AI Image Classification and Artifact Identification

A robust detection framework using LeViT and Neighboring Pixel Representation achieving 99.21% accuracy on CIFAKE.

Collaborator: Ms. Aishwarya Agarwal, Adobe Research

  • Developed a detection framework using the LeViT model and NPR to identify AI-generated images.
  • Achieved 99.21% accuracy on the CIFAKE dataset, securing the highest accuracy among competitors.
  • Extending the project with robust techniques for out-of-distribution generative models.
  • Exploring adversarial training, jigsaw-style self-supervised learning, and continual learning.