA · AMG | F1: 0.8166 -> 0.8310 | Helped |
| add entity clues (ablation) | MGCA · AMG | F1: 0.8138 -> 0.8310 | Helped |
| add event clues (ablation) | MGCA · AMG | F1: 0.7916 -> 0.8310 | Helped |
| add temporal clues (ablation) | MGCA · AMG | F1: 0.8085 -> 0.8310 | Helped |
| add visual entity module (vem) (ablation) | MGCA · AMG | F1: 0.8177 -> 0.8310 | Helped |
| add PSCC-NET manipulation features (ablation) | MGCA · AMG | F1: 0.4660 -> 0.5666 | Helped |
| add entity clues (ablation) | MGCA · AMG | F1: 0.4283 -> 0.5666 | Helped |
| add event clues (ablation) | MGCA · AMG | F1: 0.4542 -> 0.5666 | Helped |
Surprising findings
- BMR reaches about 90% accuracy on Weibo and GossipCop, while AMG detection accuracy stays below 81% for the same model family, suggesting AMG is harder.
- Attribution accuracy is high partly because the ground-truth attribution is only defined for fake news, yet many fake news posts are misclassified across categories such as entity and time inconsistency.
Limitations
- Dataset contains only post content, excluding comments and social network metadata.
- Video is handled by middle-frame extraction only, and image forgery detection via PSCC-NET covers only one manipulation type.
- Entity-inconsistency performance is limited by the accuracy of the visual entity recognition API.
Experimental setup
Datasets: AMG Twitter Weibo Weibo21
Benchmarks: AMG Twitter Weibo Weibo21
Metrics: accuracy F1 score precision recall
Baselines: MCAN CAFE BMR CLIP
Builds on: CLIP BERT PSCC-NET Compare-Net LLaVA
Closest papers by meaning
- DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025 · similarity 0.88)
- RaCMC: Residual-Aware Compensation Network with Multi-Granularity Constraints for Fake News Detection (AAAI 2025 · similarity 0.88)
- Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference (AAAI 2025 · similarity 0.88)
- External Reliable Information-enhanced Multimodal Contrastive Learning for Fake News Detection (AAAI 2025 · similarity 0.88)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.87)
- Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.