On the effect of normalization layers on Differentially Private training of deep Neural networks - Archive ouverte HAL Access content directly
Preprints, Working Papers, ... Year :

On the effect of normalization layers on Differentially Private training of deep Neural networks

(1) , (1) , (1) , (1)
1

Abstract

Differentially private stochastic gradient descent (DPSGD) is a variation of stochastic gradient descent based on the Differential Privacy (DP) paradigm, which can mitigate privacy threats that arise from the presence of sensitive information in training data. However, one major drawback of training deep neural networks with DPSGD is a reduction in the models accuracy. In this paper, we study the effect of normalization layers on the performance of DPSGD. We demonstrate that normalization layers significantly impact the utility of deep neural networks with noisy parameters and should be considered essential ingredients of training with DPSGD. In particular, we propose a novel method for integrating batch normalization with DPSGD without incurring an additional privacy loss. With our approach, we are able to train deeper networks and achieve a better utility-privacy trade-off.

Dates and versions

hal-03475600 , version 1 (11-12-2021)

Identifiers

Cite

Ali Davody, David Ifeoluwa Adelani, Thomas Kleinbauer, Dietrich Klakow. On the effect of normalization layers on Differentially Private training of deep Neural networks. 2021. ⟨hal-03475600⟩
57 View
0 Download

Altmetric

Share

Gmail Facebook Twitter LinkedIn More