Private Stochastic Convex Optimization: Optimal Rates in L1 Geometry

Hilal Asi, Vitaly Feldman, Tomer Koren, Kunal Talwar
Proceedings of the 38th International Conference on Machine Learning, PMLR 139:393-403, 2021.

Abstract

Stochastic convex optimization over an ℓ1-bounded domain is ubiquitous in machine learning applications such as LASSO but remains poorly understood when learning with differential privacy. We show that, up to logarithmic factors the optimal excess population loss of any (ϵ,δ)-differentially private optimizer is √log(d)/n+√d/ϵn. The upper bound is based on a new algorithm that combines the iterative localization approach of Feldman et al. (2020) with a new analysis of private regularized mirror descent. It applies to ℓp bounded domains for p∈[1,2] and queries at most n3/2 gradients improving over the best previously known algorithm for the ℓ2 case which needs n2 gradients. Further, we show that when the loss functions satisfy additional smoothness assumptions, the excess loss is upper bounded (up to logarithmic factors) by √log(d)/n+(log(d)/ϵn)2/3. This bound is achieved by a new variance-reduced version of the Frank-Wolfe algorithm that requires just a single pass over the data. We also show that the lower bound in this case is the minimum of the two rates mentioned above.

Cite this Paper

BibTeX
@InProceedings{pmlr-v139-asi21b, title = {Private Stochastic Convex Optimization: Optimal Rates in L1 Geometry}, author = {Asi, Hilal and Feldman, Vitaly and Koren, Tomer and Talwar, Kunal}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, pages = {393--403}, year = {2021}, editor = {Meila, Marina and Zhang, Tong}, volume = {139}, series = {Proceedings of Machine Learning Research}, month = {18--24 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v139/asi21b/asi21b.pdf}, url = {https://proceedings.mlr.press/v139/asi21b.html}, abstract = {Stochastic convex optimization over an $\ell_1$-bounded domain is ubiquitous in machine learning applications such as LASSO but remains poorly understood when learning with differential privacy. We show that, up to logarithmic factors the optimal excess population loss of any $(\epsilon,\delta)$-differentially private optimizer is $\sqrt{\log(d)/n} + \sqrt{d}/\epsilon n.$ The upper bound is based on a new algorithm that combines the iterative localization approach of Feldman et al. (2020) with a new analysis of private regularized mirror descent. It applies to $\ell_p$ bounded domains for $p\in [1,2]$ and queries at most $n^{3/2}$ gradients improving over the best previously known algorithm for the $\ell_2$ case which needs $n^2$ gradients. Further, we show that when the loss functions satisfy additional smoothness assumptions, the excess loss is upper bounded (up to logarithmic factors) by $\sqrt{\log(d)/n} + (\log(d)/\epsilon n)^{2/3}.$ This bound is achieved by a new variance-reduced version of the Frank-Wolfe algorithm that requires just a single pass over the data. We also show that the lower bound in this case is the minimum of the two rates mentioned above.} }
Endnote
%0 Conference Paper %T Private Stochastic Convex Optimization: Optimal Rates in L1 Geometry %A Hilal Asi %A Vitaly Feldman %A Tomer Koren %A Kunal Talwar %B Proceedings of the 38th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2021 %E Marina Meila %E Tong Zhang %F pmlr-v139-asi21b %I PMLR %P 393--403 %U https://proceedings.mlr.press/v139/asi21b.html %V 139 %X Stochastic convex optimization over an $\ell_1$-bounded domain is ubiquitous in machine learning applications such as LASSO but remains poorly understood when learning with differential privacy. We show that, up to logarithmic factors the optimal excess population loss of any $(\epsilon,\delta)$-differentially private optimizer is $\sqrt{\log(d)/n} + \sqrt{d}/\epsilon n.$ The upper bound is based on a new algorithm that combines the iterative localization approach of Feldman et al. (2020) with a new analysis of private regularized mirror descent. It applies to $\ell_p$ bounded domains for $p\in [1,2]$ and queries at most $n^{3/2}$ gradients improving over the best previously known algorithm for the $\ell_2$ case which needs $n^2$ gradients. Further, we show that when the loss functions satisfy additional smoothness assumptions, the excess loss is upper bounded (up to logarithmic factors) by $\sqrt{\log(d)/n} + (\log(d)/\epsilon n)^{2/3}.$ This bound is achieved by a new variance-reduced version of the Frank-Wolfe algorithm that requires just a single pass over the data. We also show that the lower bound in this case is the minimum of the two rates mentioned above.
APA
Asi, H., Feldman, V., Koren, T. & Talwar, K.. (2021). Private Stochastic Convex Optimization: Optimal Rates in L1 Geometry. Proceedings of the 38th International Conference on Machine Learning, in Proceedings of Machine Learning Research 139:393-403 Available from https://proceedings.mlr.press/v139/asi21b.html.

Related Material