Improving Fairness and Privacy in Selection Problems

This paper investigates the relation between three different notions of privacy: identifiability, differential privacy, and mutual-information privacy. Under a unified privacy-distortion framework, where the distortion is defined to be the expected Hamming distance between the input and output databases, we establish some fundamental connections between these three privacy notions. Given a maximum allowable distortion $D$ , we define the privacy-distortion functions $\epsilon _{\mathrm{ i}}^{*}(D)$ , $\epsilon _{\mathrm{ d}}^{*}(D)$ , and $\epsilon _{\mathrm{ m}}^{*}(D)$ to be the smallest (most private/best) identifiability level, differential privacy level, and mutual information between the input and the output, respectively. We characterize $\epsilon _{\mathrm{ i}}^{*}(D)$ and $\epsilon _{\mathrm{ d}}^{*}(D)$ , and prove that $\epsilon _{\mathrm{ i}}^{*}(D)-\epsilon _{X}\le \epsilon _{\mathrm{ d}}^{*}(D)\le \epsilon _{\mathrm{ i}}^{*}(D)$ for $D$ within certain range, where $\epsilon _{X}$ is a constant determined by the prior distribution of the original database $X$ , and diminishes to zero when $X$ is uniformly distributed. Furthermore, we show that $\epsilon _{\mathrm{ i}}^{*}(D)$ and $\epsilon _{\mathrm{ m}}^{*}(D)$ can be achieved by the same mechanism for $D$ within certain range, i.e., there is a mechanism that simultaneously minimizes the identifiability level and achieves the best mutual-information privacy. Based on these two connections, we prove that this mutual-information optimal mechanism satisfies $\epsilon $ -differential privacy with $\epsilon _{\mathrm{ d}}^{*}(D)\le \epsilon \le \epsilon _{\mathrm{ d}}^{*}(D)+2\epsilon _{X}$ . The results in this paper reveal some consistency between two worst case notions of privacy, namely, identifiability and differential privacy, and an average notion of privacy, mutual-information privacy.

Existing mathematical notions of fairness fail to account for the context of decision-making. We argue that moral consideration of contextual factors is an inherently human task. So we present a framework to learn context-aware mathematical formulations of fairness by eliciting people's situated fairness assessments. Our family of fairness notions corresponds to a new interpretation of economic models of Equality of Opportunity (EOP), and it includes most existing notions of fairness as special cases. Our human-in-the-loop approach is designed to learn the appropriate parameters of the EOP family by utilizing human responses to pair-wise questions about decision subjects' circumstance and deservingness, and the harm/benefit imposed on them. We illustrate our framework in a hypothetical criminal risk assessment scenario by conducting a series of human-subject experiments on Amazon Mechanical Turk. Our work takes an important initial step toward empowering stakeholders to have a voice in the formulation of fairness for Machine Learning.

SummaryIn statistical disclosure control, the goal of data analysis is twofold: the information released must provide accurate and useful statistics about the underlying population of interest, while minimizing the potential for an individual record to be identified. In recent years, the notion of differential privacy has received much attention in theoretical computer science, machine learning and statistics. It provides a rigorous and strong notion of protection for individuals’ sensitive information. A fundamental question is how to incorporate differential privacy in traditional statistical inference procedures. We study model selection in multivariate linear regression under the constraint of differential privacy. We show that model selection procedures based on penalized least squares or likelihood can be made differentially private by a combination of regularization and randomization, and we propose two algorithms to do so. We show that our privacy procedures are consistent under essentially the same conditions as the corresponding non-privacy procedures. We also find that, under differential privacy, the procedure becomes more sensitive to the tuning parameters. We illustrate and evaluate our method by using simulation studies and two real data examples.

