Scholarlink

<p>The increasing sophistication of deepfake generation techniques poses significant threats to digital media integrity, with current detection models often struggling to generalize across unseen manipulations and adversarial attacks. This study presents a novel deepfake detection framework that enhances robustness and generalization by leveraging Optimal Transport (OT) theory for representation alignment. A regularization mechanism based on the Wasserstein distance is introduced to align feature distributions of real and fake media in a geometrically meaningful latent space. This alignment promotes better separation between authentic and manipulated samples while preserving structure under compression artifacts and adversarial perturbations. To simulate real-world evasion scenarios, adversarial training is incorporated during model optimization. Experimental evaluations on benchmark datasets, including FaceForensics++, Celeb-DF, and DeepfakeTIMIT, demonstrate superior accuracy, resilience to adversarial attacks, and improved cross-domain generalization compared to conventional CNN-based methods. The proposed approach offers a mathematically grounded and scalable solution for reliable multimedia forgery 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

Optimal transport-guided framework for adversarial robust deepfake detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月22日

查看全文
  • Saswati Chatterjee

摘要

The increasing sophistication of deepfake generation techniques poses significant threats to digital media integrity, with current detection models often struggling to generalize across unseen manipulations and adversarial attacks. This study presents a novel deepfake detection framework that enhances robustness and generalization by leveraging Optimal Transport (OT) theory for representation alignment. A regularization mechanism based on the Wasserstein distance is introduced to align feature distributions of real and fake media in a geometrically meaningful latent space. This alignment promotes better separation between authentic and manipulated samples while preserving structure under compression artifacts and adversarial perturbations. To simulate real-world evasion scenarios, adversarial training is incorporated during model optimization. Experimental evaluations on benchmark datasets, including FaceForensics++, Celeb-DF, and DeepfakeTIMIT, demonstrate superior accuracy, resilience to adversarial attacks, and improved cross-domain generalization compared to conventional CNN-based methods. The proposed approach offers a mathematically grounded and scalable solution for reliable multimedia forgery detection.