A Comparative Analysis of Fake Image Detection in Generative Adversarial Networks and Variational Autoencoders

Abstract

The rise of deep fake technology has led to growing concerns about its potential misuse for propaganda, disinformation, and even cybercrimes. Deep fake detection has thus become a crucial research area to prevent the spread of fake content and protect digital authenticity. This research investigates the detection of fake human face images using deep learning models. The study utilizes a combination of the CelebA-HQ and FFHQ datasets to create real image labels. Two generative models, a Style-based generator and a Variational Autoencoder, are trained to generate GAN-fake and VAE-fake images, respectively. The deep learning models are then trained and evaluated using both real and fake images. The research focuses on two scenarios: binary net and multi-class net. In the binary net scenario, ResNet50 achieves the highest accuracy of 99.34%, along with excellent precision, recall, and F1-score. VGG16 and VGG19 also perform well in distinguishing between fake and real images. In the multi-class net scenario, ResNet50 again achieves the highest accuracy of 95.25% and balanced F1-score. VGG16 and VGG19 maintain competitive performance, while other models show slightly lower accuracy. These results provide insights into the effectiveness of different deep learning models for detecting fake human face images. ResNet50 consistently performs well in both scenarios, while VGG16 and VGG19 offer reliable alternatives.