On the Universal Adversarial Perturbations for Efficient Data-free Adversarial Detection