A black-box adversarial attack for poisoning clustering

Clustering algorithms play a fundamental role as tools in decision-making and sensible automation processes. Due to the widespread use of these applications, robustness analysis this family against adversarial noise has become imperative. To best our knowledge, however, only few works have currently addressed problem. In an attempt fill gap, work, we propose black-box attack for crafting samples test clustering algorithms. We formulate problem constrained minimization program, general its structure customizable by attacker according her capability constraints. do not assume any information about internal victim algorithm, allow query it service only. absence derivative information, perform optimization with custom approach inspired Abstract Genetic Algorithm (AGA). experimental part, demonstrate sensibility different single ensemble crafted on scenarios. Furthermore, comparison algorithm state-of-the-art showing that are able reach or even outperform performance. Finally, highlight nature generated noise, show attacks transferable supervised such SVMs, random forests neural networks.