Scholarlink

In recent years, deep learning has advanced rapidly. This has enabled the rise of deepfake face reenactment technology. It poses potential threats, including risks to personal privacy and data. For example, it can be used to forge politicians to make inappropriate statements, manipulate faces to spread false information, or fabricate videos for financial gain. To protect and respect individuals’ privacy, it is crucial to develop a model that distinguishes real videos from fake ones. This paper presents a method for detecting image and video forgeries based on RetinaFace and EfficientNet-V2 neural networks to verify video authenticity. Compared with other previously proven deepfake detection methods, this method has a maximum area under ROC curve improvement of 22.9% and achieves an average area under the curve (AUC) improvement of 15.6% on multiple datasets.

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 RetinaFace and EfficientNet-V2

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • ShuHan Cai,
  • NingFei Wang,
  • ZhengYang Cai,
  • JiaCheng Du,
  • JinJie Huang,
  • Yao Chen,
  • Bo Peng,
  • Chin Hong Wong

摘要

In recent years, deep learning has advanced rapidly. This has enabled the rise of deepfake face reenactment technology. It poses potential threats, including risks to personal privacy and data. For example, it can be used to forge politicians to make inappropriate statements, manipulate faces to spread false information, or fabricate videos for financial gain. To protect and respect individuals’ privacy, it is crucial to develop a model that distinguishes real videos from fake ones. This paper presents a method for detecting image and video forgeries based on RetinaFace and EfficientNet-V2 neural networks to verify video authenticity. Compared with other previously proven deepfake detection methods, this method has a maximum area under ROC curve improvement of 22.9% and achieves an average area under the curve (AUC) improvement of 15.6% on multiple datasets.