Scholarlink

As deepfake technology keeps improving rapidly, the demand for reliable & efficient detection methods has become more essential. This work investigates the performance of multiple deepfake detection approaches by analyzing accuracy, F1-score, precision, recall, and AUC metrics. Deepfake detection models utilizing Vision Transformer (ViT), CNN, and RESNET-50 with LSTM are implemented across three different sizes of benchmarked datasets: DF-TIMIT, Celeb-DF (v2), and FaceForensics++ (NT). The proposed work yields a maximum accuracy of 0.98 on the DF-TIMIT dataset by using Vision Transformer and 0.98, 0.93, & 0.89 on the DF-TIMIT, Celeb-DF (v2), & FaceForensics++ (NT) datasets, respectively, by using the CNN model. This study found that the CNN-based approach achieved superior performance in every dataset’s testing scenario, including the ViT model’s effectiveness in the DF-TIMIT dataset. The combination of RESNET-50 with the LSTM model fails as an adequate solution for detecting deepfake images.

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

A Detailed Performance Analysis of Deepfake Detection Approaches Across Diverse Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pawan Pandey,
  • Arun Solanki,
  • Sanjay Kumar Sharma

摘要

As deepfake technology keeps improving rapidly, the demand for reliable & efficient detection methods has become more essential. This work investigates the performance of multiple deepfake detection approaches by analyzing accuracy, F1-score, precision, recall, and AUC metrics. Deepfake detection models utilizing Vision Transformer (ViT), CNN, and RESNET-50 with LSTM are implemented across three different sizes of benchmarked datasets: DF-TIMIT, Celeb-DF (v2), and FaceForensics++ (NT). The proposed work yields a maximum accuracy of 0.98 on the DF-TIMIT dataset by using Vision Transformer and 0.98, 0.93, & 0.89 on the DF-TIMIT, Celeb-DF (v2), & FaceForensics++ (NT) datasets, respectively, by using the CNN model. This study found that the CNN-based approach achieved superior performance in every dataset’s testing scenario, including the ViT model’s effectiveness in the DF-TIMIT dataset. The combination of RESNET-50 with the LSTM model fails as an adequate solution for detecting deepfake images.