Neural Network Control Policy Verification With Persistent Adversarial Perturbation