-
Notifications
You must be signed in to change notification settings - Fork 923
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Central dp fixed clipping #2765
Conversation
) -> Tuple[Optional[Parameters], Dict[str, Scalar]]: | ||
"""Aggregate training results using unweighted aggregation.""" | ||
if failures: | ||
return None, {} |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
May I know why this DP strategy does not tolerate any failures? Is it fine to proceed with the DP algorithm with a few clients dropping out?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This makes the epsilon computation afterwards very difficult. As we need to know how many clients were selected on each round. I would say let's keep it this way for now and we can investigate those scenarios.
@@ -0,0 +1,217 @@ | |||
# Copyright 2020 Flower Labs GmbH. All Rights Reserved. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The code looks good. Just a few small things. Btw, could you get past all the CI checks?
944bc14
to
7927581
Compare
7927581
to
944bc14
Compare
Issue
Description
The implementation of enabling strategies with central differential privacy. It allows the server to clip (with fix value) the updates and add noise to them.
Related issues/PRs
N/A
Proposal
Explanation
Implement a new wrapper class
DPWrapper_fixed_clipping
(which itself is a strategy) to first it clips the updates from the clients before passing them to the inner strategy, and second, it adds noise to the aggregated results.Checklist
#contributions
)Changelog entry
The strategies now can be wrapped around with a dp class to allow them perform clipping and noise addition.
Any other comments?
N/A