Scholarlink

<p>Deepfakes, AI-generated synthetic media that realistically manipulate people, pose a&#xa0;growing security risk for companies. A&#xa0;typical example of this is the so-called CEO fraud. In a&#xa0;well-known case, a&#xa0;manager was persuaded by a&#xa0;supposed CEO to transfer millions, resulting in a&#xa0;financial loss of over $&#xa0;35&#xa0;million. Fast and reliable detection of deepfakes is therefore becoming increasingly important for companies. This article presents an approach based on Vision Transformer and an expert-based ensemble model that can quickly respond to new generation models with the help of few-shot learning and without having to perform a&#xa0;complete retraining. This approach enables companies to detect deepfakes in a&#xa0;resource-efficient, scalable, and adaptable manner. As a&#xa0;result, visual manipulations involving unknown attack patterns can be detected at an early stage, effectively reducing risks in digital communication. The developed prototype, an ensemble model based on a&#xa0;majority decision by experts, was evaluated on selected deepfake datasets and achieved an overall accuracy that increased by 57.6% compared to the baseline.</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-Erkennung auch ohne große Datensätze: Ein Prototyp für Organisationen

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月20日

查看全文
  • Jan Czemmel,
  • Faisal Karim,
  • Jochen Günther,
  • Carsten Lanquillon

摘要

Deepfakes, AI-generated synthetic media that realistically manipulate people, pose a growing security risk for companies. A typical example of this is the so-called CEO fraud. In a well-known case, a manager was persuaded by a supposed CEO to transfer millions, resulting in a financial loss of over $ 35 million. Fast and reliable detection of deepfakes is therefore becoming increasingly important for companies. This article presents an approach based on Vision Transformer and an expert-based ensemble model that can quickly respond to new generation models with the help of few-shot learning and without having to perform a complete retraining. This approach enables companies to detect deepfakes in a resource-efficient, scalable, and adaptable manner. As a result, visual manipulations involving unknown attack patterns can be detected at an early stage, effectively reducing risks in digital communication. The developed prototype, an ensemble model based on a majority decision by experts, was evaluated on selected deepfake datasets and achieved an overall accuracy that increased by 57.6% compared to the baseline.