CNN Models Compared for GAN-Generated Fake Image Detection

A recent study published on arXiv (2605.20971) investigates the performance of four pretrained CNN models—VGG16, ResNet50, EfficientNetB0, and XceptionNet—in identifying images altered by GANs. The researchers implemented a standardized preprocessing approach involving resizing, normalization, and augmentation to mitigate class imbalance. The models were assessed based on Accuracy, Precision, Recall, F1-score, and ROC-AUC. VGG16 recorded the highest accuracy at 91%, while XceptionNet, ResNet50, and EfficientNetB0 all achieved 90%. Although EfficientNetB0 demonstrated greater sensitivity to manipulated images, it exhibited diminished reliability with authentic samples, highlighting a bias due to imbalance. The study acknowledges limitations such as dataset imbalance, overfitting, and restricted interpretability that could impact robustness across different domains.