Enhancing DeepFake Classification Performance Using a CNN and XceptionNet-Based Pipeline

Deepfake technology can create a video or image that looks real but in reality it is fake.It is a real threat to our society and our digital security. The popularity of generative models like generative adversarial networks (GANs) has made it easier to produce content that we can not differentiate between fake media and authentic media. To combat this, the study introduces a dual-model deepfake detection system that combines a custom lightweight convolutional neural network (CNN) with a transfer learning-based XceptionNet. This framework is trained and tested on the 140K Real and Fake Faces dataset, which contains an similar number of real and synthetic images. The custom CNN is built from scratch, featuring optimized convolu- tional layers, ReLU activations, max-pooling, and dense layers with dropout for regularization. The XceptionNet model is fine- tuned with additional dense layers for binary classification. Both models follow the same pre-processing steps and are trained with the Adam optimizer using binary cross-entropy loss. The CNN achieves an impressive 97% accuracy, while XceptionNet reaches 91%, highlighting their strong performance and ability to generalize. Metrics like precision, recall and F1-score verify the reliability of both methods. The research suggest that despite limited data and resources, deepfake detection is feasible by utilizing efficient architectures and training techniques. This framework gives a scalable and resource efficient solution for real world deepfake forensics, particularly in environments with limited computational capacity. Future work will investigate ensemble models and real-time detection implementations.