<p>The fast development of deepfake generation methods is a major source of threat to the authenticity of data online, and its effective and efficient detection systems are required. This work contrasts with current research, which mostly targets individual-model analyses or smaller data set testing, as it attempts to offer an all-inclusive deepfake detection framework that addresses the concepts of generalization, robustness, and practicality. It compares the three state-of-the-art convolutional neural networks that are Xception, DenseNet121, and ResNet101 and an ensemble learning approach and a transformer-based architecture (Vision Transformer, ViT). It uses a large-scale balanced dataset consisting of 140,000 real and synthetic images and has an improved preprocessing pipeline that includes geometric transformations, photometric variations, and noise injection. Additionally, cross-dataset validation is presented to evaluate model generalization on unknown datasets. As experimental results reveal, ResNet101 is more effective as a standalone model (94.67% accuracy, AUC 0.987), whereas the ensemble model has a greater robustness with a higher recall (96.60%). Additionally, a Vision Transformer (ViT)-based model is incorporated to compare transformer-based feature extraction with conventional CNN architectures. The ViT model also indicates the possibility of transformer-based architectures to capture global features. The suggested framework can lay a scalable and generalizable solution to deepfake detection, which can be used in the creation of reliable digital forensics systems.</p>