Debunk and Infer: Multimodal Fake News Detection via Diffusion-Generated Evidence and LLM Reasoning

Authors: Kaiying Yan, Moyang Liu, Yukun Liu, Ruibo Fu, Zhengqi Wen, Jianhua Tao, Xuefei Liu Venue: arXiv preprint, June 2025 — arXiv:2506.21557

TL;DR

This paper introduces DIFND, a framework for multimodal fake news detection that generates refuting/authenticating evidence via diffusion models and performs reasoning over multimodal content using large language models (LLMs). The approach uses a chain-of-debunk strategy where multiple MLLMs act as collaborative agents to assess news authenticity across text, audio, and visual modalities, achieving notable improvements on FakeSV and FVC datasets.

Contributions

1. DIFND framework — A debunk-and-infer approach integrating diffusion-based evidence generation with multimodal LLM reasoning for fake news detection.
2. Debunk Diffusion mechanism — A conditional latent diffusion model that generates textual debunking evidence from multimedia content, trained to produce both refuting and authenticating statements.
3. Chain of Debunk (CoD) — A multi-agent reasoning strategy where multiple MLLMs independently analyze different modalities (text, video keyframes, audio), then synthesize their conclusions for robust veracity judgment.
4. Empirical validation — Demonstrates 2.33% and 2.88% accuracy improvements over existing multimodal methods on FakeSV and FVC datasets respectively, with ablations showing the contribution of each component.

Method

The framework consists of two main stages:

Debunk Diffusion

Rather than merely refuting false claims, the paper proposes a broader concept of debunking that encompasses both refutation and authentication of true information. A conditional latent diffusion model generates textual evidence from learned debunking feature distributions:

1. Compression Network — Transforms high-dimensional text features into compact latent space optimized for diffusion
2. Conditional Latent Diffusion — Forward process adds noise to debunk features guided by multimodal conditioning (text embeddings, vision features, audio features). Reverse process denoises iteratively to reconstruct debunking evidence.
3. Multimodal Conditioning — Uses Sinusoidal positional encoding to embed multimodal conditions so they guide the denoising process
4. Self-Refinement Network — Filters noise and refines generated debunking features

Chain of Debunk (CoD)

Multiple MLLMs are deployed as collaborative agents analyzing the news item from different perspectives:

1. Text Analysis Agent — Examines on-screen text, linguistic coherence, and factual consistency using a specialized LLM (Deepseek-R1)
2. Video Understanding Agent — Processes video keyframes and conducts visual analysis for synthetic media or manipulation detection
3. Audio Analysis Agent — Analyzes audio properties using GLM-4-Voice to capture semantic meaning and emotional tone
4. Comprehensive Reasoning — Synthesizes outputs from all agents via multi-layer fusion and attention mechanisms; integrates cross-modal consistency checks

The outputs from debunk diffusion and chain-of-debunk are fused via an attention-based mechanism to produce the final veracity judgment.

Results

Extensive experiments on two datasets show:

FakeSV dataset (Chinese fake news videos): - DIFND: 87.68% accuracy, 87.34% F1, 87.02% recall, 87.87% precision - Improvement over best baseline (Fakinrecipe): +2.33% accuracy

FVC dataset (multilingual videos with diverse content): - DIFND: 92.16% accuracy, 92.01% F1, 92.21% recall, 91.88% precision - Improvement over best baseline (MMSD): +2.88% accuracy

Ablation studies (Table II): - Diffusion Diffusion alone: 75.70% accuracy on FakeSV - Adding Chain of Debunk: 86.21% (10.51% improvement) - Full model with all components: 87.68% - Debunk data augmentation contributes 1.89% improvement, evidencing the utility of LLM-generated evidence

All components contribute meaningfully; visual and audio modalities provide complementary signals to text-based reasoning, with visual cues most influential.

Connections

• Related to Multimodal fake news detection via fusion of text, visual, and audio modalities for veracity assessment
• Builds on Large Language Models by deploying MLLMs as collaborative reasoning agents
• Extends Evidence-based fake news detection by generating debunking evidence synthetically via diffusion models rather than relying solely on external sources
• Related to Debunking and correction interventions through its focus on generating refuting and authenticating evidence
• Focuses on video-based fake news detection through explicit video frame and audio analysis

Notes

Strengths: - Novel application of diffusion models to generate debunking evidence in a controlled, multimodal setting - Thoughtful decomposition of reasoning tasks across multiple agents with specialized prompts for each modality - Comprehensive ablation studies quantifying the contribution of diffusion, each modality, and data augmentation - Consistent improvements on two diverse datasets (Chinese and multilingual) suggest generalizability

Considerations: - Relies on access to multiple specialized LLMs (Deepseek-R1, GLM-4-Voice) which may limit reproducibility and deployment scalability - Computational cost not discussed; inference time across multiple agent executions likely substantial compared to single-model baselines - Evaluation limited to two datasets; cross-dataset generalization (especially to Western platforms like Twitter/X) not demonstrated - Chain-of-debunk generates interpretable reasoning chains, but the paper does not analyze failure modes or compare explanation quality against simpler baselines