From an article [0] published in last month's CACM, describing attempts by three 2016 conferences to anonymize papers:

We find that anonymization is imperfect but fairly effective: 70%–86% of the reviews were submitted with no author guesses, and 74%–90% of reviews were submitted with no correct guesses.

[0] https://cacm.acm.org/magazines/2018/6/228027-effectiveness-o...