A Critical Review on the Use (and Misuse) of Differential Privacy in Machine Learning

Abstract

We review the use of differential privacy (DP) for privacy protection in machine learning (ML). We show that, driven by the aim of preserving the accuracy of the learned models, DP-based ML implementations are so loose that they do not offer the ex ante privacy guarantees of DP. Instead, what they deliver is basically noise addition similar to the traditional (and often criticized) statistical disclosure control approach. Due to the lack of formal privacy guarantees, the actual level of privacy offered must be experimentally assessed ex post , which is done very seldom. In this respect, we present empirical results showing that standard anti-overfitting techniques in ML can achieve a better utility/privacy/efficiency trade-off than DP.

Membership Inference Attack Against Principal Component Analysis

“…We would like to notice that this is not the usual range of values used in the literature. For example, in privacy-preserving data publishing, values of ϵ above 3 progressively seem to lose any meaningful guarantees [3]. However, for us, the fact that we will be using such large values is irrelevant, since we will empirically measure privacy leakage not through the ε itself, but through the effectiveness of an MIA.…”

Protect Your Score: Contact-Tracing with Differential Privacy Guarantees

“…We set this value to 1 1000 in all experiments. Existing literature prescribes that the δ parameter should be smaller than one divided by the data set size (Blanco-Justicia et al 2022;Hsu et al 2014;van Dijk and Nguyen 2023). Algorithm 1 uses a privacy bound for the contacts per day, and as our simulators have a max of 200 contacts per day, and an average of only fifteen contacts per day, 1 1000 is well below the recommended standard.…”

Steps Toward a Shared Infrastructure for Multi-Party Secure Private Computing in Official Statistics

“…The present contribution is scoped exclusively on Input Privacy techniques. Previous studies have exposed fundamental flaws of some recently proposed Output Privacy approaches aimed at sharing highly granular data (individual records, micro-data) and/or high-dimensional AI/ML models trained on personal data, see for instance Stadler and Troncoso (2022), Stadler et al (2022), Blanco-Justicia et al (2022), and Domingo-Ferrer et al (2021) and references therein. In the author’s opinion, the hurried adoption of these techniques, in certain business application contexts and without careful further scrutiny, may be more akin to privacywashing than to privacy enhancing.…”