FERREIRA, A. S. C.; http://lattes.cnpq.br/1104868031869926; FERREIRA, Arthur Silva Cavalcante.
Resumen:
Generative Adversarial Networks (GANs) have broad applications, ranging from image and video creation to text generation and product design. In the context of this study, synthetic face images generated by GANs will be evaluated. There are benefits to using GANs, such as external research to understand the complexity and nuances of facial images and the creation of anonymous databases for training neural networks with facial images. However, synthetic faces can be used to create false identities, leading to crimes such as identity theft and phishing, where synthetic faces are used to deceive facial recognition-based security systems. Additionally, they can also be used to create videos and fake images with malicious intent, such as defamation, misinformation, or political propaganda. In this work, a Deep Convolutional Neural Network based on the EfficientViT architecture was trained using a dataset composed of publicly available databases and synthetic images generated by the StyleGAN3 network. The results obtained indicate an accuracy rate of 99%, similar to other methods in the literature, but the databases used for training and evaluation vary beyond the number of images used in the evaluation. Furthermore, there was a search for diversified databases to mitigate bias and model fairness regarding age/ethnicity, but a separate analysis would be necessary to assess the impact of this choice of databases compared to other models already available in the literature.