Hyperparameter-Aware Evaluation of Deep Learning Architectures for AI-Generated Image Detection

Abstract

The rapid advancement of generative artificial intelligence has raised growing concerns regarding the authenticity and reliability of digital visual content. Reliable detection of AI-generated images has therefore become an important research problem in digital forensics, media verification, and content integrity. This study evaluates four deep learning architectures-ResNet-50, EfficientNetV2-S, ViT-B/16, and Xception-for binary classification of real and AI-generated images using the CIFAKE dataset. Each model is assessed under baseline and hyperparameter-optimized configurations to examine the effects of learning rate, batch size, weight decay, data augmentation, and backbone-freezing strategy. The results show that hyperparameter tuning improves all evaluated models, although the magnitude of improvement is architecturedependent. EfficientNetV2-S and ViT-B/16 benefit most from tuning, while ResNet-50 and Xception show more moderate gains. Among individual models, the tuned ViT-B/16 achieves the highest accuracy and F1-score, whereas the tuned weighted softvoting ensemble obtains the best overall performance, reaching $\mathbf{9 9. 0 7 {\%}}$ accuracy and $\mathbf{9 9. 0 7 {\%}}$ F1-score on the test set. These findings provide practical guidance for selecting and tuning deep learning models for synthetic image detection and highlight the need for further evaluation on higher-resolution datasets and diverse generative models.