Deepfake Detection: A Comparative Approach

Abstract

The rapid advancement of synthetic image generation techniques, such as Generative Adversarial Networks (GANs), has enabled the creation of realistic deepfakes, raising concerns regarding misinformation, privacy, and digital security. This work investigates automatic methods for detecting fake facial images, focusing on convolutional neural networks (CNNs) and classical techniques such as SVM and Random Forest. Experiments were conducted using a merged dataset composed of the Real and Fake Face Detection and Human Faces datasets, totaling 10,025 images. The results indicate that the tested methods achieve satisfactory performance, demonstrating the feasibility of using machine learning approaches for automatic deepfake detection. Performance and interpretability aspects are discussed, showing the superiority of CNNs in classifying fake images, achieving an average accuracy of 94.2% in the conducted tests.