Comparative Evaluation of Xception and DenseNet121 for Robust Deepfake Detection

Generative adversarial networks (GANs) have led to rapid advancement of deepfake technology, the authenticity of which poses a major challenge to digital media. In this study, two deep learning architectures, Xception and DenseNet121, are evaluated in terms of how well they can differentiate real and fake faces. The models were trained on normalized and augmented data using a curated dataset of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 4 0, 0 0 0}$</tex> images, half real faces from the Flickr dataset, the other half-synthetic faces generated by StyleGAN. With a training accuracy of 86.31 % and a test accuracy of 73.00 %, Xception model also shows strong feature extraction power. Validation accuracy (77.00 %) of DenseNet121 was slightly better than in validation accuracy (76.00 %) but more sensitive to dataset conditions. Performance graphs served to further illustrate the complementary strengths of the two models, lending support for the applicability of ensemble methods to improve detection robustness. This study highlights the benefits of advanced architectures, data preparation rigour, and ensemble approaches to combat deepfake. Future work should concentrate on improving generalization, interpretability, robustness to evolving challenges in the area of digital media authentication.