NoteNextra-origin/content/CSE5519/CSE5519_H3.md

CSE5519 Advances in Computer Vision (Topic H: 2023: Safety, Robustness, and Evaluation of CV Models)

How to backdoor diffusion models

link to paper

Tip

This is an interesting paper showing that it is possible to backdoor a diffusion model with high utility and high specificity at a low cost compared to per-training.

I wonder how this technique could possibly be used for AI watermarking and reliably detected with other AI operations?

And there are many metrics and loss functions used in this paper, I wonder what objectives they are trying to optimize, and looking for more clarification on the presentation.