Comparing privacy notions for protection against reconstruction attacks in machine learning

Within the machine learning community, reconstruction attacks are a principal concern and have been identified even in federated learning (FL), which was designed with privacy preservation in mind. In response to these threats, the privacy community recommends the use of differential privacy (DP) in the stochastic gradient descent algorithm, termed DP-SGD. However, the proliferation of variants of DP in recent years\textemdash such as metric privacy\textemdash has made it challenging to conduct a fair comparison between different mechanisms due to the different meanings of the privacy parameters $\epsilon$ and $\delta$ across different variants. Thus, interpreting the practical implications of $\epsilon$ and $\delta$ in the FL context and amongst variants of DP remains ambiguous. In this paper, we lay a foundational framework for comparing mechanisms with differing notions of privacy guarantees, namely $(\epsilon,\delta)$-DP and metric privacy. We provide two foundational means of comparison: firstly, via the well-established $(\epsilon,\delta)$-DP guarantees, made possible through the R\'enyi differential privacy framework; and secondly, via Bayes' capacity, which we identify as an appropriate measure for reconstruction threats.

Metric Privacy in Federated Learning for Medical Imaging: Improving Convergence and Preventing Client Inference Attacks

Federated learning is a distributed learning technique that allows training a global model with the participation of different data owners without the need to share raw data. This architecture is orchestrated by a central server that aggregates the local models from the clients. This server may be trusted, but not all nodes in the network. Then, differential privacy (DP) can be used to privatize the global model by adding noise. However, this may affect convergence across the rounds of the federated architecture, depending also on the aggregation strategy employed. In this work, we aim to introduce the notion of metric-privacy to mitigate the impact of classical server side global-DP on the convergence of the aggregated model. Metric-privacy is a relaxation of DP, suitable for domains provided with a notion of distance. We apply it from the server side by computing a distance for the difference between the local models. We compare our approach with standard DP by analyzing the impact on six classical aggregation strategies. The proposed methodology is applied to an example of medical imaging and different scenarios are simulated across homogeneous and non-i.i.d clients. Finally, we introduce a novel client inference attack, where a semi-honest client tries to find whether another client participated in the training and study how it can be mitigated using DP and metric-privacy. Our evaluation shows that metric-privacy can increase the performance of the model compared to standard DP, while offering similar protection against client inference attacks.

Bayes' capacity as a measure for reconstruction attacks in federated learning

Within the machine learning community, reconstruction attacks are a principal attack of concern and have been identified even in federated learning, which was designed with privacy preservation in mind. In federated learning, it has been shown that an adversary with knowledge of the machine learning architecture is able to infer the exact value of a training element given an observation of the weight updates performed during stochastic gradient descent. In response to these threats, the privacy community recommends the use of differential privacy in the stochastic gradient descent algorithm, termed DP-SGD. However, DP has not yet been formally established as an effective countermeasure against reconstruction attacks. In this paper, we formalise the reconstruction threat model using the information-theoretic framework of quantitative information flow. We show that the Bayes' capacity, related to the Sibson mutual information of order infinity, represents a tight upper bound on the leakage of the DP-SGD algorithm to an adversary interested in performing a reconstruction attack. We provide empirical results demonstrating the effectiveness of this measure for comparing mechanisms against reconstruction threats.

A Systematic and Formal Study of the Impact of Local Differential Privacy on Fairness: Preliminary Results

Machine learning (ML) algorithms rely primarily on the availability of training data, and, depending on the domain, these data may include sensitive information about the data providers, thus leading to significant privacy issues. Differential privacy (DP) is the predominant solution for privacy-preserving ML, and the local model of DP is the preferred choice when the server or the data collector are not trusted. Recent experimental studies have shown that local DP can impact ML prediction for different subgroups of individuals, thus affecting fair decision-making. However, the results are conflicting in the sense that some studies show a positive impact of privacy on fairness while others show a negative one. In this work, we conduct a systematic and formal study of the effect of local DP on fairness. Specifically, we perform a quantitative study of how the fairness of the decisions made by the ML model changes under local DP for different levels of privacy and data distributions. In particular, we provide bounds in terms of the joint distributions and the privacy level, delimiting the extent to which local DP can impact the fairness of the model. We characterize the cases in which privacy reduces discrimination and those with the opposite effect. We validate our theoretical findings on synthetic and real-world datasets. Our results are preliminary in the sense that, for now, we study only the case of one sensitive attribute, and only statistical disparity, conditional statistical disparity, and equal opportunity difference.

On the Impact of Multi-dimensional Local Differential Privacy on Fairness

Automated decision systems are increasingly used to make consequential decisions in people's lives. Due to the sensitivity of the manipulated data as well as the resulting decisions, several ethical concerns need to be addressed for the appropriate use of such technologies, in particular, fairness and privacy. Unlike previous work, which focused on centralized differential privacy (DP) or local DP (LDP) for a single sensitive attribute, in this paper, we examine the impact of LDP in the presence of several sensitive attributes (i.e., multi-dimensional data) on fairness. Detailed empirical analysis on synthetic and benchmark datasets revealed very relevant observations. In particular, (1) multi-dimensional LDP is an efficient approach to reduce disparity, (2) the multi-dimensional approach of LDP (independent vs. combined) matters only at low privacy guarantees, and (3) the outcome Y distribution has an important effect on which group is more sensitive to the obfuscation. Last, we summarize our findings in the form of recommendations to guide practitioners in adopting effective privacy-preserving practices while maintaining fairness and utility in ML applications.

