Image Classification and Explainable Identification of AI-Generated Images

Abstract

The unprecedented advancements in Artificial Intelligence (AI) over the past few years have significantly enhanced the ability of generative models to produce synthetic images that are virtually indistinguishable from real-world photographs. As the boundary between authentic and AI-generated images becomes increasingly blurred, the critical necessity for reliable methods of image authentication and data verification has emerged. In response to this growing challenge, the present study proposes a novel approach to enhancing the detection of AI-generated images through the application of computer vision techniques. The research introduces the creation of a synthetic dataset, named Artifact, which mirrors the ten classes found within the widely utilized Artifact dataset. This dataset was generated using a Latent Diffusion Model (LDM), capable of producing highly complex and photorealistic visual features, such as realistic reflections and depth of field effects. The binary classification task is then formulated, challenging the system to distinguish between real and synthetic imagery. To address this task, a Convolutional Neural Network (CNN) architecture was designed, trained, and rigorously optimized across 36 different configurations through hyperparameter tuning. The resulting model achieved an impressive classification accuracy of 94.30%, demonstrating its ability to discern subtle imperfections and anomalies often overlooked by human observers. Furthermore, the study integrates Explainable AI (XAI) techniques, employing Gradient Class Activation Mapping (Grad-CAM) to visualize the internal decision-making processes of the CNN model. These visual explanations reveal that, contrary to intuitive assumptions, the CNN does not primarily focus on the main objects within the images. Instead, it detects small background inconsistencies and imperfections that serve as key indicators of synthetic origin.