The paper recasts Wasserstein distributionally robust optimization (DRO) as an optimization over transport maps that push empirical samples to adversarial ones. It proves that optimal adversarial maps are cyclically monotone, meaning they respect a global ordering that minimises transport cost. Standard uses per-sample gradient ascent, which violates this monotonicity and wastes transport cost. The authors propose two fixes: Multi-start Particle Ascent (MPA), which alternates gradient ascent with reassignment across samples to enforce cyclical monotonicity, and parameterizing adversarial maps as gradients of input-convex neural networks (ICNNs), which guarantees cyclical monotonicity by construction. Experiments on robust regression, image classification (CIFAR-10), and robust control show both methods outperform standard adversarial training and DRO baselines in terms of robustness and generalisation under distribution shift. The ICNN method delivers the strongest overall robustness, while MPA remains competitive. The methods are complementary to existing robust training pipelines.
Cyclical monotonicity enforces efficient adversarial transport in robust learning
Authors show that optimal adversarial transport maps must be cyclically monotone, and propose two methods to enforce this property, outperforming standard adversarial training.
Big Tech
EPFL · Apple
Research Digest··3 min read
In one line
Adversarial training violates optimal transport geometry; enforcing cyclical monotonicity via MPA or ICNN improves robustness.
Before you read this
Two ideas this paper builds onoptimal transportA mathematical method for finding the cheapest way to move probability mass between distributions, used to align, denoise, or repair data representations.5 papers →adversarial trainingTraining a model on adversarially perturbed inputs to make it robust to attacks, often trading some clean accuracy for resilience.9 papers →
Read it as
The authors reformulate distributionally robust optimization (DRO) as a problem over transport maps and prove that optimal maps are cyclically monotone.
The authors study a penalized Wasserstein DRO formulation where an adversary may choose any distribution but incurs a penalty for deviating from the empirical distribution.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.