Scholarlink

<p>This paper presents a hybrid deepfake detection framework that fuses handcrafted Gabor filter-based texture features with deep semantic embeddings extracted using a pretrained Xception model enhanced with discriminatively weighted multi-head attention. To improve efficiency and interpretability, Fisher score-based feature selection is applied to retain the most discriminative features for classification. Using stratified cross-validation on the Celeb-DF dataset, the proposed Fisher-selected configuration achieves competitive accuracy of 96.99%, while the full fused representation achieves 97.35%. Cross-dataset evaluations on FaceForensics++ and UADFV show limited zero-shot transfer but demonstrate that performance can be improved through limited target-domain fine-tuning, reaching 91.3% and 86.8% accuracy, respectively, with 10% target data. These results suggest that combining complementary texture and semantic representations provides a compact and adaptable feature space for deepfake 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

Robust and efficient deepfake detection through Gabor-Xception feature fusion

  • 期刊论文

  • 发表日期：

    2026年6月26日

查看全文
  • B. N. Jyothi,
  • M. A. Jabbar

摘要

This paper presents a hybrid deepfake detection framework that fuses handcrafted Gabor filter-based texture features with deep semantic embeddings extracted using a pretrained Xception model enhanced with discriminatively weighted multi-head attention. To improve efficiency and interpretability, Fisher score-based feature selection is applied to retain the most discriminative features for classification. Using stratified cross-validation on the Celeb-DF dataset, the proposed Fisher-selected configuration achieves competitive accuracy of 96.99%, while the full fused representation achieves 97.35%. Cross-dataset evaluations on FaceForensics++ and UADFV show limited zero-shot transfer but demonstrate that performance can be improved through limited target-domain fine-tuning, reaching 91.3% and 86.8% accuracy, respectively, with 10% target data. These results suggest that combining complementary texture and semantic representations provides a compact and adaptable feature space for deepfake detection.