Scholarlink

<p>Deepfake technology is growing more and more hazardous because it makes it easier to create incredibly realistic digital content that distorts reality and spreads fabricated data. Deepfake detection is still an exceptionally challenging endeavor, and cutting-edge studies are needed to improve the precision and reliability of these detection methods. In existing CNN-based Deep Fake detection tasks, CNN may fail to detect the most realistic deep fake images, since CNN is constrained by their reliance on convolutions that focus on local neighborhoods. Usually, the transformers can detect global and local features in images, unlike CNN, whose self-attention mechanism can detect subtle discrepancies in presentation typical of deepfake manipulation, improving overall detection accuracy. As a result, a transformer variant, CSWin transformer is utilized in this research to detect deepfake images. Additionally, CSWin transformers' cross-shaped window attention mechanism contributes to a better trade-off between computing efficiency and the ability to represent long-range relationships. In this study, significant CNN models like MTCNN, InceptionV3, and Xception are implemented to showcase the superior performance of the CSWin transformer. The Deep Fake Face Detection dataset is utilized in this research to build CNN models and Transformers, where the CSWin transformer performed better, achieving an Accuracy and F1-score of 98.7% and 98.72%, respectively.</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

Building an efficient Deep Fake detection system using the recognition capabilities of convolutional neural networks and transformers

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年5月29日

查看全文
  • Atulya Prabhanjan Magesh,
  • Siva Senthil Manikkam Ramakrishnan,
  • R. Arumuga Arun,
  • N. Priyanka,
  • Mukku Nisanth Kartheek

摘要

Deepfake technology is growing more and more hazardous because it makes it easier to create incredibly realistic digital content that distorts reality and spreads fabricated data. Deepfake detection is still an exceptionally challenging endeavor, and cutting-edge studies are needed to improve the precision and reliability of these detection methods. In existing CNN-based Deep Fake detection tasks, CNN may fail to detect the most realistic deep fake images, since CNN is constrained by their reliance on convolutions that focus on local neighborhoods. Usually, the transformers can detect global and local features in images, unlike CNN, whose self-attention mechanism can detect subtle discrepancies in presentation typical of deepfake manipulation, improving overall detection accuracy. As a result, a transformer variant, CSWin transformer is utilized in this research to detect deepfake images. Additionally, CSWin transformers' cross-shaped window attention mechanism contributes to a better trade-off between computing efficiency and the ability to represent long-range relationships. In this study, significant CNN models like MTCNN, InceptionV3, and Xception are implemented to showcase the superior performance of the CSWin transformer. The Deep Fake Face Detection dataset is utilized in this research to build CNN models and Transformers, where the CSWin transformer performed better, achieving an Accuracy and F1-score of 98.7% and 98.72%, respectively.