Active Privacy-Utility Trade-Off Against A Hypothesis Testing Adversary

Abstract

We consider a user releasing her data containing some personal information in return of a service. We model user's personal information as two correlated random variables, one of them, called the secret variable, is to be kept private, while the other, called the useful variable, is to be disclosed for utility. We consider active sequential data release, where at each time step the user chooses from among a finite set of release mechanisms, each revealing some information about the user's personal information, i.e., the true hypotheses, albeit with different statistics. The user manages data release in an online fashion such that maximum amount of information is revealed about the latent useful variable, while the confidence for the sensitive variable is kept below a predefined level. For the utility, we consider both the probability of correct detection of the useful variable and the mutual information (MI) between the useful variable and released data. We formulate both problems as a Markov decision process (MDP), and numerically solve them by advantage actor-critic (A2C) deep reinforcement learning (RL).

Active Privacy-Utility Trade-off Against Inference in Time-Series Data Sharing

“…Privacy for time-series data sharing and its applications to various domains have been extensively studied [8]- [11], [15], [20]- [31]. Most of these works focus on protecting the privacy of a single data point, e.g., the current measurement [23]- [28].…”

“…For instance, the user's presence at home or favorite TV channel can be inferred from SM readings, while her sensitive daily habits can be revealed to the SP through the sensors of a wearable device. Inference privacy protects user's data from an adversary's attempt to deduce sensitive information from an underlying distribution [12], [15], [19], [32]- [36]. These techniques perform well against inference attacks, in which the adversary aims at detecting the user's underlying private information with high confidence [30].…”

“…Differential privacy (DP), k-anonymity, information theoretic metrics and the SP's error probability are commonly used as privacy measures [8], [9], [11], [15], [20]- [31], [38]. By definition, DP prevents the SP from inferring the current data of the user, even if the SP has the knowledge of all the remaining data points.…”

“…Privacy leakage is measured by the error probability of the SP in detecting the distribution of the underlying data samples. In [15], the user shares her time-series data, which intrinsically contains correlated sensitive and useful information, with an untrusted SP in an online fashion. The goal is to maximize the confidence in the true useful variable for utility, while keeping the confidence in the sensitive r.v.…”

“…Hence, privacy is an important concern for the adoption of many IoT services, and there is a growing demand from consumers to keep their personal information private against malicious attackers and/or untrusted service providers (SPs), while preserving the utility obtained from these IoT services. Privacy has been widely studied in the literature [4]- [15], and a vast number of privacy measures have been introduced, including differential privacy [4], mutual information (MI) [6]- [12], total variation distance [16], maximal leakage [17], [18], and guessing leakage [19], to count a few. In this paper, we study the privacy-utility tradeoff (PUT) of time-series data sharing in the presence of a third party which tries to infer the user's sensitive information from the released data.…”