Privacy-utility tradeoff remains as one of the fundamental issues of differentially private machine learning. This paper introduces a geometrically inspired kernel-based approach to mitigate the accuracy-loss issue in classification. In this approach, a representation of the affine hull of given data points is learned in Reproducing Kernel Hilbert Spaces (RKHS). This leads to a novel distance measure that hides privacy-sensitive information about individual data points and improves the privacy-utility tradeoff via significantly reducing the risk of membership inference attacks. The effectiveness of the approach is demonstrated through experiments on MNIST dataset, Freiburg groceries dataset, and a real biomedical dataset. It is verified that the approach remains computationally practical. The application of the approach to federated learning is considered and it is observed that the accuracy-loss due to data being distributed is either marginal or not significantly high.
%0 Journal Article
%1 Kumar2024
%A Kumar, Mohit
%A Moser, Bernhard A.
%A Fischer, Lukas
%C El Segundo, CA, USA
%D 2024
%I AI Access Foundation
%J J. Artif. Int. Res.
%K deep-learning fuzzy mappings neural-networks
%R 10.1613/jair.1.15071
%T On Mitigating the Utility-Loss in Differentially Private Learning: A New Perspective by a Geometrically Inspired Kernel Approach
%U https://doi.org/10.1613/jair.1.15071
%V 79
%X Privacy-utility tradeoff remains as one of the fundamental issues of differentially private machine learning. This paper introduces a geometrically inspired kernel-based approach to mitigate the accuracy-loss issue in classification. In this approach, a representation of the affine hull of given data points is learned in Reproducing Kernel Hilbert Spaces (RKHS). This leads to a novel distance measure that hides privacy-sensitive information about individual data points and improves the privacy-utility tradeoff via significantly reducing the risk of membership inference attacks. The effectiveness of the approach is demonstrated through experiments on MNIST dataset, Freiburg groceries dataset, and a real biomedical dataset. It is verified that the approach remains computationally practical. The application of the approach to federated learning is considered and it is observed that the accuracy-loss due to data being distributed is either marginal or not significantly high.
@article{Kumar2024,
abstract = {Privacy-utility tradeoff remains as one of the fundamental issues of differentially private machine learning. This paper introduces a geometrically inspired kernel-based approach to mitigate the accuracy-loss issue in classification. In this approach, a representation of the affine hull of given data points is learned in Reproducing Kernel Hilbert Spaces (RKHS). This leads to a novel distance measure that hides privacy-sensitive information about individual data points and improves the privacy-utility tradeoff via significantly reducing the risk of membership inference attacks. The effectiveness of the approach is demonstrated through experiments on MNIST dataset, Freiburg groceries dataset, and a real biomedical dataset. It is verified that the approach remains computationally practical. The application of the approach to federated learning is considered and it is observed that the accuracy-loss due to data being distributed is either marginal or not significantly high.},
added-at = {2024-06-18T15:02:39.000+0200},
address = {El Segundo, CA, USA},
author = {Kumar, Mohit and Moser, Bernhard A. and Fischer, Lukas},
biburl = {https://www.bibsonomy.org/bibtex/2338be78257e104852043f5236becef94/scch},
day = 11,
doi = {10.1613/jair.1.15071},
interhash = {4fadbbfb24a634ce9a3b8897b7a4e3bc},
intrahash = {338be78257e104852043f5236becef94},
issn = {1076-9757},
issue_date = {Apr 2024},
journal = {J. Artif. Int. Res.},
keywords = {deep-learning fuzzy mappings neural-networks},
month = {2},
numpages = {53},
publisher = {AI Access Foundation},
timestamp = {2024-06-18T15:02:39.000+0200},
title = {On Mitigating the Utility-Loss in Differentially Private Learning: A New Perspective by a Geometrically Inspired Kernel Approach},
url = {https://doi.org/10.1613/jair.1.15071},
volume = 79,
year = 2024
}