Abstract
Detecting fake news in short videos is a complex challenge where deceptive intent is often masked by objective-looking content. This paper introduces OpEvFake, a framework that leverages Large Language Models (LLMs) to uncover "implicit opinions" regarding credibility and utilizes a Diffusion Model to simulate the dynamic interaction and evolution of these opinions across text, audio, and video modalities. Experimental results on the FakeSV dataset demonstrate that OpEvFake achieves state-of-the-art performance, significantly outperforming existing methods.
1. Introduction
The proliferation of short video platforms has created a new frontier for misinformation. Unlike text-based fake news, short video misinformation often employs sophisticated tactics, such as pairing factual footage with misleading titles or sensationalized transcripts. Current detection methods typically rely on static feature aggregation—extracting and concatenating features from different modalities. However, these approaches often fail to capture the subtle, "implicit" deceptive intent and ignore the dynamic nature of how information from different sources should interact to resolve inconsistencies.
2. Methodology: The OpEvFake Framework
The proposed OpEvFake model addresses these limitations through a three-stage process designed to move from raw feature extraction to dynamic "opinion" formation and fusion.
2.1 Implicit Opinion Prompting
To capture deeper insights, the framework uses GPT-3.5 as a reasoning engine. By feeding the model text data (titles, transcripts, comments) via a customized Chain-of-Thought (CoT) prompt, the system generates structured "implicit opinions" on fact accuracy, source credibility, and language style. This transforms raw text into high-level reasoning features.
2.2 Intra-modal Opinion Enhancement
Before cross-modal interaction, the signal within each modality is strengthened. Multimodal Transformers are used for alignment, while Capsule Networks aggregate dispersed clues, such as identifying the most suspicious frames within a video sequence.
2.3 Cross-modal Opinion Interaction via Diffusion
The core novelty of OpEvFake lies in treating modality-specific features as "opinions" that evolve through interaction. A Diffusion Model is employed to simulate this process.
- Forward Diffusion: Gradually introduces noise into the initial single-modal opinions.
- Backward Diffusion (Denoising): The model learns to reconstruct the original opinions guided by a global "multimodal opinion." This process forces modalities to reconcile, effectively diffusing deceptive signals from one modality into the overall assessment.
3. Experimental Evaluation
The model was evaluated on the FakeSV dataset, a comprehensive benchmark containing 3,624 entries from platforms like Douyin.
3.1 Performance Results
OpEvFake achieved a state-of-the-art accuracy of 88.01% and an F1-score of 87.80%. This represents a significant improvement over previous benchmarks like SV-FEND (81.92%) and FANVM (82.84%). Notably, the framework outperformed direct classification by GPT-3.5 (~56% accuracy), highlighting the necessity of a specialized multimodal architecture.
3.2 Ablation Study
Ablation tests confirmed the importance of the LLM-generated insights. Removing the "Implicit Opinion" prompts resulted in a performance drop from 88.01% to 83.95%, proving that the reasoning provided by the LLM is a critical component of the detection process.
4. Conclusion
OpEvFake shifts the paradigm of fake news detection from static feature fusion to dynamic opinion evolution. By integrating LLM-based reasoning with a diffusion-driven interaction mechanism, the framework effectively unmasks deceptive intent in short videos. Future work will focus on improving the quality control of LLM-generated opinions and exploring more robust integration methods.
Summary updated