AI vs Reality: Classifying Synthetic Faces with a Fine-Tuned ResNet50 Neural Network

with the rise of generative technologies, distinguishing between real and AI-generated images has become increasingly challenging. Advanced generative frameworks such as Generative Adversarial Networks (GANs) and Latent Diffusion Models (LDMs) now generate highly convincing synthetic images that closely resemble genuine photographs. This phenomenon poses significant challenges for domains including cybersecurity, journalism, and social media platforms, where image authenticity verification is paramount. This study explores the application of ResNet50 deep learning architecture for distinguishing between AI-synthesized and authentic facial images. Our model underwent training using a comprehensive dataset containing 140,000 facial photographs, equally distributed between genuine and artificially generated samples. The ResNet50 architecture was enhanced through transfer learning methodologies to improve its capability in identifying subtle characteristics that differentiate authentic images from synthetic ones. Two distinct experimental approaches were employed: feature extraction methodology and comprehensive fine-tuning procedures. The optimized model demonstrated remarkable performance, achieving accuracy rates of up to 98%, validating its effectiveness in this domain. This investigation demonstrates the effectiveness of fine-tuned ResNet50 architecture in identifying AI- synthesized images. The research contributes to developing robust verification systems for image authentication, combating the proliferation of synthetic content, and maintaining the integrity of digital media platforms.