Scholarlink

<p>The emergence of deepfakes, or hyper-realistic artificial media created through the application of deep learning, needs the development of robust detection systems to check online content integrity. In this study, we propose a novel ensemble learning framework that integrates InceptionV3, VGG16, and Xception architectures to detect deepfakes with high accuracy. Using a balanced dataset comprising 70,000 real and 70,000 synthetic images, our method achieves a detection accuracy of <b>97.40%</b>, with a sensitivity of <b>97.00%</b> and specificity of <b>98.00%</b>. Compared to baseline models such as ResNet50 and EfficientNet, the proposed ensemble framework demonstrates a relative improvement of up to <b>5.6%</b> in accuracy and <b>3.0%</b> in AUC. These results highlight the effectiveness and robustness of the proposed method, making it a promising solution for combating the evolving threat of deepfake technology. We conducted extensive evaluations using a diverse dataset comprising real and synthetic images from multiple sources to ensure the robustness of our model. Furthermore, we employed Grad-CAM (Gradient-weighted Class Activation Mapping) to visualize and interpret the decision-making process of our model, providing insights into which parts of the images contribute to its predictions. Our research addresses the evolving challenges posed by sophisticated deepfake generation techniques. By combining the strengths of various pre-trained models, our ensemble approach mitigates the limitations typically associated with single-model systems. We discuss the ethical implications and highlight the necessity for ongoing research to keep pace with the rapid evolution of deepfake creation methods.</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

Deepfake detection through ensemble learning

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Marwa Ben Jabra,
  • Omar Cheikhrouhou,
  • Anouar BenAmor

摘要

The emergence of deepfakes, or hyper-realistic artificial media created through the application of deep learning, needs the development of robust detection systems to check online content integrity. In this study, we propose a novel ensemble learning framework that integrates InceptionV3, VGG16, and Xception architectures to detect deepfakes with high accuracy. Using a balanced dataset comprising 70,000 real and 70,000 synthetic images, our method achieves a detection accuracy of 97.40%, with a sensitivity of 97.00% and specificity of 98.00%. Compared to baseline models such as ResNet50 and EfficientNet, the proposed ensemble framework demonstrates a relative improvement of up to 5.6% in accuracy and 3.0% in AUC. These results highlight the effectiveness and robustness of the proposed method, making it a promising solution for combating the evolving threat of deepfake technology. We conducted extensive evaluations using a diverse dataset comprising real and synthetic images from multiple sources to ensure the robustness of our model. Furthermore, we employed Grad-CAM (Gradient-weighted Class Activation Mapping) to visualize and interpret the decision-making process of our model, providing insights into which parts of the images contribute to its predictions. Our research addresses the evolving challenges posed by sophisticated deepfake generation techniques. By combining the strengths of various pre-trained models, our ensemble approach mitigates the limitations typically associated with single-model systems. We discuss the ethical implications and highlight the necessity for ongoing research to keep pace with the rapid evolution of deepfake creation methods.