google / google/differential-privacy
Question on monotonic queries
- Dominant language
- Go
- Stars
- 3.4k
- Forks
- 434
- Avg merge
- 4d 9h
- Merged PRs (30d)
- 9
Description
Hi!
First, thanks for this amazing library! I use it heavily in my research, and it has been spotless! :)
A quick technicality: It is common knowledge (mentioned in the Algorithmic Foundations of Differential Privacy as well) that queries that are monotonic w.r.t to the neighbouring relationship enjoy a discount in sensitivity of 1/2. More precisely, in the add/remove neighbouring definition, when we add, we can only increase the output of our query; similarly, when we remove, we can only decrease the output of a query.
I assume this would also work nicely for PLDs, but I would like to know if there is support for this out of the box. I am not sure if it's enough to halve the sensitivity. Would any change to the mechanisms be required?
Contributor guide
Assessment
This issue has not been assessed yet.