<![CDATA[Purpose – Deepfake-related cybercrime is an increasingly troubling cybersecurity threat, since such content can now be produced from a single facial photo using freely available face-swapping techniques, while human ability to distinguish real from fake images remains limited (48.2–59%). This study compares a simple CNN and a fine-tuned ResNet50 within an identical, controlled framework.Methods – A controlled experiment compared a simple CNN (trained from scratch) and ResNet50 (two-phase fine-tuning) on 20,000 images from the Kaggle Deepfake and Real Images dataset (80:10:10 split, seed = 42). Candidate images were deduplicated before the data split; zero cross-split duplicates were confirmed. ResNet50 used ResNet-specific preprocessing; the CNN used inputs normalized to [0,1]. Evaluation used accuracy, precision, recall, F1-score, AUC-ROC, and specificity, with a paired McNemar's test as the primary significance measure.Findings – The CNN outperformed ResNet50 on six of seven metrics (accuracy 90.90% vs. 89.80%; AUC-ROC 97.08% vs. 96.52%), while ResNet50 achieved higher recall (93.10% vs. 91.80%). The accuracy difference was not statistically significant (p = 0.200). The CNN was 69.5 times smaller and trained faster (12.2 vs. 16.5 minutes). Neither model showed clear overfitting. Research implications – The findings rest on a single subset, a single split, and a single training run per model, limited to 20 epochs; the potential of ResNet50 with longer training has not been explored.Originality – This study combines accuracy, generalization-related diagnostics, and efficiency in a single controlled comparison, applies a paired test appropriate for a shared test set, and filters duplicates before the data split.]]>
Copyrights © 2026
