Face forgery, also known as deepfake, produces extremely real yet fake images and videos, making it impossible to distinguish between reality and fraud. It has the potential to alter expressions, swap identities, and generate new data, which may lead to fraud, identity theft, and breaches of privacy. The current research compares various pre-trained neural network models—Xception, ResNet50, EfficientNetB0, DenseNet121, and MobileNet—for detecting face forgery. The dataset, mainly collected from Flickr Face (FFHQ) and 1 million fake face sources, contains approximately 200 k face images, creating an accurate representation of both real and fake faces as the starting point for our research. The model evaluation process involves the following steps: starting with the image dataset of real and fake images, applying data augmentation techniques like rescaling, splitting the data into training, validation, and test sets, and loading the data for model training. These models were evaluated using metrics such as confusion matrix, ROC/AUC curve, classification report, accuracy, precision and recall, and F1-score, and the trained model was saved for future use. Xception was the most effective model, with a test accuracy of 99.14%. It outperformed DenseNet121, ResNet50, and MobileNet. The current research shows how transfer learning may improve the efficiency and accuracy of forgery detection. The Xception model was more effective at detecting forged images, as evidenced by its results in confusion matrix analysis and ROC-AUC curves.