PFR lets you build private algorithms that add much less noise by privately removing outliers first, rather than scaling noise to worst-case sensitivity. This works for any algorithm that's 'friendly-DP' (private when data fits well) and any measure of how well records fit the dataset.
This paper introduces Propose-Filter-Release (PFR), a method that makes machine learning models differentially private by first privately removing outlier records that would otherwise force large noise additions.