Authors:
Matús Harvan
1
;
Thomas Locher
2
;
Marta Mularczyk
3
and
Yvonne Anne Pignolet
2
Affiliations:
1
Enovos, Luxembourg
;
2
ABB Corporate Research, Switzerland
;
3
ETH Zurich, Switzerland
Keyword(s):
Machine Learning, (Linear) Regression, Cloud Computing, (Partially) Homomorphic Encryption.
Related
Ontology
Subjects/Areas/Topics:
Applied Cryptography
;
Cryptographic Techniques and Key Management
;
Data and Application Security and Privacy
;
Data Engineering
;
Databases and Data Security
;
Information and Systems Security
;
Privacy
;
Privacy Enhancing Technologies
;
Security and Privacy in the Cloud
Abstract:
There is a growing interest in leveraging the computational resources and storage capacities of remote compute
and storage infrastructures for data analysis. However, the loss of control over the data raises concerns about
data privacy. In order to remedy these concerns, data can be encrypted before transmission to the remote
infrastructure, but the use of encryption renders data analysis a challenging task. An important observation is
that it suffices to encrypt only certain parts of the data in various real-world scenarios, which makes it possible
to devise efficient algorithms for secure remote data analysis based on partially homomorphic encryption.
We present several computationally efficient algorithms for regression analysis, focusing on linear regression,
that work with partially encrypted data. Our evaluation shows that we can both train models and compute
predictions with these models quickly enough for practical use. At the expense of full data confidentiality,
ou
r algorithms outperform state-of-the-art schemes based on fully homomorphic encryption or multi-party
computation by several orders of magnitude.
(More)