Differentially Private Stochastic Gradient Descent (DP-SGD) is a key method for applying privacy in the training of deep learning models. This applies isotropic Gaussian noise to gradients during training, which can perturb these gradients in any direction, damaging utility. Metric DP, however, can provide alternative mechanisms based on arbitrary metrics that might be more suitable for preserving utility. In this paper, we apply \textit{directional privacy}, via a mechanism based on the von Mises-Fisher (VMF) distribution, to perturb gradients in terms of \textit{angular distance} so that gradient direction is broadly preserved. We show that this provides both $\epsilon$-DP and $\epsilon d$-privacy for deep learning training, rather than the $(\epsilon, \delta)$-privacy of the Gaussian mechanism; we observe that the $\epsilon d$-privacy guarantee does not require a $\delta>0$ term but degrades smoothly according to the dissimilarity of the input gradients. As $\epsilon$s between these different frameworks cannot be directly compared, we examine empirical privacy calibration mechanisms that go beyond previous work on empirically calibrating privacy within standard DP frameworks using membership inference attacks (MIA); we show that a combination of enhanced MIA and reconstruction attacks provides a suitable method for privacy calibration. Experiments on key datasets then indicate that the VMF mechanism can outperform the Gaussian in the utility-privacy trade-off. In particular, our experiments provide a direct comparison of privacy between the two approaches in terms of their ability to defend against reconstruction and membership inference.
翻译:暂无翻译