Differentially private query answering is an important problem in machine learning and data anal-ysis. As an example, the users may want to know the empirical errors of some classifiers on a training data set which contains sensitive information. In this case, the answers returned by the database holder must preserve privacy. Previous results show that answering large number of queries with differential privacy is computational-ly expensive. However, almost all the hardness results are conditional, which means that they rely on unproved conjectures. One exception is the hardness for stateless mechanisms, in which the sanitizer is assumed to have very limited memory and can only answer the queries separately. In this work, we explore the hardness for differentially private query answering mechanisms beyond the stateless restriction. Specifically, we investigate the measurement complexity of a natural class of stateful sanitizer. Here, measurement refers to the evaluation of a query (a Boolean function) on a data point. Measurement complexity is a lower bound for the running time. We prove unconditional subexponential lower bound for the measurement complexity of the class of sanitizer with ∈ -differential privacy. Possible generalization of the results is also discussed.

Differential privacy (DP) is the de facto notion of privacy both in theory and in practice. However, despite its popularity, DP imposes strict requirements which guard against strong worst-case scenarios. For example, it guards against seemingly unrealistic scenarios where an attacker has full information about all but one point in the data set, and still nothing can be learned about the remaining point. While preventing such a strong attack is desirable, many works have explored whether average-case relaxations of DP are easier to satisfy [Hall et al., 2013; Wang et al., 2016; Bassily and Freund, 2016; Liu et al., 2023]. In this work, we are motivated by the question of whether alternate, weaker notions of privacy are possible: can a weakened privacy notion still guarantee some basic level of privacy, and on the other hand, achieve privacy more efficiently and/or for a substantially broader set of tasks? Our main result shows the answer is no: even in the statistical setting, any reasonable measure of privacy satisfying nontrivial composition is equivalent to DP. To prove this, we identify a core set of four axioms or desiderata: pre-processing invariance, prohibition of blatant non-privacy, strong composition, and linear scalability. Our main theorem shows that any privacy measure satisfying our axioms is equivalent to DP, up to polynomial factors in sample complexity. We complement this result by showing our axioms are minimal: removing any one of our axioms enables ill-behaved measures of privacy.

Differential privacy is a notion that has emerged in the community of statistical databases, as a response to the problem of protecting the privacy of the database's participants when performing statistical queries. The idea is that a randomized query satisfies differential privacy if the likelihood of obtaining a certain answer for a database $x$ is not too different from the likelihood of obtaining the same answer on adjacent databases, i.e. databases which differ from $x$ for only one individual. Information flow is an area of Security concerned with the problem of controlling the leakage of confidential information in programs and protocols. Nowadays, one of the most established approaches to quantify and to reason about leakage is based on the R\'enyi min entropy version of information theory. In this paper, we analyze critically the notion of differential privacy in light of the conceptual framework provided by the R\'enyi min information theory. We show that there is a close relation between differential privacy and leakage, due to the graph symmetries induced by the adjacency relation. Furthermore, we consider the utility of the randomized answer, which measures its expected degree of accuracy. We focus on certain kinds of utility functions called "binary", which have a close correspondence with the R\'enyi min mutual information. Again, it turns out that there can be a tight correspondence between differential privacy and utility, depending on the symmetries induced by the adjacency relation and by the query. Depending on these symmetries we can also build an optimal-utility randomization mechanism while preserving the required level of differential privacy. Our main contribution is a study of the kind of structures that can be induced by the adjacency relation and the query, and how to use them to derive bounds on the leakage and achieve the optimal utility.

We present spectral differential privacy (SpDP), a novel form of differential privacy (DP) designed to protect the frequency content of time-series data that come from wide sense stationary (WSS) stochastic processes. This notion is motivated by privacy needs in applications with time-series data over unbounded time, such as smart meters. First, a notion of DP on the space of (discretized) spectral densities is introduced. A Gaussian-like mechanism for SpDP is then presented that provides DP to the spectral density. Next, a novel streaming implementation is developed to enable real-time use of the proposed mechanism. The privacy guarantee provided by SpDP is independent of the time duration over which data are collected or shared. In contrast, time-domain trajectory-level DP (TrDP) will require noise with large variance to provide privacy over an extended time duration. The technique is numerically evaluated using smart meter data from a single home to compare the utility of SpDP to that of time-domain TrDP. The noise added by SpDP is substantially smaller than that added by time-domain TrDP, particularly when privacy over long time horizons is sought by TrDP.

