DP-SGD vs PATE: Which Has Less Disparate Impact on Model Accuracy?

DP-SGD vs PATE: Which Has Less Disparate Impact on Model Accuracy?
复制标题

DP-SGD 与 PATE:哪个对模型精度的影响较小?

DOI:
--
复制
发表时间:
2021
期刊:
arXiv.org
影响因子:
--
通讯作者:
Andrew Trask
Andrew Trask
中科院分区:
--
文献类型:
--
作者:
Archit Uniyal;Rakshit Naidu;Sasikanth Kotti;Sahib Singh;Patrik Joslin Kenfack;FatemehSadat Mireshghallah;Andrew Trask

文献摘要

参考文献

被引文献

相似文献

差分隐私深度学习的最新进展表明,差分隐私的应用,特别是DP-SGD算法,对人口中的不同子群体具有不同的影响,这导致与代表性良好的子群体相比,代表性不足的子群体(少数群体)的模型效用显着较高。在这项工作中,我们的目标是比较PATE,另一种使用差分隐私训练深度学习模型的机制,与DP-SGD的公平性。我们发现,PATE确实有不同的影响,但是,它是远远低于DP-SGD严重。我们从这一观察中得出的见解可能是有前途的方向,以实现更好的公平,隐私权衡。
Recent advances in differentially private deep learning have demonstrated that application of differential privacy, specifically the DP-SGD algorithm, has a disparate impact on different sub-groups in the population, which leads to a significantly high drop-in model utility for sub-populations that are under-represented (minorities), compared to well-represented ones. In this work, we aim to compare PATE, another mechanism for training deep learning models using differential privacy, with DP-SGD in terms of fairness. We show that PATE does have a disparate impact too, however, it is much less severe than DP-SGD. We draw insights from this observation on what might be promising directions in achieving better fairness-privacy trade-offs.
DOI: 10.1162/99608f92.cfc5dd25
发表时间: 2020
期刊: Harvard data science review
影响因子: --
作者:
Bu Z;Dong J;Long Q;Su WJ
通讯作者: Su WJ
DOI: 10.1609/aaai.v35i11.17193
发表时间: 2020-09
期刊: ArXiv
影响因子: --
作者:
Cuong Tran;Ferdinando Fioretto;Pascal Van Hentenryck
通讯作者: Cuong Tran;Ferdinando Fioretto;Pascal Van Hentenryck
DOI: 10.1016/j.optlastec.2013.04.024
发表时间: 2013-12-01
影响因子: 5
作者:
Cong, Dalong;Zhou, Hong;Wang, Chuanwei
通讯作者: Wang, Chuanwei