用GPT-4o检测图文假信息，验证其效果与稳定性。
Effectiveness of Large Multimodal Models in Detecting Disinformation: Experimental Results
- 设计优化提示词与结构化分析框架，适配模型输入限制。
- 在5个数据集上测试，发现模型对图文假信息识别有较好表现但不稳定。
- 引入置信度与重复测试机制，提升评估可靠性，适合安全研究者参考。
虚假信息在融合文本与图像的多模态场景中广泛传播，给数字平台带来重大挑战。本研究探讨大型多模态模型（LMMs）在检测和缓解虚假信息方面的潜力。通过利用GPT-4o的先进能力，提出一种优化提示策略，结合先进的提示工程技术，确保评估的精确性与一致性；构建结构化多模态分析框架，包含符合模型令牌限制的图文预处理方法；定义六项具体评估标准，实现内容的细粒度分类，并引入基于置信度的自评估机制；在多个异构数据集（Gossipcop、Politifact、Fakeddit、MMFakeBench、AMMEBA）上进行全面性能分析，揭示GPT-4o在假信息检测中的优势与局限；通过重复测试考察预测变异性，评估模型分类的稳定性和可靠性；并提出基于置信度与变异性双重指标的评估方法。这些贡献构建了一个可复现、可靠的自动化多模态假信息分析框架。
原文摘要 · Abstract (English)
The proliferation of disinformation, particularly in multimodal contexts combining text and images, presents a significant challenge across digital platforms. This study investigates the potential of large multimodal models (LMMs) in detecting and mitigating false information. We propose to approach multimodal disinformation detection by leveraging the advanced capabilities of the GPT-4o model. Our contributions include: (1) the development of an optimized prompt incorporating advanced prompt engineering techniques to ensure precise and consistent evaluations; (2) the implementation of a structured framework for multimodal analysis, including a preprocessing methodology for images and text to comply with the model's token limitations; (3) the definition of six specific evaluation criteria that enable a fine-grained classification of content, complemented by a self-assessment mechanism based on confidence levels; (4) a comprehensive performance analysis of the model across multiple heterogeneous datasets Gossipcop, Politifact, Fakeddit, MMFakeBench, and AMMEBA highlighting GPT-4o's strengths and limitations in disinformation detection; (5) an investigation of prediction variability through repeated testing, evaluating the stability and reliability of the model's classifications; and (6) the introduction of confidence-level and variability-based evaluation methods. These contributions provide a robust and reproducible methodological framework for automated multimodal disinformation analysis.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。