AI Deepfake Image Classification
A transfer learning pipeline that classifies images as real or AI-generated using five pre-trained architectures, achieving strong results with ResNet-50 across multiple policy-driven evaluation scenarios.
This project utilized a Kaggle dataset of 970 images to differentiate between real and AI-generated images. The core methodology was transfer learning — leveraging five pre-trained model architectures to learn the subtle differences between authentic and synthetic imagery.
We tested each architecture to determine which performed best, with ResNet-50 emerging as our top performer. Beyond raw accuracy, we designed three distinct classification policies tailored to different real-world scenarios.
Policy 1 focused on minimizing false positives — critical in scenarios where incorrectly flagging real art as AI-generated has serious consequences. Policy 2 provided a balanced approach with equal weight on precision and recall. Policy 3 prioritized high recall, ensuring as many AI-generated images as possible are caught, even at the cost of some false positives.
Each policy was framed around practical use cases, demonstrating how the same underlying model can be tuned to serve different needs in AI art detection and content authenticity verification.