search

adap / flower

Flower: A Friendly Federated AI Framework
https://flower.ai
Apache License 2.0
5.16k stars 881 forks source link

Differential Privacy #605

Open tanertopal opened 3 years ago

tanertopal commented 3 years ago

Differential Privacy is an important element of many FL workloads. The purpose of this issue is to track interest related to it. Please upvote this issue if you need a Flower-provided Differential Privacy solution for your FL setup.

anik123 commented 3 years ago

Hi,

Any update on this?

Thanks in advance.

giorgosera commented 3 years ago

@tanertopal I am also interested in this. Is there a recommended workaround/library than can be used meanwhile?

danieljanes commented 3 years ago

@anik123 @giorgosera a first code example using DP SGD just got merged into main: https://github.com/adap/flower/tree/main/examples/dp-sgd-mnist

giorgosera commented 3 years ago

That's great news @danieljanes. Thank you!

anik123 commented 3 years ago

wow. Thanks, @danieljanes . I will give it a try.

bridgesra commented 9 months ago

It appears that the PRIVACY_LOSS variable in this code is using basic composition (summing epsilon and summing delta) across all clients. Is that right?

In many applications the clients require privacy, so one would perform privacy accounting per client. Further, for a large number of rounds, we may wish to use advanced composition theorems. Is there a privacy post-processing example that does this yet?