Detection of AI-Generated Images Using CNN

The rapid rise of modern artificial intelligence has greatly advanced text-to-image generation, enabling the creation of synthetic visuals that are often indistinguishable from real photographs. In this study, we propose a novel detection framework built on a gated-expert Convolutional Neural Network (CNN) to identify images created by various generative models, such as BigGAN, GLIDE, VQDM, and Stable Diffusion. The proposed system employs transfer learning using ResNet-50, where each expert network is individually fine-tuned for a specific generator model. The gating module assigns dynamic importance to each expert based on the characteristics of the input image. Experimental results show that the proposed approach performs better than conventional single-CNN and basic ensemble techniques, particularly when the training data are limited. Although generator-specific models perform well in their own domains, they show a decline in accuracy for unseen generators. In contrast, our gated architecture consistently maintained an accuracy above 96% across mixed datasets. This study highlights the importance of ensemble-based solutions and diverse datasets in detecting synthetic media, offering a promising direction for future research on generalizable and adaptable detection systems.