Comparative Analysis of ResNet-50 and EfficientNet Architectures in Detecting AI-Generated Images vs. Real-World Photography
Year: 2026
Type: conference-paper
Abstract: The development of artificial intelligence technology, especially generative models, has produced synthetic images that have a very high level of similarity to original photos. This condition poses new challenges in the process of verifying the authenticity of digital content and increases the need for accurate and reliable detection methods. This study aims to compare the performance of two Convolutional Neural Network (CNN) architectures, namely ResNet-50 and EfficientNet-B0, in classifying real images and AI-generated images. Testing was conducted using two datasets, namely the Artifact Dataset and the Defactify Dataset by applying the same training configuration on the Kaggle platform using the PyTorch framework. Model evaluation was carried out based on the values of *accuracy*, *precision*, *recall*, and *F1-score*. The results showed that EfficientNet-B0 consistently provided better performance than ResNet-50 on both datasets. EfficientNet-B0 achieved 97.67% accuracy on the Artifact Dataset and 97.94% on the Defactify Dataset, while ResNet-50 achieved 94.08% and 87.41% accuracy, respectively. In addition to achieving higher accuracy, EfficientNet-B0 also demonstrated better generalization capabilities in detecting AI-generated images with a high degree of realism. These findings indicate that EfficientNet-B0 is a more effective architecture for supporting AI-generated image detection systems across varying levels of data complexity.
Institution: Binus University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204759805