Scholarlink

The increase in deepfake technology through synthetic facial manipulation using deep learning raises significant privacy concerns, misinformation risks, and trust issues. This research analyzes deepfake face detection using Convolutional Neural Networks (CNNs) and Transformer based architectures. We employ VGG19, ResNet50, and DenseNet121 alongside Vision Transformer (ViT) and Swin Transformer to distinguish fake from real facial images. Performance evaluation includes accuracy, precision, recall, F1-score, and ROC AUC metrics, supported by confusion matrix analysis. While traditional CNNs showed moderate accuracy, Transformer-based models, particularly Vision Transformer and Swin Transformer, achieved exceptional results—85%–86% accuracy and AUC scores of 0.93–0.92, respectively. Attention-based models prove highly effective in capturing subtle manipulations present in deepfake content. The proposed detection system can be integrated into real-time media verifiers, content moderation tools, and forensic software to strengthen digital security and combat misinformation.

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

Deepfake Face Detection Using CNN and Transformer Architectures for Enhanced Digital Security

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Swari Patel,
  • Roma Kataria,
  • Jaiprakash Verma,
  • Sumedha Arora

摘要

The increase in deepfake technology through synthetic facial manipulation using deep learning raises significant privacy concerns, misinformation risks, and trust issues. This research analyzes deepfake face detection using Convolutional Neural Networks (CNNs) and Transformer based architectures. We employ VGG19, ResNet50, and DenseNet121 alongside Vision Transformer (ViT) and Swin Transformer to distinguish fake from real facial images. Performance evaluation includes accuracy, precision, recall, F1-score, and ROC AUC metrics, supported by confusion matrix analysis. While traditional CNNs showed moderate accuracy, Transformer-based models, particularly Vision Transformer and Swin Transformer, achieved exceptional results—85%–86% accuracy and AUC scores of 0.93–0.92, respectively. Attention-based models prove highly effective in capturing subtle manipulations present in deepfake content. The proposed detection system can be integrated into real-time media verifiers, content moderation tools, and forensic software to strengthen digital security and combat misinformation.