Scholarlink

<p>The rapid evolution of deepfake generation techniques, particularly through Generative Adversarial Networks (GANs), necessitates advanced models for accurate detection and segmentation of video forgeries for cyber forensics. Vision Transformers (ViTs) enable efficient global context understanding, capturing long-range dependencies in video sequences. Concurrently, the use of Convolutional Neural Networks (CNNs) facilitates feature extraction and hierarchical pattern recognition, allowing the model to discern subtle artifacts indicative of deepfake manipulation. Our approach aims to exploit the unique strengths of each architecture to enhance both the discriminate and spatial representation capabilities crucial for comprehensive forgery analysis. High-Resolution Networks (HRNets) are also employed to capture fine-grained details essential for precise localization of forged regions within the video frames. In this research, we propose a novel hybrid architecture comprising an accurate video pre-processing unit, a primary classifier using a ViT that uses a CNN as a patch extractor, and finally a segmentation unit where the forged area is detected using an adapted HRNet. All sub-units are trained and tested on the benchmark datasets, proving to outperform the current models in terms of accuracy, Area Under the Curve (AUC), and F1 score.</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

Hybrid deep learning architecture for comprehensive deepfake video facial forgery detection and segmentation

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Vaishnavi Raghavajosyula,
  • Digambar Pawar

摘要

The rapid evolution of deepfake generation techniques, particularly through Generative Adversarial Networks (GANs), necessitates advanced models for accurate detection and segmentation of video forgeries for cyber forensics. Vision Transformers (ViTs) enable efficient global context understanding, capturing long-range dependencies in video sequences. Concurrently, the use of Convolutional Neural Networks (CNNs) facilitates feature extraction and hierarchical pattern recognition, allowing the model to discern subtle artifacts indicative of deepfake manipulation. Our approach aims to exploit the unique strengths of each architecture to enhance both the discriminate and spatial representation capabilities crucial for comprehensive forgery analysis. High-Resolution Networks (HRNets) are also employed to capture fine-grained details essential for precise localization of forged regions within the video frames. In this research, we propose a novel hybrid architecture comprising an accurate video pre-processing unit, a primary classifier using a ViT that uses a CNN as a patch extractor, and finally a segmentation unit where the forged area is detected using an adapted HRNet. All sub-units are trained and tested on the benchmark datasets, proving to outperform the current models in terms of accuracy, Area Under the Curve (AUC), and F1 score.