The rapid expansion of digital media has sparked significant concerns regarding the swift dissemination and potential misuse of forged video content. Existing forgery detection technologies primarily focus on simple forgeries and are still evolving, resulting in a critical gap in the detection of multilevel forgeries, where one forgery is layered over another. This paper presents an innovative framework designed to address this challenge by extracting intricate features from forged frames using attention‐augmented convolutional neural networks (AACNNs). A U‐Net‐based CycleGAN is employed to accurately localize forged regions, enabling a comprehensive analysis that identifies both two‐ and three‐level forgeries by leveraging AACNN's local and global attention mechanisms. To enhance robustness and accuracy, we integrate a model‐agnostic meta‐learning approach. Our meticulously curated custom dataset, which represents complex forgery scenarios, underpins the effectiveness of our framework. In a 10‐shot scenario, the AACNN backbone achieved an impressive accuracy of 98.2%, alongside a sensitivity of 96.3%, specificity of 97.6%, and an F1‐score of 96.8%. These results represent a significant advancement in the accuracy and reliability of sophisticated video forgery detection.

中文翻译：

---

用于数字视频伪造的多级检测的高级框架


数字媒体的迅速扩张引发了人们对伪造视频内容的迅速传播和潜在滥用的严重担忧。现有的伪造检测技术主要关注简单的伪造，并且仍在不断发展，导致在检测多级伪造方面存在严重差距，其中一种伪造叠加在另一种伪造上。本文提出了一个创新框架，旨在通过使用注意力增强卷积神经网络 （AACNN） 从伪造框架中提取复杂的特征来应对这一挑战。采用基于 U-Net 的 CycleGAN 来准确定位伪造区域，从而利用 AACNN 的本地和全球关注机制进行全面分析，识别两级和三级伪造。为了提高稳健性和准确性，我们集成了一种与模型无关的元学习方法。我们精心策划的自定义数据集代表了复杂的伪造场景，支撑了我们框架的有效性。在 10 次拍摄的情况下，AACNN 主干实现了令人印象深刻的 98.2% 的准确率，同时灵敏度为 96.3%，特异性为 97.6%，F1 评分为 96.8%。这些结果代表了复杂视频伪造检测的准确性和可靠性的重大进步。