Sobel filter and linear classification for deepfake analysis of faces

Abstract

Through artificial intelligence (AI) techniques, deepfakes (deep learning + fakes) are created. Currently, almost six years after this topic's popularization, some people have been using AI algorithms to change, for example, face. In this scenario, this article proposes to identify the most discriminating characteristics, based on the multivariate statistical learning techniques, with and without the Sobel filter. Complementarily, we aim to analyze the areas that humans have visually identified as relevant to detect a deepfake. The results show that the statistical model (PCA plus MLDA) combined with the Sobel filter correctly classified most of the images, highlighting the regions that discriminate a deepfake.