Scholarlink

Synthetic media has evolved rapidly especially a deepfake technology, transforming the space of digital content creation. Using advanced deep learning models and generative techniques, deepfake enables the realistic manipulation of image, video, and audio. While these advancements give potential for creativity and innovation, including misinformation, digital identity fraud, and security breaches. Detecting fake video is quite challenging so advanced deepfake detection (DFD) technologies are necessary. In response, a novel FalsiDetect framework has been developed using federated learning to enhance generalizability and effectiveness of DFD across diverse datasets. Initially we designed an efficient module PyramidFDBNet, which is hybrid of PyramidNet and deep belief network (DBN). The FalsiDetect framework begins by converting input video data into frames, with duplicate frames reduced using a frame hashing technique to optimize the processing. Detected frames are then examined using YOLOv3-Tiny to identify facial regions. Further facial action (action units AUs) is detected using AUNet, and then key features from these detected AUs such as essential statistical features, histogram of oriented gradient (HOG), and ResNet-derived features are extracted. They are concatenated and fed into the deepfake recognition phase for classification by modified and fused PyramidFDBNet layers to detect them. In a federated learning setup, three large datasets (real, fake, and custom) are utilized as local nodes, ensuring that model generalizes well across varied media and contexts. The FalsiDetect framework attains 94.6% detection accuracy with TNR of 95.4% and TPR of 93.2%.

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

FalsiDetect: A Federated Learning Framework for Generalizable Deepfake Detection Across Diverse Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • D. Stephy Joy,
  • R. Thirumalai Selvi

摘要

Synthetic media has evolved rapidly especially a deepfake technology, transforming the space of digital content creation. Using advanced deep learning models and generative techniques, deepfake enables the realistic manipulation of image, video, and audio. While these advancements give potential for creativity and innovation, including misinformation, digital identity fraud, and security breaches. Detecting fake video is quite challenging so advanced deepfake detection (DFD) technologies are necessary. In response, a novel FalsiDetect framework has been developed using federated learning to enhance generalizability and effectiveness of DFD across diverse datasets. Initially we designed an efficient module PyramidFDBNet, which is hybrid of PyramidNet and deep belief network (DBN). The FalsiDetect framework begins by converting input video data into frames, with duplicate frames reduced using a frame hashing technique to optimize the processing. Detected frames are then examined using YOLOv3-Tiny to identify facial regions. Further facial action (action units AUs) is detected using AUNet, and then key features from these detected AUs such as essential statistical features, histogram of oriented gradient (HOG), and ResNet-derived features are extracted. They are concatenated and fed into the deepfake recognition phase for classification by modified and fused PyramidFDBNet layers to detect them. In a federated learning setup, three large datasets (real, fake, and custom) are utilized as local nodes, ensuring that model generalizes well across varied media and contexts. The FalsiDetect framework attains 94.6% detection accuracy with TNR of 95.4% and TPR of 93.2%.