Deepfake Face Detection Using Transfer Learning With EfficientNet-B4

Abstract

Deepfake media, synthetic videos generated by deep learning models, are a growing threat to information integrity, public trust, and personal privacy. This study presents a deepfake face detection system using a fine-tuned EfficientNet-B4 convolutional neural network trained on the Celeb-DF-v2 benchmark dataset. A balanced data pipeline was constructed with equal amounts of real and fake data to train, validate, and test the model. The model was trained with targeted fine tuning of the top three convolution blocks, weight decay, label smoothing, and dropout. It was finally optimized using an adaptive learning rate scheduler. The system achieves a test accuracy of 94.04% and an AUC-ROC of 0.9844 on Celeb-DF-v2, averaged across two reproducible runs. Grad-CAM visualizations confirm that the model attends to meaningful facial regions rather than artifacts in the background. To assess generalization, a second model trained on a combined Celeb-DF-v2 and FaceForensics++ set achieved 88.87% accuracy and 0.9514 AUC-ROC across both benchmarks, substantially recovering the cross-dataset performance lost by the single-dataset model. These results are competitive in comparison to published baselines on Celeb-DF-v2 and demonstrate how principled transfer learning with targeted regularization can achieve near state-of-the-art deepfake detection.