AI Image Detection Using EfficientNetB0 Architecture on Generative Adversarial Network

Abstrak

The development of Generative Adversarial Networks (GANs) has produced synthetic images that are increasingly difficult to distinguish from real photographs, driving the need for reliable automated detection systems. The core problem is that detection models frequently fail when tested on generators different from those used during training, a phenomenon known as the generalization gap. This study evaluates EfficientNetB0 in detecting AI-generated images through a cross-generator approach across five GAN architectures, namely ProGAN, StyleGAN, StyleGAN2, StyleGAN2-ADA, and StyleGAN3. The model was trained on a StyleGAN2 dataset using transfer learning from ImageNet, then evaluated on the four other generators without retraining. In-domain results showed an accuracy of 99.30%, an F1-Score of 99.30%, and an AUC of 99.99%. However, out-of-domain testing revealed an average accuracy drop of 16.64%. StyleGAN2-ADA achieved 99.33% due to its architectural similarity to StyleGAN2, suggesting that generator architecture is a more decisive factor than training strategy. In contrast, StyleGAN3 dropped to an accuracy of 63.05% because its alias-free architecture eliminates the visual patterns that the model typically relies on for detection. The model tends to recognize real images well with high specificity, but its ability to detect synthetic images declines with low sensitivity, and the false negative rate rising from 0.15% to 72.65% on StyleGAN3. These findings highlight the limitations of single-generator training and the need to explore multi-generator strategies or frequency-based feature methods.

Identifier

• DOI: https://doi.org/10.30871/jaic.v10i2.12481
• OpenAlex: https://openalex.org/W7160938625