In the rapidly advancing field of digital image manipulation, accurate detection and classification of deepfake images have become paramount due to their potential for misuse in misinformation and privacy violations. It can be used for fake news dissemination, election manipulation or even fraud. As deepfake technologies evolve, traditional methods for image classification struggle to keep pace with the increasing sophistication of synthetic media. This paper introduces a Vision Transformer (ViT) model, having 12 transformer blocks specifically designed to address these challenges, utilizing the full extensive 140k Real and Fake Faces dataset available on Kaggle, that comprises 70,000 real and 70,000 fake facial images. The results surpass that of conventional convolutional neural networks (CNNs) and other contemporary methods, showing the model’s advanced capability in discerning subtle differences between authentic and fake images. The custom trained Vision Transformer model achieves a validation accuracy of 92.76% & an Area Under Curve (AUC) score of 0.97, setting a new benchmark in the detection and classification of deepfake images.