Scholarlink

<p>The rise of complex deep learning architectures has enabled the creation of hyper-realistic media, posing significant risks to public trust. Current state-of-the-art deepfake detectors often require heavy computational resources and are limited to single-domain dataset, which limit their deployment on standard workstations and creating a mixed domain generalization gap. To overcome these challenges, we propose a Multi Stage Hybrid Lightweight Framework (MSHLF) that uses an EfficientNetB0 backbone for spatial feature extraction and a lightweight LSTM for temporal modeling. Trained on a balanced mixed-domain dataset, our model achieves accuracy of 92.48% and AUROC of 98.0%, utilizing only 5.44M parameters. This resource-aware mixed-domain model demonstrates competitive generalization performance across mixed domains settings, making it suitable for deployment on standard workstations.</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

Efficient deepfake detection: a hybrid lightweight framework for cross-domain generalization

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Harsh Vardhan Nagar,
  • Vidit Gandhi,
  • Digvijaysinh M. Rathod

摘要

The rise of complex deep learning architectures has enabled the creation of hyper-realistic media, posing significant risks to public trust. Current state-of-the-art deepfake detectors often require heavy computational resources and are limited to single-domain dataset, which limit their deployment on standard workstations and creating a mixed domain generalization gap. To overcome these challenges, we propose a Multi Stage Hybrid Lightweight Framework (MSHLF) that uses an EfficientNetB0 backbone for spatial feature extraction and a lightweight LSTM for temporal modeling. Trained on a balanced mixed-domain dataset, our model achieves accuracy of 92.48% and AUROC of 98.0%, utilizing only 5.44M parameters. This resource-aware mixed-domain model demonstrates competitive generalization performance across mixed domains settings, making it suitable for deployment on standard workstations.