Research map / Vision-language and multimodal / Human and embodied modeling
Unveiling Implicit Deceptive Patterns in Multi-Modal Fake News via Neuro-Symbolic Reasoning
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 NSLM, a neuro-symbolic latent variable model that detects multi-modal fake news while inferring three predefined deceptive patterns (image manipulation, cross-modal inconsistency, image repurposing) as explanations, weakly supervised by first-order logic rules via distillation.
Problem
Multi-modal fake news detection lacks explainability; existing explanations highlight input components or views but do not uncover the implicit deceptive patterns by which fake news is fabricated.
Approach
Model each predefined deceptive pattern's existence as a two-valued latent variable inferred by amortized variational inference with a pseudo-siamese encoder, and weakly supervise these latents with soft first-order logic rules (fake iff some pattern exists) through teacher-student distillation.
Key results
- NSLM reaches 0.925 accuracy and 0.917 macro F1 on Fakeddit, versus 0.909 accuracy for the best baseline CMC.
- On Weibo, NSLM reaches 0.885 accuracy and 0.884 macro F1, a 1.0-point accuracy gain over CMC (0.875).
- Ablations show removing any single deceptive pattern lowers accuracy, and removing two patterns hurts more than one.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NSLM instead of CMC | Fake news detection · Fakeddit | Accuracy: 0.909 -> 0.925 | Helped |
| use NSLM instead of CMC | Fake news detection · Weibo | Accuracy: 0.875 -> 0.885 | Helped |
| add image manipulation latent variable (zIM) (ablation) | NSLM · Fakeddit | Accuracy: 0.923 -> 0.925 | Helped |
| add image manipulation latent variable (zIM) (ablation) | NSLM · Weibo | Accuracy: 0.864 -> 0.885 | Helped |
| add cross-modal inconsistency latent variable (zCI) (ablation) | NSLM · Fakeddit | Accuracy: 0.922 -> 0.925 | Helped |
| add cross-modal inconsistency latent variable (zCI) (ablation) | NSLM · Weibo | Accuracy: 0.874 -> 0.885 | Helped |
| add image repurposing latent variable (zIR) (ablation) | NSLM · Fakeddit | Accuracy: 0.923 -> 0.925 | Helped |
| add image repurposing latent variable (zIR) (ablation) | NSLM · Weibo | Accuracy: 0.874 -> 0.885 | Helped |
| add image manipulation and cross-modal inconsistency latent variables (zIM, zCI) (ablation) | NSLM · Fakeddit | Accuracy: 0.918 -> 0.925 | Helped |
| add image manipulation and cross-modal inconsistency latent variables (zIM, zCI) (ablation) | NSLM · Weibo | Accuracy: 0.865 -> 0.885 | Helped |
| add image manipulation and image repurposing latent variables (zIM, zIR) (ablation) | NSLM · Fakeddit | Accuracy: 0.921 -> 0.925 | Helped |
| add image manipulation and image repurposing latent variables (zIM, zIR) (ablation) | NSLM · Weibo | Accuracy: 0.865 -> 0.885 | Helped |
Surprising findings
- Removing the cross-modal inconsistency pattern hurts most on Fakeddit while removing image manipulation hurts most on Weibo, suggesting dataset-dependent pattern distributions.
- Overall accuracy is robust to the logic-loss weight mu, while pattern-quality metrics vary strongly with it.
Limitations
- Relies on Google reverse image search for image repurposing context, which can return poor contexts and cause errors.
- Deceptive pattern set is predefined to three patterns and cannot adapt to new deception types.
- Lacks external knowledge for verifying facts, causing failures on some cases.
Experimental setup
Datasets: Fakeddit Weibo
Benchmarks: Fakeddit Weibo
Metrics: Accuracy Precision Recall Macro F1
Baselines: ResNet34 InceptionV3 RoBERTa EANN SpotFake BTIC HMCAN CAFE CMC BMR LogicDM
Builds on: variational auto-encoder logic rule distillation (Chen et al. 2022a) knowledge distillation (Hu et al. 2016) BiLinear similarity
Closest papers by meaning
- DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025 · similarity 0.86)
- Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.86)
- Text-Guided Fine-grained Counterfactual Inference for Short Video Fake News Detection (AAAI 2025 · similarity 0.86)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.86)
- Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference (AAAI 2025 · similarity 0.85)
- Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection (AAAI 2025 · similarity 0.85)
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.