When a database is protected by Differential Privacy (DP), its usability is limited in scope. In this scenario, generating a synthetic version of the data that mimics the properties of the private data allows users to perform any operation on the synthetic data, while maintaining the privacy of the original data. Therefore, multiple works have been devoted to devising systems for DP synthetic data generation. However, such systems may preserve or even magnify properties of the data that make it unfair, rendering the synthetic data unfit for use. In this work, we present PreFair, a system that allows for DP fair synthetic data generation. PreFair extends the state-of-the-art DP data generation mechanisms by incorporating a causal fairness criterion that ensures fair synthetic data. We adapt the notion of justifiable fairness to fit the synthetic data generation scenario. We further study the problem of generating DP fair synthetic data, showing its intractability and designing algorithms that are optimal under certain assumptions. We also provide an extensive experimental evaluation, showing that PreFair generates synthetic data that is significantly fairer than the data generated by leading DP data generation mechanisms, while remaining faithful to the private data.

The insurance industry is one of the industries that are digitally evolving at a high pace; therefore, harvesting and processing sensitive consumer information has become critical issues concerning privacy, security, and the moral mandate. This study explores the possibilities of combining legal frameworks and ethical requirements with technical protective measures to secure consumer information in the digital insurance providers. Conceptual and simulation-based approach was taken by applying synthetic insurance data applying four common data protection algorithms: Advanced Encryption Standard (AES), Rivest -Shamir-Adleman (RSA), k-anonymity, and differential privacy. The experimental results show that AES has a high data confidentiality rate of 99.8 per cent as well as low processing latency whereas RSA offers secure authentication rate of 98.6 per cent and attack resistance. Privacy-preserving analytics had a lower chance of re-identification of 8.5 with k-anonymity and 2.1 with differential privacy but with a slight decrease in the data utility. The outcomes of the ethical assessment showed that the highest score of the ethical compliance (9.3/10) was achieved by the methodology of differential privacy, which outcompeted the traditional security-only approaches. The comparative analysis with related work has shown that overall security, privacy, and ethical alignment scores went from 8.1, 7.4 and 7.0 to 9.2, 8.9 and 9.1 this way, due to the proposed integrated safeguarding. The results highlight the importance of a multi-layered solution that provides legal adherence, ethical governance, and advanced technical systems in order to provide credible and sustainable digital insurance systems.

PurposeThe study goes beyond the more frequent interest in information privacy to identify other notions of privacy within the workplace. The purpose of this paper is to explore how these additional notions of privacy relate to key demographic and employment characteristics and how data protection training, often instigated as a means of highlighting and addressing issues relating to privacy of customers’ data, is related to employees’ notions of their own workplace privacy.Design/methodology/approachThe study was undertaken in two telephone call centres since they offered a working environment where staff are highly monitored and hence there are likely to be issues relating to employee privacy. The study is exploratory in nature and adopts a mixed method approach based on a questionnaire survey that was followed by semi‐structured, qualitative face to face interviews.FindingsThe survey findings identified three distinct notions of privacy; the concern for personal information privacy (CfPIP), the concern for working environment privacy (CfWEP) and the concern for solitude privacy (CfSP). The findings were supported by the qualitative data provided by the interviews. CfWEP is found to be a gendered issue, with women showing a greater concern for the privacy of their working environment. Finally, the findings indicate that effective data protection training are associated with increased concern for their own privacy in the form of CfPIP, and that inclusion of data protection issues in performance reviews is associated their concern for CfWEP.Originality/valuePrevious studies of privacy in the workplace focus on the simplistic notion of information privacy. This study goes beyond such studies and provides empirically‐based evidence of multiple dimensions of privacy operant in a single, real‐world workplace setting. It also provides empirical insight to the previously unexplored issue of the association between data protection training employees’ notions of their own privacy.

