Perbandingan EfficientNetB0, MobileNetV2, dan DenseNet121 Berbasis Transfer Learning untuk Deteksi AI-Generated Images

Abstract

The rapid advancement of AI-based generative models such as GANs and Diffusion Models has produced high-quality synthetic images that are visually indistinguishable from real photographs. This poses serious challenges in digital security, media verification, and public trust in visual content. However, few studies have compared multiple CNN architectures using a two-phase transfer learning strategy on the same dataset. This study proposes a transfer learning-based approach by comparing three ImageNet-pretrained CNN architectures, namely EfficientNetB0, MobileNetV2, and DenseNet121, for detecting AI-generated images. The dataset used is the AI vs Real Images Dataset from Kaggle, consisting of two classes: AI-generated and real images. Training was conducted in two phases: the first phase with a frozen backbone to adapt pretrained weights, and the second phase involving fine-tuning the last 30 layers of the backbone. Data augmentation including rotation, shifting, zoom, shear, and brightness variation was applied to the training set, while class weighting and label smoothing were employed to address class imbalance. Evaluation was performed using accuracy, precision, recall, F1-score, and AUC-ROC metrics. Experimental results show that DenseNet121 achieved the highest performance with an accuracy of 98.00%, precision 99.10%, recall 98.21%, F1-score 98.65%, and AUC-ROC 99.72%. EfficientNetB0 ranked second with 95.33% accuracy, while MobileNetV2 recorded 94.67% accuracy with the fewest parameters, making it superior in terms of efficiency. Grad-CAM visualizations confirm that the models focus on the main objects as well as textured and color-transition regions that serve as markers of generative artifacts. This study demonstrates that a two-phase transfer learning approach with structured data augmentation is effective for AI image detection tasks.