This attack learns two triggers (temporal and spectral) to embed a backdoor into the audio-language model (ALM) during prompt learning. The ALM’s weights remain frozen, and only the learnable prompts ...