Investigating Human-Identifiable Features Hidden in Adversarial Perturbations