Choosing Epsilon for Privacy as a Service

Abstract

In many real world scenarios, terms of service allow a producer of a service to collect data from its users. Producers value data but often only compensate users for their data indirectly with reduced prices for the service. This work considers how a producer (data analyst) may offer differential privacy as a premium service for its users (data subjects), where the degree of privacy offered may itself depend on the user data. Along the way, it strengthens prior negative results for privacy markets to the pay-for-privacy setting and develops a new notion of endogenous differential privacy. A positive result for endogenous privacy is given in the form of a class of mechanisms for privacy-as-a-service markets that 1) determine ɛ using the privacy and accuracy preferences of a heterogeneous body of data subjects and a single analyst, 2) collect and distribute payments for the chosen level of privacy, and 3) privately analyze the database. These mechanisms are endogenously differentially private with respect to data subjects’ privacy preferences as well as their private data, they directly elicit data subjects’ true preferences, and they determine a level of privacy that is efficient given all parties’ preferences.

On Heuristic Models, Assumptions, and Parameters

“…(iv) social-empirical: the validity or choice of object is a social question, but one that can be placed in empirical terms by social scientists; e.g., choice of in robust ML when the modeled process is social [61,81,108], choice of in differential privacy following the work quantifying privacy in economic terms [5,67,74,75,93]; choice of fairness parameters under empirical advisement from social scientific research [27,43,103]; and (v) social: the validity or choice of object is a social question that can only be resolved by humanist analysis of its effects; e.g., all social-empirical examples when we do not have or believe in the completeness of social scientific methods.…”

“…Like any budget, if too limited it has little utilityi.e., the returned statistics will not be meaningful -while if too generous nothing is beyond acquiringi.e., little to no data privacy is actually provided. It is a knob that must be tuned to tradeoff privacy and accuracy, with (i) no immediate from-first-principles approach with which to do so, as evidenced by the breadth of techniques proposed [5,46,47,54,67,74,75,77,78,80,93], and (ii) penalties to accuracy or privacy from a flawed choice, as demonstrated by the controversy surrounding its use for disclosure avoidance in the 2020 US Census [5,9,22,24,25,54,64,68,72,100,101,112], and critical comments on deployments by Uber and Apple [52,106].…”

“…Its choice is, in the opinion of Dwork and Smith, 'essentially a social question' [47]. The problem has been studied formally through the lenses of resource allocation, economic valuations, and individual preferences [5,67,74,75,93]. These approaches read promisingly, but generally require a responsible partyan individual themselves, or a corporation or government acting on their behalf -to precisely quantify privacy valuations or preferences over data and events.…”

An Incentive Mechanism for Trading Personal Data in Data Markets

“…Several techniques for pricing data assuming a data market environment have been studied in [19][20][21][22][23][24][25][26].…”

SoK: Differential privacies

“…Different risk models, like the definitions in Section 3, are also compatible with varying the privacy parameters across inputs. For example, in [107], the author proposes endogeneous DP, which is a combination of (ε, δ)-DP and personalized DP. Similarly, pseudo-metric DP, defined in [36], is a combination of d D -privacy and (ε, δ)-DP; while extended divergence DP [97] combines d D -privacy with divergence DP.…”