Optimizing Error of High-Dimensional Statistical Queries Under Differential Privacy
In this work we describe the High-Dimensional Matrix Mechanism (HDMM), a differentially private algorithm for answering a workload of predicate counting queries. HDMM represents query workloads using a compact implicit matrix representation and exploits this representation to efficiently optimize...
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Labor Dynamics Institute
2023-08-01
|
| Series: | The Journal of Privacy and Confidentiality |
| Subjects: | |
| Online Access: | http://www.journalprivacyconfidentiality.org/index.php/jpc/article/view/791 |
| _version_ | 1797653605744902144 |
|---|---|
| author | Ryan McKenna Gerome Miklau Michael Hay Ashwin Machanavajjhala |
| author_facet | Ryan McKenna Gerome Miklau Michael Hay Ashwin Machanavajjhala |
| author_sort | Ryan McKenna |
| collection | DOAJ |
| description |
In this work we describe the High-Dimensional Matrix Mechanism (HDMM),
a differentially private algorithm for answering a workload of predicate counting queries. HDMM represents query workloads using a compact implicit matrix representation and exploits this representation to efficiently optimize over (a subset of) the space of differentially private algorithms for one that is unbiased and answers the input query workload with low expected error. HDMM can be deployed for both ϵ-differential privacy (with Laplace noise) and (ϵ, δ)-differential privacy (with Gaussian noise), although the core techniques are slightly different for each. We demonstrate empirically that HDMM can efficiently answer queries with lower expected error than state-of-the-art techniques, and in some cases, it nearly matches existing lower bounds for the particular class of mechanisms we consider.
|
| first_indexed | 2024-03-11T16:47:00Z |
| format | Article |
| id | doaj.art-2b2d3aaec79144f6bc1349ecf7066b52 |
| institution | Directory Open Access Journal |
| issn | 2575-8527 |
| language | English |
| last_indexed | 2024-03-11T16:47:00Z |
| publishDate | 2023-08-01 |
| publisher | Labor Dynamics Institute |
| record_format | Article |
| series | The Journal of Privacy and Confidentiality |
| spelling | doaj.art-2b2d3aaec79144f6bc1349ecf7066b522023-10-22T08:38:26ZengLabor Dynamics InstituteThe Journal of Privacy and Confidentiality2575-85272023-08-0113110.29012/jpc.791Optimizing Error of High-Dimensional Statistical Queries Under Differential PrivacyRyan McKenna0Gerome Miklau1Michael Hay2Ashwin Machanavajjhala3a:1:{s:5:"en_US";s:36:"University of Massachusetts, Amherst";}The University of Massachusetts, AmherstColgate UniversityDuke University In this work we describe the High-Dimensional Matrix Mechanism (HDMM), a differentially private algorithm for answering a workload of predicate counting queries. HDMM represents query workloads using a compact implicit matrix representation and exploits this representation to efficiently optimize over (a subset of) the space of differentially private algorithms for one that is unbiased and answers the input query workload with low expected error. HDMM can be deployed for both ϵ-differential privacy (with Laplace noise) and (ϵ, δ)-differential privacy (with Gaussian noise), although the core techniques are slightly different for each. We demonstrate empirically that HDMM can efficiently answer queries with lower expected error than state-of-the-art techniques, and in some cases, it nearly matches existing lower bounds for the particular class of mechanisms we consider. http://www.journalprivacyconfidentiality.org/index.php/jpc/article/view/791differential privacylinear queriesworkloadhigh-dimensionalmatrix-mechanismmarginals |
| spellingShingle | Ryan McKenna Gerome Miklau Michael Hay Ashwin Machanavajjhala Optimizing Error of High-Dimensional Statistical Queries Under Differential Privacy The Journal of Privacy and Confidentiality differential privacy linear queries workload high-dimensional matrix-mechanism marginals |
| title | Optimizing Error of High-Dimensional Statistical Queries Under Differential Privacy |
| title_full | Optimizing Error of High-Dimensional Statistical Queries Under Differential Privacy |
| title_fullStr | Optimizing Error of High-Dimensional Statistical Queries Under Differential Privacy |
| title_full_unstemmed | Optimizing Error of High-Dimensional Statistical Queries Under Differential Privacy |
| title_short | Optimizing Error of High-Dimensional Statistical Queries Under Differential Privacy |
| title_sort | optimizing error of high dimensional statistical queries under differential privacy |
| topic | differential privacy linear queries workload high-dimensional matrix-mechanism marginals |
| url | http://www.journalprivacyconfidentiality.org/index.php/jpc/article/view/791 |
| work_keys_str_mv | AT ryanmckenna optimizingerrorofhighdimensionalstatisticalqueriesunderdifferentialprivacy AT geromemiklau optimizingerrorofhighdimensionalstatisticalqueriesunderdifferentialprivacy AT michaelhay optimizingerrorofhighdimensionalstatisticalqueriesunderdifferentialprivacy AT ashwinmachanavajjhala optimizingerrorofhighdimensionalstatisticalqueriesunderdifferentialprivacy |