Vulnerability of Diffusion Models to Adversarial Attacks

Anubhooti Jain, Susim Mukul Roy, Dr. Richa Singh, Dr. Mayank Vatsa
Last updated on Nov 29, 2023

Summary

  • Conducted an extensive literature survey of diffusion models and their advantages and disadvantages.
  • Implemented an architecture which works with different pretrained DDPMs and classifiers.
  • Developed a novel adversarial attack using the Class-Activation Maps of classifiers and the predicted noise maps from the UNet model of a DDPM which gives a better attack in terms of ASR, Robust accuracy and FID.
  • Worked with datasets like CIFAR10, CelebaHQ, FFHQ etc.
Coursework CV Research Projects