Continuous publication of statistics collected from various location-based applications may compromise users’ privacy as the statistics could be procured from users’ private data. Differential Privacy (DP) is a new privacy notion that offers a strong privacy guarantee to all users who participate in the statistics. However, the existing DP mechanism for continuous publication of location statistics provides a privacy guarantee with the assumption that the data-points of users stream at consecutive timestamps are independent. In reality, users’ data-points may be temporally correlated, resulting in more privacy leakage due to an inadequate supply of privacy budget to the timestamps where the data-points are correlated. In this paper, we present a reformulated differential privacy definition to quantify the impact of temporal correlation on privacy leakage. Then, we introduce a privacy budget allocation method for allocating an adequate amount of privacy budget to each successive timestamps under the protection of differential privacy. Our solution adopts w-event privacy for continuously releasing statistics over infinite streams. The main idea is to check the dissimilarity between statistics at each timestamp and decide whether to publish current statistics or last release statistics. Finally, we evaluate the data utility of our proposed method by presenting experimental results for real and synthetic data sets.

Network data often contain sensitive relational information. One approach to protecting sensitive information while offering flexibility for network analysis is to share synthesized networks based on the information in originally observed networks. We employ differential privacy (DP) and exponential random graph models (ERGMs) and propose the DP-ERGM method to synthesize network data. We apply DP-ERGM to two real-world networks. We then compare the utility of synthesized networks generated by DP-ERGM, the DyadWise Randomized Response (DWRR) approach, and the Synthesis through Conditional distribution of Edge given nodal Attribute (SCEA) approach. In general, the results suggest that DP-ERGM preserves the original information significantly better than two other approaches in network structural statistics and inference for ERGMs and latent space models. Furthermore, DP-ERGM satisfies node DP through modeling the global network structure with ERGM, a stronger notion of privacy than the edge DP under which DWRR and SCEA operate.

The primary motivation is to address difficulties in data interpretation or a reduction in model accuracy. Although differential privacy can provide data privacy guarantees, it also creates problems. Thus, we need to consider the noise setting for differential privacy is currently inconclusive. This paper’s main contribution is finding a balance between privacy and accuracy. The training data of deep learning models may contain private or sensitive corporate information. These may be dangerous to attacks, leading to privacy data leakage for data sharing. Many strategies are for privacy protection, and differential privacy is the most widely applied one. Google proposed a federated learning technology to solve the problem of data silos in 2016. The technology can share information without exchanging original data and has made significant progress in the medical field. However, there is still the risk of data leakage in federated learning; thus, many models are now used with differential privacy mechanisms to minimize the risk. The data in the financial field are similar to medical data, which contains a substantial amount of personal data. The leakage may cause uncontrollable consequences, making data exchange and sharing difficult. Let us suppose that differential privacy applies to the financial field. Financial institutions can provide customers with higher value and personalized services and automate credit scoring and risk management. Unfortunately, the economic area rarely applies differential privacy and attains no consensus on parameter settings. This study compares data security with non-private and differential privacy financial visual models. The paper finds a balance between privacy protection with model accuracy. The results show that when the privacy loss parameter ϵ is between 12.62 and 5.41, the privacy models can protect training data, and the accuracy does not decrease too much.

Privacy preserving data mining is a substantial research area that aims at protecting the privacy of individuals while enabling to perform data mining techniques. In this study, we propose a secure protocol that fulfils the privacy restriction by combining homomorphic encryption with differential privacy and integrate this protocol into Holte’s One Rule which is a simple, but accurate and efficient classification algorithm. The proposed method allows a researcher to get the answers of his/her queries to build One Rule classifier by processing the encrypted training dataset under Paillier’s cryptosystem and also applies differential privacy to minimize the privacy leakage of individuals as much as possible in this training dataset. Therefore, both of security and privacy of the individuals in the training dataset for classification are provided thanks to our proposed method; since neither the parties, nor the researcher attain any information about the individuals in the database. Besides the One Rule classifier, we apply our proposed privacy preservation model to Naïve Bayes classification algorithm for the performance comparison, and show the efficiency of the proposed method through experiments on real data from UCI repository.

OPTDP: Towards optimal personalized trajectory differential privacy for trajectory data publishing