Authors - Pravin Game, Shubham Bhingardive Abstract - Deepfake technology, which enables manipulation of images and videos, poses serious threat to media integrity and cyber security. Existing detection models often struggle with accuracy due to data complexity and variability. This study introduces a hybrid deepfake detection model that combines MobileNetV2, EfficientNetB7 and Vision Transformer (ViT) to enhance feature extraction and classification. ViT provides strong pattern recognition, EfficientNetB7 offers scalable accuracy and MobileNetV2 ensures lightweight processing. The model is trained on a publicly available datastet from Yonsei University, consisting of real and fake facial images. Techniques such as data augmentation and image resizing improve generalization. Experimental results show that the proposed model achieves 94.64% accuracy, 93.55% precision, 96.67% sensitivity, 92.31% specificity and 95.08% F1 score outperforming precious methods. These improvements ase statistically significant (p < 0.05). The results highlight the effectiveness of multi-model fusion for robust deepfake detection offering a scalable and reliable solution for applications in digital forensics, information verification and cybersecurity.
