Scholarlink

<p>The growing sophistication of deepfake videos threatens media authenticity, requiring detectors that are accurate and robust to real-world distortions. We design and compare two deep learning architectures for video-based detection. The baseline fuses ResNeXt50 with a unidirectional LSTM, while the proposed model combines EfficientNet-B0 with a Bidirectional LSTM and temporal attention to capture bidirectional dependencies and emphasize informative frames. Both are trained and evaluated on Celeb-DF v2 under identical settings for fair comparison. Results show the proposed model outperforms the baseline, reaching 95.3% accuracy and a ROC-AUC of 0.981. These findings highlight the value of efficient convolutional encoders with temporal attention for interpretable and competitive deepfake detection, with applications in media forensics and verification.</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

Leveraging temporal attention and bidirectional modeling for robust deepfake video detection

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Marwa Ben Jabra,
  • Omar Cheikhrouhou,
  • Anouar BenAmor

摘要

The growing sophistication of deepfake videos threatens media authenticity, requiring detectors that are accurate and robust to real-world distortions. We design and compare two deep learning architectures for video-based detection. The baseline fuses ResNeXt50 with a unidirectional LSTM, while the proposed model combines EfficientNet-B0 with a Bidirectional LSTM and temporal attention to capture bidirectional dependencies and emphasize informative frames. Both are trained and evaluated on Celeb-DF v2 under identical settings for fair comparison. Results show the proposed model outperforms the baseline, reaching 95.3% accuracy and a ROC-AUC of 0.981. These findings highlight the value of efficient convolutional encoders with temporal attention for interpretable and competitive deepfake detection, with applications in media forensics and verification.