Causal Discovery Under Local Privacy

Differential privacy is a widely adopted framework designed to safeguard the sensitive information of data providers within a data set. It is based on the application of controlled noise at the interface between the server that stores and processes the data, and the data consumers. Local differential privacy is a variant that allows data providers to apply the privatization mechanism themselves on their data individually. Therefore it provides protection also in contexts in which the server, or even the data collector, cannot be trusted. The introduction of noise, however, inevitably affects the utility of the data, particularly by distorting the correlations between individual data components. This distortion can prove detrimental to tasks such as causal discovery. In this paper, we consider various well-known locally differentially private mechanisms and compare the trade-off between the privacy they provide, and the accuracy of the causal structure produced by algorithms for causal learning when applied to data obfuscated by these mechanisms. Our analysis yields valuable insights for selecting appropriate local differentially private protocols for causal discovery tasks. We foresee that our findings will aid researchers and practitioners in conducting locally private causal discovery.

Online Sensitivity Optimization in Differentially Private Learning

Training differentially private machine learning models requires constraining an individual's contribution to the optimization process. This is achieved by clipping the $2$-norm of their gradient at a predetermined threshold prior to averaging and batch sanitization. This selection adversely influences optimization in two opposing ways: it either exacerbates the bias due to excessive clipping at lower values, or augments sanitization noise at higher values. The choice significantly hinges on factors such as the dataset, model architecture, and even varies within the same optimization, demanding meticulous tuning usually accomplished through a grid search. In order to circumvent the privacy expenses incurred in hyperparameter tuning, we present a novel approach to dynamically optimize the clipping threshold. We treat this threshold as an additional learnable parameter, establishing a clean relationship between the threshold and the cost function. This allows us to optimize the former with gradient descent, with minimal repercussions on the overall privacy analysis. Our method is thoroughly assessed against alternative fixed and adaptive strategies across diverse datasets, tasks, model dimensions, and privacy levels. Our results demonstrate its comparable or superior performance in all evaluated scenarios, given the same privacy requirements.

Advancing Personalized Federated Learning: Group Privacy, Fairness, and Beyond

Federated learning (FL) is a framework for training machine learning models in a distributed and collaborative manner. During training, a set of participating clients process their data stored locally, sharing only the model updates obtained by minimizing a cost function over their local inputs. FL was proposed as a stepping-stone towards privacy-preserving machine learning, but it has been shown vulnerable to issues such as leakage of private information, lack of personalization of the model, and the possibility of having a trained model that is fairer to some groups than to others. In this paper, we address the triadic interaction among personalization, privacy guarantees, and fairness attained by models trained within the FL framework. Differential privacy and its variants have been studied and applied as cutting-edge standards for providing formal privacy guarantees. However, clients in FL often hold very diverse datasets representing heterogeneous communities, making it important to protect their sensitive information while still ensuring that the trained model upholds the aspect of fairness for the users. To attain this objective, a method is put forth that introduces group privacy assurances through the utilization of $d$-privacy (aka metric privacy). $d$-privacy represents a localized form of differential privacy that relies on a metric-oriented obfuscation approach to maintain the original data's topological distribution. This method, besides enabling personalized model training in a federated approach and providing formal privacy guarantees, possesses significantly better group fairness measured under a variety of standard metrics than a global model trained within a classical FL template. Theoretical justifications for the applicability are provided, as well as experimental validation on real-world datasets to illustrate the working of the proposed method.

On the Utility Gain of Iterative Bayesian Update for Locally Differentially Private Mechanisms

This paper investigates the utility gain of using Iterative Bayesian Update (IBU) for private discrete distribution estimation using data obfuscated with Locally Differentially Private (LDP) mechanisms. We compare the performance of IBU to Matrix Inversion (MI), a standard estimation technique, for seven LDP mechanisms designed for one-time data collection and for other seven LDP mechanisms designed for multiple data collections (e.g., RAPPOR). To broaden the scope of our study, we also varied the utility metric, the number of users n, the domain size k, and the privacy parameter {\epsilon}, using both synthetic and real-world data. Our results suggest that IBU can be a useful post-processing tool for improving the utility of LDP mechanisms in different scenarios without any additional privacy cost. For instance, our experiments show that IBU can provide better utility than MI, especially in high privacy regimes (i.e., when {\epsilon} is small). Our paper provides insights for practitioners to use IBU in conjunction with existing LDP mechanisms for more accurate and privacy-preserving data analysis. Finally, we implemented IBU for all fourteen LDP mechanisms into the state-of-the-art multi-freq-ldpy Python package (https://pypi.org/project/multi-freq-ldpy/) and open-sourced all our code used for the experiments as tutorials.

BaBE: Enhancing Fairness via Estimation of Latent Explaining Variables

We consider the problem of unfair discrimination between two groups and propose a pre-processing method to achieve fairness. Corrective methods like statistical parity usually lead to bad accuracy and do not really achieve fairness in situations where there is a correlation between the sensitive attribute S and the legitimate attribute E (explanatory variable) that should determine the decision. To overcome these drawbacks, other notions of fairness have been proposed, in particular, conditional statistical parity and equal opportunity. However, E is often not directly observable in the data, i.e., it is a latent variable. We may observe some other variable Z representing E, but the problem is that Z may also be affected by S, hence Z itself can be biased. To deal with this problem, we propose BaBE (Bayesian Bias Elimination), an approach based on a combination of Bayes inference and the Expectation-Maximization method, to estimate the most likely value of E for a given Z for each group. The decision can then be based directly on the estimated E. We show, by experiments on synthetic and real data sets, that our approach provides a good level of fairness as well as high accuracy.