protects individuals by adding noise to data analysis. Standard methods scale this noise to the worst possible influence of any single record, making results inaccurate even when most records are well-behaved and stable.
The paper introduces propose-filter-release, a method that privately filters a dataset. It uses a friendliness score to remove records that do not fit the rest, leaving only typical records. The main analysis then only needs noise scaled to the small influence of these typical records.
This turns a mechanism for typical data into a fully private one. PFR is a general primitive that enables accurate privacy-preserving statistics across many domains by focusing the privacy burden on the data that actually demands it.