Scholarlink

<p>In recent years, social media platforms have become a primary source for news consumption, often featuring articles paired with images or videos. However, the widespread dissemination of fake news on these platforms poses significant risks to individuals and society, driving the need for robust multimodal fake news detection models. These models, typically reliant on supervised learning, are highly sensitive to the size and distribution of their training datasets. A critical challenge in this domain is the prevalence of small or imbalanced datasets, particularly within fine-grained categories, which can lead to issues such as overfitting, underfitting, misclassification, and diminished accuracy. To address these challenges, we propose a novel data augmentation technique termed DA-LGS, which leverages Generative Adversarial Networks. Specifically, we employ LeakGAN for generating textual content and a combination of GAN and StackGAN for producing image samples. This approach aims to expand datasets and correct class imbalances while preserving the essential correlations between text and images. We evaluated our method using the Fakeddit multimodal, fine-grained dataset, assessing performance through precision, recall, and F1-score metrics for individual classes, alongside a macro-average for overall accuracy. Our findings demonstrate that the DA-LGS method effectively mitigates overfitting and significantly enhances classification accuracy across various classes, leading to a substantial improvement in overall model performance. Compared to existing state-of-the-art methods, our model achieved superior detection accuracy, recording a notable accuracy rate of 90.1%, which is a 2.9% improvement over the best-performing baseline model.</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 data augmentation approach based on various GAN models to address class imbalance in fine-grained multimodal fake news datasets

  • 期刊论文

  • 发表日期：

    2025年1月15日

查看全文
  • Suhaib Kh. Hamed,
  • Mohd Juzaiddin Ab Aziz,
  • Mohd Ridzwan Yaakub

摘要

In recent years, social media platforms have become a primary source for news consumption, often featuring articles paired with images or videos. However, the widespread dissemination of fake news on these platforms poses significant risks to individuals and society, driving the need for robust multimodal fake news detection models. These models, typically reliant on supervised learning, are highly sensitive to the size and distribution of their training datasets. A critical challenge in this domain is the prevalence of small or imbalanced datasets, particularly within fine-grained categories, which can lead to issues such as overfitting, underfitting, misclassification, and diminished accuracy. To address these challenges, we propose a novel data augmentation technique termed DA-LGS, which leverages Generative Adversarial Networks. Specifically, we employ LeakGAN for generating textual content and a combination of GAN and StackGAN for producing image samples. This approach aims to expand datasets and correct class imbalances while preserving the essential correlations between text and images. We evaluated our method using the Fakeddit multimodal, fine-grained dataset, assessing performance through precision, recall, and F1-score metrics for individual classes, alongside a macro-average for overall accuracy. Our findings demonstrate that the DA-LGS method effectively mitigates overfitting and significantly enhances classification accuracy across various classes, leading to a substantial improvement in overall model performance. Compared to existing state-of-the-art methods, our model achieved superior detection accuracy, recording a notable accuracy rate of 90.1%, which is a 2.9% improvement over the best-performing baseline model.