Research map / Vision-language and multimodal / Cross-modal learning
RaCMC: Residual-Aware Compensation Network with Multi-Granularity Constraints for 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.
Proposes RaCMC, a multimodal fake news detector with residual-aware cross-modal compensation interaction/fusion and multi-granularity constraints (MMD news-level and cosine feature-level), evaluated on Weibo17, Politifact and GossipCop.
Problem
Multimodal fake news detection suffers from inadequate cross-modal feature fusion and fuzzy classification because differences between real and fake news are not amplified at the news and feature levels.
Approach
Combines masked cross-attention between coarse (CLIP) and fine (BERT/ResNet) granularity features with residual connections and iAFF fusion, plus interaction consistency/exclusivity KL and Frobenius losses, and multi-granularity constraints (MMD between real and fake news and cosine-similarity pulling/pushing of image-text pairs).
Key results
- RaCMC reaches 91.5% accuracy on Weibo17, versus 90.5% for the next-best FND-CLIP/MRAN-level baselines (RaCMC claims 1.0% gain over FND-CLIP on Weibo17).
- On Politifact RaCMC achieves 92.2% accuracy, and on GossipCop 87.9% accuracy, with the gain on GossipCop described as small.
- Ablations show removing the interaction constraints drops GossipCop accuracy by 3.4% and removing the news overall constraint drops Politifact accuracy by 3.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RaCMC instead of FND-CLIP | Weibo17 fake news detection · Weibo17 | accuracy: 0.905 -> 0.915 | Helped |
| use RaCMC instead of dEFEND | Politifact fake news detection · Politifact | accuracy: 0.904 -> 0.922 | Helped |
| use RaCMC instead of FND-CLIP | GossipCop fake news detection · GossipCop | accuracy: 0.876 -> 0.879 | Helped |
| add interaction constraints (l_ic) (ablation) | RaCMC · Weibo17, Politifact, and GossipCop | accuracy: 0.909 -> 0.915; 0.911 -> 0.922; 0.845 -> 0.879 | Helped |
| add news overall constraint (l_no) (ablation) | RaCMC · Weibo17, Politifact, and GossipCop | accuracy: 0.903 -> 0.915; 0.889 -> 0.922; 0.863 -> 0.879 | Helped |
| add news internal constraint (l_ni) (ablation) | RaCMC · Weibo17, Politifact, and GossipCop | accuracy: 0.898 -> 0.915; 0.900 -> 0.922; 0.867 -> 0.879 | Helped |
| use image and text features instead of image features only (ablation) | RaCMC · Weibo17, Politifact, and GossipCop | accuracy: 0.751 -> 0.915; 0.811 -> 0.922; 0.833 -> 0.879 | Helped |
| use image and text features instead of text features only (ablation) | RaCMC · Weibo17, Politifact, and GossipCop | accuracy: 0.890 -> 0.915; 0.867 -> 0.922; 0.862 -> 0.879 | Helped |
| use MRC instead of concatenating features from different granularities (ablation) | RaCMC · Weibo17, Politifact, and GossipCop | accuracy: 0.909 -> 0.915; 0.889 -> 0.922; 0.871 -> 0.879 | Helped |
| add MGC (ablation) | RaCMC · Weibo17, Politifact, and GossipCop | accuracy: 0.903 -> 0.915; 0.900 -> 0.922; 0.862 -> 0.879 | Helped |
| use DFR instead of concatenating unimodal and multimodal features (ablation) | RaCMC · Weibo17, Politifact, and GossipCop | accuracy: 0.910 -> 0.915; 0.867 -> 0.922; 0.867 -> 0.879 | Helped |
Surprising findings
- Image-only features degrade Weibo17 accuracy by 16.4 points, yet GossipCop improvements over FND-CLIP are small because nearly 50% of its images are celebrity faces hard to distinguish.
Limitations
- Gains on GossipCop are marginal relative to FND-CLIP.
- Baseline results on Politifact and GossipCop partly reproduced by the authors, with some reported as unavailable.
- Relies on a fixed set of frozen pretrained encoders (BERT, CLIP) and translation/summarization preprocessing.
Experimental setup
Datasets: Weibo17 Politifact GossipCop
Benchmarks: Weibo17 Politifact GossipCop
Metrics: accuracy precision recall F1
Baselines: MVAE SAFE CAFE FND-CLIP MPFN MRAN CSFND dEFEND LSTM-ATT Spotfake+ GNN DistilBert
Builds on: iAFF CLIP BERT
Closest papers by meaning
- Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection (AAAI 2025 · similarity 0.88)
- DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025 · similarity 0.87)
- Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.86)
- External Reliable Information-enhanced Multimodal Contrastive Learning for Fake News Detection (AAAI 2025 · similarity 0.86)
- Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference (AAAI 2025 · similarity 0.85)
- Text-Guided Fine-grained Counterfactual Inference for Short Video Fake News Detection (AAAI 2025 · similarity 0.84)
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.