Scholarlink

<p>With the rising number of cases of financial fraud and political manipulation around, deepfake creation is getting sophisticated and exponentially increasing, researchers need to come up with remarkable innovations rapidly to detect this synthetic media. As traditional methods like CNN and ViTs fail to capture the minute difference between facial regions. There are optical inconsistencies as they ignore motion, and they only analyse rigid regions, hence struggles with pose variations. To overcome these problems, Spatio-Temporal Graph Transformer (STGT) is proposed using spatio-temporal graphs. For graph conversion, facial features are extracted dynamically along with optical flow features using Farneback algorithm. Facial region features are represented with nodes which are connected by temporal edges. The graph is sent to a 4X graph transformer, which uses dynamic attention pooling instead of traditional pooling for final prediction. The experiment is conducted using FaceForensics++ dataset consisting 1000 real and 1000 fake videos. STGT achieves 99% accuracy, which is a boost in accuracy compared to other state-of-the-art algorithm for detecting deepfake videos. This approach achieves 0.99 AUC by simultaneously analysing geometric, motion and hierarchical features.</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

STGT: Spatio-temporal graph transformer for robust deepfake detection using dynamic facial features

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Ayushi Pandey,
  • Arun Solanki

摘要

With the rising number of cases of financial fraud and political manipulation around, deepfake creation is getting sophisticated and exponentially increasing, researchers need to come up with remarkable innovations rapidly to detect this synthetic media. As traditional methods like CNN and ViTs fail to capture the minute difference between facial regions. There are optical inconsistencies as they ignore motion, and they only analyse rigid regions, hence struggles with pose variations. To overcome these problems, Spatio-Temporal Graph Transformer (STGT) is proposed using spatio-temporal graphs. For graph conversion, facial features are extracted dynamically along with optical flow features using Farneback algorithm. Facial region features are represented with nodes which are connected by temporal edges. The graph is sent to a 4X graph transformer, which uses dynamic attention pooling instead of traditional pooling for final prediction. The experiment is conducted using FaceForensics++ dataset consisting 1000 real and 1000 fake videos. STGT achieves 99% accuracy, which is a boost in accuracy compared to other state-of-the-art algorithm for detecting deepfake videos. This approach achieves 0.99 AUC by simultaneously analysing geometric, motion and hierarchical features.