Scholarlink

<p>Deepfake videos pose an escalating threat to the integrity of digital media: Advances in generative models produce highly convincing manipulations that undermine trust in audio-visual evidence. Existing detectors often trade off accuracy for scalability, struggle to generalize to diverse manipulation methods, or lack transparency about what cues drive decisions. To address these gaps, we propose a framework that leverages Vision Transformers (ViT) and patch-level self-attention to detect subtle spatial inconsistencies introduced by face manipulation. Our method fine-tunes a ViT backbone on frame-level inputs using a patch embedding scheme that enhances sensitivity to localized artifacts, then aggregates frame scores into robust video-level decisions while keeping the inference footprint suitable for near-real-time operation, and emphasizes evaluation against strong baselines and ablation analyses to isolate the contribution of each design choice. Experiments on the deepfake detection (DFD) dataset show that our model attains 98.11% accuracy, a balanced accuracy of 97.13%, and an AUC of 98.2%, outperforming several recent baselines. The proposed approach balances detection performance and practical deployability and provides a modular foundation for adding temporal modeling and explainability components in future work.</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

Deepfake video detection using Vision Transformers: exploiting self-attention for visual manipulation detection

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Muhammad Khubaib,
  • Waseem Abbasi,
  • Farhat Saeed,
  • Abeer Aljohani

摘要

Deepfake videos pose an escalating threat to the integrity of digital media: Advances in generative models produce highly convincing manipulations that undermine trust in audio-visual evidence. Existing detectors often trade off accuracy for scalability, struggle to generalize to diverse manipulation methods, or lack transparency about what cues drive decisions. To address these gaps, we propose a framework that leverages Vision Transformers (ViT) and patch-level self-attention to detect subtle spatial inconsistencies introduced by face manipulation. Our method fine-tunes a ViT backbone on frame-level inputs using a patch embedding scheme that enhances sensitivity to localized artifacts, then aggregates frame scores into robust video-level decisions while keeping the inference footprint suitable for near-real-time operation, and emphasizes evaluation against strong baselines and ablation analyses to isolate the contribution of each design choice. Experiments on the deepfake detection (DFD) dataset show that our model attains 98.11% accuracy, a balanced accuracy of 97.13%, and an AUC of 98.2%, outperforming several recent baselines. The proposed approach balances detection performance and practical deployability and provides a modular foundation for adding temporal modeling and explainability components in future work.