军事知识服务平台
基于大语言模型的双路证据驱动多模态虚假信息检测框架
A Dual-Path,Evidence-Driven Multimodal Misinformation Detection Framework Based on Large Language Models
摘要 传统多模态虚假信息检测难以提供清晰的证据链与可解释判定过程,文章旨在探索一种兼具检测性能、解释能力与落地可行性的多模态虚假信息检测方法.针对多模态虚假信息中"图文均为假""图像单模态造假""文本单模态造假"以及"图文不匹配"等四类典型形式,文章提出一种基于大语言模型的双路证据驱动多模态虚假信息检测框架.该框架依托大语言模型强大的理解能力、推理能力与生成能力,从文本真伪判定、图像真伪判定和"文本-图像"一致性匹配三个层面构建完整核查流程,通过"文本-图像"双路证据获取与联合推理实现对多模态信息的真伪判别.性能指标和案例分析结果表明,提出的框架能够在真实社交媒体场景下有效识别多模态虚假信息,并为之提供结构清晰、逻辑性强的检测解释依据.该研究为舆情监测提供了一种具备实时性、可落地的多模态虚假信息检测方法,也验证了大语言模型在多模态信息检测中的应用潜力.
Abstract:
Traditional multimodal misinformation detection struggles to provide a clear chain of evidence and an interpretable decision-making process.In light of this,this paper aims to explore a multimodal misinformation detection method that combines detection performance,interpretability and practical feasibility.Addressing four typical forms of multimodal misinformation—namely,"both text and images are fake""single-modal image forgery""single-modal text forgery"and"text-image mismatch"—this paper proposes a dual-path,evidence-driven multimodal misinformation detection framework based on large language models.Leveraging the powerful comprehension,reasoning and generation capabilities of large language models,this framework constructs a comprehensive verification process across three levels:text authenticity assessment,image authenticity assessment and"text-image"consistency matching.It achieves the authentication of multimodal information through the acquisition of"text-image"dual-path evidence and joint reasoning.Performance metrics and case study results demonstrate that the proposed framework can effectively identify multimodal misinformation in real-world social media scenarios,whilst providing clear and logically sound explanations for its detection.This research offers a real-time,practical method for detecting multimodal misinformation in public opinion monitoring,whilst also validating the application potential of large language models in multimodal information detection.
刊 名 图书馆论坛 ISTICPKU
年,卷(期): 2026, 46(11)
机标分类号: TP391 G250.2 TP181
在线出版日期: 2026年11月11日