Noise Masking Attacks and Defenses for Pretrained Speech Models