Pointwise maximal leakage (PML) is an operationally meaningful privacy measure that quantifies the amount of information leaking about a secret X to a single outcome of a related random variable Y. In this paper, we extend the notion of PML to random variables on arbitrary probability spaces. We develop two new definitions: First, we extend PML to countably infinite random variables by considering adversaries who aim to guess the value of discrete (finite or countably infinite) functions of X. Then, we consider adversaries who construct estimates of X that maximize the expected value of their corresponding gain functions. We use this latter setup to introduce a highly versatile form of PML that captures many scenarios of practical interest whose definition requires no assumptions about the underlying probability spaces.

We introduce a privacy measure called pointwise maximal leakage, generalizing the pre-existing notion of maximal leakage, which quantifies the amount of information leaking about a secret$X$by disclosing a single outcome of a (randomized) function calculated on$X$. Pointwise maximal leakage is a robust and operationally meaningful privacy measure that captures the largest amount of information leaking about$X$to adversaries seeking to guess arbitrary (possibly randomized) functions of$X$, or equivalently, aiming to maximize arbitrary gain functions. We study several properties of pointwise maximal leakage, e.g., how it composes over multiple outcomes, how it is affected by pre- and post-processing, etc. Furthermore, we propose to view information leakage as a random variable which, in turn, allows us to regard privacy guarantees as requirements imposed on different statistical properties of the information leakage random variable. We define several privacy guarantees and study how they behave under pre-processing, post-processing and composition. Finally, we examine the relationship between pointwise maximal leakage and other privacy notions such as local differential privacy, local information privacy,$f$-information, and so on. Overall, our paper constructs a robust and flexible framework for privacy risk assessment whose central notion has a strong operational meaning which can be adapted to a variety of applications and practical scenarios.

Pointwise maximal leakage (PML) is a robust and operationally meaningful privacy measure that quantifies the amount of information leaking about a secret X by disclosing a single outcome of a (randomized) function calculated on X. In this paper, we define a new privacy measure called event maximal leakage (EML), which generalizes PML by quantifying the amount of information leaking about X to arbitrary events. Then, we use our new privacy measure to define a new probabilistic privacy guarantee called (ϵ, δ)-EML. We study the data-processing and composition properties of (ϵ, δ)-EML and other privacy guarantees, where our goal is to understand whether or not they are closed under pre- and post-processing, and how they change as a result of adaptively composing privacy mechanisms.

International audience

Most methods for publishing data with privacy guarantees introduce randomness into datasets which reduces the utility of the published data. In this paper, we study the privacy-utility tradeoff by taking maximal leakage as the privacy measure and the expected Hamming distortion as the utility measure. We study three different but related problems. First, we assume that the data-generating distribution (i.e., the prior) is known, and we find the optimal privacy mechanism that achieves the smallest distortion subject to a constraint on maximal leakage. Then, we assume that the prior belongs to some set of distributions, and we formulate a min-max problem for finding the smallest distortion achievable for the worst-case prior in the set, subject to a maximal leakage constraint. Lastly, we define a partial order on privacy mechanisms based on the largest distortion they generate. Our results show that when the prior distribution is known, the optimal privacy mechanism fully discloses symbols with the largest prior probabilities, and suppresses symbols with the smallest prior probabilities. Furthermore, we show that sets of priors that contain more uniform distributions lead to larger distortion, while privacy mechanisms that distribute the privacy budget more uniformly over the symbols create smaller worst-case distortion. A full version of this paper is accessible at: https://arxiv.org/pdf/2105.01033.pdf

Machine learning models are known to memorize the unique properties of individual data points in a training set. This memorization capability can be exploited by several types of attacks to infer information about the training data, most notably, membership inference attacks. In this paper, we propose an approach based on information leakage for guaranteeing membership privacy. Specifically, we propose to use a conditional form of the notion of maximal leakage to quantify the information leaking about individual data entries in a dataset, i.e., the entrywise information leakage. We apply our privacy analysis to the Private Aggregation of Teacher Ensembles (PATE) framework for privacy-preserving classification of sensitive data and prove that the entrywise information leakage of its aggregation mechanism is Schur-concave when the injected noise has a log-concave probability density. The Schur-concavity of this leakage implies that increased consensus among teachers in labeling a query reduces its associated privacy cost. Finally, we derive upper bounds on the entrywise information leakage when the aggregation mechanism uses Laplace distributed noise.

We consider a three-node network, in which two agents wish to communicate over a noisy channel, while controlling the distribution observed by a third external agent. We use strong coordination to constrain the distribution, and we provide a complete characterization of the "remote strong coordination and reliable communication" region.

We consider a network of two nodes separated by a noisy channel with two-sided state information, in which the input and output signals have to be coordinated with the source and its reconstruction. In the case of non-causal encoding and decoding, we propose a joint source-channel coding scheme and we develop inner and outer bounds for the strong coordination region. While the inner and outer bounds do not match in general, we provide a complete characterization of the strong coordination region in three particular cases: i) when the channel is perfect; ii) when the decoder is lossless; and iii) when the random variables of the channel are independent from the random variables of the source. Through the study of these special cases, we prove that the separation principle does not hold for the joint source-channel strong coordination. Finally, in the absence of state information, we show that polar codes achieve a subset of the best known inner bound for the strong coordination region, therefore offering a constructive alternative to random binning and coding proofs.

We consider the problem of synthesizing joint distributions of signals and actions over noisy channels in the finite length regime. For a fixed blocklength n and an upper bound on the distance ε, a coding scheme is proposed such that the induced joint distribution is ε-close in L1distance to a target i.i.d. distribution. The set of achievable target distributions and rate for asymptotic strong coordination can be recovered from the main result of this paper by having n that tends to infinity.

We develop a random binning scheme for strong coordination in a network of two nodes separated by a noisy channel, in which the input and output signals have to be coordinated with the source and its reconstruction. In the case of non-causal encoding and decoding, we propose a joint source-channel coding scheme and develop inner and outer bounds for the strong coordination region. While the set of achievable target distributions is the same as for empirical coordination, we characterize the rate of common randomness required for strong coordination.

We develop a polar coding scheme for empirical coordination in a two-node network with a noisy link in which the input and output signals have to be coordinated with the source and the reconstruction. In the case of non-causal encoding and decoding, we show that polar codes achieve the best known inner bound for the empirical coordination region, provided that a vanishing rate of common randomness is available. This scheme provides a constructive alternative to random binning and coding proofs.