Discriminating of real and Generated faces using Deep Convolutional using GAN’s

Generative Adversarial Networks (GANs) has been a revolutionary methodology for generating highly realistic synthetic images, especially human faces with the rapid advancements in deep learning. Deep Convolutional GANs (DCGANs), a popular variant of GANs leveraging convolutional architectures, have demonstrated remarkable ability to produce photorealistic face images indistinguishable to the human eye. While these advances have immense applications in entertainment, art, and data augmentation, they also pose significant challenges associated to misinformation, secrecy, and safety due to the increasing prevalence of deep fakes and synthetic media. This thesis addresses the critical problem of discriminating between real and generated face images using DCGAN-based models, aiming to develop an effective and reliable classifier capable of detecting synthetic faces generated by state-of-the-art GAN architectures. The core of the methodology is centered on the Deep Convolutional GAN framework, comprising two adversarial neural networks: a generator that creates synthetic face images from random noise vectors, and a discriminator trained to organise images as actual or forged. The generator employs convolutional transpose layers and non-linear activations to produce realistic images, while the discriminator practices convolutional layers to extract discriminative features and outputs a probability of authenticity. The networks are trained iteratively using the Binary Cross Entropy loss function with the Adam optimizer, carefully tuned for stable convergence. Weight initialization and activation functions are chosen based on best practices from the literature to maximize the performance of both networks.