Is it possible to use Differential Privacy on the DoppelGANger architecture? #152
-
|
I want to train a DoppelGANger architecture and apply Differential Privacy to synthesize privacy-preserving time series data. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Hello! We do not offer differential privacy with DoppelGANger at present. You could fork the repo and swap out the optimizer in the discriminator with a differentially private one from Opacus, the DP optimizer library for torch. Note that the generator's optimizer needn't be changed since it never directly uses the training data, and would be covered by the post processing property of DP on the discriminator. Let us know how it goes! |
Beta Was this translation helpful? Give feedback.
Hello! We do not offer differential privacy with DoppelGANger at present. You could fork the repo and swap out the optimizer in the discriminator with a differentially private one from Opacus, the DP optimizer library for torch. Note that the generator's optimizer needn't be changed since it never directly uses the training data, and would be covered by the post processing property of DP on the discriminator. Let us know how it goes!