Scholarlink

<p>In the digital era, misinformation has significantly threatened public stability and collective reasoning. While multimodal misinformation detection systems have demonstrated effectiveness, their real-time applicability remains constrained due to heavy computational demands and a lack of interpretability. This paper proposes a novel explainable framework, Multimodal Transfer Learning for Fake News Detection with Explainability (MTLFND-X), which integrates attention-based multimodal fusion with visual and textual explanation modules. The model uses RoBERTa and ResNet50 for efficient text and image representation, respectively, with an attention mechanism for adaptive fusion. Additionally, explainability is embedded via Grad-CAM for images, TimeSformer for Video and token-level attention maps for text, enabling human-understandable insights into model predictions. Evaluation across benchmark datasets such as Gossipcop, Weibo, Fakeddit, and Politifact shows competitive accuracy, precision, recall, and F1-score. This model combines robustness, real-time feasibility, and interpretability, making it suitable for deployment in high-stakes environments like media monitoring and misinformation alert systems.</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

Designing Lightweight Multimodal Models for Real-Time Fake News Classification

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Sudha Patel,
  • Shivangi Surati

摘要

In the digital era, misinformation has significantly threatened public stability and collective reasoning. While multimodal misinformation detection systems have demonstrated effectiveness, their real-time applicability remains constrained due to heavy computational demands and a lack of interpretability. This paper proposes a novel explainable framework, Multimodal Transfer Learning for Fake News Detection with Explainability (MTLFND-X), which integrates attention-based multimodal fusion with visual and textual explanation modules. The model uses RoBERTa and ResNet50 for efficient text and image representation, respectively, with an attention mechanism for adaptive fusion. Additionally, explainability is embedded via Grad-CAM for images, TimeSformer for Video and token-level attention maps for text, enabling human-understandable insights into model predictions. Evaluation across benchmark datasets such as Gossipcop, Weibo, Fakeddit, and Politifact shows competitive accuracy, precision, recall, and F1-score. This model combines robustness, real-time feasibility, and interpretability, making it suitable for deployment in high-stakes environments like media monitoring and misinformation alert systems.