English

Solving Linear Programs with Differential Privacy

Data Structures and Algorithms 2025-07-16 v1

Abstract

We study the problem of solving linear programs of the form AxbAx\le b, x0x\ge0 with differential privacy. For homogeneous LPs Ax0Ax\ge0, we give an efficient (ϵ,δ)(\epsilon,\delta)-differentially private algorithm which with probability at least 1β1-\beta finds in polynomial time a solution that satisfies all but O(d2ϵlog2dδβlog1ρ0)O(\frac{d^{2}}{\epsilon}\log^{2}\frac{d}{\delta\beta}\sqrt{\log\frac{1}{\rho_{0}}}) constraints, for problems with margin ρ0>0\rho_{0}>0. This improves the bound of O(d5ϵlog1.51ρ0polylog(d,1δ,1β))O(\frac{d^{5}}{\epsilon}\log^{1.5}\frac{1}{\rho_{0}}\mathrm{poly}\log(d,\frac{1}{\delta},\frac{1}{\beta})) by [Kaplan-Mansour-Moran-Stemmer-Tur, STOC '25]. For general LPs AxbAx\le b, x0x\ge0 with potentially zero margin, we give an efficient (ϵ,δ)(\epsilon,\delta)-differentially private algorithm that w.h.p drops O(d4ϵlog2.5dδlogdU)O(\frac{d^{4}}{\epsilon}\log^{2.5}\frac{d}{\delta}\sqrt{\log dU}) constraints, where UU is an upper bound for the entries of AA and bb in absolute value. This improves the result by Kaplan et al. by at least a factor of d5d^{5}. Our techniques build upon privatizing a rescaling perceptron algorithm by [Hoberg-Rothvoss, IPCO '17] and a more refined iterative procedure for identifying equality constraints by Kaplan et al.

Keywords

Cite

@article{arxiv.2507.10946,
  title  = {Solving Linear Programs with Differential Privacy},
  author = {Alina Ene and Huy Le Nguyen and Ta Duy Nguyen and Adrian Vladu},
  journal= {arXiv preprint arXiv:2507.10946},
  year   = {2025}
}