AI Photo Detector – Deepfake Image Classification

Fine-tuned a pretrained EfficientNet-B0 CNN to classify real vs. AI-generated images on the OpenFake dataset (~448K images), reaching 91.6% test accuracy, 0.975 ROC-AUC and 0.976 precision with a production training loop (AMP mixed precision, AdamW + OneCycleLR, EMA, label smoothing, discriminative learning rates) on a Tesla T4. Publication-grade evaluation harness with 12+ metrics, ROC/PR curves, calibration and t-SNE, plus robustness testing under JPEG/blur/downscaling and GradCAM saliency explainability.