Scholarlink

<p>We introduce a unified multimodal framework for cross-platform fake news spread-chain detection. The approach jointly models multimodal content and multi-platform propagation to trace misinformation flows across social ecosystems. It integrates multimodal large models for semantic encoding under content transformations, cross-platform alignment to link related posts, typed graph construction combining in-platform and cross-platform relations, and chain-level classification via relation-aware graph reasoning. Evaluated on six benchmarks (FakeNewsNet, Fakeddit, PHEME, Twitter15/16, MCFEND) under a leave-one-platform-out protocol, our method outperforms state-of-the-art baselines by +3.5% Macro-F1, achieves robust chain detection (edge F1: 0.83, purity: 0.84), and demonstrates strong generalisation across platforms, reducing performance drop from 15% to 8%. Ablation studies validate each component’s contribution. The framework provides explainable, chain-level evidence that supports downstream moderation and forensic analysis. This work advances misinformation detection beyond single-platform, post-level analysis toward integrated, traceable, and platform-agnostic monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A unified multimodal large model framework for cross-platform fake news spread chain detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月27日

查看全文
  • Jingjing He

摘要

We introduce a unified multimodal framework for cross-platform fake news spread-chain detection. The approach jointly models multimodal content and multi-platform propagation to trace misinformation flows across social ecosystems. It integrates multimodal large models for semantic encoding under content transformations, cross-platform alignment to link related posts, typed graph construction combining in-platform and cross-platform relations, and chain-level classification via relation-aware graph reasoning. Evaluated on six benchmarks (FakeNewsNet, Fakeddit, PHEME, Twitter15/16, MCFEND) under a leave-one-platform-out protocol, our method outperforms state-of-the-art baselines by +3.5% Macro-F1, achieves robust chain detection (edge F1: 0.83, purity: 0.84), and demonstrates strong generalisation across platforms, reducing performance drop from 15% to 8%. Ablation studies validate each component’s contribution. The framework provides explainable, chain-level evidence that supports downstream moderation and forensic analysis. This work advances misinformation detection beyond single-platform, post-level analysis toward integrated, traceable, and platform-agnostic monitoring.