pytorch / opacus

Training PyTorch models with differential privacy
https://opacus.ai
Apache License 2.0
1.67k stars 332 forks source link

When using adaclip , how to calculate eps or delta #539

Closed Adsensess closed 1 year ago

Adsensess commented 1 year ago

When using adaclip , i can't find where to input my delta( privacy parameters ) . Thus, i can't control the privacy budget so that when using the paper recommendation parameters, the epsilon of the RDP is too large . when print(privacy_engine.accountant.get_privacy_spent(delta=1e-5)) , where delta=1e-5 is one of the paper recommendation parameters. I get (5088492.807283732, 1.1)

I don't know if this is a bug, but I'd love an answer. Many thanks to the Opacus team for their efforts and support.