AI-Generated Facial Image Detection Using GAN-Based Synthetic Data and ResNet-50 Transfer Learning, with Real-Time Browser Deployment

Abstract

Generative Artificial Intelligence (Gen AI) has enabled the creation of clear, photo-realistic synthetic images that, in many cases, are indistinguishable from real photographs, allowing anyone to generate them. This has posed significant challenges in the area of digital media forensics, identity protection, cybercrime and fraud prevention, misinformation detection and image authenticity verification. In this work, a deep learning-based system consisting of a CNN model is designed to classify the facial images into two classes: AI-generated and Real. The CelebA dataset is used as the source of real images, while synthetic images are produced using a custom Generative Adversarial Network (GAN). The image classification is accomplished by fine-tuning a convolutional neural network (CNN) called ResNet-50. The proposed classifier attained the accuracy of 96.97% and Area under the ROC Curve (AUC) of 0.977 on the internally balanced test set of 4384 images, and maintained the accuracy of 94.0 % on the externally independent test set of 2041 images, thereby exhibiting strong generalization. This research offers a practical and efficient tool for the detection of images and verification of the authenticity of media, which is supported by Artificial Intelligence.