Comparative Analyais of CNN Architectures for Deep Fake Detection

DeepFake technology, driven by advanced generative models, threatens online authenticity and privacy. This paper presents a comparative analysis of three convolutional neural network (CNN) architectures—VGGFace16, DenseNet-121, and a custom CNN model—for DeepFake image detection. The study utilizes the 140k Faces Dataset, comprising 70,000 real and 70,000 synthetic images, and the Real and Fake Face Detection Dataset from Yonsei University, ensuring a diverse and well-balanced training set while accounting for computational constraints. Feature extraction from the final convolutional layers, dimensionality reduction via Principal Component Analysis (PCA), and classification with a Support Vector Machine (SVM) using a polynomial kernel form the core methodology. DenseNet-121 achieved the highest accuracy (97%) on grayscaled images, while the augmented custom CNN balanced accuracy (86%) and interpretability, attaining a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) score of 0.953. PCA visualizations confirmed the models’ ability to distinguish real from fake images. The findings underscore dataset selection’s role in model performance and the necessity of resource-efficient training. Future work will expand dataset diversity, explore cross-dataset validation, and leverage advanced computational resources to enhance generalization, contributing to more robust DeepFake detection systems.