Scholarlink

<p>We propose a comprehensive system architecture designed to effectively mitigate the growing risks associated with modern DeepFake technologies. Our proposed system comprises of three primary components, each playing a vital role in the detection process. In the first component, we proposed a GAN framework with a dual discriminator, specifically developed for initial image authenticity verification. This innovative approach is essential for differentiating genuine images from those artificially manipulated. Following this, we employ a specialized form of CNN, known as a Residual Block-Based CNN, to further analyze images verified by the GANs. This deep network, consisting of 33 layers, excels in extracting and processing intricate features from the images, a critical factor in identifying complex patterns indicative of anomalies. Lastly, the system incorporates a Multiple Instance Learning (MIL) framework. The framework presented in the research significantly enhances image anomaly detection at the event level, rather than just the instant level. The system was rigorously evaluated on two benchmark datasets: FaceForensics++ (FF++) and CelebDF. It achieved a classification accuracy of 96.01% with a False Negative Rate (FNR) of 18.86% and an average computation time of 1801.10 seconds on the FF++ dataset. On the CelebDF dataset, it attained an accuracy of 94.3%, an FNR of 5.9%, and an average computation time of 110.04 seconds. These results demonstrate the proposed framework’s effectiveness and efficiency in detecting complex image forgery cases without requiring extensive manual annotations, making it suitable for real-time surveillance and media integrity verification systems.</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

Forged anomaly detection using advanced deep learning

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Nomica Choudhry,
  • Jemal Abawajy,
  • Shamsul Huda,
  • Imran Rao

摘要

We propose a comprehensive system architecture designed to effectively mitigate the growing risks associated with modern DeepFake technologies. Our proposed system comprises of three primary components, each playing a vital role in the detection process. In the first component, we proposed a GAN framework with a dual discriminator, specifically developed for initial image authenticity verification. This innovative approach is essential for differentiating genuine images from those artificially manipulated. Following this, we employ a specialized form of CNN, known as a Residual Block-Based CNN, to further analyze images verified by the GANs. This deep network, consisting of 33 layers, excels in extracting and processing intricate features from the images, a critical factor in identifying complex patterns indicative of anomalies. Lastly, the system incorporates a Multiple Instance Learning (MIL) framework. The framework presented in the research significantly enhances image anomaly detection at the event level, rather than just the instant level. The system was rigorously evaluated on two benchmark datasets: FaceForensics++ (FF++) and CelebDF. It achieved a classification accuracy of 96.01% with a False Negative Rate (FNR) of 18.86% and an average computation time of 1801.10 seconds on the FF++ dataset. On the CelebDF dataset, it attained an accuracy of 94.3%, an FNR of 5.9%, and an average computation time of 110.04 seconds. These results demonstrate the proposed framework’s effectiveness and efficiency in detecting complex image forgery cases without requiring extensive manual annotations, making it suitable for real-time surveillance and media integrity verification systems.