Scholarlink

This paper describes how the InceptionNet V3 model can detect 85% deepfakes. Deepfakes refer to synthetic media in which the image or video of a person is replaced with another, which poses grave threats in spreading misinformation and compromising security. We used the InceptionNet V3 model CNN architecture and fine-tuned it for classifying images into real and fake as a binary class. To improve models’ durability, rotation, shift, shear, zoom, and horizontal flip are implemented as data augmentation techniques. Initially, we divided the Kaggle dataset into test, validation, and training sets. We then trained our model using the Adam optimizer over 15 epochs. Accuracy, confusion matrix, and ROC AUC score are involved in evaluation criteria. The results show that InceptionNet V3 with custom layers is a good deepfake detector in the sense of achieving an 85% accuracy of the test set. Transfer learning and data augmentation can be very useful in improving the capabilities of a deepfake detector. Further studies are going to concentrate on improving the model’s functionality and extending its application to include video data.

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

Enhancing Deepfake Detection with InceptionNet V3: A Transfer Learning Approach Using CNN

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sangita Jaybhaye,
  • Swaroop Deokar,
  • Somesh Deshmukh,
  • Adityaraje Dhumal,
  • Pushpak Gadhe

摘要

This paper describes how the InceptionNet V3 model can detect 85% deepfakes. Deepfakes refer to synthetic media in which the image or video of a person is replaced with another, which poses grave threats in spreading misinformation and compromising security. We used the InceptionNet V3 model CNN architecture and fine-tuned it for classifying images into real and fake as a binary class. To improve models’ durability, rotation, shift, shear, zoom, and horizontal flip are implemented as data augmentation techniques. Initially, we divided the Kaggle dataset into test, validation, and training sets. We then trained our model using the Adam optimizer over 15 epochs. Accuracy, confusion matrix, and ROC AUC score are involved in evaluation criteria. The results show that InceptionNet V3 with custom layers is a good deepfake detector in the sense of achieving an 85% accuracy of the test set. Transfer learning and data augmentation can be very useful in improving the capabilities of a deepfake detector. Further studies are going to concentrate on improving the model’s functionality and extending its application to include video data.