English

Robust Distribution Learning with Local and Global Adversarial Corruptions

Machine Learning 2024-06-26 v2 Machine Learning

Abstract

We consider learning in an adversarial environment, where an ε\varepsilon-fraction of samples from a distribution PP are arbitrarily modified (global corruptions) and the remaining perturbations have average magnitude bounded by ρ\rho (local corruptions). Given access to nn such corrupted samples, we seek a computationally efficient estimator P^n\hat{P}_n that minimizes the Wasserstein distance W1(P^n,P)\mathsf{W}_1(\hat{P}_n,P). In fact, we attack the fine-grained task of minimizing W1(Π#P^n,Π#P)\mathsf{W}_1(\Pi_\# \hat{P}_n, \Pi_\# P) for all orthogonal projections ΠRd×d\Pi \in \mathbb{R}^{d \times d}, with performance scaling with rank(Π)=k\mathrm{rank}(\Pi) = k. This allows us to account simultaneously for mean estimation (k=1k=1), distribution estimation (k=dk=d), as well as the settings interpolating between these two extremes. We characterize the optimal population-limit risk for this task and then develop an efficient finite-sample algorithm with error bounded by εk+ρ+O~(dkn1/(k2))\sqrt{\varepsilon k} + \rho + \tilde{O}(d\sqrt{k}n^{-1/(k \lor 2)}) when PP has bounded covariance. This guarantee holds uniformly in kk and is minimax optimal up to the sub-optimality of the plug-in estimator when ρ=ε=0\rho = \varepsilon = 0. Our efficient procedure relies on a novel trace norm approximation of an ideal yet intractable 2-Wasserstein projection estimator. We apply this algorithm to robust stochastic optimization, and, in the process, uncover a new method for overcoming the curse of dimensionality in Wasserstein distributionally robust optimization.

Keywords

Cite

@article{arxiv.2406.06509,
  title  = {Robust Distribution Learning with Local and Global Adversarial Corruptions},
  author = {Sloan Nietert and Ziv Goldfeld and Soroosh Shafiee},
  journal= {arXiv preprint arXiv:2406.06509},
  year   = {2024}
}

Comments

Accepted for presentation at the Conference on Learning Theory (COLT) 2024

R2 v1 2026-06-28T17:00:00.770Z