Adversarial scratches: Deployable attacks to CNN classifiers

A growing body of work has shown that deep neural networks are susceptible to adversarial examples. These take the form small perturbations applied model's input which lead incorrect predictions. Unfortunately, most literature focuses on visually imperceivable be digital images often are, by design, impossible deployed physical targets. We present Adversarial Scratches: a novel L0 black-box attack, takes scratches in images, and possesses much greater deployability than other state-of-the-art attacks. Scratches leverage B\'ezier Curves reduce dimension search space possibly constrain attack specific location. test several scenarios, including publicly available API traffic signs. Results show that, often, our achieves higher fooling rate deployable methods, while requiring significantly fewer queries modifying very few pixels.