Scholarlink

<p>Deepfake technology generates fake images and videos using synthesized faces. Deepfakes have become a social phenomenon, often maliciously used to spread false information and perpetrate digital fraud. The detection and recognition of fake videos are now even more challenging, and it is necessary to develop an advanced Deepfake detection (DFD) technology. While DFD methods have advanced, many existing techniques suffer from reliance on large, labelled datasets, poor generalization to unseen manipulations, and limited ability to capture fine-grained facial features. Thus, the development of new DFD techniques has gathered widespread interest among researchers. To bridge this gap, an efficient module namely Pyramid Deep Belief Network (PyramidFDBNet) is proposed in this work, which is designed by the combination of PyramidNet and Deep Belief Network (DBN). Initially, the extraction of frames from the input video is done and then the extracted frames are transferred to face detection module. Here, the YOLO v3-Tiny model is utilized for detecting faces. After that, Facial Action Units (AUs) acquired the detected faces and detect the facial action units using AUNet. Thereafter, feature extraction is conducted on detected facial AUs, where essential statistical features, Histogram of Oriented Gradients (HOG), and ResNet features are extracted. The extracted features are combined and passed to the deepfake recognition phase, where the PyramidFDBNet carries out the detection. The experimental results show that the PyramidFDBNet accomplished superior levels of accuracy at 91.8%, True Positive Rate (TPR) at 90.0%, and True Negative Rate (TNR) at 92.9%.</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

An enhanced pyramid deep belief network architecture design and implementation for deepfake spotting using deep learning framework

  • 期刊论文

  • 发表日期：

    2026年1月9日

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

摘要

Deepfake technology generates fake images and videos using synthesized faces. Deepfakes have become a social phenomenon, often maliciously used to spread false information and perpetrate digital fraud. The detection and recognition of fake videos are now even more challenging, and it is necessary to develop an advanced Deepfake detection (DFD) technology. While DFD methods have advanced, many existing techniques suffer from reliance on large, labelled datasets, poor generalization to unseen manipulations, and limited ability to capture fine-grained facial features. Thus, the development of new DFD techniques has gathered widespread interest among researchers. To bridge this gap, an efficient module namely Pyramid Deep Belief Network (PyramidFDBNet) is proposed in this work, which is designed by the combination of PyramidNet and Deep Belief Network (DBN). Initially, the extraction of frames from the input video is done and then the extracted frames are transferred to face detection module. Here, the YOLO v3-Tiny model is utilized for detecting faces. After that, Facial Action Units (AUs) acquired the detected faces and detect the facial action units using AUNet. Thereafter, feature extraction is conducted on detected facial AUs, where essential statistical features, Histogram of Oriented Gradients (HOG), and ResNet features are extracted. The extracted features are combined and passed to the deepfake recognition phase, where the PyramidFDBNet carries out the detection. The experimental results show that the PyramidFDBNet accomplished superior levels of accuracy at 91.8%, True Positive Rate (TPR) at 90.0%, and True Negative Rate (TNR) at 92.9%.