Detecting Generative AI in Images

The increasing realism of AI -generated images poses a significant challenge, as they can often deceive human perception. This resulted in a pressing need to develop robust methods for detecting such synthetic imagery. This paper investigates the effectiveness of various convolutional neural network (CNN) architectures, including several adapted ResN et-50 models, in classifying real and AI -generated images. The impact of preprocessing techniques, specifically discrete Fourier transform and cross-difference filters, on enhancing classification accuracy were explored. This study focuses on fake images sourced from diverse generative models, including Stable Diffusion vl.4, Midjourney v6, Gemini, Microsoft Copilot's DALL-E 3, and BigGAN. Notably, the unexplored Midjourney v6 was focused on, which presents a recent addition to the literature. Through multiple trials, the performance of different CNN models on these datasets is assessed and the highest achieved accuracy for each is documented, highlighting the success of the adapted ResNet-50 model in several instances. The question of image sizes' impact on classification accuracy was also raised, revealing its effect in specific trials. The findings of this study provide valuable insights into effective strategies for detecting AI -generated images, thus advancing the field of image classification in the context of synthetic content.