Differential Privacy: An Economic Method for Choosing Epsilon

By Justin Hsu, Marco Gaboardi, Andreas Haeberlen, Sanjeev Khanna, Arjun Narayan, Benjamin C. Pierce, Aaron Roth. 2014

332 citations recorded by the source corpus.

Differential privacy is becoming a gold standard notion of privacy; it offers a guaranteed bound on loss of privacy due to release of query results, even under worst-case assumptions. The theory of differential privacy is an active research area, and there are now differentially private algorithms for a wide range of problems. However, the question of when differential privacy works in practice has received relatively little attention. In particular, there is still no rigorous method for choosing the key parameter ε, which controls the crucial tradeoff between the strength of the privacy guarantee and the accuracy of the published results. In this paper, we examine the role of these parameters in concrete applications, identifying the key considerations that must be addressed when choosing specific values. This choice requires balancing the interests of two parties with conflicting objectives: the data analyst, who wishes to learn something abou the data, and the prospective participant, who must decide whether to allow their data to be included in the analysis. We propose a simple model that expresses this balance as formulas over a handful of parameters, and we use our model to choose ε on a series of simple statistical studies. We also explore a surprising insight: in some circumstances, a differentially private study can be more accurate than a non-private study for the same cost, under our model. Finally, we discuss the simplifying assumptions in our model and outline a research agenda for possible refinements.

Subfields: Artificial Intelligence, Sociology and Political Science, Privacy & Security

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Sociology and Political Science

In Privacy & Security

Citations in the corpus

Privacy Loss in Apple's Implementation of Differential Privacy on MacOS 10.12 (2017, background): “…loss per datum is strictly limited to privacy budgets typically used in the literature, the daily privacy loss permitted by the implementation exceeds values typically considered acceptable by the theoretical community [12], and the overall privacy loss per device may be unbounded (Section 4).”
BLENDER: Enabling Local Search with a Hybrid Differential Privacy Model (2017, background): “01 to 10 for evaluating algorithm performance (see Table 1 in [18]).”
RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response (2014, neutral): “01 to 10 (see Table 1 in [15]).”
Preserving integrity in online social networks (2020, background): “In [46] it is shown that for a balanced privacy level epsilon of ln(3), to detect events with frequency 0.”
Cryptϵ: Crypto-Assisted Differential Privacy on Untrusted Servers (2019, methodology): “y capable of decryption in Cryptϵ. (2) Privacy Engine: Cryptϵstarts off with a total privacy budget of ϵB chosen by the data owners. The choice of value for ϵB should be guided by social prerogatives [7, 61, 69] and is currently outside the scope of Cryptϵ. For executing any program, the AS has to i”
Technical Privacy Metrics (2015, background): “ties in addition to identities, e.g. settings in a client software [50] or arbitrary strings [53]. However, the choice of the parameter ϵis difficult: values reported in the literature vary from 0.01 [68] to 100 [158]. A no-free-lunch theorem shows that differential privacy’s guarantees degrade in t”
Rademacher Observations, Private Data, and Boosting (2015, methodology): “In a first set of experiments, we have assessed the impact on learning of the feature-wise DP mechanism: on each tested domain, we have selected at random a binary feature, and then used Algorithm DP-FEAT to protect the feature for different values of DP parameter ǫ, in a range that covers usual DP ”
Open Civic Data: Of the People, For the People, By the People (2014, background): “Research on crowdsourcing [44] and data privacy & anonymization [29, 45] is highly relevant.”
On Memorization in Probabilistic Deep Generative Models (2021, background): “We expect that future work will focus on this important question, and suggest that inspiration may be taken from work on choosing ε in differential privacy [57, 58].”
Enabling Privacy-Preserving GWAS in Heterogeneous Human Populations (2016, unknown): “One can also look at exp ðeÞ from a Bayesian perspective (Hsu et al., 2014).”
Toward Distribution Estimation under Local Differential Privacy with Small Samples (2018, background): “1 and 2 in the high and middle privacy regime, respectively, since many studies used these values [24] and = 0.”
Combining Differential Privacy and Secure Multiparty Computation (2015, background): “[20] discusses how to choose for some practical ap-”
The Role of Differential Privacy in GDPR Compliance (2018, background): “Much of the theoretical literature advocates setting ϵ to be a small constant less than one, or to be diminishing in the size of the database (e.g., O ( 1 / √ n ) for a database of size n ) [10].”
Learning Games and Rademacher Observations Losses (2015, result): “ght DP protection: with plain random rados, test errors that compete with clean data can be observed for privacy budget ε ≈10−4, that is, more than a hundred times smaller than most reported studies (Hsu et al., 2014). In comparison, ADABOOST’s results, even plotted against the weak protection budge”