Please use this identifier to cite or link to this item:
http://hdl.handle.net/10609/146181
Title: | Deepfakes: creación de nuevas caras a partir de imágenes de famosos |
Author: | Sastre Toral, Maria Teresa |
Tutor: | Hernández Jiménez, Enric |
Others: | Pérez Solà, Cristina |
Keywords: | cybersecurity privacy deepfakes |
Issue Date: | Jun-2022 |
Publisher: | Universitat Oberta de Catalunya (UOC) |
Abstract: | Deepfakes techniques are used to generate manipulated videos, images or audio. Its purpose can be criminal but it can also be useful in other areas such as the film industry. In this work, the different types of deepfake techniques have been classified, explaining the purpose of each one, including several models that implement these techniques. These models have been limited to those based on GAN networks for their implementation, explaining these and the CNN networks on which they are based. In addition, a proof of concept has been implemented on the deepfakes technique for generating new faces. The chosen model has been DCGAN because it does not use as many resources as other more accurate models and the CelebA dataset for its high resolution images. The result has shown that, with a not very deep network, acceptable images are generated. This has been evaluated qualitatively, since the loss function does not ensure the convergence of the model. Other quantitative metrics have been found that could be tested in future works to see when to stop generating images more accurately. Another possible continuation of this work would be to look for detection techniques for these deepfake models and carry out some test with the generated images to check if they are false. |
Language: | Spanish |
URI: | http://hdl.handle.net/10609/146181 |
Appears in Collections: | Bachelor thesis, research projects, etc. |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
tsastretTFM0622memoria.pdf | Memoria del TFM | 9,62 MB | Adobe PDF | ![]() View/Open |
Share:


This item is licensed under a Creative Commons License