pytorch / opacus

Training PyTorch models with differential privacy
https://opacus.ai
Apache License 2.0
1.71k stars 343 forks source link

Double-charging privacy in DC-GAN example #418

Open ffuuugor opened 2 years ago

ffuuugor commented 2 years ago

tl;dr: dcgan.py example adds noise to the fake data gradients, which it doesn't need to. We should change the training pipeline (because it's more correct) and measure the impact (because we're curious).

See more details and solution proposals on the PyTorch forum thread

Koukyosyumei commented 2 years ago

@ffuuugor

Hi! I'm Hideaki, an undergrad student at the University of Tokyo and a research intern at Tsinghua University on privacy-preserving machine learning. I am currently working on this issue. I've made a PR but need approval from the maintainer to run the workflow. How should I proceed?

romovpa commented 2 years ago

@Koukyosyumei Thank you for contributing! We will review it soon. Let's discuss in your PR, I've left a comment.