GAN/CNN Image Detector

Project details

Trained multiple architectures (EfficientNetV2, StyleGAN/ProGAN discriminators) on a 10k+ image dataset to detect synthetic imagery.

• • Transfer learning, dropout, heavy augmentation
• • Evaluation with ROC/AUC, PR curves and confusion matrices

Results & impact

• • 97% acc · 0.973 F1 · 0.997 AUC (EffNetV2)