Propose-Filter-Release: Scaling Noise to Typical Sensitivity via Private Filtering
A few unusually influential records need not dictate privacy noise—if filtering them out is itself private.
The paper introduces propose-filter-release, which converts a mechanism that protects changes involving typical records into one with standard differential privacy on every input. It requires a concave score measuring how well each record fits the remaining data, plus a release mechanism that is private when the changed record fits. The privacy guarantee holds under public conditions on the number of records. The authors demonstrate applications to scalar statistics, principal subspaces, histogram divergence, graph effective resistances and treatment-rule selection.
ArXiv · AI/CL/LG's note
The paper introduces propose-filter-release, which converts a mechanism that protects changes involving typical records into one with standard differential privacy on every input. It requires a concave score measuring how well each record fits the remaining data, plus a release mechanism that is private when the changed record fits. The privacy guarantee holds under public conditions on the number of records. The authors demonstrate applications to scalar statistics, principal subspaces, histogram divergence, graph effective resistances and treatment-rule selection.
ArXiv · AI/CL/LG's note
score 5