Differential Privacy of Gradient Descent on Perturbed Objectives

We study the running time, in terms of first order oracle queries, of differentially private empirical/population risk minimization of Lipschitz convex losses. We first consider the setting where the loss is non-smooth and the optimizer interacts with a private proxy oracle, which sends only private messages about a minibatch of gradients. In this setting, we show that expected running time $Ω(\min\{\frac{\sqrt{d}}{α^2}, \frac{d}{\log(1/α)}\})$ is necessary to achieve $α$ excess risk on problems of dimension $d$ when $d \geq 1/α^2$. Upper bounds via DP-SGD show these results are tight when $d>\tildeΩ(1/α^4)$. We further show our lower bound can be strengthened to $Ω(\min\{\frac{d}{\bar{m}α^2}, \frac{d}{\log(1/α)} \})$ for algorithms which use minibatches of size at most $\bar{m} < \sqrt{d}$. We next consider smooth losses, where we relax the private oracle assumption and give lower bounds under only the condition that the optimizer is private. Here, we lower bound the expected number of first order oracle calls by $\tildeΩ\big(\frac{\sqrt{d}}α + \min\{\frac{1}{α^2}, n\}\big)$, where $n$ is the size of the dataset. Modifications to existing algorithms show this bound is nearly tight. Compared to non-private lower bounds, our results show that differentially private optimizers pay a dimension dependent runtime penalty. Finally, as a natural extension of our proof technique, we show lower bounds in the non-smooth setting for optimizers interacting with information limited oracles. Specifically, if the proxy oracle transmits at most $Γ$-bits of information about the gradients in the minibatch, then $Ω\big(\min\{\frac{d}{α^2Γ}, \frac{d}{\log(1/α)}\}\big)$ oracle calls are needed. This result shows fundamental limitations of gradient quantization techniques in optimization.

Objective perturbation is a standard mechanism in differentially private empirical risk minimization. In particular, Linear Objective Perturbation (LOP) enforces privacy by adding a random linear term, while strong convexity and stability are ensured by an additional deterministic quadratic term. However, this approach requires the strong assumption of bounded gradients of the loss function, which excludes many modern machine learning models. In this work, we introduce Quadratic Objective Perturbation (QOP), which perturbs the objective with a random quadratic form. This perturbation induces strong convexity and enforces stability of the problem through curvature, thereby enabling privacy and allowing sensitivity to be controlled through spectral properties of the perturbation rather than assumptions on the gradients. As a result, we obtain $(\varepsilon, δ)$-differential privacy under weaker assumptions, in the interpolation regime. Furthermore, we extend the analysis to account for approximate solutions, showing that privacy guarantees are preserved under inexact solves. Additionally, we derive utility guarantees in terms of empirical excess risk, and provide a theoretical and numerical comparison to LOP, highlighting the advantages of curvature-based perturbations. Finally, we discuss algorithmic aspects and show that the resulting problems can be solved efficiently using modern splitting schemes.

Understanding the relationship between generalization and privacy remains a challenge in modern machine learning theory, particularly for deep networks that are trained by variants of differentially private stochastic gradient descent (DP- SGD). In this work we make progress on this persistent open problem. First, we derive explicit upper bounds on the approximate max-information of any algorithm that fulfills $(ε, δ)$-differential privacy or Rényi differential privacy, thereby going beyond the classical results for pure $ε$-differential privacy. Subsequently, we show even stronger guarantees for two common private learning algorithms, output perturbation with the Gaussian mechanism, and streaming DP-SGD, by exploiting the structure of their internal randomization. As an application of our results, we demonstrate how to obtain non-vacuous PAC-Bayes generalization bounds for deep networks, in which the prior distribution is learned by DP-SGD instead of the classical way of choosing it in a data-independent way.