Advancing AI-Generated Image Detection: Enhanced Accuracy through CNN and Vision Transformer Models with Explainable AI Insights

In recent years, the world has witnessed remarkable progress in the field of Generative AI. These advancements have ushered in an era where we can generate images that appear extremely lifelike, making it challenging for the human eye to distinguish them from reality. Such lifelike images have the potential to be harnessed for various purposes, including manipulation and the dissemination of misinformation. Consequently, it has become increasingly crucial to detect these AI-generated images accurately to safeguard the integrity of visual information. For this purpose, we delved into the realm of Convolutional Neural Network (CNN) models and Vision Transformer models, exploring their capabilities in distinguishing between authentic and AI-generated images. We conducted our investigations using a publicly available dataset called the CIFAKE: Real and AI-Generated Synthetic Images dataset. Through training different architectures and tuning hyperparameters, we identified the optimal model for this crucial task. Our selected CNN model exhibited an impressive accuracy rate of 96.31%, signifying a substantial stride toward the accurate identification of AI-generated images. We used Grad-CAM (Gradient-weighted Class Activation Mapping) to interpret the model’s output and enhance our understanding of how it generates these outputs. This interpretative tool allowed us to elucidate the model’s decision-making process, shedding light on the discriminative features and patterns it utilizes to distinguish real from synthetic images.