Gaussian Approximation for Composition and Subsampling of Differentially Private Algorithms

Abstract

The privacy analysis of differentially private stochastic gradient descent (DP-SGD) relies on bounds for the composition of subsampled mechanisms. The privacy of such a composition is bounded by that of a pair of product distributions, whose privacy loss is a sum of independent terms and has been analyzed via a variety of notions and concentration results. Yet, existing closed-form bounds either lose a logarithmic factor or require restrictive conditions on the noise scale and delicate analysis of subsampling. Further, while it is well known that sums of independent random variables are approximately Gaussian, existing CLT-based techniques give an approximation with a large error in $δ$. We show that the privacy loss of this pair is approximately Gaussian as long as each step has a small privacy loss except with small probability. Specifically, its privacy loss distribution is close to a Gaussian with the same mean and variance, after truncating the rare large privacy losses. The closeness in both directions is up to a shift of $\varepsilon$ that grows only as $\ln(1/β)$ for an additive error $β$ in $δ$ and is a lower-order term in typical applications. In particular, the composition satisfies Gaussian differential privacy up to a lower-order shift of $\varepsilon$ and the additive error $β$. We also give a simple way to analyze the resulting variance-based bounds for subsampled mechanisms which relies on the $χ^2$ divergence of a single step. For the Gaussian mechanism at sampling rate $1/t$ and $t>\ln^{3}(1/δ)$, this gives the predicted rate for $σ\gtrsim\max\{1,\sqrt{\ln(1/δ)}/\ln(t)\}$, whereas the standard Rényi-divergence analyses need $σ\gtrsim\sqrt{t\ln(1/δ)}/\ln(t)$. We extend these results to random allocation (balls-and-bins sampling), and show that it explains the high accuracy of the connect-the-dots method.