Towards Accurate Classification of Authentic Photographs and AI-Generated Images Using Deep Learning Approaches

Controlled, matched-condition comparison: five CNN backbones (ResNet18, ResNet50, VGG16, VGG19, EfficientNetB0) evaluated under a unified transfer-learning protocol on the same 120,000-image CIFAKE benchmark, isolating architecture effects from training-setup differences.
Rigorous data-integrity pipeline: corruption checks, MD5-based cross-split duplicate detection, and blur flagging applied to CIFAKE before training, closing a common data-leakage gap in prior work on this dataset.
Fine-tuning depth as the key variable: the study shows fine-tuning strategy (full vs. frozen-backbone) matters more than raw architecture choice, with fully/partially fine-tuned models outperforming frozen ones by up to ~14 accuracy points, and frozen VGG models developing a measurable class bias (many more false positives than false negatives).
A new low-data benchmark: the custom \textit{Images} dataset (2,400 images) tests whether these findings hold when only ~1,700 training images are available, addressing a real-world constraint that CIFAKE-scale benchmarks don’t capture.
Soft-voting ensemble result: combining EfficientNetB0, ResNet, and VGG on the custom dataset eliminates all remaining classification errors (Precision = Recall = F1 = 1.0000), demonstrating that ensembling can close the residual gap left by any single fine-tuned model even in a small-data regime.
Practical deployment guidance: an accuracy-vs-efficiency comparison (params, relative inference time) identifying EfficientNetB0 as the best accuracy-per-parameter option for resource-constrained, real-world content-moderation deployment.