EfficientNetB0-Based DeepFake Image Detection: A Transfer-Learning Framework for Robust and Deployable Media Verification

Abstract

As the realism of DeepFakes has increased significantly, there is an emerging necessity to develop efficient, reliable, and robust automatic detection frameworks that are feasible to apply in the wild. This study proposes an image-based DeepFake detection system using EfficientNetB0 and transfer learning. Facial images were acquired from the FaceForensics++ and Celeb-DF datasets and preprocessed as 224 by 224 RGB images. The data were augmented to reflect realistic conditions through rotation, zoom, horizontal flipping, illumination changes, and JPEG compression. EfficientNetB0 was used as a feature extractor, followed by global average pooling, a dense layer with 128 neurons, dropout, and a two-class softmax layer. The model achieved 92.8% validation accuracy, 92.5% precision, 93.1%recall, and a 92.8% F1-score. Robustness testing produced 89.5%accuracy on images compressed at JPEG quality 70, while cross-dataset evaluation from FaceForensics++ to Celeb-DF produced 84.2% accuracy. A real-time verification system was implemented by integrating the model with a Flask RESTful API and web and mobile clients. The findings show that transfer learning provides a practical balance among accuracy, robustness, computational efficiency, and deployability. However, the reduction in cross-dataset performance indicates limited generalizability to previously unseen manipulation techniques.