Scholarlink

<p>The rapid growth of generative artificial intelligence on social media platforms has led to the spread of deepfake content. This spread is a threat to the privacy of individuals, the reputation of firms, and general trust in media. In this paper, we propose a hybrid facial deepfake detection system which integrates transformer-based embeddings with combined textural descriptors. The Swin V2 Vision Transformer applies shifted-window self-attention to analyze facial regions, which produce hierarchical embeddings that extract global semantic information. The model computes four texture descriptors, namely Local Binary Patterns (LBP), Gabor filters, Wavelet transforms, and Gray-Level Co-occurrence Matrix (GLCM) features, to detect fine-grained pixel-level anomalies which result from manipulation. A fully connected neural network processes the combined representation of transformer embeddings and textural features to produce the final classification output. The combination of these macro-level semantic information with micro-level artifacts in the system results in improved resistance to different and unanticipated deepfake content. The proposed method was evaluated using the Celeb-DF V2 and FaceForensics++ benchmark datasets. The model produced an accuracy rate of 97.44 % on Celeb-DF V2 and 96.98 % on FaceForensics++ (FF++) while achieving ROC_AUC scores of 0.9962 and 0.9929, respectively. Precision and recall maintained levels above 96 % as the combination outperformed other state of the art models such as Forgery-Domain-Supervised, Lightweight dynamic fusion network (LDFNet) and was comparable to WATCHER.</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

Hybrid Facial Deepfake Detection Framework Using Swin V2 Vision Transformers and Multi-Scale Textural Descriptors

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Shivaprakash S J,
  • Akshat Chauhan,
  • Sabireen H,
  • Abdul Quadir Md

摘要

The rapid growth of generative artificial intelligence on social media platforms has led to the spread of deepfake content. This spread is a threat to the privacy of individuals, the reputation of firms, and general trust in media. In this paper, we propose a hybrid facial deepfake detection system which integrates transformer-based embeddings with combined textural descriptors. The Swin V2 Vision Transformer applies shifted-window self-attention to analyze facial regions, which produce hierarchical embeddings that extract global semantic information. The model computes four texture descriptors, namely Local Binary Patterns (LBP), Gabor filters, Wavelet transforms, and Gray-Level Co-occurrence Matrix (GLCM) features, to detect fine-grained pixel-level anomalies which result from manipulation. A fully connected neural network processes the combined representation of transformer embeddings and textural features to produce the final classification output. The combination of these macro-level semantic information with micro-level artifacts in the system results in improved resistance to different and unanticipated deepfake content. The proposed method was evaluated using the Celeb-DF V2 and FaceForensics++ benchmark datasets. The model produced an accuracy rate of 97.44 % on Celeb-DF V2 and 96.98 % on FaceForensics++ (FF++) while achieving ROC_AUC scores of 0.9962 and 0.9929, respectively. Precision and recall maintained levels above 96 % as the combination outperformed other state of the art models such as Forgery-Domain-Supervised, Lightweight dynamic fusion network (LDFNet) and was comparable to WATCHER.