Deep Learning-Based Classification of Real and Fake Images Using Transfer Learning with EfficientNet

Abstract

Significant risks to digital security and trust are posed by the spread of deepfakes and altered images. To improve classification accuracy, this research proposes a strong deep learning framework for identifying real and fake photos via transfer learning. Various CNN designs are tested, such as VGG-16, ResNet-50, InceptionV3, and the EfficientNet family, on the Open Forensics dataset, which is a large collection of 225,000 annotated photos with a variety of occlusions, poses, and face traits. To maximize input quality, our preprocessing workflow combines face extraction, data augmentation, and normalization. EfficientNet-B7 outperforms ResNet-50 (77.15%) and InceptionV3 (78.8%) while retaining computational efficiency, achieving the greatest Top 1 accuracy of 84.4% among the tested models. The accuracy of EfficientNet-B3 has increased to 91% with additional fine-tuning, proving the usefulness of transfer learning for domain adaptation. The model's resistance to spoofing methods like 3D masks and printed pictures is confirmed by experimental findings that are verified by confusion matrices and F1-scores. By offering a scalable approach for content verification and biometric security, this study promotes automated forgery detection.