arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2211.04686cs.LGcs.CR
收藏

深度学习的方向性隐私

Directional Privacy for Deep Learning

Macquarie University(麦考瑞大学)

Pedro Faustini, Natasha Fernandes, Shakila Tonni, Annabelle McIver, Mark Dras

AI总结:

本文针对DP-SGD各向同性高斯噪声损害效用的问题,提出基于VMF分布的方向性隐私机制,通过角距离扰动梯度以保留方向,提供ε-DP和εd-隐私,在效用-隐私权衡上优于高斯机制。

摘要:

差分隐私随机梯度下降(DP-SGD)是在深度学习模型训练中应用隐私保护的关键方法。它在训练期间对梯度施加各向同性高斯噪声,这会在任何方向上扰动这些梯度,从而损害效用。然而,度量 DP 可以提供基于任意度量的替代机制,这些机制可能更适合于保留效用。在本文中,我们应用方向性隐私(基于 von Mises-Fisher (VMF) 分布的机制),根据角距离扰动梯度,从而广泛保留梯度方向。我们证明,这为深度学习训练提供了 ε-DP 和 εd-隐私,而不是高斯机制的 (ε, δ)-隐私。在关键数据集上的实验表明,VMF 机制在效用-隐私权衡方面优于高斯机制。具体而言,我们的实验在防御重构和成员推理的能力方面,直接对两种方法的隐私进行了经验性比较。

英文摘要:

Differentially Private Stochastic Gradient Descent (DP-SGD) is a key method for applying privacy in the training of deep learning models. It 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 $ε$-DP and $εd$-privacy for deep learning training, rather than the $(ε, δ)$-privacy of the Gaussian mechanism. 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 empirical comparison of privacy between the two approaches in terms of their ability to defend against reconstruction and membership inference.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