A new preprint on arXiv examines the inherent tension between adversarial robustness and geometric accuracy in binary classification. The authors study non-parametric classifiers in the presence of an adversary and prove an optimal quantitative estimate that controls geometric accuracy, measured by the Hausdorff distance to the Bayes decision boundary.
The result formalizes a tradeoff: improving robustness against adversarial perturbations necessarily degrades how closely the classifier's decision boundary matches the true Bayes boundary. The proof is described as optimal, meaning the bound cannot be improved in general.
The abstract is brief and does not specify the exact form of the bound or the assumptions beyond non-parametric binary classification. No comparisons to prior work are given in the available text, so the novelty is stated only as a new optimal estimate.