CIFAKE: IMAGE CLASSIFICATION AND EXPLAINABLE IDENTIFICATION OF AI-GENERATED SYNTHETIC IMAGES

Abstract

Recent advancements in synthetic data generation have led to AIproduced images so realistic that they are often indistinguishable from real photographs. Recognizing the importance of data authenticity, this study explores the use of computer vision to detect AI-generated images. A synthetic dataset, mimicking the ten classes of CIFAR-10, was created using Latent Diffusion Models (LDMs), producing images with complex visual details such as realistic water reflections. This dataset forms the basis for a binary classification task to determine whether an image is real or AIgenerated. By employing a Convolutional Neural Network (CNN) and optimizing 36 different network configurations, the model achieved over 90% accuracy. Gradient Class Activation Mapping (Grad-CAM) was utilized to identify key features for classification, revealing that subtle background imperfections are crucial indicators of AI-generated content.