Differentially Private Data Release under Partial Information

No ratings

Presented at PEPR'19 2019 by

David Zeber

Differential privacy (DP) is now a standard technique for releasing reports based on sensitive data. However, selecting and tuning a DP mechanism so as to obtain high utility of the privacy-protected data is often difficult without detailed knowledge of the characteristics of the sensitive dataset. We propose an applied methodology for guiding the implementation of DP data protections using only partial summary information about the private data. In this setting, candidate DP mechanisms can be evaluated across possible realizations of the sensitive dataset, the selection of which is feasibly constrained using the available partial information. We demonstrate our approach for the problem of reporting the DP-protected distribution of item frequencies from a dataset of user-item pairs.