pytorch / opacus

Training PyTorch models with differential privacy
https://opacus.ai
Apache License 2.0
1.72k stars 345 forks source link

Unable to Adjust Alpha Range for Privacy Bound in Latest Version of Opacus #670

Closed zedoul closed 2 months ago

zedoul commented 2 months ago

UserWarning: Optimal order is the largest alpha. Please consider expanding the range of alphas to get a tighter privacy bound.

According to the official Opacus documentation, the privacy_engine.make_private_with_epsilon function should allow to specify a range of alphas to obtain tighter privacy bounds as follows:

You can pass your own alphas by passing alphas=custom_alphas when calling privacy_engine.make_private_with_epsilon

However, it seems that this functionality is no longer available in the latest version.

EnayatUllah commented 2 months ago

Thanks for the pointer! You are right, the functionality is removed now. We will update the website accordingly.