Deepfake Detection using EfficientNET-B0 and Transfer Learning

Deepfake technology has advanced significantly over the last few years, enabling the development of highly realistic fake facial images, which are a threat to the security of digital information, the ability to manage the spread of misinformation, and identity verification. Therefore, a deepfake detection model based on deep learning and the efficient model, EfficientNet-B0, with transfer learning has been proposed to address the challenge of deepfake images. This model was trained and tested using a public dataset containing 29,449 real and fake facial images. Additionally, the images were preprocessed based on resizing, normalization, and augmentation for the model's robustness. The proposed model was proven to be accurate, with a value of 92.44, F1-score of 92.00, and AUC-ROC of 97.83, indicating that the model was better than a few CNN-based models. Furthermore, a web interface was created through the use of Gradio, Flask, and FastAPI to carry out a real-time test of the trained model, showing that the proposed model was efficient in the detection of deepfake images through the efficient model, EfficientNet-B0, with transfer learning.