Initializing Hard-Label Black-Box Adversarial Attacks Using Known Perturbations