Deep Learning-Based Detection of AI-Generated Images: A Comparative Evaluation
Abstract
The fast-paced evolution in the field of technology outperforms the human eye's ability to identify deep fakes, transforming automated detection from a luxury to a necessity. In this study, six deep learning models were evaluated namely MobileNetV2, ShuffleNetV2, ResNet18, DenseNet121, EfficientNetB0, and EfficientNetV2S against 79,950 labeled images to separate AI-generated imes from real photographs. A two-phase training strategy of an initial frozen-backbone warmup followed by full fine-tuning yielded F1-score gains between 8.35% and 13.5%. We have identified cross-shaped spectral patterns found in AI images using FFT analysis and used GradCAM maps to see that models were able to detect background-foreground boundaries instead of subjects. While the origin of the dataset being stock photography reveals its high separability, the results were steady across all six architectures. Among individual architectures, ShuffleNetV2 maintained a competitive 0.9849 F1-score with just 2.3M parameters, whereas DenseNet121 yielded highest individual score of 0.9935, which was finally surpassed using ensemble approach which reached 0.9953.