Reproduce “Distributionally Robust Optimization with Adversarial Data Contamination” on OpenResearch

About this paper

Shuyao Li, Ilias Diakonikolas, Jelena Diakonikolas

Researchers at the University of Wisconsin-Madison developed a framework and algorithm to make Distributionally Robust Optimization (DRO) resilient to both adversarial training data contamination and test-time distributional shifts. The approach provides the first rigorous error guarantees for generalized linear models under these combined challenges, utilizing an efficient variant of the Primal-Dual Hybrid Gradient method.

Key insights

• Explicitly separating training data contamination from test-time distributional shifts is critical for constructing theoretically sound and effective robust learning solutions.
• The inherent structure of W1-DRO for generalized linear models allows for a tractable reformulation into a regularized risk minimization problem, which is amenable to efficient optimization techniques.
• By incorporating robust mean estimation into the iterative updates of a PDHG algorithm, strong optimality guarantees can be achieved even when operating on adversarially corrupted training data.

Summary by alphaXiv.