Classifying Synthetic Facial Images: A Quantitative and Qualitative Approach with AI

Through artificial intelligence (AI) techniques, synthetic content is generated, which is an increasingly critical issue for digital security. This paper aims to classify real and synthetic facial images using a state-of-the-art AI model to analyze, both quantitatively and qualitatively, its performance and the most discriminating features across the various data transformations used for this purpose in current literature. In this scenario, the study utilized three public benchmark datasets: two for generating fake images (Stable Diffusion and StyleGAN2) and one for real images (FFHQ), employing the Xception architecture with the Grad-CAM++ viewer. Overall, the results indicate that the central region of the image is the most discriminative one, whereas random cropping produced the worst performance. Future work aims to explore other explainable Artificial Intelligence (XAI) methods, as well as other training and validation protocols, such as cross-testing between synthetic databases.