Symmetric Saliency-based Adversarial Attack To Speaker Identification