Swapping Face Images with Generative Neural Networks for Deepfake Technology – Experimental Study
Michał Zendran, Andrzej Rusiecki
Abstract
Generative neural networks are usually the most essential part of deepfake, a technique of image-to-image translation, designed to combine and overlay objects in images or videos creating deceptively realistic counterfeits. In this paper, four leading methods used for deepfake generation: autoencoders, variational autoencoders, variational autoencoders generative adversarial networks and cycle generative adversarial networks, in problem of face-to-face conversion, are analysed. We present results of experiments conducted on face-swapping task, performed on specially preprocessed data from VoxCeleb2 dataset. Due to the lack of numerical methods for deepfake comparison, a descriptive assessment method was proposed, and all obtained results were rated in a visual evaluation process. General conclusions concerning applicability of considered approaches to deepfake generation problem were formulated.