I'm aware we can quantify privacy with ε-differential privacy (ε-DP). But when we apply DP, how do we actually select the value for ε ? Are there some rule-of-thumbs? Is it decided case-by-case basis? In general, how do we decide we've enough privacy when using some algorithm that satisfies ε-DP definition?
In the 2019 paper Differential Privacy in Practice: Expose Your Epsilons!, the authors Dwork, Kohli, Mulligan summarize the state of affairs thusly:
We found no clear consensus on how to choose ε, nor agreement on how to approach this and other key implementation decisions. Given the importance of these details there is a need for shared learning amongst the differential privacy community.
They propose an "Epsilon Registry" where applications of DP can specify their choice of ε. The closest to such a registry that I've been able to find is this list of real-world uses of DP by Damien Desfontaines, from Jan 2021.
I also found this very recent blog post from NIST which says "what, exactly, does ε mean, and how should we set it? Unfortunately, we still don’t have a consensus answer to this question", but then goes on to actually give some broad numerical guidance:
- There is broad consensus that ε values in the low single digits (i.e. 0 < ε < 5) represent a conservative choice, and will provide strong privacy protection
- Increasing experience with deployed systems suggests that larger ε values (i.e. 5 < ε < 20) also provide robust privacy protection in a variety of settings
- In some contexts, even higher values of ε (i.e. ε > 20) may still provide meaningful privacy protection, but more experience is needed to understand this setting