<![CDATA[Recent advancements in generative technologies, such as Generative Adversarial Networks (GANs) and Latent Diffusion Models, have enabled the creation of AI-generated synthetic images that are increasingly indistinguishable from real ones, posing significant challenges for verifying the authenticity of visual content. This study develops a DenseNet-121 model with hyperparameter optimization and the integration of Squeeze-and-Excitation (SE) attention mechanisms at Early, Mid, and Late positions. Experiments were conducted using the CIFAKE dataset with a resolution of 32×32 pixels to compare the baseline Plain model with three SE variants. Hyperparameter optimization was applied to maximize model performance. The results demonstrate that the Plain DenseNet-121 with optimized hyperparameters achieved an accuracy of 98.52%, outperforming the standard configurations reported in previous studies. The integration of SE yielded varied outcomes, where Mid SE attained the highest accuracy of 98.56%, while Early SE (98.45%) and Late SE (98.48%) exhibited greater stability with lower standard deviations. These findings highlight that combining hyperparameter optimization with appropriate SE placement can enhance model performance for classifying real and AI-generated images. Moreover, SE placement at different positions (Early, Mid, Late) has a significant impact on feature representation and generalization in synthetic image classification, which is increasingly important given the growing difficulty of distinguishing real from AI-generated images.]]>
Copyrights © 2025
