Deep Learning Based Classification of Real and Synthetic Animal Images

This study aims to develop a convolutional neural network (CNN)-based classification framework that can distinguish between synthetic and real animal images generated by the Stable Diffusion Turbo model. Additionally, this study will evaluate the performance of different network architectures for this task. The study employed a balanced dataset of 31,995 images, including 16,000 real images and 15,995 synthetic images generated by Stable Diffusion Turbo. The dataset includes eight animal categories: dogs, cats, cows, rabbits, horses, sheep, chickens, and elephants. All images were resized to 224 by 224 pixels, and standard preprocessing techniques were applied. During the classification stage, five pretrained convolutional neural network architectures were retrained using transfer learning, including MobileNetV2, DenseNet121, DenseNet169, DenseNet201, and NASNetMobile. Model performance was evaluated using accuracy, precision, recall, the F1 score, the area under the curve of the receiver operating characteristic, confusion matrices, and training time. The experimental results demonstrate that MobileNetV2 and DenseNet201 achieved the highest classification performance, with respective accuracy rates of 99.58% and 99.56%, and perfect area under the curve values. All DenseNet variants exhibited complete sensitivity in detecting synthetic images, whereas NASNetMobile showed substantially lower performance compared to the other models. These results suggest that synthetic images produced by diffusion-based generative models can be reliably identified when appropriately designed CNN architectures and balanced datasets are used. This provides a significant methodological contribution to the discrimination of synthetic versus real images, the detection of fake content, and the verification of visual authenticity.