Scholarlink

<p>Detecting fake news has gained significant attention as misinformation spreads rapidly across social media, often supported by misleading images and texts. Prior methods mainly rely on independent concatenation of unimodal features, which fails to capture the benefits of integrated multimodal information and often lacks specialized mechanisms for feature extraction and fusion. To tackle these issues, this paper proposes an end-to-end multimodal framework that leverages BERT for textual features, ResNeXt for visual features, and BLIP and CLIP for multimodal feature extraction. A key novelty of the work lies in the use of self- and cross-attention–based fusion, which ensures deeper integration of modalities, yielding richer and more discriminative representations compared to simple concatenation. The proposed model is evaluated on Weibo, Gossipcop, and Fakeddit datasets. Results show competitive performance on Weibo and Gossipcop while achieving a notable improvement of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\sim\)</EquationSource> </InlineEquation> 7% in MCC on Fakeddit, demonstrating its effectiveness in handling diverse and realistic fake news scenarios.</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

Self and cross-modal attention based features fusion for fake news detection

  • 期刊论文

  • 发表日期：

    2026年1月20日

查看全文
  • Ajeet Kumar Gupta,
  • Maheshwari Prasad Singh

摘要

Detecting fake news has gained significant attention as misinformation spreads rapidly across social media, often supported by misleading images and texts. Prior methods mainly rely on independent concatenation of unimodal features, which fails to capture the benefits of integrated multimodal information and often lacks specialized mechanisms for feature extraction and fusion. To tackle these issues, this paper proposes an end-to-end multimodal framework that leverages BERT for textual features, ResNeXt for visual features, and BLIP and CLIP for multimodal feature extraction. A key novelty of the work lies in the use of self- and cross-attention–based fusion, which ensures deeper integration of modalities, yielding richer and more discriminative representations compared to simple concatenation. The proposed model is evaluated on Weibo, Gossipcop, and Fakeddit datasets. Results show competitive performance on Weibo and Gossipcop while achieving a notable improvement of \(\sim\) 7% in MCC on Fakeddit, demonstrating its effectiveness in handling diverse and realistic fake news scenarios.