Adversarial optimization for dictionary attacks on speaker verification