Scholarlink

<p>The growing accessibility of deepfake generation tools has introduced serious challenges to digital media security, as forged videos can realistically manipulate facial appearance, expressions, and speech. Most existing detection approaches are based on convolutional neural networks (CNNs) that analyze video frames independently, which results in two key drawbacks: repeated processing of visually similar frames and the loss of important spatial and pose-related facial information. To overcome these limitations, a novel fake face detection framework, termed SFDM-ViCapsNet, is proposed. A Statistical Fusion-Based Dynamic Margin (SFDM) is introduced as a robust face pre processing strategy that adaptively crops facial regions by fusing face area, detection confidence, and distance from image boundaries, enabling more reliable capture of manipulation-prone regions than fixed or confidence-only cropping methods. SSIM-based key-frame selection is employed to identify visually informative frames, followed by face detection using MTCNN. The SFDM-refined face crops are then processed by a VGG19 feature extractor and a capsule network with dynamic routing to preserve pose information and reveal structural inconsistencies. Experiments on FaceForensics++, DFDC, UADFV, and Celeb-DF-V1 show that emphasizing leads to competitive accuracy while reducing inference cost.</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

SFDM-ViCapsNet: using SSIM-guided frame selection and statistical fusion dynamic margin cropping to detect fake faces in videos

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • S. Immanuel,
  • M. Gethsiyal Augasta

摘要

The growing accessibility of deepfake generation tools has introduced serious challenges to digital media security, as forged videos can realistically manipulate facial appearance, expressions, and speech. Most existing detection approaches are based on convolutional neural networks (CNNs) that analyze video frames independently, which results in two key drawbacks: repeated processing of visually similar frames and the loss of important spatial and pose-related facial information. To overcome these limitations, a novel fake face detection framework, termed SFDM-ViCapsNet, is proposed. A Statistical Fusion-Based Dynamic Margin (SFDM) is introduced as a robust face pre processing strategy that adaptively crops facial regions by fusing face area, detection confidence, and distance from image boundaries, enabling more reliable capture of manipulation-prone regions than fixed or confidence-only cropping methods. SSIM-based key-frame selection is employed to identify visually informative frames, followed by face detection using MTCNN. The SFDM-refined face crops are then processed by a VGG19 feature extractor and a capsule network with dynamic routing to preserve pose information and reveal structural inconsistencies. Experiments on FaceForensics++, DFDC, UADFV, and Celeb-DF-V1 show that emphasizing leads to competitive accuracy while reducing inference cost.