Encrypted accelerated least squares regression

File Description SizeFormat 
esperanca17a.pdfPublished version405.86 kBAdobe PDFDownload
Title: Encrypted accelerated least squares regression
Author(s): Esperança, PM
Aslett, LJM
Holmes, CC
Item Type: Conference Paper
Abstract: Information that is stored in an encrypted format is, by definition, usually not amenable to statistical analysis or machine learning methods. In this paper we present detailed analysis of coordinate and accelerated gra- dient descent algorithms which are capable of fitting least squares and penalised ridge regression models, using data encrypted un- der a fully homomorphic encryption scheme. Gradient descent is shown to dominate in terms of encrypted computational speed, and theoretical results are proven to give param- eter bounds which ensure correctness of de- cryption. The characteristics of encrypted computation are empirically shown to favour a non-standard acceleration technique. This demonstrates the possibility of approximat- ing conventional statistical regression meth- ods using encrypted data without compro- mising privacy.
Publication Date: 31-Dec-2017
Date of Acceptance: 1-Jan-2017
URI: http://hdl.handle.net/10044/1/45670
Publisher: PMLR
Start Page: 334
End Page: 343
Journal / Book Title: Proceedings of Machine Learning Research
Volume: 54
Copyright Statement: © 2017 PLMR
Conference Name: Artificial Intelligence and Statistics (AISTATS)
Publication Status: Published
Start Date: 2017-04-20
Finish Date: 2017-04-22
Conference Place: Fort Lauderdale, Florida, USA
Open Access location: http://proceedings.mlr.press/v54/esperanca17a.html
Appears in Collections:Epidemiology, Public Health and Primary Care

Items in Spiral are protected by copyright, with all rights reserved, unless otherwise indicated.

Creative Commons