Scholarlink

<p>The rapid spread of deepfake technology poses significant challenges for media authenticity. In this study, we evaluate the performance of four CNN architectures —XceptionNet, VGG16, ResNet50, and DenseNet —for deepfake detection. XceptionNet achieved the highest overall accuracy, with training, validation, and test accuracies of 92.37%, 87.14%, and 87.48%, respectively; however, it displayed low precision (0.06) for detecting fake content. DenseNet attained moderate accuracy (85.61%) but exhibited the weakest classification metrics, with a precision of 0.11 and recall of 0.55 for real content. VGG16 achieved a test accuracy of 83.54%, exhibiting high recall (0.96) for genuine instances but low precision (0.11) for fraudulent content, suggesting a bias towards real data. Although test accuracy of ResNet, 82.90% is low compared to others, ResNet50 showcased its performance in terms of precision and recall,. These findings underscore the trade-offs between accuracy and precision in deepfake detection, highlighting the need for further improvements in CNNs. Overall, our results emphasize the importance of ongoing innovation in detection.</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

Deepfakes detection using convolutional neural networks

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Priyanshu Srivastva,
  • Prabadevi B

摘要

The rapid spread of deepfake technology poses significant challenges for media authenticity. In this study, we evaluate the performance of four CNN architectures —XceptionNet, VGG16, ResNet50, and DenseNet —for deepfake detection. XceptionNet achieved the highest overall accuracy, with training, validation, and test accuracies of 92.37%, 87.14%, and 87.48%, respectively; however, it displayed low precision (0.06) for detecting fake content. DenseNet attained moderate accuracy (85.61%) but exhibited the weakest classification metrics, with a precision of 0.11 and recall of 0.55 for real content. VGG16 achieved a test accuracy of 83.54%, exhibiting high recall (0.96) for genuine instances but low precision (0.11) for fraudulent content, suggesting a bias towards real data. Although test accuracy of ResNet, 82.90% is low compared to others, ResNet50 showcased its performance in terms of precision and recall,. These findings underscore the trade-offs between accuracy and precision in deepfake detection, highlighting the need for further improvements in CNNs. Overall, our results emphasize the importance of ongoing innovation in detection.