On Optimal Learning Under Targeted Data Poisoning

Consider the task of learning a hypothesis class H in the presence of an adversary that can replace up to an η fraction of the examples in the training set with arbitrary adversarial examples. The adversary aims to fail the learner on a particular target test point x which is known to the adversary but not to the learner. In this work we aim to characterize the smallest achievable error ε=ε(η) by the learner in the presence of such an adversary in both realizable and agnostic settings. We fully achieve this in the realizable setting, proving that ε=Theta(mathttVC(H)· η), where mathttVC(H) is the VC dimension of H. Remarkably, we show that the upper bound can be attained by a deterministic learner. In the agnostic setting we reveal a more elaborate landscape: we devise a deterministic learner with a multiplicative regret guarantee of ε ≤ C·mathttOPT + O(mathttVC(H)· η), where C > 1 is a universal numerical constant. We complement this by showing that for any deterministic learner there is an attack which worsens its error to at least 2· mathttOPT. This implies that a multiplicative deterioration in the regret is unavoidable in this case. Finally, the algorithms we develop for achieving the optimal rates are inherently improper. Nevertheless, we show that for a variety of natural concept classes, such as linear classifiers, it is possible to retain the dependence ε=ThetaH(η) by a proper algorithm in the realizable setting. Here ThetaH conceals a polynomial dependence on mathttVC(H).